4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, 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 <linux/libcfs/libcfs.h>
49 # include <linux/module.h>
52 #include <linux/libcfs/libcfs_hash.h>
53 #include <obd_class.h>
54 #include <obd_support.h>
55 #include <lustre_disk.h>
56 #include <lustre_fid.h>
57 #include <lu_object.h>
59 #include <linux/list.h>
61 static void lu_object_free(const struct lu_env
*env
, struct lu_object
*o
);
64 * Decrease reference counter on object. If last reference is freed, return
65 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
66 * case, free object immediately.
68 void lu_object_put(const struct lu_env
*env
, struct lu_object
*o
)
70 struct lu_site_bkt_data
*bkt
;
71 struct lu_object_header
*top
;
73 struct lu_object
*orig
;
75 const struct lu_fid
*fid
;
78 site
= o
->lo_dev
->ld_site
;
82 * till we have full fids-on-OST implemented anonymous objects
83 * are possible in OSP. such an object isn't listed in the site
84 * so we should not remove it from the site.
86 fid
= lu_object_fid(o
);
87 if (fid_is_zero(fid
)) {
88 LASSERT(top
->loh_hash
.next
== NULL
89 && top
->loh_hash
.pprev
== NULL
);
90 LASSERT(list_empty(&top
->loh_lru
));
91 if (!atomic_dec_and_test(&top
->loh_ref
))
93 list_for_each_entry_reverse(o
, &top
->loh_layers
, lo_linkage
) {
94 if (o
->lo_ops
->loo_object_release
!= NULL
)
95 o
->lo_ops
->loo_object_release(env
, o
);
97 lu_object_free(env
, orig
);
101 cfs_hash_bd_get(site
->ls_obj_hash
, &top
->loh_fid
, &bd
);
102 bkt
= cfs_hash_bd_extra_get(site
->ls_obj_hash
, &bd
);
104 if (!cfs_hash_bd_dec_and_lock(site
->ls_obj_hash
, &bd
, &top
->loh_ref
)) {
105 if (lu_object_is_dying(top
)) {
108 * somebody may be waiting for this, currently only
109 * used for cl_object, see cl_object_put_last().
111 wake_up_all(&bkt
->lsb_marche_funebre
);
116 LASSERT(bkt
->lsb_busy
> 0);
119 * When last reference is released, iterate over object
120 * layers, and notify them that object is no longer busy.
122 list_for_each_entry_reverse(o
, &top
->loh_layers
, lo_linkage
) {
123 if (o
->lo_ops
->loo_object_release
!= NULL
)
124 o
->lo_ops
->loo_object_release(env
, o
);
127 if (!lu_object_is_dying(top
)) {
128 LASSERT(list_empty(&top
->loh_lru
));
129 list_add_tail(&top
->loh_lru
, &bkt
->lsb_lru
);
130 cfs_hash_bd_unlock(site
->ls_obj_hash
, &bd
, 1);
135 * If object is dying (will not be cached), removed it
136 * from hash table and LRU.
138 * This is done with hash table and LRU lists locked. As the only
139 * way to acquire first reference to previously unreferenced
140 * object is through hash-table lookup (lu_object_find()),
141 * or LRU scanning (lu_site_purge()), that are done under hash-table
142 * and LRU lock, no race with concurrent object lookup is possible
143 * and we can safely destroy object below.
145 if (!test_and_set_bit(LU_OBJECT_UNHASHED
, &top
->loh_flags
))
146 cfs_hash_bd_del_locked(site
->ls_obj_hash
, &bd
, &top
->loh_hash
);
147 cfs_hash_bd_unlock(site
->ls_obj_hash
, &bd
, 1);
149 * Object was already removed from hash and lru above, can
152 lu_object_free(env
, orig
);
154 EXPORT_SYMBOL(lu_object_put
);
157 * Put object and don't keep in cache. This is temporary solution for
158 * multi-site objects when its layering is not constant.
160 void lu_object_put_nocache(const struct lu_env
*env
, struct lu_object
*o
)
162 set_bit(LU_OBJECT_HEARD_BANSHEE
, &o
->lo_header
->loh_flags
);
163 return lu_object_put(env
, o
);
165 EXPORT_SYMBOL(lu_object_put_nocache
);
168 * Kill the object and take it out of LRU cache.
169 * Currently used by client code for layout change.
171 void lu_object_unhash(const struct lu_env
*env
, struct lu_object
*o
)
173 struct lu_object_header
*top
;
176 set_bit(LU_OBJECT_HEARD_BANSHEE
, &top
->loh_flags
);
177 if (!test_and_set_bit(LU_OBJECT_UNHASHED
, &top
->loh_flags
)) {
178 cfs_hash_t
*obj_hash
= o
->lo_dev
->ld_site
->ls_obj_hash
;
181 cfs_hash_bd_get_and_lock(obj_hash
, &top
->loh_fid
, &bd
, 1);
182 list_del_init(&top
->loh_lru
);
183 cfs_hash_bd_del_locked(obj_hash
, &bd
, &top
->loh_hash
);
184 cfs_hash_bd_unlock(obj_hash
, &bd
, 1);
187 EXPORT_SYMBOL(lu_object_unhash
);
190 * Allocate new object.
192 * This follows object creation protocol, described in the comment within
193 * struct lu_device_operations definition.
195 static struct lu_object
*lu_object_alloc(const struct lu_env
*env
,
196 struct lu_device
*dev
,
197 const struct lu_fid
*f
,
198 const struct lu_object_conf
*conf
)
200 struct lu_object
*scan
;
201 struct lu_object
*top
;
202 struct list_head
*layers
;
208 * Create top-level object slice. This will also create
211 top
= dev
->ld_ops
->ldo_object_alloc(env
, NULL
, dev
);
213 RETURN(ERR_PTR(-ENOMEM
));
217 * This is the only place where object fid is assigned. It's constant
220 top
->lo_header
->loh_fid
= *f
;
221 layers
= &top
->lo_header
->loh_layers
;
224 * Call ->loo_object_init() repeatedly, until no more new
225 * object slices are created.
228 list_for_each_entry(scan
, layers
, lo_linkage
) {
229 if (scan
->lo_flags
& LU_OBJECT_ALLOCATED
)
232 scan
->lo_header
= top
->lo_header
;
233 result
= scan
->lo_ops
->loo_object_init(env
, scan
, conf
);
235 lu_object_free(env
, top
);
236 RETURN(ERR_PTR(result
));
238 scan
->lo_flags
|= LU_OBJECT_ALLOCATED
;
242 list_for_each_entry_reverse(scan
, layers
, lo_linkage
) {
243 if (scan
->lo_ops
->loo_object_start
!= NULL
) {
244 result
= scan
->lo_ops
->loo_object_start(env
, scan
);
246 lu_object_free(env
, top
);
247 RETURN(ERR_PTR(result
));
252 lprocfs_counter_incr(dev
->ld_site
->ls_stats
, LU_SS_CREATED
);
259 static void lu_object_free(const struct lu_env
*env
, struct lu_object
*o
)
261 struct lu_site_bkt_data
*bkt
;
262 struct lu_site
*site
;
263 struct lu_object
*scan
;
264 struct list_head
*layers
;
265 struct list_head splice
;
267 site
= o
->lo_dev
->ld_site
;
268 layers
= &o
->lo_header
->loh_layers
;
269 bkt
= lu_site_bkt_from_fid(site
, &o
->lo_header
->loh_fid
);
271 * First call ->loo_object_delete() method to release all resources.
273 list_for_each_entry_reverse(scan
, layers
, lo_linkage
) {
274 if (scan
->lo_ops
->loo_object_delete
!= NULL
)
275 scan
->lo_ops
->loo_object_delete(env
, scan
);
279 * Then, splice object layers into stand-alone list, and call
280 * ->loo_object_free() on all layers to free memory. Splice is
281 * necessary, because lu_object_header is freed together with the
284 INIT_LIST_HEAD(&splice
);
285 list_splice_init(layers
, &splice
);
286 while (!list_empty(&splice
)) {
288 * Free layers in bottom-to-top order, so that object header
289 * lives as long as possible and ->loo_object_free() methods
290 * can look at its contents.
292 o
= container_of0(splice
.prev
, struct lu_object
, lo_linkage
);
293 list_del_init(&o
->lo_linkage
);
294 LASSERT(o
->lo_ops
->loo_object_free
!= NULL
);
295 o
->lo_ops
->loo_object_free(env
, o
);
298 if (waitqueue_active(&bkt
->lsb_marche_funebre
))
299 wake_up_all(&bkt
->lsb_marche_funebre
);
303 * Free \a nr objects from the cold end of the site LRU list.
305 int lu_site_purge(const struct lu_env
*env
, struct lu_site
*s
, int nr
)
307 struct lu_object_header
*h
;
308 struct lu_object_header
*temp
;
309 struct lu_site_bkt_data
*bkt
;
312 struct list_head dispose
;
319 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU
))
322 INIT_LIST_HEAD(&dispose
);
324 * Under LRU list lock, scan LRU list and move unreferenced objects to
325 * the dispose list, removing them from LRU and hash table.
327 start
= s
->ls_purge_start
;
328 bnr
= (nr
== ~0) ? -1 : nr
/ CFS_HASH_NBKT(s
->ls_obj_hash
) + 1;
331 cfs_hash_for_each_bucket(s
->ls_obj_hash
, &bd
, i
) {
335 cfs_hash_bd_lock(s
->ls_obj_hash
, &bd
, 1);
336 bkt
= cfs_hash_bd_extra_get(s
->ls_obj_hash
, &bd
);
338 list_for_each_entry_safe(h
, temp
, &bkt
->lsb_lru
, loh_lru
) {
339 LASSERT(atomic_read(&h
->loh_ref
) == 0);
341 cfs_hash_bd_get(s
->ls_obj_hash
, &h
->loh_fid
, &bd2
);
342 LASSERT(bd
.bd_bucket
== bd2
.bd_bucket
);
344 cfs_hash_bd_del_locked(s
->ls_obj_hash
,
346 list_move(&h
->loh_lru
, &dispose
);
350 if (nr
!= ~0 && --nr
== 0)
353 if (count
> 0 && --count
== 0)
357 cfs_hash_bd_unlock(s
->ls_obj_hash
, &bd
, 1);
360 * Free everything on the dispose list. This is safe against
361 * races due to the reasons described in lu_object_put().
363 while (!list_empty(&dispose
)) {
364 h
= container_of0(dispose
.next
,
365 struct lu_object_header
, loh_lru
);
366 list_del_init(&h
->loh_lru
);
367 lu_object_free(env
, lu_object_top(h
));
368 lprocfs_counter_incr(s
->ls_stats
, LU_SS_LRU_PURGED
);
375 if (nr
!= 0 && did_sth
&& start
!= 0) {
376 start
= 0; /* restart from the first bucket */
379 /* race on s->ls_purge_start, but nobody cares */
380 s
->ls_purge_start
= i
% CFS_HASH_NBKT(s
->ls_obj_hash
);
384 EXPORT_SYMBOL(lu_site_purge
);
389 * Code below has to jump through certain loops to output object description
390 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
391 * composes object description from strings that are parts of _lines_ of
392 * output (i.e., strings that are not terminated by newline). This doesn't fit
393 * very well into libcfs_debug_msg() interface that assumes that each message
394 * supplied to it is a self-contained output line.
396 * To work around this, strings are collected in a temporary buffer
397 * (implemented as a value of lu_cdebug_key key), until terminating newline
398 * character is detected.
406 * XXX overflow is not handled correctly.
411 struct lu_cdebug_data
{
415 char lck_area
[LU_CDEBUG_LINE
];
418 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
419 LU_KEY_INIT_FINI(lu_global
, struct lu_cdebug_data
);
422 * Key, holding temporary buffer. This key is registered very early by
425 struct lu_context_key lu_global_key
= {
426 .lct_tags
= LCT_MD_THREAD
| LCT_DT_THREAD
|
427 LCT_MG_THREAD
| LCT_CL_THREAD
,
428 .lct_init
= lu_global_key_init
,
429 .lct_fini
= lu_global_key_fini
433 * Printer function emitting messages through libcfs_debug_msg().
435 int lu_cdebug_printer(const struct lu_env
*env
,
436 void *cookie
, const char *format
, ...)
438 struct libcfs_debug_msg_data
*msgdata
= cookie
;
439 struct lu_cdebug_data
*key
;
444 va_start(args
, format
);
446 key
= lu_context_key_get(&env
->le_ctx
, &lu_global_key
);
447 LASSERT(key
!= NULL
);
449 used
= strlen(key
->lck_area
);
450 complete
= format
[strlen(format
) - 1] == '\n';
452 * Append new chunk to the buffer.
454 vsnprintf(key
->lck_area
+ used
,
455 ARRAY_SIZE(key
->lck_area
) - used
, format
, args
);
457 if (cfs_cdebug_show(msgdata
->msg_mask
, msgdata
->msg_subsys
))
458 libcfs_debug_msg(msgdata
, "%s", key
->lck_area
);
459 key
->lck_area
[0] = 0;
464 EXPORT_SYMBOL(lu_cdebug_printer
);
467 * Print object header.
469 void lu_object_header_print(const struct lu_env
*env
, void *cookie
,
470 lu_printer_t printer
,
471 const struct lu_object_header
*hdr
)
473 (*printer
)(env
, cookie
, "header@%p[%#lx, %d, "DFID
"%s%s%s]",
474 hdr
, hdr
->loh_flags
, atomic_read(&hdr
->loh_ref
),
476 hlist_unhashed(&hdr
->loh_hash
) ? "" : " hash",
477 list_empty((struct list_head
*)&hdr
->loh_lru
) ? \
479 hdr
->loh_attr
& LOHA_EXISTS
? " exist":"");
481 EXPORT_SYMBOL(lu_object_header_print
);
484 * Print human readable representation of the \a o to the \a printer.
486 void lu_object_print(const struct lu_env
*env
, void *cookie
,
487 lu_printer_t printer
, const struct lu_object
*o
)
489 static const char ruler
[] = "........................................";
490 struct lu_object_header
*top
;
494 lu_object_header_print(env
, cookie
, printer
, top
);
495 (*printer
)(env
, cookie
, "{ \n");
496 list_for_each_entry(o
, &top
->loh_layers
, lo_linkage
) {
497 depth
= o
->lo_depth
+ 4;
500 * print `.' \a depth times followed by type name and address
502 (*printer
)(env
, cookie
, "%*.*s%s@%p", depth
, depth
, ruler
,
503 o
->lo_dev
->ld_type
->ldt_name
, o
);
504 if (o
->lo_ops
->loo_object_print
!= NULL
)
505 o
->lo_ops
->loo_object_print(env
, cookie
, printer
, o
);
506 (*printer
)(env
, cookie
, "\n");
508 (*printer
)(env
, cookie
, "} header@%p\n", top
);
510 EXPORT_SYMBOL(lu_object_print
);
513 * Check object consistency.
515 int lu_object_invariant(const struct lu_object
*o
)
517 struct lu_object_header
*top
;
520 list_for_each_entry(o
, &top
->loh_layers
, lo_linkage
) {
521 if (o
->lo_ops
->loo_object_invariant
!= NULL
&&
522 !o
->lo_ops
->loo_object_invariant(o
))
527 EXPORT_SYMBOL(lu_object_invariant
);
529 static struct lu_object
*htable_lookup(struct lu_site
*s
,
531 const struct lu_fid
*f
,
532 wait_queue_t
*waiter
,
535 struct lu_site_bkt_data
*bkt
;
536 struct lu_object_header
*h
;
537 struct hlist_node
*hnode
;
538 __u64 ver
= cfs_hash_bd_version_get(bd
);
541 return ERR_PTR(-ENOENT
);
544 bkt
= cfs_hash_bd_extra_get(s
->ls_obj_hash
, bd
);
545 /* cfs_hash_bd_peek_locked is a somehow "internal" function
546 * of cfs_hash, it doesn't add refcount on object. */
547 hnode
= cfs_hash_bd_peek_locked(s
->ls_obj_hash
, bd
, (void *)f
);
549 lprocfs_counter_incr(s
->ls_stats
, LU_SS_CACHE_MISS
);
550 return ERR_PTR(-ENOENT
);
553 h
= container_of0(hnode
, struct lu_object_header
, loh_hash
);
554 if (likely(!lu_object_is_dying(h
))) {
555 cfs_hash_get(s
->ls_obj_hash
, hnode
);
556 lprocfs_counter_incr(s
->ls_stats
, LU_SS_CACHE_HIT
);
557 list_del_init(&h
->loh_lru
);
558 return lu_object_top(h
);
562 * Lookup found an object being destroyed this object cannot be
563 * returned (to assure that references to dying objects are eventually
564 * drained), and moreover, lookup has to wait until object is freed.
567 init_waitqueue_entry_current(waiter
);
568 add_wait_queue(&bkt
->lsb_marche_funebre
, waiter
);
569 set_current_state(TASK_UNINTERRUPTIBLE
);
570 lprocfs_counter_incr(s
->ls_stats
, LU_SS_CACHE_DEATH_RACE
);
571 return ERR_PTR(-EAGAIN
);
575 * Search cache for an object with the fid \a f. If such object is found,
576 * return it. Otherwise, create new object, insert it into cache and return
577 * it. In any case, additional reference is acquired on the returned object.
579 struct lu_object
*lu_object_find(const struct lu_env
*env
,
580 struct lu_device
*dev
, const struct lu_fid
*f
,
581 const struct lu_object_conf
*conf
)
583 return lu_object_find_at(env
, dev
->ld_site
->ls_top_dev
, f
, conf
);
585 EXPORT_SYMBOL(lu_object_find
);
587 static struct lu_object
*lu_object_new(const struct lu_env
*env
,
588 struct lu_device
*dev
,
589 const struct lu_fid
*f
,
590 const struct lu_object_conf
*conf
)
595 struct lu_site_bkt_data
*bkt
;
597 o
= lu_object_alloc(env
, dev
, f
, conf
);
598 if (unlikely(IS_ERR(o
)))
601 hs
= dev
->ld_site
->ls_obj_hash
;
602 cfs_hash_bd_get_and_lock(hs
, (void *)f
, &bd
, 1);
603 bkt
= cfs_hash_bd_extra_get(hs
, &bd
);
604 cfs_hash_bd_add_locked(hs
, &bd
, &o
->lo_header
->loh_hash
);
606 cfs_hash_bd_unlock(hs
, &bd
, 1);
611 * Core logic of lu_object_find*() functions.
613 static struct lu_object
*lu_object_find_try(const struct lu_env
*env
,
614 struct lu_device
*dev
,
615 const struct lu_fid
*f
,
616 const struct lu_object_conf
*conf
,
617 wait_queue_t
*waiter
)
620 struct lu_object
*shadow
;
627 * This uses standard index maintenance protocol:
629 * - search index under lock, and return object if found;
630 * - otherwise, unlock index, allocate new object;
631 * - lock index and search again;
632 * - if nothing is found (usual case), insert newly created
634 * - otherwise (race: other thread inserted object), free
635 * object just allocated.
639 * For "LOC_F_NEW" case, we are sure the object is new established.
640 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
641 * just alloc and insert directly.
643 * If dying object is found during index search, add @waiter to the
644 * site wait-queue and return ERR_PTR(-EAGAIN).
646 if (conf
!= NULL
&& conf
->loc_flags
& LOC_F_NEW
)
647 return lu_object_new(env
, dev
, f
, conf
);
651 cfs_hash_bd_get_and_lock(hs
, (void *)f
, &bd
, 1);
652 o
= htable_lookup(s
, &bd
, f
, waiter
, &version
);
653 cfs_hash_bd_unlock(hs
, &bd
, 1);
654 if (!IS_ERR(o
) || PTR_ERR(o
) != -ENOENT
)
658 * Allocate new object. This may result in rather complicated
659 * operations, including fld queries, inode loading, etc.
661 o
= lu_object_alloc(env
, dev
, f
, conf
);
662 if (unlikely(IS_ERR(o
)))
665 LASSERT(lu_fid_eq(lu_object_fid(o
), f
));
667 cfs_hash_bd_lock(hs
, &bd
, 1);
669 shadow
= htable_lookup(s
, &bd
, f
, waiter
, &version
);
670 if (likely(IS_ERR(shadow
) && PTR_ERR(shadow
) == -ENOENT
)) {
671 struct lu_site_bkt_data
*bkt
;
673 bkt
= cfs_hash_bd_extra_get(hs
, &bd
);
674 cfs_hash_bd_add_locked(hs
, &bd
, &o
->lo_header
->loh_hash
);
676 cfs_hash_bd_unlock(hs
, &bd
, 1);
680 lprocfs_counter_incr(s
->ls_stats
, LU_SS_CACHE_RACE
);
681 cfs_hash_bd_unlock(hs
, &bd
, 1);
682 lu_object_free(env
, o
);
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
)
696 struct lu_site_bkt_data
*bkt
;
697 struct lu_object
*obj
;
701 obj
= lu_object_find_try(env
, dev
, f
, conf
, &wait
);
702 if (obj
!= ERR_PTR(-EAGAIN
))
705 * lu_object_find_try() already added waiter into the
708 waitq_wait(&wait
, TASK_UNINTERRUPTIBLE
);
709 bkt
= lu_site_bkt_from_fid(dev
->ld_site
, (void *)f
);
710 remove_wait_queue(&bkt
->lsb_marche_funebre
, &wait
);
713 EXPORT_SYMBOL(lu_object_find_at
);
716 * Find object with given fid, and return its slice belonging to given device.
718 struct lu_object
*lu_object_find_slice(const struct lu_env
*env
,
719 struct lu_device
*dev
,
720 const struct lu_fid
*f
,
721 const struct lu_object_conf
*conf
)
723 struct lu_object
*top
;
724 struct lu_object
*obj
;
726 top
= lu_object_find(env
, dev
, f
, conf
);
728 obj
= lu_object_locate(top
->lo_header
, dev
->ld_type
);
730 lu_object_put(env
, top
);
735 EXPORT_SYMBOL(lu_object_find_slice
);
738 * Global list of all device types.
740 static LIST_HEAD(lu_device_types
);
742 int lu_device_type_init(struct lu_device_type
*ldt
)
746 INIT_LIST_HEAD(&ldt
->ldt_linkage
);
747 if (ldt
->ldt_ops
->ldto_init
)
748 result
= ldt
->ldt_ops
->ldto_init(ldt
);
750 list_add(&ldt
->ldt_linkage
, &lu_device_types
);
753 EXPORT_SYMBOL(lu_device_type_init
);
755 void lu_device_type_fini(struct lu_device_type
*ldt
)
757 list_del_init(&ldt
->ldt_linkage
);
758 if (ldt
->ldt_ops
->ldto_fini
)
759 ldt
->ldt_ops
->ldto_fini(ldt
);
761 EXPORT_SYMBOL(lu_device_type_fini
);
763 void lu_types_stop(void)
765 struct lu_device_type
*ldt
;
767 list_for_each_entry(ldt
, &lu_device_types
, ldt_linkage
) {
768 if (ldt
->ldt_device_nr
== 0 && ldt
->ldt_ops
->ldto_stop
)
769 ldt
->ldt_ops
->ldto_stop(ldt
);
772 EXPORT_SYMBOL(lu_types_stop
);
775 * Global list of all sites on this node
777 static LIST_HEAD(lu_sites
);
778 static DEFINE_MUTEX(lu_sites_guard
);
781 * Global environment used by site shrinker.
783 static struct lu_env lu_shrink_env
;
785 struct lu_site_print_arg
{
786 struct lu_env
*lsp_env
;
788 lu_printer_t lsp_printer
;
792 lu_site_obj_print(cfs_hash_t
*hs
, cfs_hash_bd_t
*bd
,
793 struct hlist_node
*hnode
, void *data
)
795 struct lu_site_print_arg
*arg
= (struct lu_site_print_arg
*)data
;
796 struct lu_object_header
*h
;
798 h
= hlist_entry(hnode
, struct lu_object_header
, loh_hash
);
799 if (!list_empty(&h
->loh_layers
)) {
800 const struct lu_object
*o
;
802 o
= lu_object_top(h
);
803 lu_object_print(arg
->lsp_env
, arg
->lsp_cookie
,
804 arg
->lsp_printer
, o
);
806 lu_object_header_print(arg
->lsp_env
, arg
->lsp_cookie
,
807 arg
->lsp_printer
, h
);
813 * Print all objects in \a s.
815 void lu_site_print(const struct lu_env
*env
, struct lu_site
*s
, void *cookie
,
816 lu_printer_t printer
)
818 struct lu_site_print_arg arg
= {
819 .lsp_env
= (struct lu_env
*)env
,
820 .lsp_cookie
= cookie
,
821 .lsp_printer
= printer
,
824 cfs_hash_for_each(s
->ls_obj_hash
, lu_site_obj_print
, &arg
);
826 EXPORT_SYMBOL(lu_site_print
);
829 LU_CACHE_PERCENT_MAX
= 50,
830 LU_CACHE_PERCENT_DEFAULT
= 20
833 static unsigned int lu_cache_percent
= LU_CACHE_PERCENT_DEFAULT
;
834 CFS_MODULE_PARM(lu_cache_percent
, "i", int, 0644,
835 "Percentage of memory to be used as lu_object cache");
838 * Return desired hash table order.
840 static int lu_htable_order(void)
842 unsigned long cache_size
;
846 * Calculate hash table size, assuming that we want reasonable
847 * performance when 20% of total memory is occupied by cache of
850 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
852 cache_size
= totalram_pages
;
854 #if BITS_PER_LONG == 32
855 /* limit hashtable size for lowmem systems to low RAM */
856 if (cache_size
> 1 << (30 - PAGE_CACHE_SHIFT
))
857 cache_size
= 1 << (30 - PAGE_CACHE_SHIFT
) * 3 / 4;
860 /* clear off unreasonable cache setting. */
861 if (lu_cache_percent
== 0 || lu_cache_percent
> LU_CACHE_PERCENT_MAX
) {
862 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
863 " the range of (0, %u]. Will use default value: %u.\n",
864 lu_cache_percent
, LU_CACHE_PERCENT_MAX
,
865 LU_CACHE_PERCENT_DEFAULT
);
867 lu_cache_percent
= LU_CACHE_PERCENT_DEFAULT
;
869 cache_size
= cache_size
/ 100 * lu_cache_percent
*
870 (PAGE_CACHE_SIZE
/ 1024);
872 for (bits
= 1; (1 << bits
) < cache_size
; ++bits
) {
878 static unsigned lu_obj_hop_hash(cfs_hash_t
*hs
,
879 const void *key
, unsigned mask
)
881 struct lu_fid
*fid
= (struct lu_fid
*)key
;
884 hash
= fid_flatten32(fid
);
885 hash
+= (hash
>> 4) + (hash
<< 12); /* mixing oid and seq */
886 hash
= cfs_hash_long(hash
, hs
->hs_bkt_bits
);
888 /* give me another random factor */
889 hash
-= cfs_hash_long((unsigned long)hs
, fid_oid(fid
) % 11 + 3);
891 hash
<<= hs
->hs_cur_bits
- hs
->hs_bkt_bits
;
892 hash
|= (fid_seq(fid
) + fid_oid(fid
)) & (CFS_HASH_NBKT(hs
) - 1);
897 static void *lu_obj_hop_object(struct hlist_node
*hnode
)
899 return hlist_entry(hnode
, struct lu_object_header
, loh_hash
);
902 static void *lu_obj_hop_key(struct hlist_node
*hnode
)
904 struct lu_object_header
*h
;
906 h
= hlist_entry(hnode
, struct lu_object_header
, loh_hash
);
910 static int lu_obj_hop_keycmp(const void *key
, struct hlist_node
*hnode
)
912 struct lu_object_header
*h
;
914 h
= hlist_entry(hnode
, struct lu_object_header
, loh_hash
);
915 return lu_fid_eq(&h
->loh_fid
, (struct lu_fid
*)key
);
918 static void lu_obj_hop_get(cfs_hash_t
*hs
, struct hlist_node
*hnode
)
920 struct lu_object_header
*h
;
922 h
= hlist_entry(hnode
, struct lu_object_header
, loh_hash
);
923 if (atomic_add_return(1, &h
->loh_ref
) == 1) {
924 struct lu_site_bkt_data
*bkt
;
927 cfs_hash_bd_get(hs
, &h
->loh_fid
, &bd
);
928 bkt
= cfs_hash_bd_extra_get(hs
, &bd
);
933 static void lu_obj_hop_put_locked(cfs_hash_t
*hs
, struct hlist_node
*hnode
)
935 LBUG(); /* we should never called it */
938 cfs_hash_ops_t lu_site_hash_ops
= {
939 .hs_hash
= lu_obj_hop_hash
,
940 .hs_key
= lu_obj_hop_key
,
941 .hs_keycmp
= lu_obj_hop_keycmp
,
942 .hs_object
= lu_obj_hop_object
,
943 .hs_get
= lu_obj_hop_get
,
944 .hs_put_locked
= lu_obj_hop_put_locked
,
947 void lu_dev_add_linkage(struct lu_site
*s
, struct lu_device
*d
)
949 spin_lock(&s
->ls_ld_lock
);
950 if (list_empty(&d
->ld_linkage
))
951 list_add(&d
->ld_linkage
, &s
->ls_ld_linkage
);
952 spin_unlock(&s
->ls_ld_lock
);
954 EXPORT_SYMBOL(lu_dev_add_linkage
);
956 void lu_dev_del_linkage(struct lu_site
*s
, struct lu_device
*d
)
958 spin_lock(&s
->ls_ld_lock
);
959 list_del_init(&d
->ld_linkage
);
960 spin_unlock(&s
->ls_ld_lock
);
962 EXPORT_SYMBOL(lu_dev_del_linkage
);
965 * Initialize site \a s, with \a d as the top level device.
967 #define LU_SITE_BITS_MIN 12
968 #define LU_SITE_BITS_MAX 24
970 * total 256 buckets, we don't want too many buckets because:
971 * - consume too much memory
972 * - avoid unbalanced LRU list
974 #define LU_SITE_BKT_BITS 8
976 int lu_site_init(struct lu_site
*s
, struct lu_device
*top
)
978 struct lu_site_bkt_data
*bkt
;
985 memset(s
, 0, sizeof *s
);
986 bits
= lu_htable_order();
987 snprintf(name
, 16, "lu_site_%s", top
->ld_type
->ldt_name
);
988 for (bits
= min(max(LU_SITE_BITS_MIN
, bits
), LU_SITE_BITS_MAX
);
989 bits
>= LU_SITE_BITS_MIN
; bits
--) {
990 s
->ls_obj_hash
= cfs_hash_create(name
, bits
, bits
,
991 bits
- LU_SITE_BKT_BITS
,
994 CFS_HASH_SPIN_BKTLOCK
|
995 CFS_HASH_NO_ITEMREF
|
997 CFS_HASH_ASSERT_EMPTY
);
998 if (s
->ls_obj_hash
!= NULL
)
1002 if (s
->ls_obj_hash
== NULL
) {
1003 CERROR("failed to create lu_site hash with bits: %d\n", bits
);
1007 cfs_hash_for_each_bucket(s
->ls_obj_hash
, &bd
, i
) {
1008 bkt
= cfs_hash_bd_extra_get(s
->ls_obj_hash
, &bd
);
1009 INIT_LIST_HEAD(&bkt
->lsb_lru
);
1010 init_waitqueue_head(&bkt
->lsb_marche_funebre
);
1013 s
->ls_stats
= lprocfs_alloc_stats(LU_SS_LAST_STAT
, 0);
1014 if (s
->ls_stats
== NULL
) {
1015 cfs_hash_putref(s
->ls_obj_hash
);
1016 s
->ls_obj_hash
= NULL
;
1020 lprocfs_counter_init(s
->ls_stats
, LU_SS_CREATED
,
1021 0, "created", "created");
1022 lprocfs_counter_init(s
->ls_stats
, LU_SS_CACHE_HIT
,
1023 0, "cache_hit", "cache_hit");
1024 lprocfs_counter_init(s
->ls_stats
, LU_SS_CACHE_MISS
,
1025 0, "cache_miss", "cache_miss");
1026 lprocfs_counter_init(s
->ls_stats
, LU_SS_CACHE_RACE
,
1027 0, "cache_race", "cache_race");
1028 lprocfs_counter_init(s
->ls_stats
, LU_SS_CACHE_DEATH_RACE
,
1029 0, "cache_death_race", "cache_death_race");
1030 lprocfs_counter_init(s
->ls_stats
, LU_SS_LRU_PURGED
,
1031 0, "lru_purged", "lru_purged");
1033 INIT_LIST_HEAD(&s
->ls_linkage
);
1034 s
->ls_top_dev
= top
;
1037 lu_ref_add(&top
->ld_reference
, "site-top", s
);
1039 INIT_LIST_HEAD(&s
->ls_ld_linkage
);
1040 spin_lock_init(&s
->ls_ld_lock
);
1042 lu_dev_add_linkage(s
, top
);
1046 EXPORT_SYMBOL(lu_site_init
);
1049 * Finalize \a s and release its resources.
1051 void lu_site_fini(struct lu_site
*s
)
1053 mutex_lock(&lu_sites_guard
);
1054 list_del_init(&s
->ls_linkage
);
1055 mutex_unlock(&lu_sites_guard
);
1057 if (s
->ls_obj_hash
!= NULL
) {
1058 cfs_hash_putref(s
->ls_obj_hash
);
1059 s
->ls_obj_hash
= NULL
;
1062 if (s
->ls_top_dev
!= NULL
) {
1063 s
->ls_top_dev
->ld_site
= NULL
;
1064 lu_ref_del(&s
->ls_top_dev
->ld_reference
, "site-top", s
);
1065 lu_device_put(s
->ls_top_dev
);
1066 s
->ls_top_dev
= NULL
;
1069 if (s
->ls_stats
!= NULL
)
1070 lprocfs_free_stats(&s
->ls_stats
);
1072 EXPORT_SYMBOL(lu_site_fini
);
1075 * Called when initialization of stack for this site is completed.
1077 int lu_site_init_finish(struct lu_site
*s
)
1080 mutex_lock(&lu_sites_guard
);
1081 result
= lu_context_refill(&lu_shrink_env
.le_ctx
);
1083 list_add(&s
->ls_linkage
, &lu_sites
);
1084 mutex_unlock(&lu_sites_guard
);
1087 EXPORT_SYMBOL(lu_site_init_finish
);
1090 * Acquire additional reference on device \a d
1092 void lu_device_get(struct lu_device
*d
)
1094 atomic_inc(&d
->ld_ref
);
1096 EXPORT_SYMBOL(lu_device_get
);
1099 * Release reference on device \a d.
1101 void lu_device_put(struct lu_device
*d
)
1103 LASSERT(atomic_read(&d
->ld_ref
) > 0);
1104 atomic_dec(&d
->ld_ref
);
1106 EXPORT_SYMBOL(lu_device_put
);
1109 * Initialize device \a d of type \a t.
1111 int lu_device_init(struct lu_device
*d
, struct lu_device_type
*t
)
1113 if (t
->ldt_device_nr
++ == 0 && t
->ldt_ops
->ldto_start
!= NULL
)
1114 t
->ldt_ops
->ldto_start(t
);
1115 memset(d
, 0, sizeof *d
);
1116 atomic_set(&d
->ld_ref
, 0);
1118 lu_ref_init(&d
->ld_reference
);
1119 INIT_LIST_HEAD(&d
->ld_linkage
);
1122 EXPORT_SYMBOL(lu_device_init
);
1125 * Finalize device \a d.
1127 void lu_device_fini(struct lu_device
*d
)
1129 struct lu_device_type
*t
;
1132 if (d
->ld_obd
!= NULL
) {
1133 d
->ld_obd
->obd_lu_dev
= NULL
;
1137 lu_ref_fini(&d
->ld_reference
);
1138 LASSERTF(atomic_read(&d
->ld_ref
) == 0,
1139 "Refcount is %u\n", atomic_read(&d
->ld_ref
));
1140 LASSERT(t
->ldt_device_nr
> 0);
1141 if (--t
->ldt_device_nr
== 0 && t
->ldt_ops
->ldto_stop
!= NULL
)
1142 t
->ldt_ops
->ldto_stop(t
);
1144 EXPORT_SYMBOL(lu_device_fini
);
1147 * Initialize object \a o that is part of compound object \a h and was created
1150 int lu_object_init(struct lu_object
*o
, struct lu_object_header
*h
,
1151 struct lu_device
*d
)
1153 memset(o
, 0, sizeof(*o
));
1157 lu_ref_add_at(&d
->ld_reference
, &o
->lo_dev_ref
, "lu_object", o
);
1158 INIT_LIST_HEAD(&o
->lo_linkage
);
1162 EXPORT_SYMBOL(lu_object_init
);
1165 * Finalize object and release its resources.
1167 void lu_object_fini(struct lu_object
*o
)
1169 struct lu_device
*dev
= o
->lo_dev
;
1171 LASSERT(list_empty(&o
->lo_linkage
));
1174 lu_ref_del_at(&dev
->ld_reference
, &o
->lo_dev_ref
,
1180 EXPORT_SYMBOL(lu_object_fini
);
1183 * Add object \a o as first layer of compound object \a h
1185 * This is typically called by the ->ldo_object_alloc() method of top-level
1188 void lu_object_add_top(struct lu_object_header
*h
, struct lu_object
*o
)
1190 list_move(&o
->lo_linkage
, &h
->loh_layers
);
1192 EXPORT_SYMBOL(lu_object_add_top
);
1195 * Add object \a o as a layer of compound object, going after \a before.
1197 * This is typically called by the ->ldo_object_alloc() method of \a
1200 void lu_object_add(struct lu_object
*before
, struct lu_object
*o
)
1202 list_move(&o
->lo_linkage
, &before
->lo_linkage
);
1204 EXPORT_SYMBOL(lu_object_add
);
1207 * Initialize compound object.
1209 int lu_object_header_init(struct lu_object_header
*h
)
1211 memset(h
, 0, sizeof *h
);
1212 atomic_set(&h
->loh_ref
, 1);
1213 INIT_HLIST_NODE(&h
->loh_hash
);
1214 INIT_LIST_HEAD(&h
->loh_lru
);
1215 INIT_LIST_HEAD(&h
->loh_layers
);
1216 lu_ref_init(&h
->loh_reference
);
1219 EXPORT_SYMBOL(lu_object_header_init
);
1222 * Finalize compound object.
1224 void lu_object_header_fini(struct lu_object_header
*h
)
1226 LASSERT(list_empty(&h
->loh_layers
));
1227 LASSERT(list_empty(&h
->loh_lru
));
1228 LASSERT(hlist_unhashed(&h
->loh_hash
));
1229 lu_ref_fini(&h
->loh_reference
);
1231 EXPORT_SYMBOL(lu_object_header_fini
);
1234 * Given a compound object, find its slice, corresponding to the device type
1237 struct lu_object
*lu_object_locate(struct lu_object_header
*h
,
1238 const struct lu_device_type
*dtype
)
1240 struct lu_object
*o
;
1242 list_for_each_entry(o
, &h
->loh_layers
, lo_linkage
) {
1243 if (o
->lo_dev
->ld_type
== dtype
)
1248 EXPORT_SYMBOL(lu_object_locate
);
1253 * Finalize and free devices in the device stack.
1255 * Finalize device stack by purging object cache, and calling
1256 * lu_device_type_operations::ldto_device_fini() and
1257 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1259 void lu_stack_fini(const struct lu_env
*env
, struct lu_device
*top
)
1261 struct lu_site
*site
= top
->ld_site
;
1262 struct lu_device
*scan
;
1263 struct lu_device
*next
;
1265 lu_site_purge(env
, site
, ~0);
1266 for (scan
= top
; scan
!= NULL
; scan
= next
) {
1267 next
= scan
->ld_type
->ldt_ops
->ldto_device_fini(env
, scan
);
1268 lu_ref_del(&scan
->ld_reference
, "lu-stack", &lu_site_init
);
1269 lu_device_put(scan
);
1273 lu_site_purge(env
, site
, ~0);
1275 for (scan
= top
; scan
!= NULL
; scan
= next
) {
1276 const struct lu_device_type
*ldt
= scan
->ld_type
;
1277 struct obd_type
*type
;
1279 next
= ldt
->ldt_ops
->ldto_device_free(env
, scan
);
1280 type
= ldt
->ldt_obd_type
;
1283 class_put_type(type
);
1287 EXPORT_SYMBOL(lu_stack_fini
);
1291 * Maximal number of tld slots.
1293 LU_CONTEXT_KEY_NR
= 40
1296 static struct lu_context_key
*lu_keys
[LU_CONTEXT_KEY_NR
] = { NULL
, };
1298 static DEFINE_SPINLOCK(lu_keys_guard
);
1301 * Global counter incremented whenever key is registered, unregistered,
1302 * revived or quiesced. This is used to void unnecessary calls to
1303 * lu_context_refill(). No locking is provided, as initialization and shutdown
1304 * are supposed to be externally serialized.
1306 static unsigned key_set_version
= 0;
1311 int lu_context_key_register(struct lu_context_key
*key
)
1316 LASSERT(key
->lct_init
!= NULL
);
1317 LASSERT(key
->lct_fini
!= NULL
);
1318 LASSERT(key
->lct_tags
!= 0);
1321 spin_lock(&lu_keys_guard
);
1322 for (i
= 0; i
< ARRAY_SIZE(lu_keys
); ++i
) {
1323 if (lu_keys
[i
] == NULL
) {
1325 atomic_set(&key
->lct_used
, 1);
1327 lu_ref_init(&key
->lct_reference
);
1333 spin_unlock(&lu_keys_guard
);
1336 EXPORT_SYMBOL(lu_context_key_register
);
1338 static void key_fini(struct lu_context
*ctx
, int index
)
1340 if (ctx
->lc_value
!= NULL
&& ctx
->lc_value
[index
] != NULL
) {
1341 struct lu_context_key
*key
;
1343 key
= lu_keys
[index
];
1344 LASSERT(key
!= NULL
);
1345 LASSERT(key
->lct_fini
!= NULL
);
1346 LASSERT(atomic_read(&key
->lct_used
) > 1);
1348 key
->lct_fini(ctx
, key
, ctx
->lc_value
[index
]);
1349 lu_ref_del(&key
->lct_reference
, "ctx", ctx
);
1350 atomic_dec(&key
->lct_used
);
1352 if ((ctx
->lc_tags
& LCT_NOREF
) == 0) {
1353 #ifdef CONFIG_MODULE_UNLOAD
1354 LINVRNT(module_refcount(key
->lct_owner
) > 0);
1356 module_put(key
->lct_owner
);
1358 ctx
->lc_value
[index
] = NULL
;
1365 void lu_context_key_degister(struct lu_context_key
*key
)
1367 LASSERT(atomic_read(&key
->lct_used
) >= 1);
1368 LINVRNT(0 <= key
->lct_index
&& key
->lct_index
< ARRAY_SIZE(lu_keys
));
1370 lu_context_key_quiesce(key
);
1373 spin_lock(&lu_keys_guard
);
1374 key_fini(&lu_shrink_env
.le_ctx
, key
->lct_index
);
1375 if (lu_keys
[key
->lct_index
]) {
1376 lu_keys
[key
->lct_index
] = NULL
;
1377 lu_ref_fini(&key
->lct_reference
);
1379 spin_unlock(&lu_keys_guard
);
1381 LASSERTF(atomic_read(&key
->lct_used
) == 1,
1382 "key has instances: %d\n",
1383 atomic_read(&key
->lct_used
));
1385 EXPORT_SYMBOL(lu_context_key_degister
);
1388 * Register a number of keys. This has to be called after all keys have been
1389 * initialized by a call to LU_CONTEXT_KEY_INIT().
1391 int lu_context_key_register_many(struct lu_context_key
*k
, ...)
1393 struct lu_context_key
*key
= k
;
1399 result
= lu_context_key_register(key
);
1402 key
= va_arg(args
, struct lu_context_key
*);
1403 } while (key
!= NULL
);
1409 lu_context_key_degister(k
);
1410 k
= va_arg(args
, struct lu_context_key
*);
1417 EXPORT_SYMBOL(lu_context_key_register_many
);
1420 * De-register a number of keys. This is a dual to
1421 * lu_context_key_register_many().
1423 void lu_context_key_degister_many(struct lu_context_key
*k
, ...)
1429 lu_context_key_degister(k
);
1430 k
= va_arg(args
, struct lu_context_key
*);
1431 } while (k
!= NULL
);
1434 EXPORT_SYMBOL(lu_context_key_degister_many
);
1437 * Revive a number of keys.
1439 void lu_context_key_revive_many(struct lu_context_key
*k
, ...)
1445 lu_context_key_revive(k
);
1446 k
= va_arg(args
, struct lu_context_key
*);
1447 } while (k
!= NULL
);
1450 EXPORT_SYMBOL(lu_context_key_revive_many
);
1453 * Quiescent a number of keys.
1455 void lu_context_key_quiesce_many(struct lu_context_key
*k
, ...)
1461 lu_context_key_quiesce(k
);
1462 k
= va_arg(args
, struct lu_context_key
*);
1463 } while (k
!= NULL
);
1466 EXPORT_SYMBOL(lu_context_key_quiesce_many
);
1469 * Return value associated with key \a key in context \a ctx.
1471 void *lu_context_key_get(const struct lu_context
*ctx
,
1472 const struct lu_context_key
*key
)
1474 LINVRNT(ctx
->lc_state
== LCS_ENTERED
);
1475 LINVRNT(0 <= key
->lct_index
&& key
->lct_index
< ARRAY_SIZE(lu_keys
));
1476 LASSERT(lu_keys
[key
->lct_index
] == key
);
1477 return ctx
->lc_value
[key
->lct_index
];
1479 EXPORT_SYMBOL(lu_context_key_get
);
1482 * List of remembered contexts. XXX document me.
1484 static LIST_HEAD(lu_context_remembered
);
1487 * Destroy \a key in all remembered contexts. This is used to destroy key
1488 * values in "shared" contexts (like service threads), when a module owning
1489 * the key is about to be unloaded.
1491 void lu_context_key_quiesce(struct lu_context_key
*key
)
1493 struct lu_context
*ctx
;
1495 if (!(key
->lct_tags
& LCT_QUIESCENT
)) {
1497 * XXX layering violation.
1499 key
->lct_tags
|= LCT_QUIESCENT
;
1501 * XXX memory barrier has to go here.
1503 spin_lock(&lu_keys_guard
);
1504 list_for_each_entry(ctx
, &lu_context_remembered
,
1506 key_fini(ctx
, key
->lct_index
);
1507 spin_unlock(&lu_keys_guard
);
1511 EXPORT_SYMBOL(lu_context_key_quiesce
);
1513 void lu_context_key_revive(struct lu_context_key
*key
)
1515 key
->lct_tags
&= ~LCT_QUIESCENT
;
1518 EXPORT_SYMBOL(lu_context_key_revive
);
1520 static void keys_fini(struct lu_context
*ctx
)
1524 if (ctx
->lc_value
== NULL
)
1527 for (i
= 0; i
< ARRAY_SIZE(lu_keys
); ++i
)
1530 OBD_FREE(ctx
->lc_value
, ARRAY_SIZE(lu_keys
) * sizeof ctx
->lc_value
[0]);
1531 ctx
->lc_value
= NULL
;
1534 static int keys_fill(struct lu_context
*ctx
)
1538 LINVRNT(ctx
->lc_value
!= NULL
);
1539 for (i
= 0; i
< ARRAY_SIZE(lu_keys
); ++i
) {
1540 struct lu_context_key
*key
;
1543 if (ctx
->lc_value
[i
] == NULL
&& key
!= NULL
&&
1544 (key
->lct_tags
& ctx
->lc_tags
) &&
1546 * Don't create values for a LCT_QUIESCENT key, as this
1547 * will pin module owning a key.
1549 !(key
->lct_tags
& LCT_QUIESCENT
)) {
1552 LINVRNT(key
->lct_init
!= NULL
);
1553 LINVRNT(key
->lct_index
== i
);
1555 value
= key
->lct_init(ctx
, key
);
1556 if (unlikely(IS_ERR(value
)))
1557 return PTR_ERR(value
);
1559 if (!(ctx
->lc_tags
& LCT_NOREF
))
1560 try_module_get(key
->lct_owner
);
1561 lu_ref_add_atomic(&key
->lct_reference
, "ctx", ctx
);
1562 atomic_inc(&key
->lct_used
);
1564 * This is the only place in the code, where an
1565 * element of ctx->lc_value[] array is set to non-NULL
1568 ctx
->lc_value
[i
] = value
;
1569 if (key
->lct_exit
!= NULL
)
1570 ctx
->lc_tags
|= LCT_HAS_EXIT
;
1572 ctx
->lc_version
= key_set_version
;
1577 static int keys_init(struct lu_context
*ctx
)
1579 OBD_ALLOC(ctx
->lc_value
, ARRAY_SIZE(lu_keys
) * sizeof ctx
->lc_value
[0]);
1580 if (likely(ctx
->lc_value
!= NULL
))
1581 return keys_fill(ctx
);
1587 * Initialize context data-structure. Create values for all keys.
1589 int lu_context_init(struct lu_context
*ctx
, __u32 tags
)
1593 memset(ctx
, 0, sizeof *ctx
);
1594 ctx
->lc_state
= LCS_INITIALIZED
;
1595 ctx
->lc_tags
= tags
;
1596 if (tags
& LCT_REMEMBER
) {
1597 spin_lock(&lu_keys_guard
);
1598 list_add(&ctx
->lc_remember
, &lu_context_remembered
);
1599 spin_unlock(&lu_keys_guard
);
1601 INIT_LIST_HEAD(&ctx
->lc_remember
);
1604 rc
= keys_init(ctx
);
1606 lu_context_fini(ctx
);
1610 EXPORT_SYMBOL(lu_context_init
);
1613 * Finalize context data-structure. Destroy key values.
1615 void lu_context_fini(struct lu_context
*ctx
)
1617 LINVRNT(ctx
->lc_state
== LCS_INITIALIZED
|| ctx
->lc_state
== LCS_LEFT
);
1618 ctx
->lc_state
= LCS_FINALIZED
;
1620 if ((ctx
->lc_tags
& LCT_REMEMBER
) == 0) {
1621 LASSERT(list_empty(&ctx
->lc_remember
));
1624 } else { /* could race with key degister */
1625 spin_lock(&lu_keys_guard
);
1627 list_del_init(&ctx
->lc_remember
);
1628 spin_unlock(&lu_keys_guard
);
1631 EXPORT_SYMBOL(lu_context_fini
);
1634 * Called before entering context.
1636 void lu_context_enter(struct lu_context
*ctx
)
1638 LINVRNT(ctx
->lc_state
== LCS_INITIALIZED
|| ctx
->lc_state
== LCS_LEFT
);
1639 ctx
->lc_state
= LCS_ENTERED
;
1641 EXPORT_SYMBOL(lu_context_enter
);
1644 * Called after exiting from \a ctx
1646 void lu_context_exit(struct lu_context
*ctx
)
1650 LINVRNT(ctx
->lc_state
== LCS_ENTERED
);
1651 ctx
->lc_state
= LCS_LEFT
;
1652 if (ctx
->lc_tags
& LCT_HAS_EXIT
&& ctx
->lc_value
!= NULL
) {
1653 for (i
= 0; i
< ARRAY_SIZE(lu_keys
); ++i
) {
1654 if (ctx
->lc_value
[i
] != NULL
) {
1655 struct lu_context_key
*key
;
1658 LASSERT(key
!= NULL
);
1659 if (key
->lct_exit
!= NULL
)
1661 key
, ctx
->lc_value
[i
]);
1666 EXPORT_SYMBOL(lu_context_exit
);
1669 * Allocate for context all missing keys that were registered after context
1670 * creation. key_set_version is only changed in rare cases when modules
1671 * are loaded and removed.
1673 int lu_context_refill(struct lu_context
*ctx
)
1675 return likely(ctx
->lc_version
== key_set_version
) ? 0 : keys_fill(ctx
);
1677 EXPORT_SYMBOL(lu_context_refill
);
1680 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1681 * obd being added. Currently, this is only used on client side, specifically
1682 * for echo device client, for other stack (like ptlrpc threads), context are
1683 * predefined when the lu_device type are registered, during the module probe
1686 __u32 lu_context_tags_default
= 0;
1687 __u32 lu_session_tags_default
= 0;
1689 void lu_context_tags_update(__u32 tags
)
1691 spin_lock(&lu_keys_guard
);
1692 lu_context_tags_default
|= tags
;
1694 spin_unlock(&lu_keys_guard
);
1696 EXPORT_SYMBOL(lu_context_tags_update
);
1698 void lu_context_tags_clear(__u32 tags
)
1700 spin_lock(&lu_keys_guard
);
1701 lu_context_tags_default
&= ~tags
;
1703 spin_unlock(&lu_keys_guard
);
1705 EXPORT_SYMBOL(lu_context_tags_clear
);
1707 void lu_session_tags_update(__u32 tags
)
1709 spin_lock(&lu_keys_guard
);
1710 lu_session_tags_default
|= tags
;
1712 spin_unlock(&lu_keys_guard
);
1714 EXPORT_SYMBOL(lu_session_tags_update
);
1716 void lu_session_tags_clear(__u32 tags
)
1718 spin_lock(&lu_keys_guard
);
1719 lu_session_tags_default
&= ~tags
;
1721 spin_unlock(&lu_keys_guard
);
1723 EXPORT_SYMBOL(lu_session_tags_clear
);
1725 int lu_env_init(struct lu_env
*env
, __u32 tags
)
1730 result
= lu_context_init(&env
->le_ctx
, tags
);
1731 if (likely(result
== 0))
1732 lu_context_enter(&env
->le_ctx
);
1735 EXPORT_SYMBOL(lu_env_init
);
1737 void lu_env_fini(struct lu_env
*env
)
1739 lu_context_exit(&env
->le_ctx
);
1740 lu_context_fini(&env
->le_ctx
);
1743 EXPORT_SYMBOL(lu_env_fini
);
1745 int lu_env_refill(struct lu_env
*env
)
1749 result
= lu_context_refill(&env
->le_ctx
);
1750 if (result
== 0 && env
->le_ses
!= NULL
)
1751 result
= lu_context_refill(env
->le_ses
);
1754 EXPORT_SYMBOL(lu_env_refill
);
1757 * Currently, this API will only be used by echo client.
1758 * Because echo client and normal lustre client will share
1759 * same cl_env cache. So echo client needs to refresh
1760 * the env context after it get one from the cache, especially
1761 * when normal client and echo client co-exist in the same client.
1763 int lu_env_refill_by_tags(struct lu_env
*env
, __u32 ctags
,
1768 if ((env
->le_ctx
.lc_tags
& ctags
) != ctags
) {
1769 env
->le_ctx
.lc_version
= 0;
1770 env
->le_ctx
.lc_tags
|= ctags
;
1773 if (env
->le_ses
&& (env
->le_ses
->lc_tags
& stags
) != stags
) {
1774 env
->le_ses
->lc_version
= 0;
1775 env
->le_ses
->lc_tags
|= stags
;
1778 result
= lu_env_refill(env
);
1782 EXPORT_SYMBOL(lu_env_refill_by_tags
);
1784 static struct shrinker
*lu_site_shrinker
= NULL
;
1786 typedef struct lu_site_stats
{
1787 unsigned lss_populated
;
1788 unsigned lss_max_search
;
1793 static void lu_site_stats_get(cfs_hash_t
*hs
,
1794 lu_site_stats_t
*stats
, int populated
)
1799 cfs_hash_for_each_bucket(hs
, &bd
, i
) {
1800 struct lu_site_bkt_data
*bkt
= cfs_hash_bd_extra_get(hs
, &bd
);
1801 struct hlist_head
*hhead
;
1803 cfs_hash_bd_lock(hs
, &bd
, 1);
1804 stats
->lss_busy
+= bkt
->lsb_busy
;
1805 stats
->lss_total
+= cfs_hash_bd_count_get(&bd
);
1806 stats
->lss_max_search
= max((int)stats
->lss_max_search
,
1807 cfs_hash_bd_depmax_get(&bd
));
1809 cfs_hash_bd_unlock(hs
, &bd
, 1);
1813 cfs_hash_bd_for_each_hlist(hs
, &bd
, hhead
) {
1814 if (!hlist_empty(hhead
))
1815 stats
->lss_populated
++;
1817 cfs_hash_bd_unlock(hs
, &bd
, 1);
1823 * There exists a potential lock inversion deadlock scenario when using
1824 * Lustre on top of ZFS. This occurs between one of ZFS's
1825 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1826 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1827 * while thread B will take the ht_lock and sleep on the lu_sites_guard
1828 * lock. Obviously neither thread will wake and drop their respective hold
1831 * To prevent this from happening we must ensure the lu_sites_guard lock is
1832 * not taken while down this code path. ZFS reliably does not set the
1833 * __GFP_FS bit in its code paths, so this can be used to determine if it
1834 * is safe to take the lu_sites_guard lock.
1836 * Ideally we should accurately return the remaining number of cached
1837 * objects without taking the lu_sites_guard lock, but this is not
1838 * possible in the current implementation.
1840 static int lu_cache_shrink(SHRINKER_ARGS(sc
, nr_to_scan
, gfp_mask
))
1842 lu_site_stats_t stats
;
1844 struct lu_site
*tmp
;
1846 int remain
= shrink_param(sc
, nr_to_scan
);
1849 if (!(shrink_param(sc
, gfp_mask
) & __GFP_FS
)) {
1853 /* We must not take the lu_sites_guard lock when
1854 * __GFP_FS is *not* set because of the deadlock
1855 * possibility detailed above. Additionally,
1856 * since we cannot determine the number of
1857 * objects in the cache without taking this
1858 * lock, we're in a particularly tough spot. As
1859 * a result, we'll just lie and say our cache is
1860 * empty. This _should_ be ok, as we can't
1861 * reclaim objects when __GFP_FS is *not* set
1867 CDEBUG(D_INODE
, "Shrink %d objects\n", remain
);
1869 mutex_lock(&lu_sites_guard
);
1870 list_for_each_entry_safe(s
, tmp
, &lu_sites
, ls_linkage
) {
1871 if (shrink_param(sc
, nr_to_scan
) != 0) {
1872 remain
= lu_site_purge(&lu_shrink_env
, s
, remain
);
1874 * Move just shrunk site to the tail of site list to
1875 * assure shrinking fairness.
1877 list_move_tail(&s
->ls_linkage
, &splice
);
1880 memset(&stats
, 0, sizeof(stats
));
1881 lu_site_stats_get(s
->ls_obj_hash
, &stats
, 0);
1882 cached
+= stats
.lss_total
- stats
.lss_busy
;
1883 if (shrink_param(sc
, nr_to_scan
) && remain
<= 0)
1886 list_splice(&splice
, lu_sites
.prev
);
1887 mutex_unlock(&lu_sites_guard
);
1889 cached
= (cached
/ 100) * sysctl_vfs_cache_pressure
;
1890 if (shrink_param(sc
, nr_to_scan
) == 0)
1891 CDEBUG(D_INODE
, "%d objects cached\n", cached
);
1900 * Environment to be used in debugger, contains all tags.
1902 struct lu_env lu_debugging_env
;
1905 * Debugging printer function using printk().
1907 int lu_printk_printer(const struct lu_env
*env
,
1908 void *unused
, const char *format
, ...)
1912 va_start(args
, format
);
1913 vprintk(format
, args
);
1919 * Initialization of global lu_* data.
1921 int lu_global_init(void)
1925 CDEBUG(D_INFO
, "Lustre LU module (%p).\n", &lu_keys
);
1927 result
= lu_ref_global_init();
1931 LU_CONTEXT_KEY_INIT(&lu_global_key
);
1932 result
= lu_context_key_register(&lu_global_key
);
1937 * At this level, we don't know what tags are needed, so allocate them
1938 * conservatively. This should not be too bad, because this
1939 * environment is global.
1941 mutex_lock(&lu_sites_guard
);
1942 result
= lu_env_init(&lu_shrink_env
, LCT_SHRINKER
);
1943 mutex_unlock(&lu_sites_guard
);
1948 * seeks estimation: 3 seeks to read a record from oi, one to read
1949 * inode, one for ea. Unfortunately setting this high value results in
1950 * lu_object/inode cache consuming all the memory.
1952 lu_site_shrinker
= set_shrinker(DEFAULT_SEEKS
, lu_cache_shrink
);
1953 if (lu_site_shrinker
== NULL
)
1960 * Dual to lu_global_init().
1962 void lu_global_fini(void)
1964 if (lu_site_shrinker
!= NULL
) {
1965 remove_shrinker(lu_site_shrinker
);
1966 lu_site_shrinker
= NULL
;
1969 lu_context_key_degister(&lu_global_key
);
1972 * Tear shrinker environment down _after_ de-registering
1973 * lu_global_key, because the latter has a value in the former.
1975 mutex_lock(&lu_sites_guard
);
1976 lu_env_fini(&lu_shrink_env
);
1977 mutex_unlock(&lu_sites_guard
);
1979 lu_ref_global_fini();
1982 static __u32
ls_stats_read(struct lprocfs_stats
*stats
, int idx
)
1985 struct lprocfs_counter ret
;
1987 lprocfs_stats_collect(stats
, idx
, &ret
);
1988 return (__u32
)ret
.lc_count
;
1995 * Output site statistical counters into a buffer. Suitable for
1996 * lprocfs_rd_*()-style functions.
1998 int lu_site_stats_print(const struct lu_site
*s
, struct seq_file
*m
)
2000 lu_site_stats_t stats
;
2002 memset(&stats
, 0, sizeof(stats
));
2003 lu_site_stats_get(s
->ls_obj_hash
, &stats
, 1);
2005 return seq_printf(m
, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2008 stats
.lss_populated
,
2009 CFS_HASH_NHLIST(s
->ls_obj_hash
),
2010 stats
.lss_max_search
,
2011 ls_stats_read(s
->ls_stats
, LU_SS_CREATED
),
2012 ls_stats_read(s
->ls_stats
, LU_SS_CACHE_HIT
),
2013 ls_stats_read(s
->ls_stats
, LU_SS_CACHE_MISS
),
2014 ls_stats_read(s
->ls_stats
, LU_SS_CACHE_RACE
),
2015 ls_stats_read(s
->ls_stats
, LU_SS_CACHE_DEATH_RACE
),
2016 ls_stats_read(s
->ls_stats
, LU_SS_LRU_PURGED
));
2018 EXPORT_SYMBOL(lu_site_stats_print
);
2021 * Helper function to initialize a number of kmem slab caches at once.
2023 int lu_kmem_init(struct lu_kmem_descr
*caches
)
2026 struct lu_kmem_descr
*iter
= caches
;
2028 for (result
= 0; iter
->ckd_cache
!= NULL
; ++iter
) {
2029 *iter
->ckd_cache
= kmem_cache_create(iter
->ckd_name
,
2032 if (*iter
->ckd_cache
== NULL
) {
2034 /* free all previously allocated caches */
2035 lu_kmem_fini(caches
);
2041 EXPORT_SYMBOL(lu_kmem_init
);
2044 * Helper function to finalize a number of kmem slab cached at once. Dual to
2047 void lu_kmem_fini(struct lu_kmem_descr
*caches
)
2049 for (; caches
->ckd_cache
!= NULL
; ++caches
) {
2050 if (*caches
->ckd_cache
!= NULL
) {
2051 kmem_cache_destroy(*caches
->ckd_cache
);
2052 *caches
->ckd_cache
= NULL
;
2056 EXPORT_SYMBOL(lu_kmem_fini
);
2059 * Temporary solution to be able to assign fid in ->do_create()
2060 * till we have fully-functional OST fids
2062 void lu_object_assign_fid(const struct lu_env
*env
, struct lu_object
*o
,
2063 const struct lu_fid
*fid
)
2065 struct lu_site
*s
= o
->lo_dev
->ld_site
;
2066 struct lu_fid
*old
= &o
->lo_header
->loh_fid
;
2067 struct lu_site_bkt_data
*bkt
;
2068 struct lu_object
*shadow
;
2069 wait_queue_t waiter
;
2074 LASSERT(fid_is_zero(old
));
2076 hs
= s
->ls_obj_hash
;
2077 cfs_hash_bd_get_and_lock(hs
, (void *)fid
, &bd
, 1);
2078 shadow
= htable_lookup(s
, &bd
, fid
, &waiter
, &version
);
2079 /* supposed to be unique */
2080 LASSERT(IS_ERR(shadow
) && PTR_ERR(shadow
) == -ENOENT
);
2082 bkt
= cfs_hash_bd_extra_get(hs
, &bd
);
2083 cfs_hash_bd_add_locked(hs
, &bd
, &o
->lo_header
->loh_hash
);
2085 cfs_hash_bd_unlock(hs
, &bd
, 1);
2087 EXPORT_SYMBOL(lu_object_assign_fid
);
2090 * allocates object with 0 (non-assiged) fid
2091 * XXX: temporary solution to be able to assign fid in ->do_create()
2092 * till we have fully-functional OST fids
2094 struct lu_object
*lu_object_anon(const struct lu_env
*env
,
2095 struct lu_device
*dev
,
2096 const struct lu_object_conf
*conf
)
2099 struct lu_object
*o
;
2102 o
= lu_object_alloc(env
, dev
, &fid
, conf
);
2106 EXPORT_SYMBOL(lu_object_anon
);
2108 struct lu_buf LU_BUF_NULL
= {
2112 EXPORT_SYMBOL(LU_BUF_NULL
);
2114 void lu_buf_free(struct lu_buf
*buf
)
2118 LASSERT(buf
->lb_len
> 0);
2119 OBD_FREE_LARGE(buf
->lb_buf
, buf
->lb_len
);
2124 EXPORT_SYMBOL(lu_buf_free
);
2126 void lu_buf_alloc(struct lu_buf
*buf
, int size
)
2129 LASSERT(buf
->lb_buf
== NULL
);
2130 LASSERT(buf
->lb_len
== 0);
2131 OBD_ALLOC_LARGE(buf
->lb_buf
, size
);
2132 if (likely(buf
->lb_buf
))
2135 EXPORT_SYMBOL(lu_buf_alloc
);
2137 void lu_buf_realloc(struct lu_buf
*buf
, int size
)
2140 lu_buf_alloc(buf
, size
);
2142 EXPORT_SYMBOL(lu_buf_realloc
);
2144 struct lu_buf
*lu_buf_check_and_alloc(struct lu_buf
*buf
, int len
)
2146 if (buf
->lb_buf
== NULL
&& buf
->lb_len
== 0)
2147 lu_buf_alloc(buf
, len
);
2149 if ((len
> buf
->lb_len
) && (buf
->lb_buf
!= NULL
))
2150 lu_buf_realloc(buf
, len
);
2154 EXPORT_SYMBOL(lu_buf_check_and_alloc
);
2157 * Increase the size of the \a buf.
2158 * preserves old data in buffer
2159 * old buffer remains unchanged on error
2160 * \retval 0 or -ENOMEM
2162 int lu_buf_check_and_grow(struct lu_buf
*buf
, int len
)
2166 if (len
<= buf
->lb_len
)
2169 OBD_ALLOC_LARGE(ptr
, len
);
2173 /* Free the old buf */
2174 if (buf
->lb_buf
!= NULL
) {
2175 memcpy(ptr
, buf
->lb_buf
, buf
->lb_len
);
2176 OBD_FREE_LARGE(buf
->lb_buf
, buf
->lb_len
);
2183 EXPORT_SYMBOL(lu_buf_check_and_grow
);