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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/obdclass/lu_object.c
35 * These are the only exported functions, they provide some generic
36 * infrastructure for managing object devices
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_CLASS
43 #include <linux/module.h>
44 #include <linux/list.h>
45 #include <libcfs/libcfs.h>
46 #include <libcfs/libcfs_hash.h> /* hash_long() */
47 #include <libcfs/linux/linux-mem.h>
48 #include <obd_class.h>
49 #include <obd_support.h>
50 #include <lustre_disk.h>
51 #include <lustre_fid.h>
52 #include <lu_object.h>
56 LU_CACHE_PERCENT_MAX = 50,
57 LU_CACHE_PERCENT_DEFAULT = 20
60 #define LU_CACHE_NR_MAX_ADJUST 512
61 #define LU_CACHE_NR_UNLIMITED -1
62 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
63 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
64 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
65 #define LU_CACHE_NR_ZFS_LIMIT 10240
67 #define LU_SITE_BITS_MIN 12
68 #define LU_SITE_BITS_MAX 24
69 #define LU_SITE_BITS_MAX_CL 19
71 * total 256 buckets, we don't want too many buckets because:
72 * - consume too much memory
73 * - avoid unbalanced LRU list
75 #define LU_SITE_BKT_BITS 8
78 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
79 module_param(lu_cache_percent, int, 0644);
80 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
82 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
83 module_param(lu_cache_nr, long, 0644);
84 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
86 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
87 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
90 * Decrease reference counter on object. If last reference is freed, return
91 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
92 * case, free object immediately.
94 void lu_object_put(const struct lu_env *env, struct lu_object *o)
96 struct lu_site_bkt_data *bkt;
97 struct lu_object_header *top;
99 struct lu_object *orig;
100 struct cfs_hash_bd bd;
101 const struct lu_fid *fid;
104 site = o->lo_dev->ld_site;
108 * till we have full fids-on-OST implemented anonymous objects
109 * are possible in OSP. such an object isn't listed in the site
110 * so we should not remove it from the site.
112 fid = lu_object_fid(o);
113 if (fid_is_zero(fid)) {
114 LASSERT(top->loh_hash.next == NULL
115 && top->loh_hash.pprev == NULL);
116 LASSERT(list_empty(&top->loh_lru));
117 if (!atomic_dec_and_test(&top->loh_ref))
119 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
120 if (o->lo_ops->loo_object_release != NULL)
121 o->lo_ops->loo_object_release(env, o);
123 lu_object_free(env, orig);
127 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
128 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
130 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
131 if (lu_object_is_dying(top)) {
133 * somebody may be waiting for this, currently only
134 * used for cl_object, see cl_object_put_last().
136 wake_up_all(&bkt->lsb_marche_funebre);
142 * When last reference is released, iterate over object
143 * layers, and notify them that object is no longer busy.
145 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
146 if (o->lo_ops->loo_object_release != NULL)
147 o->lo_ops->loo_object_release(env, o);
150 if (!lu_object_is_dying(top) &&
151 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
152 LASSERT(list_empty(&top->loh_lru));
153 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
155 percpu_counter_inc(&site->ls_lru_len_counter);
156 CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, "
158 o, site->ls_obj_hash, bkt, bkt->lsb_lru_len);
159 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
164 * If object is dying (will not be cached) then remove it
165 * from hash table and LRU.
167 * This is done with hash table and LRU lists locked. As the only
168 * way to acquire first reference to previously unreferenced
169 * object is through hash-table lookup (lu_object_find()),
170 * or LRU scanning (lu_site_purge()), that are done under hash-table
171 * and LRU lock, no race with concurrent object lookup is possible
172 * and we can safely destroy object below.
174 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
175 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
176 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
178 * Object was already removed from hash and lru above, can
181 lu_object_free(env, orig);
183 EXPORT_SYMBOL(lu_object_put);
186 * Put object and don't keep in cache. This is temporary solution for
187 * multi-site objects when its layering is not constant.
189 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
191 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
192 return lu_object_put(env, o);
194 EXPORT_SYMBOL(lu_object_put_nocache);
197 * Kill the object and take it out of LRU cache.
198 * Currently used by client code for layout change.
200 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
202 struct lu_object_header *top;
205 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
206 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
207 struct lu_site *site = o->lo_dev->ld_site;
208 struct cfs_hash *obj_hash = site->ls_obj_hash;
209 struct cfs_hash_bd bd;
211 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
212 if (!list_empty(&top->loh_lru)) {
213 struct lu_site_bkt_data *bkt;
215 list_del_init(&top->loh_lru);
216 bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
218 percpu_counter_dec(&site->ls_lru_len_counter);
220 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
221 cfs_hash_bd_unlock(obj_hash, &bd, 1);
224 EXPORT_SYMBOL(lu_object_unhash);
227 * Allocate new object.
229 * This follows object creation protocol, described in the comment within
230 * struct lu_device_operations definition.
232 static struct lu_object *lu_object_alloc(const struct lu_env *env,
233 struct lu_device *dev,
234 const struct lu_fid *f,
235 const struct lu_object_conf *conf)
237 struct lu_object *scan;
238 struct lu_object *top;
239 struct list_head *layers;
240 unsigned int init_mask = 0;
241 unsigned int init_flag;
247 * Create top-level object slice. This will also create
250 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
252 RETURN(ERR_PTR(-ENOMEM));
256 * This is the only place where object fid is assigned. It's constant
259 top->lo_header->loh_fid = *f;
260 layers = &top->lo_header->loh_layers;
264 * Call ->loo_object_init() repeatedly, until no more new
265 * object slices are created.
269 list_for_each_entry(scan, layers, lo_linkage) {
270 if (init_mask & init_flag)
273 scan->lo_header = top->lo_header;
274 result = scan->lo_ops->loo_object_init(env, scan, conf);
276 lu_object_free(env, top);
277 RETURN(ERR_PTR(result));
279 init_mask |= init_flag;
285 list_for_each_entry_reverse(scan, layers, lo_linkage) {
286 if (scan->lo_ops->loo_object_start != NULL) {
287 result = scan->lo_ops->loo_object_start(env, scan);
289 lu_object_free(env, top);
290 RETURN(ERR_PTR(result));
295 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
302 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
304 struct lu_site_bkt_data *bkt;
305 struct lu_site *site;
306 struct lu_object *scan;
307 struct list_head *layers;
308 struct list_head splice;
310 site = o->lo_dev->ld_site;
311 layers = &o->lo_header->loh_layers;
312 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
314 * First call ->loo_object_delete() method to release all resources.
316 list_for_each_entry_reverse(scan, layers, lo_linkage) {
317 if (scan->lo_ops->loo_object_delete != NULL)
318 scan->lo_ops->loo_object_delete(env, scan);
322 * Then, splice object layers into stand-alone list, and call
323 * ->loo_object_free() on all layers to free memory. Splice is
324 * necessary, because lu_object_header is freed together with the
327 INIT_LIST_HEAD(&splice);
328 list_splice_init(layers, &splice);
329 while (!list_empty(&splice)) {
331 * Free layers in bottom-to-top order, so that object header
332 * lives as long as possible and ->loo_object_free() methods
333 * can look at its contents.
335 o = container_of0(splice.prev, struct lu_object, lo_linkage);
336 list_del_init(&o->lo_linkage);
337 LASSERT(o->lo_ops->loo_object_free != NULL);
338 o->lo_ops->loo_object_free(env, o);
341 if (waitqueue_active(&bkt->lsb_marche_funebre))
342 wake_up_all(&bkt->lsb_marche_funebre);
346 * Free \a nr objects from the cold end of the site LRU list.
347 * if canblock is 0, then don't block awaiting for another
348 * instance of lu_site_purge() to complete
350 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
351 int nr, int canblock)
353 struct lu_object_header *h;
354 struct lu_object_header *temp;
355 struct lu_site_bkt_data *bkt;
356 struct cfs_hash_bd bd;
357 struct cfs_hash_bd bd2;
358 struct list_head dispose;
360 unsigned int start = 0;
365 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
368 INIT_LIST_HEAD(&dispose);
370 * Under LRU list lock, scan LRU list and move unreferenced objects to
371 * the dispose list, removing them from LRU and hash table.
374 start = s->ls_purge_start;
375 bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
378 * It doesn't make any sense to make purge threads parallel, that can
379 * only bring troubles to us. See LU-5331.
382 mutex_lock(&s->ls_purge_mutex);
383 else if (mutex_trylock(&s->ls_purge_mutex) == 0)
387 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
391 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
392 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
394 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
395 LASSERT(atomic_read(&h->loh_ref) == 0);
397 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
398 LASSERT(bd.bd_bucket == bd2.bd_bucket);
400 cfs_hash_bd_del_locked(s->ls_obj_hash,
402 list_move(&h->loh_lru, &dispose);
404 percpu_counter_dec(&s->ls_lru_len_counter);
408 if (nr != ~0 && --nr == 0)
411 if (count > 0 && --count == 0)
415 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
418 * Free everything on the dispose list. This is safe against
419 * races due to the reasons described in lu_object_put().
421 while (!list_empty(&dispose)) {
422 h = container_of0(dispose.next,
423 struct lu_object_header, loh_lru);
424 list_del_init(&h->loh_lru);
425 lu_object_free(env, lu_object_top(h));
426 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
432 mutex_unlock(&s->ls_purge_mutex);
434 if (nr != 0 && did_sth && start != 0) {
435 start = 0; /* restart from the first bucket */
438 /* race on s->ls_purge_start, but nobody cares */
439 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
444 EXPORT_SYMBOL(lu_site_purge_objects);
449 * Code below has to jump through certain loops to output object description
450 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
451 * composes object description from strings that are parts of _lines_ of
452 * output (i.e., strings that are not terminated by newline). This doesn't fit
453 * very well into libcfs_debug_msg() interface that assumes that each message
454 * supplied to it is a self-contained output line.
456 * To work around this, strings are collected in a temporary buffer
457 * (implemented as a value of lu_cdebug_key key), until terminating newline
458 * character is detected.
466 * XXX overflow is not handled correctly.
471 struct lu_cdebug_data {
475 char lck_area[LU_CDEBUG_LINE];
478 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
479 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
482 * Key, holding temporary buffer. This key is registered very early by
485 static struct lu_context_key lu_global_key = {
486 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
487 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
488 .lct_init = lu_global_key_init,
489 .lct_fini = lu_global_key_fini
493 * Printer function emitting messages through libcfs_debug_msg().
495 int lu_cdebug_printer(const struct lu_env *env,
496 void *cookie, const char *format, ...)
498 struct libcfs_debug_msg_data *msgdata = cookie;
499 struct lu_cdebug_data *key;
504 va_start(args, format);
506 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
507 LASSERT(key != NULL);
509 used = strlen(key->lck_area);
510 complete = format[strlen(format) - 1] == '\n';
512 * Append new chunk to the buffer.
514 vsnprintf(key->lck_area + used,
515 ARRAY_SIZE(key->lck_area) - used, format, args);
517 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
518 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
519 key->lck_area[0] = 0;
524 EXPORT_SYMBOL(lu_cdebug_printer);
527 * Print object header.
529 void lu_object_header_print(const struct lu_env *env, void *cookie,
530 lu_printer_t printer,
531 const struct lu_object_header *hdr)
533 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
534 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
536 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
537 list_empty((struct list_head *)&hdr->loh_lru) ? \
539 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
541 EXPORT_SYMBOL(lu_object_header_print);
544 * Print human readable representation of the \a o to the \a printer.
546 void lu_object_print(const struct lu_env *env, void *cookie,
547 lu_printer_t printer, const struct lu_object *o)
549 static const char ruler[] = "........................................";
550 struct lu_object_header *top;
554 lu_object_header_print(env, cookie, printer, top);
555 (*printer)(env, cookie, "{\n");
557 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
559 * print `.' \a depth times followed by type name and address
561 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
562 o->lo_dev->ld_type->ldt_name, o);
564 if (o->lo_ops->loo_object_print != NULL)
565 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
567 (*printer)(env, cookie, "\n");
570 (*printer)(env, cookie, "} header@%p\n", top);
572 EXPORT_SYMBOL(lu_object_print);
575 * Check object consistency.
577 int lu_object_invariant(const struct lu_object *o)
579 struct lu_object_header *top;
582 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
583 if (o->lo_ops->loo_object_invariant != NULL &&
584 !o->lo_ops->loo_object_invariant(o))
590 static struct lu_object *htable_lookup(struct lu_site *s,
591 struct cfs_hash_bd *bd,
592 const struct lu_fid *f,
595 struct lu_site_bkt_data *bkt;
596 struct lu_object_header *h;
597 struct hlist_node *hnode;
598 __u64 ver = cfs_hash_bd_version_get(bd);
601 return ERR_PTR(-ENOENT);
604 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
605 /* cfs_hash_bd_peek_locked is a somehow "internal" function
606 * of cfs_hash, it doesn't add refcount on object. */
607 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
609 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
610 return ERR_PTR(-ENOENT);
613 h = container_of0(hnode, struct lu_object_header, loh_hash);
614 cfs_hash_get(s->ls_obj_hash, hnode);
615 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
616 if (!list_empty(&h->loh_lru)) {
617 list_del_init(&h->loh_lru);
619 percpu_counter_dec(&s->ls_lru_len_counter);
621 return lu_object_top(h);
625 * Search cache for an object with the fid \a f. If such object is found,
626 * return it. Otherwise, create new object, insert it into cache and return
627 * it. In any case, additional reference is acquired on the returned object.
629 struct lu_object *lu_object_find(const struct lu_env *env,
630 struct lu_device *dev, const struct lu_fid *f,
631 const struct lu_object_conf *conf)
633 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
635 EXPORT_SYMBOL(lu_object_find);
638 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
639 * the calculation for the number of objects to reclaim is not covered by
640 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
641 * This ensures that many concurrent threads will not accidentally purge
644 static void lu_object_limit(const struct lu_env *env,
645 struct lu_device *dev)
649 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
652 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
653 nr = (__u64)lu_cache_nr;
657 lu_site_purge_objects(env, dev->ld_site,
658 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
661 static struct lu_object *lu_object_new(const struct lu_env *env,
662 struct lu_device *dev,
663 const struct lu_fid *f,
664 const struct lu_object_conf *conf)
668 struct cfs_hash_bd bd;
670 o = lu_object_alloc(env, dev, f, conf);
671 if (unlikely(IS_ERR(o)))
674 hs = dev->ld_site->ls_obj_hash;
675 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
676 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
677 cfs_hash_bd_unlock(hs, &bd, 1);
679 lu_object_limit(env, dev);
685 * Core logic of lu_object_find*() functions.
687 * Much like lu_object_find(), but top level device of object is specifically
688 * \a dev rather than top level device of the site. This interface allows
689 * objects of different "stacking" to be created within the same site.
691 struct lu_object *lu_object_find_at(const struct lu_env *env,
692 struct lu_device *dev,
693 const struct lu_fid *f,
694 const struct lu_object_conf *conf)
697 struct lu_object *shadow;
700 struct cfs_hash_bd bd;
704 * This uses standard index maintenance protocol:
706 * - search index under lock, and return object if found;
707 * - otherwise, unlock index, allocate new object;
708 * - lock index and search again;
709 * - if nothing is found (usual case), insert newly created
711 * - otherwise (race: other thread inserted object), free
712 * object just allocated.
716 * For "LOC_F_NEW" case, we are sure the object is new established.
717 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
718 * just alloc and insert directly.
720 * If dying object is found during index search, add @waiter to the
721 * site wait-queue and return ERR_PTR(-EAGAIN).
723 if (conf && conf->loc_flags & LOC_F_NEW)
724 return lu_object_new(env, dev, f, conf);
728 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
729 o = htable_lookup(s, &bd, f, &version);
730 cfs_hash_bd_unlock(hs, &bd, 1);
731 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
735 * Allocate new object. This may result in rather complicated
736 * operations, including fld queries, inode loading, etc.
738 o = lu_object_alloc(env, dev, f, conf);
739 if (unlikely(IS_ERR(o)))
742 LASSERT(lu_fid_eq(lu_object_fid(o), f));
744 cfs_hash_bd_lock(hs, &bd, 1);
746 shadow = htable_lookup(s, &bd, f, &version);
747 if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
748 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
749 cfs_hash_bd_unlock(hs, &bd, 1);
751 lu_object_limit(env, dev);
756 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
757 cfs_hash_bd_unlock(hs, &bd, 1);
758 lu_object_free(env, o);
761 EXPORT_SYMBOL(lu_object_find_at);
764 * Find object with given fid, and return its slice belonging to given device.
766 struct lu_object *lu_object_find_slice(const struct lu_env *env,
767 struct lu_device *dev,
768 const struct lu_fid *f,
769 const struct lu_object_conf *conf)
771 struct lu_object *top;
772 struct lu_object *obj;
774 top = lu_object_find(env, dev, f, conf);
778 obj = lu_object_locate(top->lo_header, dev->ld_type);
779 if (unlikely(obj == NULL)) {
780 lu_object_put(env, top);
781 obj = ERR_PTR(-ENOENT);
786 EXPORT_SYMBOL(lu_object_find_slice);
788 int lu_device_type_init(struct lu_device_type *ldt)
792 atomic_set(&ldt->ldt_device_nr, 0);
793 if (ldt->ldt_ops->ldto_init)
794 result = ldt->ldt_ops->ldto_init(ldt);
798 EXPORT_SYMBOL(lu_device_type_init);
800 void lu_device_type_fini(struct lu_device_type *ldt)
802 if (ldt->ldt_ops->ldto_fini)
803 ldt->ldt_ops->ldto_fini(ldt);
805 EXPORT_SYMBOL(lu_device_type_fini);
808 * Global list of all sites on this node
810 static LIST_HEAD(lu_sites);
811 static DECLARE_RWSEM(lu_sites_guard);
814 * Global environment used by site shrinker.
816 static struct lu_env lu_shrink_env;
818 struct lu_site_print_arg {
819 struct lu_env *lsp_env;
821 lu_printer_t lsp_printer;
825 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
826 struct hlist_node *hnode, void *data)
828 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
829 struct lu_object_header *h;
831 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
832 if (!list_empty(&h->loh_layers)) {
833 const struct lu_object *o;
835 o = lu_object_top(h);
836 lu_object_print(arg->lsp_env, arg->lsp_cookie,
837 arg->lsp_printer, o);
839 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
840 arg->lsp_printer, h);
846 * Print all objects in \a s.
848 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
849 lu_printer_t printer)
851 struct lu_site_print_arg arg = {
852 .lsp_env = (struct lu_env *)env,
853 .lsp_cookie = cookie,
854 .lsp_printer = printer,
857 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
859 EXPORT_SYMBOL(lu_site_print);
862 * Return desired hash table order.
864 static unsigned long lu_htable_order(struct lu_device *top)
866 unsigned long cache_size;
868 unsigned long bits_max = LU_SITE_BITS_MAX;
871 * For ZFS based OSDs the cache should be disabled by default. This
872 * allows the ZFS ARC maximum flexibility in determining what buffers
873 * to cache. If Lustre has objects or buffer which it wants to ensure
874 * always stay cached it must maintain a hold on them.
876 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
877 lu_cache_percent = 1;
878 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
879 return LU_SITE_BITS_MIN;
882 if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
883 bits_max = LU_SITE_BITS_MAX_CL;
886 * Calculate hash table size, assuming that we want reasonable
887 * performance when 20% of total memory is occupied by cache of
890 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
892 cache_size = totalram_pages;
894 #if BITS_PER_LONG == 32
895 /* limit hashtable size for lowmem systems to low RAM */
896 if (cache_size > 1 << (30 - PAGE_SHIFT))
897 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
900 /* clear off unreasonable cache setting. */
901 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
902 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
903 " the range of (0, %u]. Will use default value: %u.\n",
904 lu_cache_percent, LU_CACHE_PERCENT_MAX,
905 LU_CACHE_PERCENT_DEFAULT);
907 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
909 cache_size = cache_size / 100 * lu_cache_percent *
912 for (bits = 1; (1 << bits) < cache_size; ++bits) {
916 return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
919 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
920 const void *key, unsigned mask)
922 struct lu_fid *fid = (struct lu_fid *)key;
925 hash = fid_flatten32(fid);
926 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
927 hash = hash_long(hash, hs->hs_bkt_bits);
929 /* give me another random factor */
930 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
932 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
933 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
938 static void *lu_obj_hop_object(struct hlist_node *hnode)
940 return hlist_entry(hnode, struct lu_object_header, loh_hash);
943 static void *lu_obj_hop_key(struct hlist_node *hnode)
945 struct lu_object_header *h;
947 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
951 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
953 struct lu_object_header *h;
955 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
956 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
959 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
961 struct lu_object_header *h;
963 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
964 atomic_inc(&h->loh_ref);
967 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
969 LBUG(); /* we should never called it */
972 static struct cfs_hash_ops lu_site_hash_ops = {
973 .hs_hash = lu_obj_hop_hash,
974 .hs_key = lu_obj_hop_key,
975 .hs_keycmp = lu_obj_hop_keycmp,
976 .hs_object = lu_obj_hop_object,
977 .hs_get = lu_obj_hop_get,
978 .hs_put_locked = lu_obj_hop_put_locked,
981 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
983 spin_lock(&s->ls_ld_lock);
984 if (list_empty(&d->ld_linkage))
985 list_add(&d->ld_linkage, &s->ls_ld_linkage);
986 spin_unlock(&s->ls_ld_lock);
988 EXPORT_SYMBOL(lu_dev_add_linkage);
990 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
992 spin_lock(&s->ls_ld_lock);
993 list_del_init(&d->ld_linkage);
994 spin_unlock(&s->ls_ld_lock);
996 EXPORT_SYMBOL(lu_dev_del_linkage);
999 * Initialize site \a s, with \a d as the top level device.
1001 int lu_site_init(struct lu_site *s, struct lu_device *top)
1003 struct lu_site_bkt_data *bkt;
1004 struct cfs_hash_bd bd;
1011 memset(s, 0, sizeof *s);
1012 mutex_init(&s->ls_purge_mutex);
1014 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1015 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1017 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1022 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1023 for (bits = lu_htable_order(top);
1024 bits >= LU_SITE_BITS_MIN; bits--) {
1025 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1026 bits - LU_SITE_BKT_BITS,
1029 CFS_HASH_SPIN_BKTLOCK |
1030 CFS_HASH_NO_ITEMREF |
1032 CFS_HASH_ASSERT_EMPTY |
1034 if (s->ls_obj_hash != NULL)
1038 if (s->ls_obj_hash == NULL) {
1039 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1043 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1044 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1045 INIT_LIST_HEAD(&bkt->lsb_lru);
1046 init_waitqueue_head(&bkt->lsb_marche_funebre);
1049 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1050 if (s->ls_stats == NULL) {
1051 cfs_hash_putref(s->ls_obj_hash);
1052 s->ls_obj_hash = NULL;
1056 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1057 0, "created", "created");
1058 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1059 0, "cache_hit", "cache_hit");
1060 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1061 0, "cache_miss", "cache_miss");
1062 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1063 0, "cache_race", "cache_race");
1064 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1065 0, "cache_death_race", "cache_death_race");
1066 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1067 0, "lru_purged", "lru_purged");
1069 INIT_LIST_HEAD(&s->ls_linkage);
1070 s->ls_top_dev = top;
1073 lu_ref_add(&top->ld_reference, "site-top", s);
1075 INIT_LIST_HEAD(&s->ls_ld_linkage);
1076 spin_lock_init(&s->ls_ld_lock);
1078 lu_dev_add_linkage(s, top);
1082 EXPORT_SYMBOL(lu_site_init);
1085 * Finalize \a s and release its resources.
1087 void lu_site_fini(struct lu_site *s)
1089 down_write(&lu_sites_guard);
1090 list_del_init(&s->ls_linkage);
1091 up_write(&lu_sites_guard);
1093 percpu_counter_destroy(&s->ls_lru_len_counter);
1095 if (s->ls_obj_hash != NULL) {
1096 cfs_hash_putref(s->ls_obj_hash);
1097 s->ls_obj_hash = NULL;
1100 if (s->ls_top_dev != NULL) {
1101 s->ls_top_dev->ld_site = NULL;
1102 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1103 lu_device_put(s->ls_top_dev);
1104 s->ls_top_dev = NULL;
1107 if (s->ls_stats != NULL)
1108 lprocfs_free_stats(&s->ls_stats);
1110 EXPORT_SYMBOL(lu_site_fini);
1113 * Called when initialization of stack for this site is completed.
1115 int lu_site_init_finish(struct lu_site *s)
1118 down_write(&lu_sites_guard);
1119 result = lu_context_refill(&lu_shrink_env.le_ctx);
1121 list_add(&s->ls_linkage, &lu_sites);
1122 up_write(&lu_sites_guard);
1125 EXPORT_SYMBOL(lu_site_init_finish);
1128 * Acquire additional reference on device \a d
1130 void lu_device_get(struct lu_device *d)
1132 atomic_inc(&d->ld_ref);
1134 EXPORT_SYMBOL(lu_device_get);
1137 * Release reference on device \a d.
1139 void lu_device_put(struct lu_device *d)
1141 LASSERT(atomic_read(&d->ld_ref) > 0);
1142 atomic_dec(&d->ld_ref);
1144 EXPORT_SYMBOL(lu_device_put);
1147 * Initialize device \a d of type \a t.
1149 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1151 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1152 t->ldt_ops->ldto_start != NULL)
1153 t->ldt_ops->ldto_start(t);
1155 memset(d, 0, sizeof *d);
1157 lu_ref_init(&d->ld_reference);
1158 INIT_LIST_HEAD(&d->ld_linkage);
1162 EXPORT_SYMBOL(lu_device_init);
1165 * Finalize device \a d.
1167 void lu_device_fini(struct lu_device *d)
1169 struct lu_device_type *t = d->ld_type;
1171 if (d->ld_obd != NULL) {
1172 d->ld_obd->obd_lu_dev = NULL;
1176 lu_ref_fini(&d->ld_reference);
1177 LASSERTF(atomic_read(&d->ld_ref) == 0,
1178 "Refcount is %u\n", atomic_read(&d->ld_ref));
1179 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1181 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1182 t->ldt_ops->ldto_stop != NULL)
1183 t->ldt_ops->ldto_stop(t);
1185 EXPORT_SYMBOL(lu_device_fini);
1188 * Initialize object \a o that is part of compound object \a h and was created
1191 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1192 struct lu_device *d)
1194 memset(o, 0, sizeof(*o));
1198 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1199 INIT_LIST_HEAD(&o->lo_linkage);
1203 EXPORT_SYMBOL(lu_object_init);
1206 * Finalize object and release its resources.
1208 void lu_object_fini(struct lu_object *o)
1210 struct lu_device *dev = o->lo_dev;
1212 LASSERT(list_empty(&o->lo_linkage));
1215 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1221 EXPORT_SYMBOL(lu_object_fini);
1224 * Add object \a o as first layer of compound object \a h
1226 * This is typically called by the ->ldo_object_alloc() method of top-level
1229 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1231 list_move(&o->lo_linkage, &h->loh_layers);
1233 EXPORT_SYMBOL(lu_object_add_top);
1236 * Add object \a o as a layer of compound object, going after \a before.
1238 * This is typically called by the ->ldo_object_alloc() method of \a
1241 void lu_object_add(struct lu_object *before, struct lu_object *o)
1243 list_move(&o->lo_linkage, &before->lo_linkage);
1245 EXPORT_SYMBOL(lu_object_add);
1248 * Initialize compound object.
1250 int lu_object_header_init(struct lu_object_header *h)
1252 memset(h, 0, sizeof *h);
1253 atomic_set(&h->loh_ref, 1);
1254 INIT_HLIST_NODE(&h->loh_hash);
1255 INIT_LIST_HEAD(&h->loh_lru);
1256 INIT_LIST_HEAD(&h->loh_layers);
1257 lu_ref_init(&h->loh_reference);
1260 EXPORT_SYMBOL(lu_object_header_init);
1263 * Finalize compound object.
1265 void lu_object_header_fini(struct lu_object_header *h)
1267 LASSERT(list_empty(&h->loh_layers));
1268 LASSERT(list_empty(&h->loh_lru));
1269 LASSERT(hlist_unhashed(&h->loh_hash));
1270 lu_ref_fini(&h->loh_reference);
1272 EXPORT_SYMBOL(lu_object_header_fini);
1275 * Given a compound object, find its slice, corresponding to the device type
1278 struct lu_object *lu_object_locate(struct lu_object_header *h,
1279 const struct lu_device_type *dtype)
1281 struct lu_object *o;
1283 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1284 if (o->lo_dev->ld_type == dtype)
1289 EXPORT_SYMBOL(lu_object_locate);
1292 * Finalize and free devices in the device stack.
1294 * Finalize device stack by purging object cache, and calling
1295 * lu_device_type_operations::ldto_device_fini() and
1296 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1298 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1300 struct lu_site *site = top->ld_site;
1301 struct lu_device *scan;
1302 struct lu_device *next;
1304 lu_site_purge(env, site, ~0);
1305 for (scan = top; scan != NULL; scan = next) {
1306 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1307 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1308 lu_device_put(scan);
1312 lu_site_purge(env, site, ~0);
1314 for (scan = top; scan != NULL; scan = next) {
1315 const struct lu_device_type *ldt = scan->ld_type;
1316 struct obd_type *type;
1318 next = ldt->ldt_ops->ldto_device_free(env, scan);
1319 type = ldt->ldt_obd_type;
1322 class_put_type(type);
1329 * Maximal number of tld slots.
1331 LU_CONTEXT_KEY_NR = 40
1334 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1336 DEFINE_RWLOCK(lu_keys_guard);
1337 static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
1340 * Global counter incremented whenever key is registered, unregistered,
1341 * revived or quiesced. This is used to void unnecessary calls to
1342 * lu_context_refill(). No locking is provided, as initialization and shutdown
1343 * are supposed to be externally serialized.
1345 static unsigned key_set_version = 0;
1350 int lu_context_key_register(struct lu_context_key *key)
1355 LASSERT(key->lct_init != NULL);
1356 LASSERT(key->lct_fini != NULL);
1357 LASSERT(key->lct_tags != 0);
1358 LASSERT(key->lct_owner != NULL);
1361 write_lock(&lu_keys_guard);
1362 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1363 if (lu_keys[i] == NULL) {
1365 atomic_set(&key->lct_used, 1);
1367 lu_ref_init(&key->lct_reference);
1373 write_unlock(&lu_keys_guard);
1376 EXPORT_SYMBOL(lu_context_key_register);
1378 static void key_fini(struct lu_context *ctx, int index)
1380 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1381 struct lu_context_key *key;
1383 key = lu_keys[index];
1384 LASSERT(key != NULL);
1385 LASSERT(key->lct_fini != NULL);
1386 LASSERT(atomic_read(&key->lct_used) > 1);
1388 key->lct_fini(ctx, key, ctx->lc_value[index]);
1389 lu_ref_del(&key->lct_reference, "ctx", ctx);
1390 atomic_dec(&key->lct_used);
1392 LASSERT(key->lct_owner != NULL);
1393 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1394 LINVRNT(module_refcount(key->lct_owner) > 0);
1395 module_put(key->lct_owner);
1397 ctx->lc_value[index] = NULL;
1404 void lu_context_key_degister(struct lu_context_key *key)
1406 LASSERT(atomic_read(&key->lct_used) >= 1);
1407 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1409 lu_context_key_quiesce(key);
1411 write_lock(&lu_keys_guard);
1413 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1416 * Wait until all transient contexts referencing this key have
1417 * run lu_context_key::lct_fini() method.
1419 while (atomic_read(&key->lct_used) > 1) {
1420 write_unlock(&lu_keys_guard);
1421 CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
1422 key->lct_owner ? key->lct_owner->name : "", key,
1423 atomic_read(&key->lct_used));
1425 write_lock(&lu_keys_guard);
1427 if (lu_keys[key->lct_index]) {
1428 lu_keys[key->lct_index] = NULL;
1429 lu_ref_fini(&key->lct_reference);
1431 write_unlock(&lu_keys_guard);
1433 LASSERTF(atomic_read(&key->lct_used) == 1,
1434 "key has instances: %d\n",
1435 atomic_read(&key->lct_used));
1437 EXPORT_SYMBOL(lu_context_key_degister);
1440 * Register a number of keys. This has to be called after all keys have been
1441 * initialized by a call to LU_CONTEXT_KEY_INIT().
1443 int lu_context_key_register_many(struct lu_context_key *k, ...)
1445 struct lu_context_key *key = k;
1451 result = lu_context_key_register(key);
1454 key = va_arg(args, struct lu_context_key *);
1455 } while (key != NULL);
1461 lu_context_key_degister(k);
1462 k = va_arg(args, struct lu_context_key *);
1469 EXPORT_SYMBOL(lu_context_key_register_many);
1472 * De-register a number of keys. This is a dual to
1473 * lu_context_key_register_many().
1475 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1481 lu_context_key_degister(k);
1482 k = va_arg(args, struct lu_context_key*);
1483 } while (k != NULL);
1486 EXPORT_SYMBOL(lu_context_key_degister_many);
1489 * Revive a number of keys.
1491 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1497 lu_context_key_revive(k);
1498 k = va_arg(args, struct lu_context_key*);
1499 } while (k != NULL);
1502 EXPORT_SYMBOL(lu_context_key_revive_many);
1505 * Quiescent a number of keys.
1507 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1513 lu_context_key_quiesce(k);
1514 k = va_arg(args, struct lu_context_key*);
1515 } while (k != NULL);
1518 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1521 * Return value associated with key \a key in context \a ctx.
1523 void *lu_context_key_get(const struct lu_context *ctx,
1524 const struct lu_context_key *key)
1526 LINVRNT(ctx->lc_state == LCS_ENTERED);
1527 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1528 LASSERT(lu_keys[key->lct_index] == key);
1529 return ctx->lc_value[key->lct_index];
1531 EXPORT_SYMBOL(lu_context_key_get);
1534 * List of remembered contexts. XXX document me.
1536 static LIST_HEAD(lu_context_remembered);
1539 * Destroy \a key in all remembered contexts. This is used to destroy key
1540 * values in "shared" contexts (like service threads), when a module owning
1541 * the key is about to be unloaded.
1543 void lu_context_key_quiesce(struct lu_context_key *key)
1545 struct lu_context *ctx;
1547 if (!(key->lct_tags & LCT_QUIESCENT)) {
1549 * XXX memory barrier has to go here.
1551 write_lock(&lu_keys_guard);
1552 key->lct_tags |= LCT_QUIESCENT;
1555 * Wait until all lu_context_key::lct_init() methods
1558 while (atomic_read(&lu_key_initing_cnt) > 0) {
1559 write_unlock(&lu_keys_guard);
1560 CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
1562 key->lct_owner ? key->lct_owner->name : "",
1563 key, atomic_read(&key->lct_used),
1564 atomic_read(&lu_key_initing_cnt));
1566 write_lock(&lu_keys_guard);
1569 list_for_each_entry(ctx, &lu_context_remembered,
1571 key_fini(ctx, key->lct_index);
1574 write_unlock(&lu_keys_guard);
1578 void lu_context_key_revive(struct lu_context_key *key)
1580 write_lock(&lu_keys_guard);
1581 key->lct_tags &= ~LCT_QUIESCENT;
1583 write_unlock(&lu_keys_guard);
1586 static void keys_fini(struct lu_context *ctx)
1590 if (ctx->lc_value == NULL)
1593 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1596 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1597 ctx->lc_value = NULL;
1600 static int keys_fill(struct lu_context *ctx)
1603 unsigned pre_version;
1606 * A serialisation with lu_context_key_quiesce() is needed, but some
1607 * "key->lct_init()" are calling kernel memory allocation routine and
1608 * can't be called while holding a spin_lock.
1609 * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
1610 * to ensure the start of the serialisation.
1611 * An atomic_t variable is still used, in order not to reacquire the
1612 * lock when decrementing the counter.
1614 read_lock(&lu_keys_guard);
1615 atomic_inc(&lu_key_initing_cnt);
1616 pre_version = key_set_version;
1617 read_unlock(&lu_keys_guard);
1620 LINVRNT(ctx->lc_value != NULL);
1621 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1622 struct lu_context_key *key;
1625 if (ctx->lc_value[i] == NULL && key != NULL &&
1626 (key->lct_tags & ctx->lc_tags) &&
1628 * Don't create values for a LCT_QUIESCENT key, as this
1629 * will pin module owning a key.
1631 !(key->lct_tags & LCT_QUIESCENT)) {
1634 LINVRNT(key->lct_init != NULL);
1635 LINVRNT(key->lct_index == i);
1637 LASSERT(key->lct_owner != NULL);
1638 if (!(ctx->lc_tags & LCT_NOREF) &&
1639 try_module_get(key->lct_owner) == 0) {
1640 /* module is unloading, skip this key */
1644 value = key->lct_init(ctx, key);
1645 if (unlikely(IS_ERR(value))) {
1646 atomic_dec(&lu_key_initing_cnt);
1647 return PTR_ERR(value);
1650 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1651 atomic_inc(&key->lct_used);
1653 * This is the only place in the code, where an
1654 * element of ctx->lc_value[] array is set to non-NULL
1657 ctx->lc_value[i] = value;
1658 if (key->lct_exit != NULL)
1659 ctx->lc_tags |= LCT_HAS_EXIT;
1663 read_lock(&lu_keys_guard);
1664 if (pre_version != key_set_version) {
1665 pre_version = key_set_version;
1666 read_unlock(&lu_keys_guard);
1670 ctx->lc_version = key_set_version;
1672 atomic_dec(&lu_key_initing_cnt);
1673 read_unlock(&lu_keys_guard);
1677 static int keys_init(struct lu_context *ctx)
1679 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1680 if (likely(ctx->lc_value != NULL))
1681 return keys_fill(ctx);
1687 * Initialize context data-structure. Create values for all keys.
1689 int lu_context_init(struct lu_context *ctx, __u32 tags)
1693 memset(ctx, 0, sizeof *ctx);
1694 ctx->lc_state = LCS_INITIALIZED;
1695 ctx->lc_tags = tags;
1696 if (tags & LCT_REMEMBER) {
1697 write_lock(&lu_keys_guard);
1698 list_add(&ctx->lc_remember, &lu_context_remembered);
1699 write_unlock(&lu_keys_guard);
1701 INIT_LIST_HEAD(&ctx->lc_remember);
1704 rc = keys_init(ctx);
1706 lu_context_fini(ctx);
1710 EXPORT_SYMBOL(lu_context_init);
1713 * Finalize context data-structure. Destroy key values.
1715 void lu_context_fini(struct lu_context *ctx)
1717 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1718 ctx->lc_state = LCS_FINALIZED;
1720 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1721 LASSERT(list_empty(&ctx->lc_remember));
1724 } else { /* could race with key degister */
1725 write_lock(&lu_keys_guard);
1727 list_del_init(&ctx->lc_remember);
1728 write_unlock(&lu_keys_guard);
1731 EXPORT_SYMBOL(lu_context_fini);
1734 * Called before entering context.
1736 void lu_context_enter(struct lu_context *ctx)
1738 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1739 ctx->lc_state = LCS_ENTERED;
1741 EXPORT_SYMBOL(lu_context_enter);
1744 * Called after exiting from \a ctx
1746 void lu_context_exit(struct lu_context *ctx)
1750 LINVRNT(ctx->lc_state == LCS_ENTERED);
1751 ctx->lc_state = LCS_LEFT;
1752 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1753 /* could race with key quiescency */
1754 if (ctx->lc_tags & LCT_REMEMBER)
1755 read_lock(&lu_keys_guard);
1757 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1758 if (ctx->lc_value[i] != NULL) {
1759 struct lu_context_key *key;
1762 LASSERT(key != NULL);
1763 if (key->lct_exit != NULL)
1765 key, ctx->lc_value[i]);
1769 if (ctx->lc_tags & LCT_REMEMBER)
1770 read_unlock(&lu_keys_guard);
1773 EXPORT_SYMBOL(lu_context_exit);
1776 * Allocate for context all missing keys that were registered after context
1777 * creation. key_set_version is only changed in rare cases when modules
1778 * are loaded and removed.
1780 int lu_context_refill(struct lu_context *ctx)
1782 read_lock(&lu_keys_guard);
1783 if (likely(ctx->lc_version == key_set_version)) {
1784 read_unlock(&lu_keys_guard);
1788 read_unlock(&lu_keys_guard);
1789 return keys_fill(ctx);
1793 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1794 * obd being added. Currently, this is only used on client side, specifically
1795 * for echo device client, for other stack (like ptlrpc threads), context are
1796 * predefined when the lu_device type are registered, during the module probe
1799 __u32 lu_context_tags_default = 0;
1800 __u32 lu_session_tags_default = 0;
1802 void lu_context_tags_update(__u32 tags)
1804 write_lock(&lu_keys_guard);
1805 lu_context_tags_default |= tags;
1807 write_unlock(&lu_keys_guard);
1809 EXPORT_SYMBOL(lu_context_tags_update);
1811 void lu_context_tags_clear(__u32 tags)
1813 write_lock(&lu_keys_guard);
1814 lu_context_tags_default &= ~tags;
1816 write_unlock(&lu_keys_guard);
1818 EXPORT_SYMBOL(lu_context_tags_clear);
1820 void lu_session_tags_update(__u32 tags)
1822 write_lock(&lu_keys_guard);
1823 lu_session_tags_default |= tags;
1825 write_unlock(&lu_keys_guard);
1827 EXPORT_SYMBOL(lu_session_tags_update);
1829 void lu_session_tags_clear(__u32 tags)
1831 write_lock(&lu_keys_guard);
1832 lu_session_tags_default &= ~tags;
1834 write_unlock(&lu_keys_guard);
1836 EXPORT_SYMBOL(lu_session_tags_clear);
1838 int lu_env_init(struct lu_env *env, __u32 tags)
1843 result = lu_context_init(&env->le_ctx, tags);
1844 if (likely(result == 0))
1845 lu_context_enter(&env->le_ctx);
1848 EXPORT_SYMBOL(lu_env_init);
1850 void lu_env_fini(struct lu_env *env)
1852 lu_context_exit(&env->le_ctx);
1853 lu_context_fini(&env->le_ctx);
1856 EXPORT_SYMBOL(lu_env_fini);
1858 int lu_env_refill(struct lu_env *env)
1862 result = lu_context_refill(&env->le_ctx);
1863 if (result == 0 && env->le_ses != NULL)
1864 result = lu_context_refill(env->le_ses);
1867 EXPORT_SYMBOL(lu_env_refill);
1870 * Currently, this API will only be used by echo client.
1871 * Because echo client and normal lustre client will share
1872 * same cl_env cache. So echo client needs to refresh
1873 * the env context after it get one from the cache, especially
1874 * when normal client and echo client co-exist in the same client.
1876 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1881 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1882 env->le_ctx.lc_version = 0;
1883 env->le_ctx.lc_tags |= ctags;
1886 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1887 env->le_ses->lc_version = 0;
1888 env->le_ses->lc_tags |= stags;
1891 result = lu_env_refill(env);
1895 EXPORT_SYMBOL(lu_env_refill_by_tags);
1897 static struct shrinker *lu_site_shrinker;
1899 typedef struct lu_site_stats{
1900 unsigned lss_populated;
1901 unsigned lss_max_search;
1906 static void lu_site_stats_get(struct cfs_hash *hs,
1907 lu_site_stats_t *stats, int populated)
1909 struct cfs_hash_bd bd;
1912 cfs_hash_for_each_bucket(hs, &bd, i) {
1913 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1914 struct hlist_head *hhead;
1916 cfs_hash_bd_lock(hs, &bd, 1);
1918 cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
1919 stats->lss_total += cfs_hash_bd_count_get(&bd);
1920 stats->lss_max_search = max((int)stats->lss_max_search,
1921 cfs_hash_bd_depmax_get(&bd));
1923 cfs_hash_bd_unlock(hs, &bd, 1);
1927 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1928 if (!hlist_empty(hhead))
1929 stats->lss_populated++;
1931 cfs_hash_bd_unlock(hs, &bd, 1);
1937 * lu_cache_shrink_count() returns an approximate number of cached objects
1938 * that can be freed by shrink_slab(). A counter, which tracks the
1939 * number of items in the site's lru, is maintained in a percpu_counter
1940 * for each site. The percpu values are incremented and decremented as
1941 * objects are added or removed from the lru. The percpu values are summed
1942 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
1943 * summed value at any given time may not accurately reflect the current
1944 * lru length. But this value is sufficiently accurate for the needs of
1947 * Using a per cpu counter is a compromise solution to concurrent access:
1948 * lu_object_put() can update the counter without locking the site and
1949 * lu_cache_shrink_count can sum the counters without locking each
1950 * ls_obj_hash bucket.
1952 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1953 struct shrink_control *sc)
1956 struct lu_site *tmp;
1957 unsigned long cached = 0;
1959 if (!(sc->gfp_mask & __GFP_FS))
1962 down_read(&lu_sites_guard);
1963 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
1964 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
1965 up_read(&lu_sites_guard);
1967 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1968 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
1969 cached, sysctl_vfs_cache_pressure);
1974 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1975 struct shrink_control *sc)
1978 struct lu_site *tmp;
1979 unsigned long remain = sc->nr_to_scan;
1982 if (!(sc->gfp_mask & __GFP_FS))
1983 /* We must not take the lu_sites_guard lock when
1984 * __GFP_FS is *not* set because of the deadlock
1985 * possibility detailed above. Additionally,
1986 * since we cannot determine the number of
1987 * objects in the cache without taking this
1988 * lock, we're in a particularly tough spot. As
1989 * a result, we'll just lie and say our cache is
1990 * empty. This _should_ be ok, as we can't
1991 * reclaim objects when __GFP_FS is *not* set
1996 down_write(&lu_sites_guard);
1997 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1998 remain = lu_site_purge(&lu_shrink_env, s, remain);
2000 * Move just shrunk site to the tail of site list to
2001 * assure shrinking fairness.
2003 list_move_tail(&s->ls_linkage, &splice);
2005 list_splice(&splice, lu_sites.prev);
2006 up_write(&lu_sites_guard);
2008 return sc->nr_to_scan - remain;
2011 #ifndef HAVE_SHRINKER_COUNT
2013 * There exists a potential lock inversion deadlock scenario when using
2014 * Lustre on top of ZFS. This occurs between one of ZFS's
2015 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2016 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2017 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2018 * lock. Obviously neither thread will wake and drop their respective hold
2021 * To prevent this from happening we must ensure the lu_sites_guard lock is
2022 * not taken while down this code path. ZFS reliably does not set the
2023 * __GFP_FS bit in its code paths, so this can be used to determine if it
2024 * is safe to take the lu_sites_guard lock.
2026 * Ideally we should accurately return the remaining number of cached
2027 * objects without taking the lu_sites_guard lock, but this is not
2028 * possible in the current implementation.
2030 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2033 struct shrink_control scv = {
2034 .nr_to_scan = shrink_param(sc, nr_to_scan),
2035 .gfp_mask = shrink_param(sc, gfp_mask)
2037 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2038 struct shrinker* shrinker = NULL;
2042 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2044 if (scv.nr_to_scan != 0)
2045 lu_cache_shrink_scan(shrinker, &scv);
2047 cached = lu_cache_shrink_count(shrinker, &scv);
2051 #endif /* HAVE_SHRINKER_COUNT */
2059 * Environment to be used in debugger, contains all tags.
2061 static struct lu_env lu_debugging_env;
2064 * Debugging printer function using printk().
2066 int lu_printk_printer(const struct lu_env *env,
2067 void *unused, const char *format, ...)
2071 va_start(args, format);
2072 vprintk(format, args);
2077 int lu_debugging_setup(void)
2079 return lu_env_init(&lu_debugging_env, ~0);
2082 void lu_context_keys_dump(void)
2086 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2087 struct lu_context_key *key;
2091 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2092 i, key, key->lct_tags,
2093 key->lct_init, key->lct_fini, key->lct_exit,
2094 key->lct_index, atomic_read(&key->lct_used),
2095 key->lct_owner ? key->lct_owner->name : "",
2097 lu_ref_print(&key->lct_reference);
2103 * Initialization of global lu_* data.
2105 int lu_global_init(void)
2108 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2109 lu_cache_shrink_count, lu_cache_shrink_scan);
2111 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2113 result = lu_ref_global_init();
2117 LU_CONTEXT_KEY_INIT(&lu_global_key);
2118 result = lu_context_key_register(&lu_global_key);
2123 * At this level, we don't know what tags are needed, so allocate them
2124 * conservatively. This should not be too bad, because this
2125 * environment is global.
2127 down_write(&lu_sites_guard);
2128 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2129 up_write(&lu_sites_guard);
2134 * seeks estimation: 3 seeks to read a record from oi, one to read
2135 * inode, one for ea. Unfortunately setting this high value results in
2136 * lu_object/inode cache consuming all the memory.
2138 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2139 if (lu_site_shrinker == NULL)
2146 * Dual to lu_global_init().
2148 void lu_global_fini(void)
2150 if (lu_site_shrinker != NULL) {
2151 remove_shrinker(lu_site_shrinker);
2152 lu_site_shrinker = NULL;
2155 lu_context_key_degister(&lu_global_key);
2158 * Tear shrinker environment down _after_ de-registering
2159 * lu_global_key, because the latter has a value in the former.
2161 down_write(&lu_sites_guard);
2162 lu_env_fini(&lu_shrink_env);
2163 up_write(&lu_sites_guard);
2165 lu_ref_global_fini();
2168 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2170 #ifdef CONFIG_PROC_FS
2171 struct lprocfs_counter ret;
2173 lprocfs_stats_collect(stats, idx, &ret);
2174 return (__u32)ret.lc_count;
2181 * Output site statistical counters into a buffer. Suitable for
2182 * lprocfs_rd_*()-style functions.
2184 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2186 lu_site_stats_t stats;
2188 memset(&stats, 0, sizeof(stats));
2189 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2191 seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2194 stats.lss_populated,
2195 CFS_HASH_NHLIST(s->ls_obj_hash),
2196 stats.lss_max_search,
2197 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2198 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2199 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2200 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2201 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2202 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2205 EXPORT_SYMBOL(lu_site_stats_seq_print);
2208 * Helper function to initialize a number of kmem slab caches at once.
2210 int lu_kmem_init(struct lu_kmem_descr *caches)
2213 struct lu_kmem_descr *iter = caches;
2215 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2216 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2219 if (*iter->ckd_cache == NULL) {
2221 /* free all previously allocated caches */
2222 lu_kmem_fini(caches);
2228 EXPORT_SYMBOL(lu_kmem_init);
2231 * Helper function to finalize a number of kmem slab cached at once. Dual to
2234 void lu_kmem_fini(struct lu_kmem_descr *caches)
2236 for (; caches->ckd_cache != NULL; ++caches) {
2237 if (*caches->ckd_cache != NULL) {
2238 kmem_cache_destroy(*caches->ckd_cache);
2239 *caches->ckd_cache = NULL;
2243 EXPORT_SYMBOL(lu_kmem_fini);
2246 * Temporary solution to be able to assign fid in ->do_create()
2247 * till we have fully-functional OST fids
2249 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2250 const struct lu_fid *fid)
2252 struct lu_site *s = o->lo_dev->ld_site;
2253 struct lu_fid *old = &o->lo_header->loh_fid;
2254 struct cfs_hash *hs;
2255 struct cfs_hash_bd bd;
2257 LASSERT(fid_is_zero(old));
2259 /* supposed to be unique */
2260 hs = s->ls_obj_hash;
2261 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2262 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2265 struct lu_object *shadow;
2267 shadow = htable_lookup(s, &bd, fid, &version);
2268 /* supposed to be unique */
2269 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2273 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2274 cfs_hash_bd_unlock(hs, &bd, 1);
2276 EXPORT_SYMBOL(lu_object_assign_fid);
2279 * allocates object with 0 (non-assiged) fid
2280 * XXX: temporary solution to be able to assign fid in ->do_create()
2281 * till we have fully-functional OST fids
2283 struct lu_object *lu_object_anon(const struct lu_env *env,
2284 struct lu_device *dev,
2285 const struct lu_object_conf *conf)
2288 struct lu_object *o;
2291 o = lu_object_alloc(env, dev, &fid, conf);
2295 EXPORT_SYMBOL(lu_object_anon);
2297 struct lu_buf LU_BUF_NULL = {
2301 EXPORT_SYMBOL(LU_BUF_NULL);
2303 void lu_buf_free(struct lu_buf *buf)
2307 LASSERT(buf->lb_len > 0);
2308 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2313 EXPORT_SYMBOL(lu_buf_free);
2315 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2318 LASSERT(buf->lb_buf == NULL);
2319 LASSERT(buf->lb_len == 0);
2320 OBD_ALLOC_LARGE(buf->lb_buf, size);
2321 if (likely(buf->lb_buf))
2324 EXPORT_SYMBOL(lu_buf_alloc);
2326 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2329 lu_buf_alloc(buf, size);
2331 EXPORT_SYMBOL(lu_buf_realloc);
2333 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2335 if (buf->lb_buf == NULL && buf->lb_len == 0)
2336 lu_buf_alloc(buf, len);
2338 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2339 lu_buf_realloc(buf, len);
2343 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2346 * Increase the size of the \a buf.
2347 * preserves old data in buffer
2348 * old buffer remains unchanged on error
2349 * \retval 0 or -ENOMEM
2351 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2355 if (len <= buf->lb_len)
2358 OBD_ALLOC_LARGE(ptr, len);
2362 /* Free the old buf */
2363 if (buf->lb_buf != NULL) {
2364 memcpy(ptr, buf->lb_buf, buf->lb_len);
2365 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2372 EXPORT_SYMBOL(lu_buf_check_and_grow);