4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/lu_object.c
39 * These are the only exported functions, they provide some generic
40 * infrastructure for managing object devices
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
45 #define DEBUG_SUBSYSTEM S_CLASS
47 #include <libcfs/libcfs.h>
48 #include <linux/module.h>
49 #include <libcfs/libcfs_hash.h> /* hash_long() */
50 #include <obd_class.h>
51 #include <obd_support.h>
52 #include <lustre_disk.h>
53 #include <lustre_fid.h>
54 #include <lu_object.h>
56 #include <libcfs/list.h>
59 LU_CACHE_PERCENT_MAX = 50,
60 LU_CACHE_PERCENT_DEFAULT = 20
63 #define LU_CACHE_NR_MAX_ADJUST 128
64 #define LU_CACHE_NR_UNLIMITED -1
65 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
66 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
67 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
68 #define LU_CACHE_NR_ZFS_LIMIT 10240
70 #define LU_SITE_BITS_MIN 12
71 #define LU_SITE_BITS_MAX 24
73 * total 256 buckets, we don't want too many buckets because:
74 * - consume too much memory
75 * - avoid unbalanced LRU list
77 #define LU_SITE_BKT_BITS 8
80 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
81 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
82 "Percentage of memory to be used as lu_object cache");
84 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
85 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
86 "Maximum number of objects in lu_object cache");
88 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
89 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
92 * Decrease reference counter on object. If last reference is freed, return
93 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
94 * case, free object immediately.
96 void lu_object_put(const struct lu_env *env, struct lu_object *o)
98 struct lu_site_bkt_data *bkt;
99 struct lu_object_header *top;
100 struct lu_site *site;
101 struct lu_object *orig;
102 struct cfs_hash_bd bd;
103 const struct lu_fid *fid;
106 site = o->lo_dev->ld_site;
110 * till we have full fids-on-OST implemented anonymous objects
111 * are possible in OSP. such an object isn't listed in the site
112 * so we should not remove it from the site.
114 fid = lu_object_fid(o);
115 if (fid_is_zero(fid)) {
116 LASSERT(top->loh_hash.next == NULL
117 && top->loh_hash.pprev == NULL);
118 LASSERT(list_empty(&top->loh_lru));
119 if (!atomic_dec_and_test(&top->loh_ref))
121 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
122 if (o->lo_ops->loo_object_release != NULL)
123 o->lo_ops->loo_object_release(env, o);
125 lu_object_free(env, orig);
129 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
130 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
132 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
133 if (lu_object_is_dying(top)) {
136 * somebody may be waiting for this, currently only
137 * used for cl_object, see cl_object_put_last().
139 wake_up_all(&bkt->lsb_marche_funebre);
145 * When last reference is released, iterate over object
146 * layers, and notify them that object is no longer busy.
148 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
149 if (o->lo_ops->loo_object_release != NULL)
150 o->lo_ops->loo_object_release(env, o);
153 if (!lu_object_is_dying(top) &&
154 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
155 LASSERT(list_empty(&top->loh_lru));
156 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
158 lprocfs_counter_incr(site->ls_stats, LU_SS_LRU_LEN);
159 CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, "
161 o, site->ls_obj_hash, bkt, bkt->lsb_lru_len);
162 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
167 * If object is dying (will not be cached) then remove it
168 * from hash table and LRU.
170 * This is done with hash table and LRU lists locked. As the only
171 * way to acquire first reference to previously unreferenced
172 * object is through hash-table lookup (lu_object_find()),
173 * or LRU scanning (lu_site_purge()), that are done under hash-table
174 * and LRU lock, no race with concurrent object lookup is possible
175 * and we can safely destroy object below.
177 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
178 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
179 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
181 * Object was already removed from hash and lru above, can
184 lu_object_free(env, orig);
186 EXPORT_SYMBOL(lu_object_put);
189 * Put object and don't keep in cache. This is temporary solution for
190 * multi-site objects when its layering is not constant.
192 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
194 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
195 return lu_object_put(env, o);
197 EXPORT_SYMBOL(lu_object_put_nocache);
200 * Kill the object and take it out of LRU cache.
201 * Currently used by client code for layout change.
203 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
205 struct lu_object_header *top;
208 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
209 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
210 struct lu_site *site = o->lo_dev->ld_site;
211 struct cfs_hash *obj_hash = site->ls_obj_hash;
212 struct cfs_hash_bd bd;
214 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
215 if (!list_empty(&top->loh_lru)) {
216 struct lu_site_bkt_data *bkt;
218 list_del_init(&top->loh_lru);
219 bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
221 lprocfs_counter_decr(site->ls_stats, LU_SS_LRU_LEN);
223 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
224 cfs_hash_bd_unlock(obj_hash, &bd, 1);
227 EXPORT_SYMBOL(lu_object_unhash);
230 * Allocate new object.
232 * This follows object creation protocol, described in the comment within
233 * struct lu_device_operations definition.
235 static struct lu_object *lu_object_alloc(const struct lu_env *env,
236 struct lu_device *dev,
237 const struct lu_fid *f,
238 const struct lu_object_conf *conf)
240 struct lu_object *scan;
241 struct lu_object *top;
242 struct list_head *layers;
243 unsigned int init_mask = 0;
244 unsigned int init_flag;
250 * Create top-level object slice. This will also create
253 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
255 RETURN(ERR_PTR(-ENOMEM));
259 * This is the only place where object fid is assigned. It's constant
262 top->lo_header->loh_fid = *f;
263 layers = &top->lo_header->loh_layers;
267 * Call ->loo_object_init() repeatedly, until no more new
268 * object slices are created.
272 list_for_each_entry(scan, layers, lo_linkage) {
273 if (init_mask & init_flag)
276 scan->lo_header = top->lo_header;
277 result = scan->lo_ops->loo_object_init(env, scan, conf);
279 lu_object_free(env, top);
280 RETURN(ERR_PTR(result));
282 init_mask |= init_flag;
288 list_for_each_entry_reverse(scan, layers, lo_linkage) {
289 if (scan->lo_ops->loo_object_start != NULL) {
290 result = scan->lo_ops->loo_object_start(env, scan);
292 lu_object_free(env, top);
293 RETURN(ERR_PTR(result));
298 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
305 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
307 struct lu_site_bkt_data *bkt;
308 struct lu_site *site;
309 struct lu_object *scan;
310 struct list_head *layers;
311 struct list_head splice;
313 site = o->lo_dev->ld_site;
314 layers = &o->lo_header->loh_layers;
315 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
317 * First call ->loo_object_delete() method to release all resources.
319 list_for_each_entry_reverse(scan, layers, lo_linkage) {
320 if (scan->lo_ops->loo_object_delete != NULL)
321 scan->lo_ops->loo_object_delete(env, scan);
325 * Then, splice object layers into stand-alone list, and call
326 * ->loo_object_free() on all layers to free memory. Splice is
327 * necessary, because lu_object_header is freed together with the
330 INIT_LIST_HEAD(&splice);
331 list_splice_init(layers, &splice);
332 while (!list_empty(&splice)) {
334 * Free layers in bottom-to-top order, so that object header
335 * lives as long as possible and ->loo_object_free() methods
336 * can look at its contents.
338 o = container_of0(splice.prev, struct lu_object, lo_linkage);
339 list_del_init(&o->lo_linkage);
340 LASSERT(o->lo_ops->loo_object_free != NULL);
341 o->lo_ops->loo_object_free(env, o);
344 if (waitqueue_active(&bkt->lsb_marche_funebre))
345 wake_up_all(&bkt->lsb_marche_funebre);
349 * Free \a nr objects from the cold end of the site LRU list.
351 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
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.
381 mutex_lock(&s->ls_purge_mutex);
383 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
387 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
388 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
390 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
391 LASSERT(atomic_read(&h->loh_ref) == 0);
393 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
394 LASSERT(bd.bd_bucket == bd2.bd_bucket);
396 cfs_hash_bd_del_locked(s->ls_obj_hash,
398 list_move(&h->loh_lru, &dispose);
400 lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
404 if (nr != ~0 && --nr == 0)
407 if (count > 0 && --count == 0)
411 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
414 * Free everything on the dispose list. This is safe against
415 * races due to the reasons described in lu_object_put().
417 while (!list_empty(&dispose)) {
418 h = container_of0(dispose.next,
419 struct lu_object_header, loh_lru);
420 list_del_init(&h->loh_lru);
421 lu_object_free(env, lu_object_top(h));
422 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
428 mutex_unlock(&s->ls_purge_mutex);
430 if (nr != 0 && did_sth && start != 0) {
431 start = 0; /* restart from the first bucket */
434 /* race on s->ls_purge_start, but nobody cares */
435 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
439 EXPORT_SYMBOL(lu_site_purge);
444 * Code below has to jump through certain loops to output object description
445 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
446 * composes object description from strings that are parts of _lines_ of
447 * output (i.e., strings that are not terminated by newline). This doesn't fit
448 * very well into libcfs_debug_msg() interface that assumes that each message
449 * supplied to it is a self-contained output line.
451 * To work around this, strings are collected in a temporary buffer
452 * (implemented as a value of lu_cdebug_key key), until terminating newline
453 * character is detected.
461 * XXX overflow is not handled correctly.
466 struct lu_cdebug_data {
470 char lck_area[LU_CDEBUG_LINE];
473 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
474 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
477 * Key, holding temporary buffer. This key is registered very early by
480 static struct lu_context_key lu_global_key = {
481 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
482 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
483 .lct_init = lu_global_key_init,
484 .lct_fini = lu_global_key_fini
488 * Printer function emitting messages through libcfs_debug_msg().
490 int lu_cdebug_printer(const struct lu_env *env,
491 void *cookie, const char *format, ...)
493 struct libcfs_debug_msg_data *msgdata = cookie;
494 struct lu_cdebug_data *key;
499 va_start(args, format);
501 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
502 LASSERT(key != NULL);
504 used = strlen(key->lck_area);
505 complete = format[strlen(format) - 1] == '\n';
507 * Append new chunk to the buffer.
509 vsnprintf(key->lck_area + used,
510 ARRAY_SIZE(key->lck_area) - used, format, args);
512 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
513 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
514 key->lck_area[0] = 0;
519 EXPORT_SYMBOL(lu_cdebug_printer);
522 * Print object header.
524 void lu_object_header_print(const struct lu_env *env, void *cookie,
525 lu_printer_t printer,
526 const struct lu_object_header *hdr)
528 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
529 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
531 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
532 list_empty((struct list_head *)&hdr->loh_lru) ? \
534 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
536 EXPORT_SYMBOL(lu_object_header_print);
539 * Print human readable representation of the \a o to the \a printer.
541 void lu_object_print(const struct lu_env *env, void *cookie,
542 lu_printer_t printer, const struct lu_object *o)
544 static const char ruler[] = "........................................";
545 struct lu_object_header *top;
549 lu_object_header_print(env, cookie, printer, top);
550 (*printer)(env, cookie, "{\n");
552 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
554 * print `.' \a depth times followed by type name and address
556 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
557 o->lo_dev->ld_type->ldt_name, o);
559 if (o->lo_ops->loo_object_print != NULL)
560 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
562 (*printer)(env, cookie, "\n");
565 (*printer)(env, cookie, "} header@%p\n", top);
567 EXPORT_SYMBOL(lu_object_print);
570 * Check object consistency.
572 int lu_object_invariant(const struct lu_object *o)
574 struct lu_object_header *top;
577 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
578 if (o->lo_ops->loo_object_invariant != NULL &&
579 !o->lo_ops->loo_object_invariant(o))
585 static struct lu_object *htable_lookup(struct lu_site *s,
586 struct cfs_hash_bd *bd,
587 const struct lu_fid *f,
588 wait_queue_t *waiter,
591 struct lu_site_bkt_data *bkt;
592 struct lu_object_header *h;
593 struct hlist_node *hnode;
594 __u64 ver = cfs_hash_bd_version_get(bd);
597 return ERR_PTR(-ENOENT);
600 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
601 /* cfs_hash_bd_peek_locked is a somehow "internal" function
602 * of cfs_hash, it doesn't add refcount on object. */
603 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
605 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
606 return ERR_PTR(-ENOENT);
609 h = container_of0(hnode, struct lu_object_header, loh_hash);
610 if (likely(!lu_object_is_dying(h))) {
611 cfs_hash_get(s->ls_obj_hash, hnode);
612 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
613 if (!list_empty(&h->loh_lru)) {
614 list_del_init(&h->loh_lru);
616 lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
618 return lu_object_top(h);
622 * Lookup found an object being destroyed this object cannot be
623 * returned (to assure that references to dying objects are eventually
624 * drained), and moreover, lookup has to wait until object is freed.
627 if (likely(waiter != NULL)) {
628 init_waitqueue_entry(waiter, current);
629 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
630 set_current_state(TASK_UNINTERRUPTIBLE);
631 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
634 return ERR_PTR(-EAGAIN);
638 * Search cache for an object with the fid \a f. If such object is found,
639 * return it. Otherwise, create new object, insert it into cache and return
640 * it. In any case, additional reference is acquired on the returned object.
642 struct lu_object *lu_object_find(const struct lu_env *env,
643 struct lu_device *dev, const struct lu_fid *f,
644 const struct lu_object_conf *conf)
646 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
648 EXPORT_SYMBOL(lu_object_find);
651 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
652 * the calculation for the number of objects to reclaim is not covered by
653 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
654 * This ensures that many concurrent threads will not accidentally purge
657 static void lu_object_limit(const struct lu_env *env,
658 struct lu_device *dev)
662 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
665 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
666 nr = (__u64)lu_cache_nr;
668 lu_site_purge(env, dev->ld_site,
669 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
674 static struct lu_object *lu_object_new(const struct lu_env *env,
675 struct lu_device *dev,
676 const struct lu_fid *f,
677 const struct lu_object_conf *conf)
681 struct cfs_hash_bd bd;
683 o = lu_object_alloc(env, dev, f, conf);
684 if (unlikely(IS_ERR(o)))
687 hs = dev->ld_site->ls_obj_hash;
688 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
689 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
690 cfs_hash_bd_unlock(hs, &bd, 1);
692 lu_object_limit(env, dev);
698 * Core logic of lu_object_find*() functions.
700 static struct lu_object *lu_object_find_try(const struct lu_env *env,
701 struct lu_device *dev,
702 const struct lu_fid *f,
703 const struct lu_object_conf *conf,
704 wait_queue_t *waiter)
707 struct lu_object *shadow;
710 struct cfs_hash_bd bd;
714 * This uses standard index maintenance protocol:
716 * - search index under lock, and return object if found;
717 * - otherwise, unlock index, allocate new object;
718 * - lock index and search again;
719 * - if nothing is found (usual case), insert newly created
721 * - otherwise (race: other thread inserted object), free
722 * object just allocated.
726 * For "LOC_F_NEW" case, we are sure the object is new established.
727 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
728 * just alloc and insert directly.
730 * If dying object is found during index search, add @waiter to the
731 * site wait-queue and return ERR_PTR(-EAGAIN).
733 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
734 return lu_object_new(env, dev, f, conf);
738 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
739 o = htable_lookup(s, &bd, f, waiter, &version);
740 cfs_hash_bd_unlock(hs, &bd, 1);
741 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
745 * Allocate new object. This may result in rather complicated
746 * operations, including fld queries, inode loading, etc.
748 o = lu_object_alloc(env, dev, f, conf);
749 if (unlikely(IS_ERR(o)))
752 LASSERT(lu_fid_eq(lu_object_fid(o), f));
754 cfs_hash_bd_lock(hs, &bd, 1);
756 shadow = htable_lookup(s, &bd, f, waiter, &version);
757 if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
758 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
759 cfs_hash_bd_unlock(hs, &bd, 1);
761 lu_object_limit(env, dev);
766 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
767 cfs_hash_bd_unlock(hs, &bd, 1);
768 lu_object_free(env, o);
773 * Much like lu_object_find(), but top level device of object is specifically
774 * \a dev rather than top level device of the site. This interface allows
775 * objects of different "stacking" to be created within the same site.
777 struct lu_object *lu_object_find_at(const struct lu_env *env,
778 struct lu_device *dev,
779 const struct lu_fid *f,
780 const struct lu_object_conf *conf)
782 struct lu_site_bkt_data *bkt;
783 struct lu_object *obj;
786 if (conf != NULL && conf->loc_flags & LOC_F_NOWAIT)
787 return lu_object_find_try(env, dev, f, conf, NULL);
790 obj = lu_object_find_try(env, dev, f, conf, &wait);
791 if (obj != ERR_PTR(-EAGAIN))
794 * lu_object_find_try() already added waiter into the
798 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
799 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
802 EXPORT_SYMBOL(lu_object_find_at);
805 * Find object with given fid, and return its slice belonging to given device.
807 struct lu_object *lu_object_find_slice(const struct lu_env *env,
808 struct lu_device *dev,
809 const struct lu_fid *f,
810 const struct lu_object_conf *conf)
812 struct lu_object *top;
813 struct lu_object *obj;
815 top = lu_object_find(env, dev, f, conf);
819 obj = lu_object_locate(top->lo_header, dev->ld_type);
820 if (unlikely(obj == NULL)) {
821 lu_object_put(env, top);
822 obj = ERR_PTR(-ENOENT);
827 EXPORT_SYMBOL(lu_object_find_slice);
830 * Global list of all device types.
832 static struct list_head lu_device_types;
834 int lu_device_type_init(struct lu_device_type *ldt)
838 atomic_set(&ldt->ldt_device_nr, 0);
839 INIT_LIST_HEAD(&ldt->ldt_linkage);
840 if (ldt->ldt_ops->ldto_init)
841 result = ldt->ldt_ops->ldto_init(ldt);
844 spin_lock(&obd_types_lock);
845 list_add(&ldt->ldt_linkage, &lu_device_types);
846 spin_unlock(&obd_types_lock);
851 EXPORT_SYMBOL(lu_device_type_init);
853 void lu_device_type_fini(struct lu_device_type *ldt)
855 spin_lock(&obd_types_lock);
856 list_del_init(&ldt->ldt_linkage);
857 spin_unlock(&obd_types_lock);
858 if (ldt->ldt_ops->ldto_fini)
859 ldt->ldt_ops->ldto_fini(ldt);
861 EXPORT_SYMBOL(lu_device_type_fini);
864 * Global list of all sites on this node
866 static struct list_head lu_sites;
867 static DEFINE_MUTEX(lu_sites_guard);
870 * Global environment used by site shrinker.
872 static struct lu_env lu_shrink_env;
874 struct lu_site_print_arg {
875 struct lu_env *lsp_env;
877 lu_printer_t lsp_printer;
881 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
882 struct hlist_node *hnode, void *data)
884 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
885 struct lu_object_header *h;
887 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
888 if (!list_empty(&h->loh_layers)) {
889 const struct lu_object *o;
891 o = lu_object_top(h);
892 lu_object_print(arg->lsp_env, arg->lsp_cookie,
893 arg->lsp_printer, o);
895 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
896 arg->lsp_printer, h);
902 * Print all objects in \a s.
904 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
905 lu_printer_t printer)
907 struct lu_site_print_arg arg = {
908 .lsp_env = (struct lu_env *)env,
909 .lsp_cookie = cookie,
910 .lsp_printer = printer,
913 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
915 EXPORT_SYMBOL(lu_site_print);
918 * Return desired hash table order.
920 static unsigned long lu_htable_order(struct lu_device *top)
922 unsigned long cache_size;
926 * For ZFS based OSDs the cache should be disabled by default. This
927 * allows the ZFS ARC maximum flexibility in determining what buffers
928 * to cache. If Lustre has objects or buffer which it wants to ensure
929 * always stay cached it must maintain a hold on them.
931 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
932 lu_cache_percent = 1;
933 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
934 return LU_SITE_BITS_MIN;
938 * Calculate hash table size, assuming that we want reasonable
939 * performance when 20% of total memory is occupied by cache of
942 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
944 cache_size = totalram_pages;
946 #if BITS_PER_LONG == 32
947 /* limit hashtable size for lowmem systems to low RAM */
948 if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
949 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
952 /* clear off unreasonable cache setting. */
953 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
954 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
955 " the range of (0, %u]. Will use default value: %u.\n",
956 lu_cache_percent, LU_CACHE_PERCENT_MAX,
957 LU_CACHE_PERCENT_DEFAULT);
959 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
961 cache_size = cache_size / 100 * lu_cache_percent *
962 (PAGE_CACHE_SIZE / 1024);
964 for (bits = 1; (1 << bits) < cache_size; ++bits) {
970 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
971 const void *key, unsigned mask)
973 struct lu_fid *fid = (struct lu_fid *)key;
976 hash = fid_flatten32(fid);
977 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
978 hash = hash_long(hash, hs->hs_bkt_bits);
980 /* give me another random factor */
981 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
983 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
984 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
989 static void *lu_obj_hop_object(struct hlist_node *hnode)
991 return hlist_entry(hnode, struct lu_object_header, loh_hash);
994 static void *lu_obj_hop_key(struct hlist_node *hnode)
996 struct lu_object_header *h;
998 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1002 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1004 struct lu_object_header *h;
1006 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1007 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1010 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
1012 struct lu_object_header *h;
1014 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1015 atomic_inc(&h->loh_ref);
1018 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
1020 LBUG(); /* we should never called it */
1023 static struct cfs_hash_ops lu_site_hash_ops = {
1024 .hs_hash = lu_obj_hop_hash,
1025 .hs_key = lu_obj_hop_key,
1026 .hs_keycmp = lu_obj_hop_keycmp,
1027 .hs_object = lu_obj_hop_object,
1028 .hs_get = lu_obj_hop_get,
1029 .hs_put_locked = lu_obj_hop_put_locked,
1032 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1034 spin_lock(&s->ls_ld_lock);
1035 if (list_empty(&d->ld_linkage))
1036 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1037 spin_unlock(&s->ls_ld_lock);
1039 EXPORT_SYMBOL(lu_dev_add_linkage);
1041 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1043 spin_lock(&s->ls_ld_lock);
1044 list_del_init(&d->ld_linkage);
1045 spin_unlock(&s->ls_ld_lock);
1047 EXPORT_SYMBOL(lu_dev_del_linkage);
1050 * Initialize site \a s, with \a d as the top level device.
1052 int lu_site_init(struct lu_site *s, struct lu_device *top)
1054 struct lu_site_bkt_data *bkt;
1055 struct cfs_hash_bd bd;
1061 memset(s, 0, sizeof *s);
1062 mutex_init(&s->ls_purge_mutex);
1063 bits = lu_htable_order(top);
1064 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1065 for (bits = clamp_t(typeof(bits), bits,
1066 LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
1067 bits >= LU_SITE_BITS_MIN; bits--) {
1068 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1069 bits - LU_SITE_BKT_BITS,
1072 CFS_HASH_SPIN_BKTLOCK |
1073 CFS_HASH_NO_ITEMREF |
1075 CFS_HASH_ASSERT_EMPTY |
1077 if (s->ls_obj_hash != NULL)
1081 if (s->ls_obj_hash == NULL) {
1082 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1086 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1087 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1088 INIT_LIST_HEAD(&bkt->lsb_lru);
1089 init_waitqueue_head(&bkt->lsb_marche_funebre);
1092 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1093 if (s->ls_stats == NULL) {
1094 cfs_hash_putref(s->ls_obj_hash);
1095 s->ls_obj_hash = NULL;
1099 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1100 0, "created", "created");
1101 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1102 0, "cache_hit", "cache_hit");
1103 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1104 0, "cache_miss", "cache_miss");
1105 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1106 0, "cache_race", "cache_race");
1107 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1108 0, "cache_death_race", "cache_death_race");
1109 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1110 0, "lru_purged", "lru_purged");
1112 * Unlike other counters, lru_len can be decremented so
1113 * need lc_sum instead of just lc_count
1115 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_LEN,
1116 LPROCFS_CNTR_AVGMINMAX, "lru_len", "lru_len");
1118 INIT_LIST_HEAD(&s->ls_linkage);
1119 s->ls_top_dev = top;
1122 lu_ref_add(&top->ld_reference, "site-top", s);
1124 INIT_LIST_HEAD(&s->ls_ld_linkage);
1125 spin_lock_init(&s->ls_ld_lock);
1127 lu_dev_add_linkage(s, top);
1131 EXPORT_SYMBOL(lu_site_init);
1134 * Finalize \a s and release its resources.
1136 void lu_site_fini(struct lu_site *s)
1138 mutex_lock(&lu_sites_guard);
1139 list_del_init(&s->ls_linkage);
1140 mutex_unlock(&lu_sites_guard);
1142 if (s->ls_obj_hash != NULL) {
1143 cfs_hash_putref(s->ls_obj_hash);
1144 s->ls_obj_hash = NULL;
1147 if (s->ls_top_dev != NULL) {
1148 s->ls_top_dev->ld_site = NULL;
1149 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1150 lu_device_put(s->ls_top_dev);
1151 s->ls_top_dev = NULL;
1154 if (s->ls_stats != NULL)
1155 lprocfs_free_stats(&s->ls_stats);
1157 EXPORT_SYMBOL(lu_site_fini);
1160 * Called when initialization of stack for this site is completed.
1162 int lu_site_init_finish(struct lu_site *s)
1165 mutex_lock(&lu_sites_guard);
1166 result = lu_context_refill(&lu_shrink_env.le_ctx);
1168 list_add(&s->ls_linkage, &lu_sites);
1169 mutex_unlock(&lu_sites_guard);
1172 EXPORT_SYMBOL(lu_site_init_finish);
1175 * Acquire additional reference on device \a d
1177 void lu_device_get(struct lu_device *d)
1179 atomic_inc(&d->ld_ref);
1181 EXPORT_SYMBOL(lu_device_get);
1184 * Release reference on device \a d.
1186 void lu_device_put(struct lu_device *d)
1188 LASSERT(atomic_read(&d->ld_ref) > 0);
1189 atomic_dec(&d->ld_ref);
1191 EXPORT_SYMBOL(lu_device_put);
1194 * Initialize device \a d of type \a t.
1196 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1198 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1199 t->ldt_ops->ldto_start != NULL)
1200 t->ldt_ops->ldto_start(t);
1202 memset(d, 0, sizeof *d);
1204 lu_ref_init(&d->ld_reference);
1205 INIT_LIST_HEAD(&d->ld_linkage);
1209 EXPORT_SYMBOL(lu_device_init);
1212 * Finalize device \a d.
1214 void lu_device_fini(struct lu_device *d)
1216 struct lu_device_type *t = d->ld_type;
1218 if (d->ld_obd != NULL) {
1219 d->ld_obd->obd_lu_dev = NULL;
1223 lu_ref_fini(&d->ld_reference);
1224 LASSERTF(atomic_read(&d->ld_ref) == 0,
1225 "Refcount is %u\n", atomic_read(&d->ld_ref));
1226 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1228 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1229 t->ldt_ops->ldto_stop != NULL)
1230 t->ldt_ops->ldto_stop(t);
1232 EXPORT_SYMBOL(lu_device_fini);
1235 * Initialize object \a o that is part of compound object \a h and was created
1238 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1239 struct lu_device *d)
1241 memset(o, 0, sizeof(*o));
1245 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1246 INIT_LIST_HEAD(&o->lo_linkage);
1250 EXPORT_SYMBOL(lu_object_init);
1253 * Finalize object and release its resources.
1255 void lu_object_fini(struct lu_object *o)
1257 struct lu_device *dev = o->lo_dev;
1259 LASSERT(list_empty(&o->lo_linkage));
1262 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1268 EXPORT_SYMBOL(lu_object_fini);
1271 * Add object \a o as first layer of compound object \a h
1273 * This is typically called by the ->ldo_object_alloc() method of top-level
1276 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1278 list_move(&o->lo_linkage, &h->loh_layers);
1280 EXPORT_SYMBOL(lu_object_add_top);
1283 * Add object \a o as a layer of compound object, going after \a before.
1285 * This is typically called by the ->ldo_object_alloc() method of \a
1288 void lu_object_add(struct lu_object *before, struct lu_object *o)
1290 list_move(&o->lo_linkage, &before->lo_linkage);
1292 EXPORT_SYMBOL(lu_object_add);
1295 * Initialize compound object.
1297 int lu_object_header_init(struct lu_object_header *h)
1299 memset(h, 0, sizeof *h);
1300 atomic_set(&h->loh_ref, 1);
1301 INIT_HLIST_NODE(&h->loh_hash);
1302 INIT_LIST_HEAD(&h->loh_lru);
1303 INIT_LIST_HEAD(&h->loh_layers);
1304 lu_ref_init(&h->loh_reference);
1307 EXPORT_SYMBOL(lu_object_header_init);
1310 * Finalize compound object.
1312 void lu_object_header_fini(struct lu_object_header *h)
1314 LASSERT(list_empty(&h->loh_layers));
1315 LASSERT(list_empty(&h->loh_lru));
1316 LASSERT(hlist_unhashed(&h->loh_hash));
1317 lu_ref_fini(&h->loh_reference);
1319 EXPORT_SYMBOL(lu_object_header_fini);
1322 * Given a compound object, find its slice, corresponding to the device type
1325 struct lu_object *lu_object_locate(struct lu_object_header *h,
1326 const struct lu_device_type *dtype)
1328 struct lu_object *o;
1330 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1331 if (o->lo_dev->ld_type == dtype)
1336 EXPORT_SYMBOL(lu_object_locate);
1339 * Finalize and free devices in the device stack.
1341 * Finalize device stack by purging object cache, and calling
1342 * lu_device_type_operations::ldto_device_fini() and
1343 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1345 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1347 struct lu_site *site = top->ld_site;
1348 struct lu_device *scan;
1349 struct lu_device *next;
1351 lu_site_purge(env, site, ~0);
1352 for (scan = top; scan != NULL; scan = next) {
1353 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1354 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1355 lu_device_put(scan);
1359 lu_site_purge(env, site, ~0);
1361 for (scan = top; scan != NULL; scan = next) {
1362 const struct lu_device_type *ldt = scan->ld_type;
1363 struct obd_type *type;
1365 next = ldt->ldt_ops->ldto_device_free(env, scan);
1366 type = ldt->ldt_obd_type;
1369 class_put_type(type);
1376 * Maximal number of tld slots.
1378 LU_CONTEXT_KEY_NR = 40
1381 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1383 DEFINE_RWLOCK(lu_keys_guard);
1384 static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
1387 * Global counter incremented whenever key is registered, unregistered,
1388 * revived or quiesced. This is used to void unnecessary calls to
1389 * lu_context_refill(). No locking is provided, as initialization and shutdown
1390 * are supposed to be externally serialized.
1392 static unsigned key_set_version = 0;
1397 int lu_context_key_register(struct lu_context_key *key)
1402 LASSERT(key->lct_init != NULL);
1403 LASSERT(key->lct_fini != NULL);
1404 LASSERT(key->lct_tags != 0);
1405 LASSERT(key->lct_owner != NULL);
1408 write_lock(&lu_keys_guard);
1409 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1410 if (lu_keys[i] == NULL) {
1412 atomic_set(&key->lct_used, 1);
1414 lu_ref_init(&key->lct_reference);
1420 write_unlock(&lu_keys_guard);
1423 EXPORT_SYMBOL(lu_context_key_register);
1425 static void key_fini(struct lu_context *ctx, int index)
1427 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1428 struct lu_context_key *key;
1430 key = lu_keys[index];
1431 LASSERT(key != NULL);
1432 LASSERT(key->lct_fini != NULL);
1433 LASSERT(atomic_read(&key->lct_used) > 1);
1435 key->lct_fini(ctx, key, ctx->lc_value[index]);
1436 lu_ref_del(&key->lct_reference, "ctx", ctx);
1437 atomic_dec(&key->lct_used);
1439 LASSERT(key->lct_owner != NULL);
1440 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1441 LINVRNT(module_refcount(key->lct_owner) > 0);
1442 module_put(key->lct_owner);
1444 ctx->lc_value[index] = NULL;
1451 void lu_context_key_degister(struct lu_context_key *key)
1453 LASSERT(atomic_read(&key->lct_used) >= 1);
1454 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1456 lu_context_key_quiesce(key);
1459 write_lock(&lu_keys_guard);
1460 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1463 * Wait until all transient contexts referencing this key have
1464 * run lu_context_key::lct_fini() method.
1466 while (atomic_read(&key->lct_used) > 1) {
1467 write_unlock(&lu_keys_guard);
1468 CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
1469 key->lct_owner ? key->lct_owner->name : "", key,
1470 atomic_read(&key->lct_used));
1472 write_lock(&lu_keys_guard);
1474 if (lu_keys[key->lct_index]) {
1475 lu_keys[key->lct_index] = NULL;
1476 lu_ref_fini(&key->lct_reference);
1478 write_unlock(&lu_keys_guard);
1480 LASSERTF(atomic_read(&key->lct_used) == 1,
1481 "key has instances: %d\n",
1482 atomic_read(&key->lct_used));
1484 EXPORT_SYMBOL(lu_context_key_degister);
1487 * Register a number of keys. This has to be called after all keys have been
1488 * initialized by a call to LU_CONTEXT_KEY_INIT().
1490 int lu_context_key_register_many(struct lu_context_key *k, ...)
1492 struct lu_context_key *key = k;
1498 result = lu_context_key_register(key);
1501 key = va_arg(args, struct lu_context_key *);
1502 } while (key != NULL);
1508 lu_context_key_degister(k);
1509 k = va_arg(args, struct lu_context_key *);
1516 EXPORT_SYMBOL(lu_context_key_register_many);
1519 * De-register a number of keys. This is a dual to
1520 * lu_context_key_register_many().
1522 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1528 lu_context_key_degister(k);
1529 k = va_arg(args, struct lu_context_key*);
1530 } while (k != NULL);
1533 EXPORT_SYMBOL(lu_context_key_degister_many);
1536 * Revive a number of keys.
1538 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1544 lu_context_key_revive(k);
1545 k = va_arg(args, struct lu_context_key*);
1546 } while (k != NULL);
1549 EXPORT_SYMBOL(lu_context_key_revive_many);
1552 * Quiescent a number of keys.
1554 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1560 lu_context_key_quiesce(k);
1561 k = va_arg(args, struct lu_context_key*);
1562 } while (k != NULL);
1565 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1568 * Return value associated with key \a key in context \a ctx.
1570 void *lu_context_key_get(const struct lu_context *ctx,
1571 const struct lu_context_key *key)
1573 LINVRNT(ctx->lc_state == LCS_ENTERED);
1574 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1575 LASSERT(lu_keys[key->lct_index] == key);
1576 return ctx->lc_value[key->lct_index];
1578 EXPORT_SYMBOL(lu_context_key_get);
1581 * List of remembered contexts. XXX document me.
1583 static struct list_head lu_context_remembered;
1586 * Destroy \a key in all remembered contexts. This is used to destroy key
1587 * values in "shared" contexts (like service threads), when a module owning
1588 * the key is about to be unloaded.
1590 void lu_context_key_quiesce(struct lu_context_key *key)
1592 struct lu_context *ctx;
1593 extern unsigned cl_env_cache_purge(unsigned nr);
1595 if (!(key->lct_tags & LCT_QUIESCENT)) {
1597 * XXX layering violation.
1599 cl_env_cache_purge(~0);
1601 * XXX memory barrier has to go here.
1603 write_lock(&lu_keys_guard);
1604 key->lct_tags |= LCT_QUIESCENT;
1607 * Wait until all lu_context_key::lct_init() methods
1610 while (atomic_read(&lu_key_initing_cnt) > 0) {
1611 write_unlock(&lu_keys_guard);
1612 CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
1614 key->lct_owner ? key->lct_owner->name : "",
1615 key, atomic_read(&key->lct_used),
1616 atomic_read(&lu_key_initing_cnt));
1618 write_lock(&lu_keys_guard);
1621 list_for_each_entry(ctx, &lu_context_remembered,
1623 key_fini(ctx, key->lct_index);
1624 write_unlock(&lu_keys_guard);
1629 void lu_context_key_revive(struct lu_context_key *key)
1631 key->lct_tags &= ~LCT_QUIESCENT;
1635 static void keys_fini(struct lu_context *ctx)
1639 if (ctx->lc_value == NULL)
1642 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1645 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1646 ctx->lc_value = NULL;
1649 static int keys_fill(struct lu_context *ctx)
1654 * A serialisation with lu_context_key_quiesce() is needed, but some
1655 * "key->lct_init()" are calling kernel memory allocation routine and
1656 * can't be called while holding a spin_lock.
1657 * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
1658 * to ensure the start of the serialisation.
1659 * An atomic_t variable is still used, in order not to reacquire the
1660 * lock when decrementing the counter.
1662 read_lock(&lu_keys_guard);
1663 atomic_inc(&lu_key_initing_cnt);
1664 read_unlock(&lu_keys_guard);
1666 LINVRNT(ctx->lc_value != NULL);
1667 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1668 struct lu_context_key *key;
1671 if (ctx->lc_value[i] == NULL && key != NULL &&
1672 (key->lct_tags & ctx->lc_tags) &&
1674 * Don't create values for a LCT_QUIESCENT key, as this
1675 * will pin module owning a key.
1677 !(key->lct_tags & LCT_QUIESCENT)) {
1680 LINVRNT(key->lct_init != NULL);
1681 LINVRNT(key->lct_index == i);
1683 LASSERT(key->lct_owner != NULL);
1684 if (!(ctx->lc_tags & LCT_NOREF) &&
1685 try_module_get(key->lct_owner) == 0) {
1686 /* module is unloading, skip this key */
1690 value = key->lct_init(ctx, key);
1691 if (unlikely(IS_ERR(value))) {
1692 atomic_dec(&lu_key_initing_cnt);
1693 return PTR_ERR(value);
1696 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1697 atomic_inc(&key->lct_used);
1699 * This is the only place in the code, where an
1700 * element of ctx->lc_value[] array is set to non-NULL
1703 ctx->lc_value[i] = value;
1704 if (key->lct_exit != NULL)
1705 ctx->lc_tags |= LCT_HAS_EXIT;
1707 ctx->lc_version = key_set_version;
1709 atomic_dec(&lu_key_initing_cnt);
1713 static int keys_init(struct lu_context *ctx)
1715 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1716 if (likely(ctx->lc_value != NULL))
1717 return keys_fill(ctx);
1723 * Initialize context data-structure. Create values for all keys.
1725 int lu_context_init(struct lu_context *ctx, __u32 tags)
1729 memset(ctx, 0, sizeof *ctx);
1730 ctx->lc_state = LCS_INITIALIZED;
1731 ctx->lc_tags = tags;
1732 if (tags & LCT_REMEMBER) {
1733 write_lock(&lu_keys_guard);
1734 list_add(&ctx->lc_remember, &lu_context_remembered);
1735 write_unlock(&lu_keys_guard);
1737 INIT_LIST_HEAD(&ctx->lc_remember);
1740 rc = keys_init(ctx);
1742 lu_context_fini(ctx);
1746 EXPORT_SYMBOL(lu_context_init);
1749 * Finalize context data-structure. Destroy key values.
1751 void lu_context_fini(struct lu_context *ctx)
1753 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1754 ctx->lc_state = LCS_FINALIZED;
1756 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1757 LASSERT(list_empty(&ctx->lc_remember));
1760 } else { /* could race with key degister */
1761 write_lock(&lu_keys_guard);
1763 list_del_init(&ctx->lc_remember);
1764 write_unlock(&lu_keys_guard);
1767 EXPORT_SYMBOL(lu_context_fini);
1770 * Called before entering context.
1772 void lu_context_enter(struct lu_context *ctx)
1774 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1775 ctx->lc_state = LCS_ENTERED;
1777 EXPORT_SYMBOL(lu_context_enter);
1780 * Called after exiting from \a ctx
1782 void lu_context_exit(struct lu_context *ctx)
1786 LINVRNT(ctx->lc_state == LCS_ENTERED);
1787 ctx->lc_state = LCS_LEFT;
1788 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1789 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1790 /* could race with key quiescency */
1791 if (ctx->lc_tags & LCT_REMEMBER)
1792 read_lock(&lu_keys_guard);
1793 if (ctx->lc_value[i] != NULL) {
1794 struct lu_context_key *key;
1797 LASSERT(key != NULL);
1798 if (key->lct_exit != NULL)
1800 key, ctx->lc_value[i]);
1802 if (ctx->lc_tags & LCT_REMEMBER)
1803 read_unlock(&lu_keys_guard);
1807 EXPORT_SYMBOL(lu_context_exit);
1810 * Allocate for context all missing keys that were registered after context
1811 * creation. key_set_version is only changed in rare cases when modules
1812 * are loaded and removed.
1814 int lu_context_refill(struct lu_context *ctx)
1816 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1820 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1821 * obd being added. Currently, this is only used on client side, specifically
1822 * for echo device client, for other stack (like ptlrpc threads), context are
1823 * predefined when the lu_device type are registered, during the module probe
1826 __u32 lu_context_tags_default = 0;
1827 __u32 lu_session_tags_default = 0;
1829 void lu_context_tags_update(__u32 tags)
1831 write_lock(&lu_keys_guard);
1832 lu_context_tags_default |= tags;
1834 write_unlock(&lu_keys_guard);
1836 EXPORT_SYMBOL(lu_context_tags_update);
1838 void lu_context_tags_clear(__u32 tags)
1840 write_lock(&lu_keys_guard);
1841 lu_context_tags_default &= ~tags;
1843 write_unlock(&lu_keys_guard);
1845 EXPORT_SYMBOL(lu_context_tags_clear);
1847 void lu_session_tags_update(__u32 tags)
1849 write_lock(&lu_keys_guard);
1850 lu_session_tags_default |= tags;
1852 write_unlock(&lu_keys_guard);
1854 EXPORT_SYMBOL(lu_session_tags_update);
1856 void lu_session_tags_clear(__u32 tags)
1858 write_lock(&lu_keys_guard);
1859 lu_session_tags_default &= ~tags;
1861 write_unlock(&lu_keys_guard);
1863 EXPORT_SYMBOL(lu_session_tags_clear);
1865 int lu_env_init(struct lu_env *env, __u32 tags)
1870 result = lu_context_init(&env->le_ctx, tags);
1871 if (likely(result == 0))
1872 lu_context_enter(&env->le_ctx);
1875 EXPORT_SYMBOL(lu_env_init);
1877 void lu_env_fini(struct lu_env *env)
1879 lu_context_exit(&env->le_ctx);
1880 lu_context_fini(&env->le_ctx);
1883 EXPORT_SYMBOL(lu_env_fini);
1885 int lu_env_refill(struct lu_env *env)
1889 result = lu_context_refill(&env->le_ctx);
1890 if (result == 0 && env->le_ses != NULL)
1891 result = lu_context_refill(env->le_ses);
1894 EXPORT_SYMBOL(lu_env_refill);
1897 * Currently, this API will only be used by echo client.
1898 * Because echo client and normal lustre client will share
1899 * same cl_env cache. So echo client needs to refresh
1900 * the env context after it get one from the cache, especially
1901 * when normal client and echo client co-exist in the same client.
1903 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1908 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1909 env->le_ctx.lc_version = 0;
1910 env->le_ctx.lc_tags |= ctags;
1913 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1914 env->le_ses->lc_version = 0;
1915 env->le_ses->lc_tags |= stags;
1918 result = lu_env_refill(env);
1922 EXPORT_SYMBOL(lu_env_refill_by_tags);
1924 static struct shrinker *lu_site_shrinker;
1926 typedef struct lu_site_stats{
1927 unsigned lss_populated;
1928 unsigned lss_max_search;
1933 static void lu_site_stats_get(struct cfs_hash *hs,
1934 lu_site_stats_t *stats, int populated)
1936 struct cfs_hash_bd bd;
1939 cfs_hash_for_each_bucket(hs, &bd, i) {
1940 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1941 struct hlist_head *hhead;
1943 cfs_hash_bd_lock(hs, &bd, 1);
1945 cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
1946 stats->lss_total += cfs_hash_bd_count_get(&bd);
1947 stats->lss_max_search = max((int)stats->lss_max_search,
1948 cfs_hash_bd_depmax_get(&bd));
1950 cfs_hash_bd_unlock(hs, &bd, 1);
1954 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1955 if (!hlist_empty(hhead))
1956 stats->lss_populated++;
1958 cfs_hash_bd_unlock(hs, &bd, 1);
1964 * lu_cache_shrink_count returns the number of cached objects that are
1965 * candidates to be freed by shrink_slab(). A counter, which tracks
1966 * the number of items in the site's lru, is maintained in the per cpu
1967 * stats of each site. The counter is incremented when an object is added
1968 * to a site's lru and decremented when one is removed. The number of
1969 * free-able objects is the sum of all per cpu counters for all sites.
1971 * Using a per cpu counter is a compromise solution to concurrent access:
1972 * lu_object_put() can update the counter without locking the site and
1973 * lu_cache_shrink_count can sum the counters without locking each
1974 * ls_obj_hash bucket.
1976 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1977 struct shrink_control *sc)
1980 struct lu_site *tmp;
1981 unsigned long cached = 0;
1983 if (!(sc->gfp_mask & __GFP_FS))
1986 mutex_lock(&lu_sites_guard);
1987 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1988 cached += ls_stats_read(s->ls_stats, LU_SS_LRU_LEN);
1990 mutex_unlock(&lu_sites_guard);
1992 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1993 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
1994 cached, sysctl_vfs_cache_pressure);
1999 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2000 struct shrink_control *sc)
2003 struct lu_site *tmp;
2004 unsigned long remain = sc->nr_to_scan;
2007 if (!(sc->gfp_mask & __GFP_FS))
2008 /* We must not take the lu_sites_guard lock when
2009 * __GFP_FS is *not* set because of the deadlock
2010 * possibility detailed above. Additionally,
2011 * since we cannot determine the number of
2012 * objects in the cache without taking this
2013 * lock, we're in a particularly tough spot. As
2014 * a result, we'll just lie and say our cache is
2015 * empty. This _should_ be ok, as we can't
2016 * reclaim objects when __GFP_FS is *not* set
2021 mutex_lock(&lu_sites_guard);
2022 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2023 remain = lu_site_purge(&lu_shrink_env, s, remain);
2025 * Move just shrunk site to the tail of site list to
2026 * assure shrinking fairness.
2028 list_move_tail(&s->ls_linkage, &splice);
2030 list_splice(&splice, lu_sites.prev);
2031 mutex_unlock(&lu_sites_guard);
2033 return sc->nr_to_scan - remain;
2036 #ifndef HAVE_SHRINKER_COUNT
2038 * There exists a potential lock inversion deadlock scenario when using
2039 * Lustre on top of ZFS. This occurs between one of ZFS's
2040 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2041 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2042 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2043 * lock. Obviously neither thread will wake and drop their respective hold
2046 * To prevent this from happening we must ensure the lu_sites_guard lock is
2047 * not taken while down this code path. ZFS reliably does not set the
2048 * __GFP_FS bit in its code paths, so this can be used to determine if it
2049 * is safe to take the lu_sites_guard lock.
2051 * Ideally we should accurately return the remaining number of cached
2052 * objects without taking the lu_sites_guard lock, but this is not
2053 * possible in the current implementation.
2055 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2058 struct shrink_control scv = {
2059 .nr_to_scan = shrink_param(sc, nr_to_scan),
2060 .gfp_mask = shrink_param(sc, gfp_mask)
2062 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2063 struct shrinker* shrinker = NULL;
2067 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2069 if (scv.nr_to_scan != 0)
2070 lu_cache_shrink_scan(shrinker, &scv);
2072 cached = lu_cache_shrink_count(shrinker, &scv);
2076 #endif /* HAVE_SHRINKER_COUNT */
2084 * Environment to be used in debugger, contains all tags.
2086 static struct lu_env lu_debugging_env;
2089 * Debugging printer function using printk().
2091 int lu_printk_printer(const struct lu_env *env,
2092 void *unused, const char *format, ...)
2096 va_start(args, format);
2097 vprintk(format, args);
2102 int lu_debugging_setup(void)
2104 return lu_env_init(&lu_debugging_env, ~0);
2107 void lu_context_keys_dump(void)
2111 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2112 struct lu_context_key *key;
2116 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2117 i, key, key->lct_tags,
2118 key->lct_init, key->lct_fini, key->lct_exit,
2119 key->lct_index, atomic_read(&key->lct_used),
2120 key->lct_owner ? key->lct_owner->name : "",
2122 lu_ref_print(&key->lct_reference);
2128 * Initialization of global lu_* data.
2130 int lu_global_init(void)
2133 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2134 lu_cache_shrink_count, lu_cache_shrink_scan);
2136 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2138 INIT_LIST_HEAD(&lu_device_types);
2139 INIT_LIST_HEAD(&lu_context_remembered);
2140 INIT_LIST_HEAD(&lu_sites);
2142 result = lu_ref_global_init();
2146 LU_CONTEXT_KEY_INIT(&lu_global_key);
2147 result = lu_context_key_register(&lu_global_key);
2152 * At this level, we don't know what tags are needed, so allocate them
2153 * conservatively. This should not be too bad, because this
2154 * environment is global.
2156 mutex_lock(&lu_sites_guard);
2157 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2158 mutex_unlock(&lu_sites_guard);
2163 * seeks estimation: 3 seeks to read a record from oi, one to read
2164 * inode, one for ea. Unfortunately setting this high value results in
2165 * lu_object/inode cache consuming all the memory.
2167 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2168 if (lu_site_shrinker == NULL)
2175 * Dual to lu_global_init().
2177 void lu_global_fini(void)
2179 if (lu_site_shrinker != NULL) {
2180 remove_shrinker(lu_site_shrinker);
2181 lu_site_shrinker = NULL;
2184 lu_context_key_degister(&lu_global_key);
2187 * Tear shrinker environment down _after_ de-registering
2188 * lu_global_key, because the latter has a value in the former.
2190 mutex_lock(&lu_sites_guard);
2191 lu_env_fini(&lu_shrink_env);
2192 mutex_unlock(&lu_sites_guard);
2194 lu_ref_global_fini();
2197 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2199 #ifdef CONFIG_PROC_FS
2200 struct lprocfs_counter ret;
2202 lprocfs_stats_collect(stats, idx, &ret);
2203 if (idx == LU_SS_LRU_LEN)
2205 * protect against counter on cpu A being decremented
2206 * before counter is incremented on cpu B; unlikely
2208 return (__u32)((ret.lc_sum > 0) ? ret.lc_sum : 0);
2210 return (__u32)ret.lc_count;
2217 * Output site statistical counters into a buffer. Suitable for
2218 * lprocfs_rd_*()-style functions.
2220 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2222 lu_site_stats_t stats;
2224 memset(&stats, 0, sizeof(stats));
2225 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2227 return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d %d\n",
2230 stats.lss_populated,
2231 CFS_HASH_NHLIST(s->ls_obj_hash),
2232 stats.lss_max_search,
2233 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2234 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2235 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2236 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2237 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2238 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED),
2239 ls_stats_read(s->ls_stats, LU_SS_LRU_LEN));
2241 EXPORT_SYMBOL(lu_site_stats_seq_print);
2244 * Helper function to initialize a number of kmem slab caches at once.
2246 int lu_kmem_init(struct lu_kmem_descr *caches)
2249 struct lu_kmem_descr *iter = caches;
2251 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2252 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2255 if (*iter->ckd_cache == NULL) {
2257 /* free all previously allocated caches */
2258 lu_kmem_fini(caches);
2264 EXPORT_SYMBOL(lu_kmem_init);
2267 * Helper function to finalize a number of kmem slab cached at once. Dual to
2270 void lu_kmem_fini(struct lu_kmem_descr *caches)
2272 for (; caches->ckd_cache != NULL; ++caches) {
2273 if (*caches->ckd_cache != NULL) {
2274 kmem_cache_destroy(*caches->ckd_cache);
2275 *caches->ckd_cache = NULL;
2279 EXPORT_SYMBOL(lu_kmem_fini);
2282 * Temporary solution to be able to assign fid in ->do_create()
2283 * till we have fully-functional OST fids
2285 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2286 const struct lu_fid *fid)
2288 struct lu_site *s = o->lo_dev->ld_site;
2289 struct lu_fid *old = &o->lo_header->loh_fid;
2290 struct lu_object *shadow;
2291 wait_queue_t waiter;
2292 struct cfs_hash *hs;
2293 struct cfs_hash_bd bd;
2296 LASSERT(fid_is_zero(old));
2298 hs = s->ls_obj_hash;
2299 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2300 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2301 /* supposed to be unique */
2302 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2304 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2305 cfs_hash_bd_unlock(hs, &bd, 1);
2307 EXPORT_SYMBOL(lu_object_assign_fid);
2310 * allocates object with 0 (non-assiged) fid
2311 * XXX: temporary solution to be able to assign fid in ->do_create()
2312 * till we have fully-functional OST fids
2314 struct lu_object *lu_object_anon(const struct lu_env *env,
2315 struct lu_device *dev,
2316 const struct lu_object_conf *conf)
2319 struct lu_object *o;
2322 o = lu_object_alloc(env, dev, &fid, conf);
2326 EXPORT_SYMBOL(lu_object_anon);
2328 struct lu_buf LU_BUF_NULL = {
2332 EXPORT_SYMBOL(LU_BUF_NULL);
2334 void lu_buf_free(struct lu_buf *buf)
2338 LASSERT(buf->lb_len > 0);
2339 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2344 EXPORT_SYMBOL(lu_buf_free);
2346 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2349 LASSERT(buf->lb_buf == NULL);
2350 LASSERT(buf->lb_len == 0);
2351 OBD_ALLOC_LARGE(buf->lb_buf, size);
2352 if (likely(buf->lb_buf))
2355 EXPORT_SYMBOL(lu_buf_alloc);
2357 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2360 lu_buf_alloc(buf, size);
2362 EXPORT_SYMBOL(lu_buf_realloc);
2364 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2366 if (buf->lb_buf == NULL && buf->lb_len == 0)
2367 lu_buf_alloc(buf, len);
2369 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2370 lu_buf_realloc(buf, len);
2374 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2377 * Increase the size of the \a buf.
2378 * preserves old data in buffer
2379 * old buffer remains unchanged on error
2380 * \retval 0 or -ENOMEM
2382 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2386 if (len <= buf->lb_len)
2389 OBD_ALLOC_LARGE(ptr, len);
2393 /* Free the old buf */
2394 if (buf->lb_buf != NULL) {
2395 memcpy(ptr, buf->lb_buf, buf->lb_len);
2396 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);