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.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/obdclass/lu_object.c
41 * These are the only exported functions, they provide some generic
42 * infrastructure for managing object devices
44 * Author: Nikita Danilov <nikita.danilov@sun.com>
47 #define DEBUG_SUBSYSTEM S_CLASS
49 # define EXPORT_SYMTAB
52 #include <libcfs/libcfs.h>
55 # include <linux/module.h>
59 #include <libcfs/libcfs_hash.h>
60 #include <obd_class.h>
61 #include <obd_support.h>
62 #include <lustre_disk.h>
63 #include <lustre_fid.h>
64 #include <lu_object.h>
65 #include <libcfs/list.h>
66 /* lu_time_global_{init,fini}() */
69 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
72 * Decrease reference counter on object. If last reference is freed, return
73 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
74 * case, free object immediately.
76 void lu_object_put(const struct lu_env *env, struct lu_object *o)
78 struct lu_site_bkt_data *bkt;
79 struct lu_object_header *top;
81 struct lu_object *orig;
85 site = o->lo_dev->ld_site;
88 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
89 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
91 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
92 if (lu_object_is_dying(top)) {
95 * somebody may be waiting for this, currently only
96 * used for cl_object, see cl_object_put_last().
98 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
103 LASSERT(bkt->lsb_busy > 0);
106 * When last reference is released, iterate over object
107 * layers, and notify them that object is no longer busy.
109 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
110 if (o->lo_ops->loo_object_release != NULL)
111 o->lo_ops->loo_object_release(env, o);
114 if (!lu_object_is_dying(top)) {
115 LASSERT(cfs_list_empty(&top->loh_lru));
116 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
117 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
122 * If object is dying (will not be cached), removed it
123 * from hash table and LRU.
125 * This is done with hash table and LRU lists locked. As the only
126 * way to acquire first reference to previously unreferenced
127 * object is through hash-table lookup (lu_object_find()),
128 * or LRU scanning (lu_site_purge()), that are done under hash-table
129 * and LRU lock, no race with concurrent object lookup is possible
130 * and we can safely destroy object below.
132 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
133 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
135 * Object was already removed from hash and lru above, can
138 lu_object_free(env, orig);
140 EXPORT_SYMBOL(lu_object_put);
143 * Allocate new object.
145 * This follows object creation protocol, described in the comment within
146 * struct lu_device_operations definition.
148 static struct lu_object *lu_object_alloc(const struct lu_env *env,
149 struct lu_device *dev,
150 const struct lu_fid *f,
151 const struct lu_object_conf *conf)
153 struct lu_object *scan;
154 struct lu_object *top;
161 * Create top-level object slice. This will also create
164 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
166 RETURN(ERR_PTR(-ENOMEM));
168 * This is the only place where object fid is assigned. It's constant
171 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
172 top->lo_header->loh_fid = *f;
173 layers = &top->lo_header->loh_layers;
176 * Call ->loo_object_init() repeatedly, until no more new
177 * object slices are created.
180 cfs_list_for_each_entry(scan, layers, lo_linkage) {
181 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
184 scan->lo_header = top->lo_header;
185 result = scan->lo_ops->loo_object_init(env, scan, conf);
187 lu_object_free(env, top);
188 RETURN(ERR_PTR(result));
190 scan->lo_flags |= LU_OBJECT_ALLOCATED;
194 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
195 if (scan->lo_ops->loo_object_start != NULL) {
196 result = scan->lo_ops->loo_object_start(env, scan);
198 lu_object_free(env, top);
199 RETURN(ERR_PTR(result));
204 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
211 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
213 struct lu_site_bkt_data *bkt;
214 struct lu_site *site;
215 struct lu_object *scan;
219 site = o->lo_dev->ld_site;
220 layers = &o->lo_header->loh_layers;
221 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
223 * First call ->loo_object_delete() method to release all resources.
225 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
226 if (scan->lo_ops->loo_object_delete != NULL)
227 scan->lo_ops->loo_object_delete(env, scan);
231 * Then, splice object layers into stand-alone list, and call
232 * ->loo_object_free() on all layers to free memory. Splice is
233 * necessary, because lu_object_header is freed together with the
236 CFS_INIT_LIST_HEAD(&splice);
237 cfs_list_splice_init(layers, &splice);
238 while (!cfs_list_empty(&splice)) {
240 * Free layers in bottom-to-top order, so that object header
241 * lives as long as possible and ->loo_object_free() methods
242 * can look at its contents.
244 o = container_of0(splice.prev, struct lu_object, lo_linkage);
245 cfs_list_del_init(&o->lo_linkage);
246 LASSERT(o->lo_ops->loo_object_free != NULL);
247 o->lo_ops->loo_object_free(env, o);
250 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
251 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
255 * Free \a nr objects from the cold end of the site LRU list.
257 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
259 struct lu_object_header *h;
260 struct lu_object_header *temp;
261 struct lu_site_bkt_data *bkt;
271 CFS_INIT_LIST_HEAD(&dispose);
273 * Under LRU list lock, scan LRU list and move unreferenced objects to
274 * the dispose list, removing them from LRU and hash table.
276 start = s->ls_purge_start;
277 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
280 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
284 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
285 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
287 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
288 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
290 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
291 LASSERT(bd.bd_bucket == bd2.bd_bucket);
293 cfs_hash_bd_del_locked(s->ls_obj_hash,
295 cfs_list_move(&h->loh_lru, &dispose);
299 if (nr != ~0 && --nr == 0)
302 if (count > 0 && --count == 0)
306 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
309 * Free everything on the dispose list. This is safe against
310 * races due to the reasons described in lu_object_put().
312 while (!cfs_list_empty(&dispose)) {
313 h = container_of0(dispose.next,
314 struct lu_object_header, loh_lru);
315 cfs_list_del_init(&h->loh_lru);
316 lu_object_free(env, lu_object_top(h));
317 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
324 if (nr != 0 && did_sth && start != 0) {
325 start = 0; /* restart from the first bucket */
328 /* race on s->ls_purge_start, but nobody cares */
329 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
333 EXPORT_SYMBOL(lu_site_purge);
338 * Code below has to jump through certain loops to output object description
339 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
340 * composes object description from strings that are parts of _lines_ of
341 * output (i.e., strings that are not terminated by newline). This doesn't fit
342 * very well into libcfs_debug_msg() interface that assumes that each message
343 * supplied to it is a self-contained output line.
345 * To work around this, strings are collected in a temporary buffer
346 * (implemented as a value of lu_cdebug_key key), until terminating newline
347 * character is detected.
355 * XXX overflow is not handled correctly.
360 struct lu_cdebug_data {
364 char lck_area[LU_CDEBUG_LINE];
367 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
368 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
371 * Key, holding temporary buffer. This key is registered very early by
374 struct lu_context_key lu_global_key = {
375 .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
376 .lct_init = lu_global_key_init,
377 .lct_fini = lu_global_key_fini
381 * Printer function emitting messages through libcfs_debug_msg().
383 int lu_cdebug_printer(const struct lu_env *env,
384 void *cookie, const char *format, ...)
386 struct lu_cdebug_print_info *info = cookie;
387 struct lu_cdebug_data *key;
392 va_start(args, format);
394 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
395 LASSERT(key != NULL);
397 used = strlen(key->lck_area);
398 complete = format[strlen(format) - 1] == '\n';
400 * Append new chunk to the buffer.
402 vsnprintf(key->lck_area + used,
403 ARRAY_SIZE(key->lck_area) - used, format, args);
405 if (cfs_cdebug_show(info->lpi_mask, info->lpi_subsys))
406 libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
407 (char *)info->lpi_file, info->lpi_fn,
408 info->lpi_line, "%s", key->lck_area);
409 key->lck_area[0] = 0;
414 EXPORT_SYMBOL(lu_cdebug_printer);
417 * Print object header.
419 void lu_object_header_print(const struct lu_env *env, void *cookie,
420 lu_printer_t printer,
421 const struct lu_object_header *hdr)
423 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
424 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
426 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
427 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
429 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
431 EXPORT_SYMBOL(lu_object_header_print);
434 * Print human readable representation of the \a o to the \a printer.
436 void lu_object_print(const struct lu_env *env, void *cookie,
437 lu_printer_t printer, const struct lu_object *o)
439 static const char ruler[] = "........................................";
440 struct lu_object_header *top;
444 lu_object_header_print(env, cookie, printer, top);
445 (*printer)(env, cookie, "{ \n");
446 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
447 depth = o->lo_depth + 4;
450 * print `.' \a depth times followed by type name and address
452 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
453 o->lo_dev->ld_type->ldt_name, o);
454 if (o->lo_ops->loo_object_print != NULL)
455 o->lo_ops->loo_object_print(env, cookie, printer, o);
456 (*printer)(env, cookie, "\n");
458 (*printer)(env, cookie, "} header@%p\n", top);
460 EXPORT_SYMBOL(lu_object_print);
463 * Check object consistency.
465 int lu_object_invariant(const struct lu_object *o)
467 struct lu_object_header *top;
470 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
471 if (o->lo_ops->loo_object_invariant != NULL &&
472 !o->lo_ops->loo_object_invariant(o))
477 EXPORT_SYMBOL(lu_object_invariant);
479 static struct lu_object *htable_lookup(struct lu_site *s,
481 const struct lu_fid *f,
482 cfs_waitlink_t *waiter,
485 struct lu_site_bkt_data *bkt;
486 struct lu_object_header *h;
487 cfs_hlist_node_t *hnode;
488 __u64 ver = cfs_hash_bd_version_get(bd);
494 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
495 /* cfs_hash_bd_lookup_intent is a somehow "internal" function
496 * of cfs_hash, but we don't want refcount on object right now */
497 hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
499 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
503 h = container_of0(hnode, struct lu_object_header, loh_hash);
504 if (likely(!lu_object_is_dying(h))) {
505 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
506 cfs_list_del_init(&h->loh_lru);
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 cfs_hash_bd_unlock(hs, &bd, 1);
608 * Allocate new object. This may result in rather complicated
609 * operations, including fld queries, inode loading, etc.
611 o = lu_object_alloc(env, dev, f, conf);
612 if (unlikely(IS_ERR(o)))
615 LASSERT(lu_fid_eq(lu_object_fid(o), f));
617 cfs_hash_bd_lock(hs, &bd, 1);
619 shadow = htable_lookup(s, &bd, f, waiter, &version);
620 if (likely(shadow == NULL)) {
621 struct lu_site_bkt_data *bkt;
623 bkt = cfs_hash_bd_extra_get(hs, &bd);
624 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
626 cfs_hash_bd_unlock(hs, &bd, 1);
630 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
631 cfs_hash_bd_unlock(hs, &bd, 1);
632 lu_object_free(env, o);
637 * Much like lu_object_find(), but top level device of object is specifically
638 * \a dev rather than top level device of the site. This interface allows
639 * objects of different "stacking" to be created within the same site.
641 struct lu_object *lu_object_find_at(const struct lu_env *env,
642 struct lu_device *dev,
643 const struct lu_fid *f,
644 const struct lu_object_conf *conf)
646 struct lu_site_bkt_data *bkt;
647 struct lu_object *obj;
651 obj = lu_object_find_try(env, dev, f, conf, &wait);
652 if (obj != ERR_PTR(-EAGAIN))
655 * lu_object_find_try() already added waiter into the
658 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
659 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
660 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
663 EXPORT_SYMBOL(lu_object_find_at);
666 * Find object with given fid, and return its slice belonging to given device.
668 struct lu_object *lu_object_find_slice(const struct lu_env *env,
669 struct lu_device *dev,
670 const struct lu_fid *f,
671 const struct lu_object_conf *conf)
673 struct lu_object *top;
674 struct lu_object *obj;
676 top = lu_object_find(env, dev, f, conf);
678 obj = lu_object_locate(top->lo_header, dev->ld_type);
680 lu_object_put(env, top);
685 EXPORT_SYMBOL(lu_object_find_slice);
688 * Global list of all device types.
690 static CFS_LIST_HEAD(lu_device_types);
692 int lu_device_type_init(struct lu_device_type *ldt)
696 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
697 result = ldt->ldt_ops->ldto_init(ldt);
699 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
702 EXPORT_SYMBOL(lu_device_type_init);
704 void lu_device_type_fini(struct lu_device_type *ldt)
706 cfs_list_del_init(&ldt->ldt_linkage);
707 ldt->ldt_ops->ldto_fini(ldt);
709 EXPORT_SYMBOL(lu_device_type_fini);
711 void lu_types_stop(void)
713 struct lu_device_type *ldt;
715 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
716 if (ldt->ldt_device_nr == 0)
717 ldt->ldt_ops->ldto_stop(ldt);
720 EXPORT_SYMBOL(lu_types_stop);
723 * Global list of all sites on this node
725 static CFS_LIST_HEAD(lu_sites);
726 static CFS_DECLARE_MUTEX(lu_sites_guard);
729 * Global environment used by site shrinker.
731 static struct lu_env lu_shrink_env;
733 struct lu_site_print_arg {
734 struct lu_env *lsp_env;
736 lu_printer_t lsp_printer;
740 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
741 cfs_hlist_node_t *hnode, void *data)
743 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
744 struct lu_object_header *h;
746 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
747 if (!cfs_list_empty(&h->loh_layers)) {
748 const struct lu_object *o;
750 o = lu_object_top(h);
751 lu_object_print(arg->lsp_env, arg->lsp_cookie,
752 arg->lsp_printer, o);
754 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
755 arg->lsp_printer, h);
761 * Print all objects in \a s.
763 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
764 lu_printer_t printer)
766 struct lu_site_print_arg arg = {
767 .lsp_env = (struct lu_env *)env,
768 .lsp_cookie = cookie,
769 .lsp_printer = printer,
772 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
774 EXPORT_SYMBOL(lu_site_print);
777 LU_CACHE_PERCENT_MAX = 50,
778 LU_CACHE_PERCENT_DEFAULT = 20
781 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
782 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
783 "Percentage of memory to be used as lu_object cache");
786 * Return desired hash table order.
788 static int lu_htable_order(void)
790 unsigned long cache_size;
794 * Calculate hash table size, assuming that we want reasonable
795 * performance when 20% of total memory is occupied by cache of
798 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
800 cache_size = cfs_num_physpages;
802 #if BITS_PER_LONG == 32
803 /* limit hashtable size for lowmem systems to low RAM */
804 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
805 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
808 /* clear off unreasonable cache setting. */
809 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
810 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
811 " the range of (0, %u]. Will use default value: %u.\n",
812 lu_cache_percent, LU_CACHE_PERCENT_MAX,
813 LU_CACHE_PERCENT_DEFAULT);
815 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
817 cache_size = cache_size / 100 * lu_cache_percent *
818 (CFS_PAGE_SIZE / 1024);
820 for (bits = 1; (1 << bits) < cache_size; ++bits) {
826 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
827 const void *key, unsigned mask)
829 struct lu_fid *fid = (struct lu_fid *)key;
832 hash = fid_flatten32(fid);
833 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
834 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
836 /* give me another random factor */
837 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
839 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
840 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
845 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
847 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
850 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
852 struct lu_object_header *h;
854 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
858 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
860 struct lu_object_header *h;
862 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
863 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
866 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
868 struct lu_object_header *h;
870 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
871 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
872 struct lu_site_bkt_data *bkt;
875 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
876 bkt = cfs_hash_bd_extra_get(hs, &bd);
881 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
883 LBUG(); /* we should never called it */
886 cfs_hash_ops_t lu_site_hash_ops = {
887 .hs_hash = lu_obj_hop_hash,
888 .hs_key = lu_obj_hop_key,
889 .hs_keycmp = lu_obj_hop_keycmp,
890 .hs_object = lu_obj_hop_object,
891 .hs_get = lu_obj_hop_get,
892 .hs_put_locked = lu_obj_hop_put_locked,
896 * Initialize site \a s, with \a d as the top level device.
898 #define LU_SITE_BITS_MIN 12
899 #define LU_SITE_BITS_MAX 24
901 * total 256 buckets, we don't want too many buckets because:
902 * - consume too much memory
903 * - avoid unbalanced LRU list
905 #define LU_SITE_BKT_BITS 8
907 int lu_site_init(struct lu_site *s, struct lu_device *top)
909 struct lu_site_bkt_data *bkt;
916 memset(s, 0, sizeof *s);
917 bits = lu_htable_order();
918 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
919 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
920 bits >= LU_SITE_BITS_MIN; bits--) {
921 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
922 bits - LU_SITE_BKT_BITS,
925 CFS_HASH_SPIN_BKTLOCK |
926 CFS_HASH_NO_ITEMREF |
928 CFS_HASH_ASSERT_EMPTY);
929 if (s->ls_obj_hash != NULL)
933 if (s->ls_obj_hash == NULL) {
934 CERROR("failed to create lu_site hash with bits: %d\n", bits);
938 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
939 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
940 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
941 cfs_waitq_init(&bkt->lsb_marche_funebre);
944 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
945 if (s->ls_stats == NULL) {
946 cfs_hash_putref(s->ls_obj_hash);
947 s->ls_obj_hash = NULL;
951 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
952 0, "created", "created");
953 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
954 0, "cache_hit", "cache_hit");
955 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
956 0, "cache_miss", "cache_miss");
957 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
958 0, "cache_race", "cache_race");
959 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
960 0, "cache_death_race", "cache_death_race");
961 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
962 0, "lru_purged", "lru_purged");
964 CFS_INIT_LIST_HEAD(&s->ls_linkage);
968 lu_ref_add(&top->ld_reference, "site-top", s);
970 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
971 cfs_spin_lock_init(&s->ls_ld_lock);
973 cfs_spin_lock(&s->ls_ld_lock);
974 cfs_list_add(&top->ld_linkage, &s->ls_ld_linkage);
975 cfs_spin_unlock(&s->ls_ld_lock);
979 EXPORT_SYMBOL(lu_site_init);
982 * Finalize \a s and release its resources.
984 void lu_site_fini(struct lu_site *s)
986 cfs_down(&lu_sites_guard);
987 cfs_list_del_init(&s->ls_linkage);
988 cfs_up(&lu_sites_guard);
990 if (s->ls_obj_hash != NULL) {
991 cfs_hash_putref(s->ls_obj_hash);
992 s->ls_obj_hash = NULL;
995 if (s->ls_top_dev != NULL) {
996 s->ls_top_dev->ld_site = NULL;
997 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
998 lu_device_put(s->ls_top_dev);
999 s->ls_top_dev = NULL;
1002 if (s->ls_stats != NULL)
1003 lprocfs_free_stats(&s->ls_stats);
1005 EXPORT_SYMBOL(lu_site_fini);
1008 * Called when initialization of stack for this site is completed.
1010 int lu_site_init_finish(struct lu_site *s)
1013 cfs_down(&lu_sites_guard);
1014 result = lu_context_refill(&lu_shrink_env.le_ctx);
1016 cfs_list_add(&s->ls_linkage, &lu_sites);
1017 cfs_up(&lu_sites_guard);
1020 EXPORT_SYMBOL(lu_site_init_finish);
1023 * Acquire additional reference on device \a d
1025 void lu_device_get(struct lu_device *d)
1027 cfs_atomic_inc(&d->ld_ref);
1029 EXPORT_SYMBOL(lu_device_get);
1032 * Release reference on device \a d.
1034 void lu_device_put(struct lu_device *d)
1036 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1037 cfs_atomic_dec(&d->ld_ref);
1039 EXPORT_SYMBOL(lu_device_put);
1042 * Initialize device \a d of type \a t.
1044 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1046 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1047 t->ldt_ops->ldto_start(t);
1048 memset(d, 0, sizeof *d);
1049 cfs_atomic_set(&d->ld_ref, 0);
1051 lu_ref_init(&d->ld_reference);
1052 CFS_INIT_LIST_HEAD(&d->ld_linkage);
1055 EXPORT_SYMBOL(lu_device_init);
1058 * Finalize device \a d.
1060 void lu_device_fini(struct lu_device *d)
1062 struct lu_device_type *t;
1065 if (d->ld_obd != NULL) {
1066 d->ld_obd->obd_lu_dev = NULL;
1070 lu_ref_fini(&d->ld_reference);
1071 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1072 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1073 LASSERT(t->ldt_device_nr > 0);
1074 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1075 t->ldt_ops->ldto_stop(t);
1077 EXPORT_SYMBOL(lu_device_fini);
1080 * Initialize object \a o that is part of compound object \a h and was created
1083 int lu_object_init(struct lu_object *o,
1084 struct lu_object_header *h, struct lu_device *d)
1086 memset(o, 0, sizeof *o);
1090 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1091 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1094 EXPORT_SYMBOL(lu_object_init);
1097 * Finalize object and release its resources.
1099 void lu_object_fini(struct lu_object *o)
1101 struct lu_device *dev = o->lo_dev;
1103 LASSERT(cfs_list_empty(&o->lo_linkage));
1106 lu_ref_del_at(&dev->ld_reference,
1107 o->lo_dev_ref , "lu_object", o);
1112 EXPORT_SYMBOL(lu_object_fini);
1115 * Add object \a o as first layer of compound object \a h
1117 * This is typically called by the ->ldo_object_alloc() method of top-level
1120 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1122 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1124 EXPORT_SYMBOL(lu_object_add_top);
1127 * Add object \a o as a layer of compound object, going after \a before.
1129 * This is typically called by the ->ldo_object_alloc() method of \a
1132 void lu_object_add(struct lu_object *before, struct lu_object *o)
1134 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1136 EXPORT_SYMBOL(lu_object_add);
1139 * Initialize compound object.
1141 int lu_object_header_init(struct lu_object_header *h)
1143 memset(h, 0, sizeof *h);
1144 cfs_atomic_set(&h->loh_ref, 1);
1145 CFS_INIT_HLIST_NODE(&h->loh_hash);
1146 CFS_INIT_LIST_HEAD(&h->loh_lru);
1147 CFS_INIT_LIST_HEAD(&h->loh_layers);
1148 lu_ref_init(&h->loh_reference);
1151 EXPORT_SYMBOL(lu_object_header_init);
1154 * Finalize compound object.
1156 void lu_object_header_fini(struct lu_object_header *h)
1158 LASSERT(cfs_list_empty(&h->loh_layers));
1159 LASSERT(cfs_list_empty(&h->loh_lru));
1160 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1161 lu_ref_fini(&h->loh_reference);
1163 EXPORT_SYMBOL(lu_object_header_fini);
1166 * Given a compound object, find its slice, corresponding to the device type
1169 struct lu_object *lu_object_locate(struct lu_object_header *h,
1170 const struct lu_device_type *dtype)
1172 struct lu_object *o;
1174 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1175 if (o->lo_dev->ld_type == dtype)
1180 EXPORT_SYMBOL(lu_object_locate);
1185 * Finalize and free devices in the device stack.
1187 * Finalize device stack by purging object cache, and calling
1188 * lu_device_type_operations::ldto_device_fini() and
1189 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1191 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1193 struct lu_site *site = top->ld_site;
1194 struct lu_device *scan;
1195 struct lu_device *next;
1197 lu_site_purge(env, site, ~0);
1198 for (scan = top; scan != NULL; scan = next) {
1199 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1200 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1201 lu_device_put(scan);
1205 lu_site_purge(env, site, ~0);
1207 if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1209 * Uh-oh, objects still exist.
1211 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
1213 lu_site_print(env, site, &cookie, lu_cdebug_printer);
1216 for (scan = top; scan != NULL; scan = next) {
1217 const struct lu_device_type *ldt = scan->ld_type;
1218 struct obd_type *type;
1220 next = ldt->ldt_ops->ldto_device_free(env, scan);
1221 type = ldt->ldt_obd_type;
1224 class_put_type(type);
1228 EXPORT_SYMBOL(lu_stack_fini);
1232 * Maximal number of tld slots.
1234 LU_CONTEXT_KEY_NR = 32
1237 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1239 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1242 * Global counter incremented whenever key is registered, unregistered,
1243 * revived or quiesced. This is used to void unnecessary calls to
1244 * lu_context_refill(). No locking is provided, as initialization and shutdown
1245 * are supposed to be externally serialized.
1247 static unsigned key_set_version = 0;
1252 int lu_context_key_register(struct lu_context_key *key)
1257 LASSERT(key->lct_init != NULL);
1258 LASSERT(key->lct_fini != NULL);
1259 LASSERT(key->lct_tags != 0);
1260 LASSERT(key->lct_owner != NULL);
1263 cfs_spin_lock(&lu_keys_guard);
1264 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1265 if (lu_keys[i] == NULL) {
1267 cfs_atomic_set(&key->lct_used, 1);
1269 lu_ref_init(&key->lct_reference);
1275 cfs_spin_unlock(&lu_keys_guard);
1278 EXPORT_SYMBOL(lu_context_key_register);
1280 static void key_fini(struct lu_context *ctx, int index)
1282 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1283 struct lu_context_key *key;
1285 key = lu_keys[index];
1286 LASSERT(key != NULL);
1287 LASSERT(key->lct_fini != NULL);
1288 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1290 key->lct_fini(ctx, key, ctx->lc_value[index]);
1291 lu_ref_del(&key->lct_reference, "ctx", ctx);
1292 cfs_atomic_dec(&key->lct_used);
1293 LASSERT(key->lct_owner != NULL);
1294 if (!(ctx->lc_tags & LCT_NOREF)) {
1295 LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1296 cfs_module_put(key->lct_owner);
1298 ctx->lc_value[index] = NULL;
1305 void lu_context_key_degister(struct lu_context_key *key)
1307 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1308 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1310 lu_context_key_quiesce(key);
1313 cfs_spin_lock(&lu_keys_guard);
1314 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1315 if (lu_keys[key->lct_index]) {
1316 lu_keys[key->lct_index] = NULL;
1317 lu_ref_fini(&key->lct_reference);
1319 cfs_spin_unlock(&lu_keys_guard);
1321 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1322 "key has instances: %d\n",
1323 cfs_atomic_read(&key->lct_used));
1325 EXPORT_SYMBOL(lu_context_key_degister);
1328 * Register a number of keys. This has to be called after all keys have been
1329 * initialized by a call to LU_CONTEXT_KEY_INIT().
1331 int lu_context_key_register_many(struct lu_context_key *k, ...)
1333 struct lu_context_key *key = k;
1339 result = lu_context_key_register(key);
1342 key = va_arg(args, struct lu_context_key *);
1343 } while (key != NULL);
1349 lu_context_key_degister(k);
1350 k = va_arg(args, struct lu_context_key *);
1357 EXPORT_SYMBOL(lu_context_key_register_many);
1360 * De-register a number of keys. This is a dual to
1361 * lu_context_key_register_many().
1363 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1369 lu_context_key_degister(k);
1370 k = va_arg(args, struct lu_context_key*);
1371 } while (k != NULL);
1374 EXPORT_SYMBOL(lu_context_key_degister_many);
1377 * Revive a number of keys.
1379 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1385 lu_context_key_revive(k);
1386 k = va_arg(args, struct lu_context_key*);
1387 } while (k != NULL);
1390 EXPORT_SYMBOL(lu_context_key_revive_many);
1393 * Quiescent a number of keys.
1395 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1401 lu_context_key_quiesce(k);
1402 k = va_arg(args, struct lu_context_key*);
1403 } while (k != NULL);
1406 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1409 * Return value associated with key \a key in context \a ctx.
1411 void *lu_context_key_get(const struct lu_context *ctx,
1412 const struct lu_context_key *key)
1414 LINVRNT(ctx->lc_state == LCS_ENTERED);
1415 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1416 LASSERT(lu_keys[key->lct_index] == key);
1417 return ctx->lc_value[key->lct_index];
1419 EXPORT_SYMBOL(lu_context_key_get);
1422 * List of remembered contexts. XXX document me.
1424 static CFS_LIST_HEAD(lu_context_remembered);
1427 * Destroy \a key in all remembered contexts. This is used to destroy key
1428 * values in "shared" contexts (like service threads), when a module owning
1429 * the key is about to be unloaded.
1431 void lu_context_key_quiesce(struct lu_context_key *key)
1433 struct lu_context *ctx;
1434 extern unsigned cl_env_cache_purge(unsigned nr);
1436 if (!(key->lct_tags & LCT_QUIESCENT)) {
1438 * XXX layering violation.
1440 cl_env_cache_purge(~0);
1441 key->lct_tags |= LCT_QUIESCENT;
1443 * XXX memory barrier has to go here.
1445 cfs_spin_lock(&lu_keys_guard);
1446 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1448 key_fini(ctx, key->lct_index);
1449 cfs_spin_unlock(&lu_keys_guard);
1453 EXPORT_SYMBOL(lu_context_key_quiesce);
1455 void lu_context_key_revive(struct lu_context_key *key)
1457 key->lct_tags &= ~LCT_QUIESCENT;
1460 EXPORT_SYMBOL(lu_context_key_revive);
1462 static void keys_fini(struct lu_context *ctx)
1466 cfs_spin_lock(&lu_keys_guard);
1467 if (ctx->lc_value != NULL) {
1468 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1470 OBD_FREE(ctx->lc_value,
1471 ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1472 ctx->lc_value = NULL;
1474 cfs_spin_unlock(&lu_keys_guard);
1477 static int keys_fill(struct lu_context *ctx)
1481 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1482 struct lu_context_key *key;
1485 if (ctx->lc_value[i] == NULL && key != NULL &&
1486 (key->lct_tags & ctx->lc_tags) &&
1488 * Don't create values for a LCT_QUIESCENT key, as this
1489 * will pin module owning a key.
1491 !(key->lct_tags & LCT_QUIESCENT)) {
1494 LINVRNT(key->lct_init != NULL);
1495 LINVRNT(key->lct_index == i);
1497 value = key->lct_init(ctx, key);
1498 if (unlikely(IS_ERR(value)))
1499 return PTR_ERR(value);
1501 LASSERT(key->lct_owner != NULL);
1502 if (!(ctx->lc_tags & LCT_NOREF))
1503 cfs_try_module_get(key->lct_owner);
1504 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1505 cfs_atomic_inc(&key->lct_used);
1507 * This is the only place in the code, where an
1508 * element of ctx->lc_value[] array is set to non-NULL
1511 ctx->lc_value[i] = value;
1512 if (key->lct_exit != NULL)
1513 ctx->lc_tags |= LCT_HAS_EXIT;
1515 ctx->lc_version = key_set_version;
1520 static int keys_init(struct lu_context *ctx)
1524 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1525 if (likely(ctx->lc_value != NULL))
1526 result = keys_fill(ctx);
1536 * Initialize context data-structure. Create values for all keys.
1538 int lu_context_init(struct lu_context *ctx, __u32 tags)
1540 memset(ctx, 0, sizeof *ctx);
1541 ctx->lc_state = LCS_INITIALIZED;
1542 ctx->lc_tags = tags;
1543 if (tags & LCT_REMEMBER) {
1544 cfs_spin_lock(&lu_keys_guard);
1545 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1546 cfs_spin_unlock(&lu_keys_guard);
1548 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1549 return keys_init(ctx);
1551 EXPORT_SYMBOL(lu_context_init);
1554 * Finalize context data-structure. Destroy key values.
1556 void lu_context_fini(struct lu_context *ctx)
1558 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1559 ctx->lc_state = LCS_FINALIZED;
1561 cfs_spin_lock(&lu_keys_guard);
1562 cfs_list_del_init(&ctx->lc_remember);
1563 cfs_spin_unlock(&lu_keys_guard);
1565 EXPORT_SYMBOL(lu_context_fini);
1568 * Called before entering context.
1570 void lu_context_enter(struct lu_context *ctx)
1572 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1573 ctx->lc_state = LCS_ENTERED;
1575 EXPORT_SYMBOL(lu_context_enter);
1578 * Called after exiting from \a ctx
1580 void lu_context_exit(struct lu_context *ctx)
1584 LINVRNT(ctx->lc_state == LCS_ENTERED);
1585 ctx->lc_state = LCS_LEFT;
1586 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1587 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1588 if (ctx->lc_value[i] != NULL) {
1589 struct lu_context_key *key;
1592 LASSERT(key != NULL);
1593 if (key->lct_exit != NULL)
1595 key, ctx->lc_value[i]);
1600 EXPORT_SYMBOL(lu_context_exit);
1603 * Allocate for context all missing keys that were registered after context
1606 int lu_context_refill(struct lu_context *ctx)
1608 LINVRNT(ctx->lc_value != NULL);
1609 return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1611 EXPORT_SYMBOL(lu_context_refill);
1614 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1615 * obd being added. Currently, this is only used on client side, specifically
1616 * for echo device client, for other stack (like ptlrpc threads), context are
1617 * predefined when the lu_device type are registered, during the module probe
1620 __u32 lu_context_tags_default = 0;
1621 __u32 lu_session_tags_default = 0;
1623 void lu_context_tags_update(__u32 tags)
1625 cfs_spin_lock(&lu_keys_guard);
1626 lu_context_tags_default |= tags;
1628 cfs_spin_unlock(&lu_keys_guard);
1630 EXPORT_SYMBOL(lu_context_tags_update);
1632 void lu_context_tags_clear(__u32 tags)
1634 cfs_spin_lock(&lu_keys_guard);
1635 lu_context_tags_default &= ~tags;
1637 cfs_spin_unlock(&lu_keys_guard);
1639 EXPORT_SYMBOL(lu_context_tags_clear);
1641 void lu_session_tags_update(__u32 tags)
1643 cfs_spin_lock(&lu_keys_guard);
1644 lu_session_tags_default |= tags;
1646 cfs_spin_unlock(&lu_keys_guard);
1648 EXPORT_SYMBOL(lu_session_tags_update);
1650 void lu_session_tags_clear(__u32 tags)
1652 cfs_spin_lock(&lu_keys_guard);
1653 lu_session_tags_default &= ~tags;
1655 cfs_spin_unlock(&lu_keys_guard);
1657 EXPORT_SYMBOL(lu_session_tags_clear);
1659 int lu_env_init(struct lu_env *env, __u32 tags)
1664 result = lu_context_init(&env->le_ctx, tags);
1665 if (likely(result == 0))
1666 lu_context_enter(&env->le_ctx);
1669 EXPORT_SYMBOL(lu_env_init);
1671 void lu_env_fini(struct lu_env *env)
1673 lu_context_exit(&env->le_ctx);
1674 lu_context_fini(&env->le_ctx);
1677 EXPORT_SYMBOL(lu_env_fini);
1679 int lu_env_refill(struct lu_env *env)
1683 result = lu_context_refill(&env->le_ctx);
1684 if (result == 0 && env->le_ses != NULL)
1685 result = lu_context_refill(env->le_ses);
1688 EXPORT_SYMBOL(lu_env_refill);
1691 * Currently, this API will only be used by echo client.
1692 * Because echo client and normal lustre client will share
1693 * same cl_env cache. So echo client needs to refresh
1694 * the env context after it get one from the cache, especially
1695 * when normal client and echo client co-exist in the same client.
1697 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1702 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1703 env->le_ctx.lc_version = 0;
1704 env->le_ctx.lc_tags |= ctags;
1707 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1708 env->le_ses->lc_version = 0;
1709 env->le_ses->lc_tags |= stags;
1712 result = lu_env_refill(env);
1716 EXPORT_SYMBOL(lu_env_refill_by_tags);
1718 static struct cfs_shrinker *lu_site_shrinker = NULL;
1720 typedef struct lu_site_stats{
1721 unsigned lss_populated;
1722 unsigned lss_max_search;
1727 static void lu_site_stats_get(cfs_hash_t *hs,
1728 lu_site_stats_t *stats, int populated)
1733 cfs_hash_for_each_bucket(hs, &bd, i) {
1734 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1735 cfs_hlist_head_t *hhead;
1737 cfs_hash_bd_lock(hs, &bd, 1);
1738 stats->lss_busy += bkt->lsb_busy;
1739 stats->lss_total += cfs_hash_bd_count_get(&bd);
1740 stats->lss_max_search = max((int)stats->lss_max_search,
1741 cfs_hash_bd_depmax_get(&bd));
1743 cfs_hash_bd_unlock(hs, &bd, 1);
1747 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1748 if (!cfs_hlist_empty(hhead))
1749 stats->lss_populated++;
1751 cfs_hash_bd_unlock(hs, &bd, 1);
1757 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1759 lu_site_stats_t stats;
1761 struct lu_site *tmp;
1763 int remain = shrink_param(sc, nr_to_scan);
1764 CFS_LIST_HEAD(splice);
1767 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1769 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1772 cfs_down(&lu_sites_guard);
1773 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1774 if (shrink_param(sc, nr_to_scan) != 0) {
1775 remain = lu_site_purge(&lu_shrink_env, s, remain);
1777 * Move just shrunk site to the tail of site list to
1778 * assure shrinking fairness.
1780 cfs_list_move_tail(&s->ls_linkage, &splice);
1783 memset(&stats, 0, sizeof(stats));
1784 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1785 cached += stats.lss_total - stats.lss_busy;
1786 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1789 cfs_list_splice(&splice, lu_sites.prev);
1790 cfs_up(&lu_sites_guard);
1792 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1793 if (shrink_param(sc, nr_to_scan) == 0)
1794 CDEBUG(D_INODE, "%d objects cached\n", cached);
1803 * Environment to be used in debugger, contains all tags.
1805 struct lu_env lu_debugging_env;
1808 * Debugging printer function using printk().
1810 int lu_printk_printer(const struct lu_env *env,
1811 void *unused, const char *format, ...)
1815 va_start(args, format);
1816 vprintk(format, args);
1821 void lu_debugging_setup(void)
1823 lu_env_init(&lu_debugging_env, ~0);
1826 void lu_context_keys_dump(void)
1830 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1831 struct lu_context_key *key;
1835 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1836 i, key, key->lct_tags,
1837 key->lct_init, key->lct_fini, key->lct_exit,
1838 key->lct_index, cfs_atomic_read(&key->lct_used),
1839 key->lct_owner ? key->lct_owner->name : "",
1841 lu_ref_print(&key->lct_reference);
1845 EXPORT_SYMBOL(lu_context_keys_dump);
1846 #else /* !__KERNEL__ */
1847 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1851 #endif /* __KERNEL__ */
1853 int cl_global_init(void);
1854 void cl_global_fini(void);
1855 int lu_ref_global_init(void);
1856 void lu_ref_global_fini(void);
1858 int dt_global_init(void);
1859 void dt_global_fini(void);
1861 int llo_global_init(void);
1862 void llo_global_fini(void);
1865 * Initialization of global lu_* data.
1867 int lu_global_init(void)
1871 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1873 result = lu_ref_global_init();
1877 LU_CONTEXT_KEY_INIT(&lu_global_key);
1878 result = lu_context_key_register(&lu_global_key);
1882 * At this level, we don't know what tags are needed, so allocate them
1883 * conservatively. This should not be too bad, because this
1884 * environment is global.
1886 cfs_down(&lu_sites_guard);
1887 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1888 cfs_up(&lu_sites_guard);
1893 * seeks estimation: 3 seeks to read a record from oi, one to read
1894 * inode, one for ea. Unfortunately setting this high value results in
1895 * lu_object/inode cache consuming all the memory.
1897 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1898 if (lu_site_shrinker == NULL)
1901 result = lu_time_global_init();
1906 result = dt_global_init();
1910 result = llo_global_init();
1914 result = cl_global_init();
1921 * Dual to lu_global_init().
1923 void lu_global_fini(void)
1930 lu_time_global_fini();
1931 if (lu_site_shrinker != NULL) {
1932 cfs_remove_shrinker(lu_site_shrinker);
1933 lu_site_shrinker = NULL;
1936 lu_context_key_degister(&lu_global_key);
1939 * Tear shrinker environment down _after_ de-registering
1940 * lu_global_key, because the latter has a value in the former.
1942 cfs_down(&lu_sites_guard);
1943 lu_env_fini(&lu_shrink_env);
1944 cfs_up(&lu_sites_guard);
1946 lu_ref_global_fini();
1949 struct lu_buf LU_BUF_NULL = {
1953 EXPORT_SYMBOL(LU_BUF_NULL);
1955 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1958 struct lprocfs_counter ret;
1960 lprocfs_stats_collect(stats, idx, &ret);
1961 return (__u32)ret.lc_count;
1968 * Output site statistical counters into a buffer. Suitable for
1969 * lprocfs_rd_*()-style functions.
1971 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1973 lu_site_stats_t stats;
1975 memset(&stats, 0, sizeof(stats));
1976 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
1978 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
1981 stats.lss_populated,
1982 CFS_HASH_NHLIST(s->ls_obj_hash),
1983 stats.lss_max_search,
1984 ls_stats_read(s->ls_stats, LU_SS_CREATED),
1985 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
1986 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
1987 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
1988 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
1989 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
1991 EXPORT_SYMBOL(lu_site_stats_print);
1993 const char *lu_time_names[LU_TIME_NR] = {
1994 [LU_TIME_FIND_LOOKUP] = "find_lookup",
1995 [LU_TIME_FIND_ALLOC] = "find_alloc",
1996 [LU_TIME_FIND_INSERT] = "find_insert"
1998 EXPORT_SYMBOL(lu_time_names);
2001 * Helper function to initialize a number of kmem slab caches at once.
2003 int lu_kmem_init(struct lu_kmem_descr *caches)
2006 struct lu_kmem_descr *iter = caches;
2008 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2009 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2012 if (*iter->ckd_cache == NULL) {
2014 /* free all previously allocated caches */
2015 lu_kmem_fini(caches);
2021 EXPORT_SYMBOL(lu_kmem_init);
2024 * Helper function to finalize a number of kmem slab cached at once. Dual to
2027 void lu_kmem_fini(struct lu_kmem_descr *caches)
2031 for (; caches->ckd_cache != NULL; ++caches) {
2032 if (*caches->ckd_cache != NULL) {
2033 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2034 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2036 *caches->ckd_cache = NULL;
2040 EXPORT_SYMBOL(lu_kmem_fini);