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, 2013, 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>
50 # include <linux/module.h>
54 #include <libcfs/libcfs_hash.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
57 #include <lustre_disk.h>
58 #include <lustre_fid.h>
59 #include <lu_object.h>
61 #include <libcfs/list.h>
63 extern spinlock_t obd_types_lock;
66 LU_CACHE_PERCENT_MAX = 50,
67 LU_CACHE_PERCENT_DEFAULT = 20
70 #define LU_CACHE_NR_MAX_ADJUST 128
71 #define LU_CACHE_NR_UNLIMITED -1
72 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
73 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
74 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
75 #define LU_CACHE_NR_ZFS_LIMIT 10240
77 #define LU_SITE_BITS_MIN 12
78 #define LU_SITE_BITS_MAX 24
80 * total 256 buckets, we don't want too many buckets because:
81 * - consume too much memory
82 * - avoid unbalanced LRU list
84 #define LU_SITE_BKT_BITS 8
87 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
88 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
89 "Percentage of memory to be used as lu_object cache");
91 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
92 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
93 "Maximum number of objects in lu_object cache");
95 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
98 * Decrease reference counter on object. If last reference is freed, return
99 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
100 * case, free object immediately.
102 void lu_object_put(const struct lu_env *env, struct lu_object *o)
104 struct lu_site_bkt_data *bkt;
105 struct lu_object_header *top;
106 struct lu_site *site;
107 struct lu_object *orig;
109 const struct lu_fid *fid;
112 site = o->lo_dev->ld_site;
116 * till we have full fids-on-OST implemented anonymous objects
117 * are possible in OSP. such an object isn't listed in the site
118 * so we should not remove it from the site.
120 fid = lu_object_fid(o);
121 if (fid_is_zero(fid)) {
122 LASSERT(top->loh_hash.next == NULL
123 && top->loh_hash.pprev == NULL);
124 LASSERT(cfs_list_empty(&top->loh_lru));
125 if (!atomic_dec_and_test(&top->loh_ref))
127 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
128 if (o->lo_ops->loo_object_release != NULL)
129 o->lo_ops->loo_object_release(env, o);
131 lu_object_free(env, orig);
135 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
136 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
138 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
139 if (lu_object_is_dying(top)) {
142 * somebody may be waiting for this, currently only
143 * used for cl_object, see cl_object_put_last().
145 wake_up_all(&bkt->lsb_marche_funebre);
150 LASSERT(bkt->lsb_busy > 0);
153 * When last reference is released, iterate over object
154 * layers, and notify them that object is no longer busy.
156 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
157 if (o->lo_ops->loo_object_release != NULL)
158 o->lo_ops->loo_object_release(env, o);
161 if (!lu_object_is_dying(top)) {
162 LASSERT(cfs_list_empty(&top->loh_lru));
163 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
164 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
169 * If object is dying (will not be cached), removed it
170 * from hash table and LRU.
172 * This is done with hash table and LRU lists locked. As the only
173 * way to acquire first reference to previously unreferenced
174 * object is through hash-table lookup (lu_object_find()),
175 * or LRU scanning (lu_site_purge()), that are done under hash-table
176 * and LRU lock, no race with concurrent object lookup is possible
177 * and we can safely destroy object below.
179 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
180 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
181 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
183 * Object was already removed from hash and lru above, can
186 lu_object_free(env, orig);
188 EXPORT_SYMBOL(lu_object_put);
191 * Put object and don't keep in cache. This is temporary solution for
192 * multi-site objects when its layering is not constant.
194 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
196 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
197 return lu_object_put(env, o);
199 EXPORT_SYMBOL(lu_object_put_nocache);
202 * Kill the object and take it out of LRU cache.
203 * Currently used by client code for layout change.
205 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
207 struct lu_object_header *top;
210 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
211 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
212 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
215 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
216 cfs_list_del_init(&top->loh_lru);
217 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
218 cfs_hash_bd_unlock(obj_hash, &bd, 1);
221 EXPORT_SYMBOL(lu_object_unhash);
224 * Allocate new object.
226 * This follows object creation protocol, described in the comment within
227 * struct lu_device_operations definition.
229 static struct lu_object *lu_object_alloc(const struct lu_env *env,
230 struct lu_device *dev,
231 const struct lu_fid *f,
232 const struct lu_object_conf *conf)
234 struct lu_object *scan;
235 struct lu_object *top;
237 unsigned int init_mask = 0;
238 unsigned int init_flag;
244 * Create top-level object slice. This will also create
247 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
249 RETURN(ERR_PTR(-ENOMEM));
253 * This is the only place where object fid is assigned. It's constant
256 top->lo_header->loh_fid = *f;
257 layers = &top->lo_header->loh_layers;
261 * Call ->loo_object_init() repeatedly, until no more new
262 * object slices are created.
266 cfs_list_for_each_entry(scan, layers, lo_linkage) {
267 if (init_mask & init_flag)
270 scan->lo_header = top->lo_header;
271 result = scan->lo_ops->loo_object_init(env, scan, conf);
273 lu_object_free(env, top);
274 RETURN(ERR_PTR(result));
276 init_mask |= init_flag;
282 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
283 if (scan->lo_ops->loo_object_start != NULL) {
284 result = scan->lo_ops->loo_object_start(env, scan);
286 lu_object_free(env, top);
287 RETURN(ERR_PTR(result));
292 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
299 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
301 struct lu_site_bkt_data *bkt;
302 struct lu_site *site;
303 struct lu_object *scan;
307 site = o->lo_dev->ld_site;
308 layers = &o->lo_header->loh_layers;
309 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
311 * First call ->loo_object_delete() method to release all resources.
313 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
314 if (scan->lo_ops->loo_object_delete != NULL)
315 scan->lo_ops->loo_object_delete(env, scan);
319 * Then, splice object layers into stand-alone list, and call
320 * ->loo_object_free() on all layers to free memory. Splice is
321 * necessary, because lu_object_header is freed together with the
324 CFS_INIT_LIST_HEAD(&splice);
325 cfs_list_splice_init(layers, &splice);
326 while (!cfs_list_empty(&splice)) {
328 * Free layers in bottom-to-top order, so that object header
329 * lives as long as possible and ->loo_object_free() methods
330 * can look at its contents.
332 o = container_of0(splice.prev, struct lu_object, lo_linkage);
333 cfs_list_del_init(&o->lo_linkage);
334 LASSERT(o->lo_ops->loo_object_free != NULL);
335 o->lo_ops->loo_object_free(env, o);
338 if (waitqueue_active(&bkt->lsb_marche_funebre))
339 wake_up_all(&bkt->lsb_marche_funebre);
343 * Free \a nr objects from the cold end of the site LRU list.
345 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
347 struct lu_object_header *h;
348 struct lu_object_header *temp;
349 struct lu_site_bkt_data *bkt;
359 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
362 CFS_INIT_LIST_HEAD(&dispose);
364 * Under LRU list lock, scan LRU list and move unreferenced objects to
365 * the dispose list, removing them from LRU and hash table.
367 start = s->ls_purge_start;
368 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
371 * It doesn't make any sense to make purge threads parallel, that can
372 * only bring troubles to us. See LU-5331.
374 mutex_lock(&s->ls_purge_mutex);
376 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
380 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
381 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
383 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
384 LASSERT(atomic_read(&h->loh_ref) == 0);
386 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
387 LASSERT(bd.bd_bucket == bd2.bd_bucket);
389 cfs_hash_bd_del_locked(s->ls_obj_hash,
391 cfs_list_move(&h->loh_lru, &dispose);
395 if (nr != ~0 && --nr == 0)
398 if (count > 0 && --count == 0)
402 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
405 * Free everything on the dispose list. This is safe against
406 * races due to the reasons described in lu_object_put().
408 while (!cfs_list_empty(&dispose)) {
409 h = container_of0(dispose.next,
410 struct lu_object_header, loh_lru);
411 cfs_list_del_init(&h->loh_lru);
412 lu_object_free(env, lu_object_top(h));
413 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
419 mutex_unlock(&s->ls_purge_mutex);
421 if (nr != 0 && did_sth && start != 0) {
422 start = 0; /* restart from the first bucket */
425 /* race on s->ls_purge_start, but nobody cares */
426 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
430 EXPORT_SYMBOL(lu_site_purge);
435 * Code below has to jump through certain loops to output object description
436 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
437 * composes object description from strings that are parts of _lines_ of
438 * output (i.e., strings that are not terminated by newline). This doesn't fit
439 * very well into libcfs_debug_msg() interface that assumes that each message
440 * supplied to it is a self-contained output line.
442 * To work around this, strings are collected in a temporary buffer
443 * (implemented as a value of lu_cdebug_key key), until terminating newline
444 * character is detected.
452 * XXX overflow is not handled correctly.
457 struct lu_cdebug_data {
461 char lck_area[LU_CDEBUG_LINE];
464 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
465 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
468 * Key, holding temporary buffer. This key is registered very early by
471 struct lu_context_key lu_global_key = {
472 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
473 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
474 .lct_init = lu_global_key_init,
475 .lct_fini = lu_global_key_fini
479 * Printer function emitting messages through libcfs_debug_msg().
481 int lu_cdebug_printer(const struct lu_env *env,
482 void *cookie, const char *format, ...)
484 struct libcfs_debug_msg_data *msgdata = cookie;
485 struct lu_cdebug_data *key;
490 va_start(args, format);
492 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
493 LASSERT(key != NULL);
495 used = strlen(key->lck_area);
496 complete = format[strlen(format) - 1] == '\n';
498 * Append new chunk to the buffer.
500 vsnprintf(key->lck_area + used,
501 ARRAY_SIZE(key->lck_area) - used, format, args);
503 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
504 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
505 key->lck_area[0] = 0;
510 EXPORT_SYMBOL(lu_cdebug_printer);
513 * Print object header.
515 void lu_object_header_print(const struct lu_env *env, void *cookie,
516 lu_printer_t printer,
517 const struct lu_object_header *hdr)
519 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
520 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
522 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
523 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
525 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
527 EXPORT_SYMBOL(lu_object_header_print);
530 * Print human readable representation of the \a o to the \a printer.
532 void lu_object_print(const struct lu_env *env, void *cookie,
533 lu_printer_t printer, const struct lu_object *o)
535 static const char ruler[] = "........................................";
536 struct lu_object_header *top;
540 lu_object_header_print(env, cookie, printer, top);
541 (*printer)(env, cookie, "{\n");
543 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
545 * print `.' \a depth times followed by type name and address
547 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
548 o->lo_dev->ld_type->ldt_name, o);
550 if (o->lo_ops->loo_object_print != NULL)
551 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
553 (*printer)(env, cookie, "\n");
556 (*printer)(env, cookie, "} header@%p\n", top);
558 EXPORT_SYMBOL(lu_object_print);
561 * Check object consistency.
563 int lu_object_invariant(const struct lu_object *o)
565 struct lu_object_header *top;
568 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
569 if (o->lo_ops->loo_object_invariant != NULL &&
570 !o->lo_ops->loo_object_invariant(o))
575 EXPORT_SYMBOL(lu_object_invariant);
577 static struct lu_object *htable_lookup(struct lu_site *s,
579 const struct lu_fid *f,
580 wait_queue_t *waiter,
583 struct lu_site_bkt_data *bkt;
584 struct lu_object_header *h;
585 cfs_hlist_node_t *hnode;
586 __u64 ver = cfs_hash_bd_version_get(bd);
589 return ERR_PTR(-ENOENT);
592 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
593 /* cfs_hash_bd_peek_locked is a somehow "internal" function
594 * of cfs_hash, it doesn't add refcount on object. */
595 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
597 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
598 return ERR_PTR(-ENOENT);
601 h = container_of0(hnode, struct lu_object_header, loh_hash);
602 if (likely(!lu_object_is_dying(h))) {
603 cfs_hash_get(s->ls_obj_hash, hnode);
604 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
605 cfs_list_del_init(&h->loh_lru);
606 return lu_object_top(h);
610 * Lookup found an object being destroyed this object cannot be
611 * returned (to assure that references to dying objects are eventually
612 * drained), and moreover, lookup has to wait until object is freed.
615 init_waitqueue_entry_current(waiter);
616 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
617 set_current_state(TASK_UNINTERRUPTIBLE);
618 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
619 return ERR_PTR(-EAGAIN);
622 static struct lu_object *htable_lookup_nowait(struct lu_site *s,
624 const struct lu_fid *f)
626 cfs_hlist_node_t *hnode;
627 struct lu_object_header *h;
629 /* cfs_hash_bd_peek_locked is a somehow "internal" function
630 * of cfs_hash, it doesn't add refcount on object. */
631 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
633 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
634 return ERR_PTR(-ENOENT);
637 h = container_of0(hnode, struct lu_object_header, loh_hash);
638 if (unlikely(lu_object_is_dying(h)))
639 return ERR_PTR(-ENOENT);
641 cfs_hash_get(s->ls_obj_hash, hnode);
642 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
643 cfs_list_del_init(&h->loh_lru);
644 return lu_object_top(h);
648 * Search cache for an object with the fid \a f. If such object is found,
649 * return it. Otherwise, create new object, insert it into cache and return
650 * it. In any case, additional reference is acquired on the returned object.
652 struct lu_object *lu_object_find(const struct lu_env *env,
653 struct lu_device *dev, const struct lu_fid *f,
654 const struct lu_object_conf *conf)
656 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
658 EXPORT_SYMBOL(lu_object_find);
661 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
662 * the calculation for the number of objects to reclaim is not covered by
663 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
664 * This ensures that many concurrent threads will not accidentally purge
667 static void lu_object_limit(const struct lu_env *env,
668 struct lu_device *dev)
672 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
675 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
676 nr = (__u64)lu_cache_nr;
678 lu_site_purge(env, dev->ld_site,
679 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
684 static struct lu_object *lu_object_new(const struct lu_env *env,
685 struct lu_device *dev,
686 const struct lu_fid *f,
687 const struct lu_object_conf *conf)
692 struct lu_site_bkt_data *bkt;
694 o = lu_object_alloc(env, dev, f, conf);
695 if (unlikely(IS_ERR(o)))
698 hs = dev->ld_site->ls_obj_hash;
699 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
700 bkt = cfs_hash_bd_extra_get(hs, &bd);
701 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
703 cfs_hash_bd_unlock(hs, &bd, 1);
705 lu_object_limit(env, dev);
711 * Core logic of lu_object_find*() functions.
713 static struct lu_object *lu_object_find_try(const struct lu_env *env,
714 struct lu_device *dev,
715 const struct lu_fid *f,
716 const struct lu_object_conf *conf,
717 wait_queue_t *waiter)
720 struct lu_object *shadow;
727 * This uses standard index maintenance protocol:
729 * - search index under lock, and return object if found;
730 * - otherwise, unlock index, allocate new object;
731 * - lock index and search again;
732 * - if nothing is found (usual case), insert newly created
734 * - otherwise (race: other thread inserted object), free
735 * object just allocated.
739 * For "LOC_F_NEW" case, we are sure the object is new established.
740 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
741 * just alloc and insert directly.
743 * If dying object is found during index search, add @waiter to the
744 * site wait-queue and return ERR_PTR(-EAGAIN).
746 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
747 return lu_object_new(env, dev, f, conf);
751 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
752 o = htable_lookup(s, &bd, f, waiter, &version);
753 cfs_hash_bd_unlock(hs, &bd, 1);
754 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
758 * Allocate new object. This may result in rather complicated
759 * operations, including fld queries, inode loading, etc.
761 o = lu_object_alloc(env, dev, f, conf);
762 if (unlikely(IS_ERR(o)))
765 LASSERT(lu_fid_eq(lu_object_fid(o), f));
767 cfs_hash_bd_lock(hs, &bd, 1);
769 shadow = htable_lookup(s, &bd, f, waiter, &version);
770 if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
771 struct lu_site_bkt_data *bkt;
773 bkt = cfs_hash_bd_extra_get(hs, &bd);
774 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
776 cfs_hash_bd_unlock(hs, &bd, 1);
778 lu_object_limit(env, dev);
783 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
784 cfs_hash_bd_unlock(hs, &bd, 1);
785 lu_object_free(env, o);
790 * Much like lu_object_find(), but top level device of object is specifically
791 * \a dev rather than top level device of the site. This interface allows
792 * objects of different "stacking" to be created within the same site.
794 struct lu_object *lu_object_find_at(const struct lu_env *env,
795 struct lu_device *dev,
796 const struct lu_fid *f,
797 const struct lu_object_conf *conf)
799 struct lu_site_bkt_data *bkt;
800 struct lu_object *obj;
804 obj = lu_object_find_try(env, dev, f, conf, &wait);
805 if (obj != ERR_PTR(-EAGAIN))
808 * lu_object_find_try() already added waiter into the
811 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
812 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
813 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
816 EXPORT_SYMBOL(lu_object_find_at);
819 * Try to find the object in cache without waiting for the dead object
820 * to be released nor allocating object if no cached one was found.
822 * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
824 void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
825 const struct lu_fid *f)
827 struct lu_site *s = dev->ld_site;
828 cfs_hash_t *hs = s->ls_obj_hash;
832 cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
833 o = htable_lookup_nowait(s, &bd, f);
834 cfs_hash_bd_unlock(hs, &bd, 1);
836 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
837 lu_object_put(env, o);
840 EXPORT_SYMBOL(lu_object_purge);
843 * Find object with given fid, and return its slice belonging to given device.
845 struct lu_object *lu_object_find_slice(const struct lu_env *env,
846 struct lu_device *dev,
847 const struct lu_fid *f,
848 const struct lu_object_conf *conf)
850 struct lu_object *top;
851 struct lu_object *obj;
853 top = lu_object_find(env, dev, f, conf);
855 obj = lu_object_locate(top->lo_header, dev->ld_type);
857 lu_object_put(env, top);
862 EXPORT_SYMBOL(lu_object_find_slice);
865 * Global list of all device types.
867 static CFS_LIST_HEAD(lu_device_types);
869 int lu_device_type_init(struct lu_device_type *ldt)
873 atomic_set(&ldt->ldt_device_nr, 0);
874 INIT_LIST_HEAD(&ldt->ldt_linkage);
875 if (ldt->ldt_ops->ldto_init)
876 result = ldt->ldt_ops->ldto_init(ldt);
879 spin_lock(&obd_types_lock);
880 list_add(&ldt->ldt_linkage, &lu_device_types);
881 spin_unlock(&obd_types_lock);
886 EXPORT_SYMBOL(lu_device_type_init);
888 void lu_device_type_fini(struct lu_device_type *ldt)
890 spin_lock(&obd_types_lock);
891 list_del_init(&ldt->ldt_linkage);
892 spin_unlock(&obd_types_lock);
893 if (ldt->ldt_ops->ldto_fini)
894 ldt->ldt_ops->ldto_fini(ldt);
896 EXPORT_SYMBOL(lu_device_type_fini);
899 * Global list of all sites on this node
901 static CFS_LIST_HEAD(lu_sites);
902 static DEFINE_MUTEX(lu_sites_guard);
905 * Global environment used by site shrinker.
907 static struct lu_env lu_shrink_env;
909 struct lu_site_print_arg {
910 struct lu_env *lsp_env;
912 lu_printer_t lsp_printer;
916 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
917 cfs_hlist_node_t *hnode, void *data)
919 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
920 struct lu_object_header *h;
922 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
923 if (!cfs_list_empty(&h->loh_layers)) {
924 const struct lu_object *o;
926 o = lu_object_top(h);
927 lu_object_print(arg->lsp_env, arg->lsp_cookie,
928 arg->lsp_printer, o);
930 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
931 arg->lsp_printer, h);
937 * Print all objects in \a s.
939 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
940 lu_printer_t printer)
942 struct lu_site_print_arg arg = {
943 .lsp_env = (struct lu_env *)env,
944 .lsp_cookie = cookie,
945 .lsp_printer = printer,
948 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
950 EXPORT_SYMBOL(lu_site_print);
953 * Return desired hash table order.
955 static int lu_htable_order(struct lu_device *top)
957 unsigned long cache_size;
961 * For ZFS based OSDs the cache should be disabled by default. This
962 * allows the ZFS ARC maximum flexibility in determining what buffers
963 * to cache. If Lustre has objects or buffer which it wants to ensure
964 * always stay cached it must maintain a hold on them.
966 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
967 lu_cache_percent = 1;
968 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
969 return LU_SITE_BITS_MIN;
973 * Calculate hash table size, assuming that we want reasonable
974 * performance when 20% of total memory is occupied by cache of
977 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
979 cache_size = totalram_pages;
981 #if BITS_PER_LONG == 32
982 /* limit hashtable size for lowmem systems to low RAM */
983 if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
984 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
987 /* clear off unreasonable cache setting. */
988 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
989 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
990 " the range of (0, %u]. Will use default value: %u.\n",
991 lu_cache_percent, LU_CACHE_PERCENT_MAX,
992 LU_CACHE_PERCENT_DEFAULT);
994 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
996 cache_size = cache_size / 100 * lu_cache_percent *
997 (PAGE_CACHE_SIZE / 1024);
999 for (bits = 1; (1 << bits) < cache_size; ++bits) {
1005 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
1006 const void *key, unsigned mask)
1008 struct lu_fid *fid = (struct lu_fid *)key;
1011 hash = fid_flatten32(fid);
1012 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1013 hash = hash_long(hash, hs->hs_bkt_bits);
1015 /* give me another random factor */
1016 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1018 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1019 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1024 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
1026 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
1029 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
1031 struct lu_object_header *h;
1033 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
1037 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
1039 struct lu_object_header *h;
1041 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
1042 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1045 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1047 struct lu_object_header *h;
1049 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
1050 if (atomic_add_return(1, &h->loh_ref) == 1) {
1051 struct lu_site_bkt_data *bkt;
1054 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
1055 bkt = cfs_hash_bd_extra_get(hs, &bd);
1060 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1062 LBUG(); /* we should never called it */
1065 cfs_hash_ops_t lu_site_hash_ops = {
1066 .hs_hash = lu_obj_hop_hash,
1067 .hs_key = lu_obj_hop_key,
1068 .hs_keycmp = lu_obj_hop_keycmp,
1069 .hs_object = lu_obj_hop_object,
1070 .hs_get = lu_obj_hop_get,
1071 .hs_put_locked = lu_obj_hop_put_locked,
1074 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1076 spin_lock(&s->ls_ld_lock);
1077 if (cfs_list_empty(&d->ld_linkage))
1078 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
1079 spin_unlock(&s->ls_ld_lock);
1081 EXPORT_SYMBOL(lu_dev_add_linkage);
1083 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1085 spin_lock(&s->ls_ld_lock);
1086 cfs_list_del_init(&d->ld_linkage);
1087 spin_unlock(&s->ls_ld_lock);
1089 EXPORT_SYMBOL(lu_dev_del_linkage);
1092 * Initialize site \a s, with \a d as the top level device.
1094 int lu_site_init(struct lu_site *s, struct lu_device *top)
1096 struct lu_site_bkt_data *bkt;
1103 memset(s, 0, sizeof *s);
1104 mutex_init(&s->ls_purge_mutex);
1105 bits = lu_htable_order(top);
1106 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
1107 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
1108 bits >= LU_SITE_BITS_MIN; bits--) {
1109 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1110 bits - LU_SITE_BKT_BITS,
1113 CFS_HASH_SPIN_BKTLOCK |
1114 CFS_HASH_NO_ITEMREF |
1116 CFS_HASH_ASSERT_EMPTY |
1118 if (s->ls_obj_hash != NULL)
1122 if (s->ls_obj_hash == NULL) {
1123 CERROR("failed to create lu_site hash with bits: %d\n", bits);
1127 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1128 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1129 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
1130 init_waitqueue_head(&bkt->lsb_marche_funebre);
1133 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1134 if (s->ls_stats == NULL) {
1135 cfs_hash_putref(s->ls_obj_hash);
1136 s->ls_obj_hash = NULL;
1140 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1141 0, "created", "created");
1142 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1143 0, "cache_hit", "cache_hit");
1144 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1145 0, "cache_miss", "cache_miss");
1146 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1147 0, "cache_race", "cache_race");
1148 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1149 0, "cache_death_race", "cache_death_race");
1150 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1151 0, "lru_purged", "lru_purged");
1153 CFS_INIT_LIST_HEAD(&s->ls_linkage);
1154 s->ls_top_dev = top;
1157 lu_ref_add(&top->ld_reference, "site-top", s);
1159 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1160 spin_lock_init(&s->ls_ld_lock);
1162 lu_dev_add_linkage(s, top);
1166 EXPORT_SYMBOL(lu_site_init);
1169 * Finalize \a s and release its resources.
1171 void lu_site_fini(struct lu_site *s)
1173 mutex_lock(&lu_sites_guard);
1174 cfs_list_del_init(&s->ls_linkage);
1175 mutex_unlock(&lu_sites_guard);
1177 if (s->ls_obj_hash != NULL) {
1178 cfs_hash_putref(s->ls_obj_hash);
1179 s->ls_obj_hash = NULL;
1182 if (s->ls_top_dev != NULL) {
1183 s->ls_top_dev->ld_site = NULL;
1184 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1185 lu_device_put(s->ls_top_dev);
1186 s->ls_top_dev = NULL;
1189 if (s->ls_stats != NULL)
1190 lprocfs_free_stats(&s->ls_stats);
1192 EXPORT_SYMBOL(lu_site_fini);
1195 * Called when initialization of stack for this site is completed.
1197 int lu_site_init_finish(struct lu_site *s)
1200 mutex_lock(&lu_sites_guard);
1201 result = lu_context_refill(&lu_shrink_env.le_ctx);
1203 cfs_list_add(&s->ls_linkage, &lu_sites);
1204 mutex_unlock(&lu_sites_guard);
1207 EXPORT_SYMBOL(lu_site_init_finish);
1210 * Acquire additional reference on device \a d
1212 void lu_device_get(struct lu_device *d)
1214 atomic_inc(&d->ld_ref);
1216 EXPORT_SYMBOL(lu_device_get);
1219 * Release reference on device \a d.
1221 void lu_device_put(struct lu_device *d)
1223 LASSERT(atomic_read(&d->ld_ref) > 0);
1224 atomic_dec(&d->ld_ref);
1226 EXPORT_SYMBOL(lu_device_put);
1229 * Initialize device \a d of type \a t.
1231 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1233 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1234 t->ldt_ops->ldto_start != NULL)
1235 t->ldt_ops->ldto_start(t);
1237 memset(d, 0, sizeof *d);
1239 lu_ref_init(&d->ld_reference);
1240 INIT_LIST_HEAD(&d->ld_linkage);
1244 EXPORT_SYMBOL(lu_device_init);
1247 * Finalize device \a d.
1249 void lu_device_fini(struct lu_device *d)
1251 struct lu_device_type *t = d->ld_type;
1253 if (d->ld_obd != NULL) {
1254 d->ld_obd->obd_lu_dev = NULL;
1258 lu_ref_fini(&d->ld_reference);
1259 LASSERTF(atomic_read(&d->ld_ref) == 0,
1260 "Refcount is %u\n", atomic_read(&d->ld_ref));
1261 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1263 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1264 t->ldt_ops->ldto_stop != NULL)
1265 t->ldt_ops->ldto_stop(t);
1267 EXPORT_SYMBOL(lu_device_fini);
1270 * Initialize object \a o that is part of compound object \a h and was created
1273 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1274 struct lu_device *d)
1276 memset(o, 0, sizeof(*o));
1280 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1281 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1285 EXPORT_SYMBOL(lu_object_init);
1288 * Finalize object and release its resources.
1290 void lu_object_fini(struct lu_object *o)
1292 struct lu_device *dev = o->lo_dev;
1294 LASSERT(cfs_list_empty(&o->lo_linkage));
1297 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1303 EXPORT_SYMBOL(lu_object_fini);
1306 * Add object \a o as first layer of compound object \a h
1308 * This is typically called by the ->ldo_object_alloc() method of top-level
1311 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1313 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1315 EXPORT_SYMBOL(lu_object_add_top);
1318 * Add object \a o as a layer of compound object, going after \a before.
1320 * This is typically called by the ->ldo_object_alloc() method of \a
1323 void lu_object_add(struct lu_object *before, struct lu_object *o)
1325 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1327 EXPORT_SYMBOL(lu_object_add);
1330 * Initialize compound object.
1332 int lu_object_header_init(struct lu_object_header *h)
1334 memset(h, 0, sizeof *h);
1335 atomic_set(&h->loh_ref, 1);
1336 CFS_INIT_HLIST_NODE(&h->loh_hash);
1337 CFS_INIT_LIST_HEAD(&h->loh_lru);
1338 CFS_INIT_LIST_HEAD(&h->loh_layers);
1339 lu_ref_init(&h->loh_reference);
1342 EXPORT_SYMBOL(lu_object_header_init);
1345 * Finalize compound object.
1347 void lu_object_header_fini(struct lu_object_header *h)
1349 LASSERT(cfs_list_empty(&h->loh_layers));
1350 LASSERT(cfs_list_empty(&h->loh_lru));
1351 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1352 lu_ref_fini(&h->loh_reference);
1354 EXPORT_SYMBOL(lu_object_header_fini);
1357 * Given a compound object, find its slice, corresponding to the device type
1360 struct lu_object *lu_object_locate(struct lu_object_header *h,
1361 const struct lu_device_type *dtype)
1363 struct lu_object *o;
1365 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1366 if (o->lo_dev->ld_type == dtype)
1371 EXPORT_SYMBOL(lu_object_locate);
1376 * Finalize and free devices in the device stack.
1378 * Finalize device stack by purging object cache, and calling
1379 * lu_device_type_operations::ldto_device_fini() and
1380 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1382 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1384 struct lu_site *site = top->ld_site;
1385 struct lu_device *scan;
1386 struct lu_device *next;
1388 lu_site_purge(env, site, ~0);
1389 for (scan = top; scan != NULL; scan = next) {
1390 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1391 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1392 lu_device_put(scan);
1396 lu_site_purge(env, site, ~0);
1398 for (scan = top; scan != NULL; scan = next) {
1399 const struct lu_device_type *ldt = scan->ld_type;
1400 struct obd_type *type;
1402 next = ldt->ldt_ops->ldto_device_free(env, scan);
1403 type = ldt->ldt_obd_type;
1406 class_put_type(type);
1410 EXPORT_SYMBOL(lu_stack_fini);
1414 * Maximal number of tld slots.
1416 LU_CONTEXT_KEY_NR = 40
1419 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1421 static DEFINE_SPINLOCK(lu_keys_guard);
1424 * Global counter incremented whenever key is registered, unregistered,
1425 * revived or quiesced. This is used to void unnecessary calls to
1426 * lu_context_refill(). No locking is provided, as initialization and shutdown
1427 * are supposed to be externally serialized.
1429 static unsigned key_set_version = 0;
1434 int lu_context_key_register(struct lu_context_key *key)
1439 LASSERT(key->lct_init != NULL);
1440 LASSERT(key->lct_fini != NULL);
1441 LASSERT(key->lct_tags != 0);
1442 LASSERT(key->lct_owner != NULL);
1445 spin_lock(&lu_keys_guard);
1446 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1447 if (lu_keys[i] == NULL) {
1449 atomic_set(&key->lct_used, 1);
1451 lu_ref_init(&key->lct_reference);
1457 spin_unlock(&lu_keys_guard);
1460 EXPORT_SYMBOL(lu_context_key_register);
1462 static void key_fini(struct lu_context *ctx, int index)
1464 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1465 struct lu_context_key *key;
1467 key = lu_keys[index];
1468 LASSERT(key != NULL);
1469 LASSERT(key->lct_fini != NULL);
1470 LASSERT(atomic_read(&key->lct_used) > 1);
1472 key->lct_fini(ctx, key, ctx->lc_value[index]);
1473 lu_ref_del(&key->lct_reference, "ctx", ctx);
1474 atomic_dec(&key->lct_used);
1476 LASSERT(key->lct_owner != NULL);
1477 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1478 LINVRNT(module_refcount(key->lct_owner) > 0);
1479 module_put(key->lct_owner);
1481 ctx->lc_value[index] = NULL;
1488 void lu_context_key_degister(struct lu_context_key *key)
1490 LASSERT(atomic_read(&key->lct_used) >= 1);
1491 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1493 lu_context_key_quiesce(key);
1496 spin_lock(&lu_keys_guard);
1497 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1498 if (lu_keys[key->lct_index]) {
1499 lu_keys[key->lct_index] = NULL;
1500 lu_ref_fini(&key->lct_reference);
1502 spin_unlock(&lu_keys_guard);
1504 LASSERTF(atomic_read(&key->lct_used) == 1,
1505 "key has instances: %d\n",
1506 atomic_read(&key->lct_used));
1508 EXPORT_SYMBOL(lu_context_key_degister);
1511 * Register a number of keys. This has to be called after all keys have been
1512 * initialized by a call to LU_CONTEXT_KEY_INIT().
1514 int lu_context_key_register_many(struct lu_context_key *k, ...)
1516 struct lu_context_key *key = k;
1522 result = lu_context_key_register(key);
1525 key = va_arg(args, struct lu_context_key *);
1526 } while (key != NULL);
1532 lu_context_key_degister(k);
1533 k = va_arg(args, struct lu_context_key *);
1540 EXPORT_SYMBOL(lu_context_key_register_many);
1543 * De-register a number of keys. This is a dual to
1544 * lu_context_key_register_many().
1546 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1552 lu_context_key_degister(k);
1553 k = va_arg(args, struct lu_context_key*);
1554 } while (k != NULL);
1557 EXPORT_SYMBOL(lu_context_key_degister_many);
1560 * Revive a number of keys.
1562 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1568 lu_context_key_revive(k);
1569 k = va_arg(args, struct lu_context_key*);
1570 } while (k != NULL);
1573 EXPORT_SYMBOL(lu_context_key_revive_many);
1576 * Quiescent a number of keys.
1578 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1584 lu_context_key_quiesce(k);
1585 k = va_arg(args, struct lu_context_key*);
1586 } while (k != NULL);
1589 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1592 * Return value associated with key \a key in context \a ctx.
1594 void *lu_context_key_get(const struct lu_context *ctx,
1595 const struct lu_context_key *key)
1597 LINVRNT(ctx->lc_state == LCS_ENTERED);
1598 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1599 LASSERT(lu_keys[key->lct_index] == key);
1600 return ctx->lc_value[key->lct_index];
1602 EXPORT_SYMBOL(lu_context_key_get);
1605 * List of remembered contexts. XXX document me.
1607 static CFS_LIST_HEAD(lu_context_remembered);
1610 * Destroy \a key in all remembered contexts. This is used to destroy key
1611 * values in "shared" contexts (like service threads), when a module owning
1612 * the key is about to be unloaded.
1614 void lu_context_key_quiesce(struct lu_context_key *key)
1616 struct lu_context *ctx;
1617 extern unsigned cl_env_cache_purge(unsigned nr);
1619 if (!(key->lct_tags & LCT_QUIESCENT)) {
1621 * XXX layering violation.
1623 cl_env_cache_purge(~0);
1624 key->lct_tags |= LCT_QUIESCENT;
1626 * XXX memory barrier has to go here.
1628 spin_lock(&lu_keys_guard);
1629 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1631 key_fini(ctx, key->lct_index);
1632 spin_unlock(&lu_keys_guard);
1636 EXPORT_SYMBOL(lu_context_key_quiesce);
1638 void lu_context_key_revive(struct lu_context_key *key)
1640 key->lct_tags &= ~LCT_QUIESCENT;
1643 EXPORT_SYMBOL(lu_context_key_revive);
1645 static void keys_fini(struct lu_context *ctx)
1649 if (ctx->lc_value == NULL)
1652 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1655 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1656 ctx->lc_value = NULL;
1659 static int keys_fill(struct lu_context *ctx)
1663 LINVRNT(ctx->lc_value != NULL);
1664 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1665 struct lu_context_key *key;
1668 if (ctx->lc_value[i] == NULL && key != NULL &&
1669 (key->lct_tags & ctx->lc_tags) &&
1671 * Don't create values for a LCT_QUIESCENT key, as this
1672 * will pin module owning a key.
1674 !(key->lct_tags & LCT_QUIESCENT)) {
1677 LINVRNT(key->lct_init != NULL);
1678 LINVRNT(key->lct_index == i);
1680 value = key->lct_init(ctx, key);
1681 if (unlikely(IS_ERR(value)))
1682 return PTR_ERR(value);
1684 LASSERT(key->lct_owner != NULL);
1685 if (!(ctx->lc_tags & LCT_NOREF))
1686 try_module_get(key->lct_owner);
1687 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1688 atomic_inc(&key->lct_used);
1690 * This is the only place in the code, where an
1691 * element of ctx->lc_value[] array is set to non-NULL
1694 ctx->lc_value[i] = value;
1695 if (key->lct_exit != NULL)
1696 ctx->lc_tags |= LCT_HAS_EXIT;
1698 ctx->lc_version = key_set_version;
1703 static int keys_init(struct lu_context *ctx)
1705 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1706 if (likely(ctx->lc_value != NULL))
1707 return keys_fill(ctx);
1713 * Initialize context data-structure. Create values for all keys.
1715 int lu_context_init(struct lu_context *ctx, __u32 tags)
1719 memset(ctx, 0, sizeof *ctx);
1720 ctx->lc_state = LCS_INITIALIZED;
1721 ctx->lc_tags = tags;
1722 if (tags & LCT_REMEMBER) {
1723 spin_lock(&lu_keys_guard);
1724 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1725 spin_unlock(&lu_keys_guard);
1727 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1730 rc = keys_init(ctx);
1732 lu_context_fini(ctx);
1736 EXPORT_SYMBOL(lu_context_init);
1739 * Finalize context data-structure. Destroy key values.
1741 void lu_context_fini(struct lu_context *ctx)
1743 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1744 ctx->lc_state = LCS_FINALIZED;
1746 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1747 LASSERT(cfs_list_empty(&ctx->lc_remember));
1750 } else { /* could race with key degister */
1751 spin_lock(&lu_keys_guard);
1753 cfs_list_del_init(&ctx->lc_remember);
1754 spin_unlock(&lu_keys_guard);
1757 EXPORT_SYMBOL(lu_context_fini);
1760 * Called before entering context.
1762 void lu_context_enter(struct lu_context *ctx)
1764 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1765 ctx->lc_state = LCS_ENTERED;
1767 EXPORT_SYMBOL(lu_context_enter);
1770 * Called after exiting from \a ctx
1772 void lu_context_exit(struct lu_context *ctx)
1776 LINVRNT(ctx->lc_state == LCS_ENTERED);
1777 ctx->lc_state = LCS_LEFT;
1778 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1779 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1780 if (ctx->lc_value[i] != NULL) {
1781 struct lu_context_key *key;
1784 LASSERT(key != NULL);
1785 if (key->lct_exit != NULL)
1787 key, ctx->lc_value[i]);
1792 EXPORT_SYMBOL(lu_context_exit);
1795 * Allocate for context all missing keys that were registered after context
1796 * creation. key_set_version is only changed in rare cases when modules
1797 * are loaded and removed.
1799 int lu_context_refill(struct lu_context *ctx)
1801 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1803 EXPORT_SYMBOL(lu_context_refill);
1806 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1807 * obd being added. Currently, this is only used on client side, specifically
1808 * for echo device client, for other stack (like ptlrpc threads), context are
1809 * predefined when the lu_device type are registered, during the module probe
1812 __u32 lu_context_tags_default = 0;
1813 __u32 lu_session_tags_default = 0;
1815 void lu_context_tags_update(__u32 tags)
1817 spin_lock(&lu_keys_guard);
1818 lu_context_tags_default |= tags;
1820 spin_unlock(&lu_keys_guard);
1822 EXPORT_SYMBOL(lu_context_tags_update);
1824 void lu_context_tags_clear(__u32 tags)
1826 spin_lock(&lu_keys_guard);
1827 lu_context_tags_default &= ~tags;
1829 spin_unlock(&lu_keys_guard);
1831 EXPORT_SYMBOL(lu_context_tags_clear);
1833 void lu_session_tags_update(__u32 tags)
1835 spin_lock(&lu_keys_guard);
1836 lu_session_tags_default |= tags;
1838 spin_unlock(&lu_keys_guard);
1840 EXPORT_SYMBOL(lu_session_tags_update);
1842 void lu_session_tags_clear(__u32 tags)
1844 spin_lock(&lu_keys_guard);
1845 lu_session_tags_default &= ~tags;
1847 spin_unlock(&lu_keys_guard);
1849 EXPORT_SYMBOL(lu_session_tags_clear);
1851 int lu_env_init(struct lu_env *env, __u32 tags)
1856 result = lu_context_init(&env->le_ctx, tags);
1857 if (likely(result == 0))
1858 lu_context_enter(&env->le_ctx);
1861 EXPORT_SYMBOL(lu_env_init);
1863 void lu_env_fini(struct lu_env *env)
1865 lu_context_exit(&env->le_ctx);
1866 lu_context_fini(&env->le_ctx);
1869 EXPORT_SYMBOL(lu_env_fini);
1871 int lu_env_refill(struct lu_env *env)
1875 result = lu_context_refill(&env->le_ctx);
1876 if (result == 0 && env->le_ses != NULL)
1877 result = lu_context_refill(env->le_ses);
1880 EXPORT_SYMBOL(lu_env_refill);
1883 * Currently, this API will only be used by echo client.
1884 * Because echo client and normal lustre client will share
1885 * same cl_env cache. So echo client needs to refresh
1886 * the env context after it get one from the cache, especially
1887 * when normal client and echo client co-exist in the same client.
1889 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1894 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1895 env->le_ctx.lc_version = 0;
1896 env->le_ctx.lc_tags |= ctags;
1899 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1900 env->le_ses->lc_version = 0;
1901 env->le_ses->lc_tags |= stags;
1904 result = lu_env_refill(env);
1908 EXPORT_SYMBOL(lu_env_refill_by_tags);
1910 static struct shrinker *lu_site_shrinker;
1912 typedef struct lu_site_stats{
1913 unsigned lss_populated;
1914 unsigned lss_max_search;
1919 static void lu_site_stats_get(cfs_hash_t *hs,
1920 lu_site_stats_t *stats, int populated)
1925 cfs_hash_for_each_bucket(hs, &bd, i) {
1926 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1927 cfs_hlist_head_t *hhead;
1929 cfs_hash_bd_lock(hs, &bd, 1);
1930 stats->lss_busy += bkt->lsb_busy;
1931 stats->lss_total += cfs_hash_bd_count_get(&bd);
1932 stats->lss_max_search = max((int)stats->lss_max_search,
1933 cfs_hash_bd_depmax_get(&bd));
1935 cfs_hash_bd_unlock(hs, &bd, 1);
1939 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1940 if (!cfs_hlist_empty(hhead))
1941 stats->lss_populated++;
1943 cfs_hash_bd_unlock(hs, &bd, 1);
1949 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1950 struct shrink_control *sc)
1952 lu_site_stats_t stats;
1954 struct lu_site *tmp;
1955 unsigned long cached = 0;
1957 if (!(sc->gfp_mask & __GFP_FS))
1960 mutex_lock(&lu_sites_guard);
1961 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1962 memset(&stats, 0, sizeof(stats));
1963 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1964 cached += stats.lss_total - stats.lss_busy;
1966 mutex_unlock(&lu_sites_guard);
1968 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1969 CDEBUG(D_INODE, "%ld objects cached\n", cached);
1973 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1974 struct shrink_control *sc)
1977 struct lu_site *tmp;
1978 unsigned long remain = sc->nr_to_scan;
1981 if (!(sc->gfp_mask & __GFP_FS))
1982 /* We must not take the lu_sites_guard lock when
1983 * __GFP_FS is *not* set because of the deadlock
1984 * possibility detailed above. Additionally,
1985 * since we cannot determine the number of
1986 * objects in the cache without taking this
1987 * lock, we're in a particularly tough spot. As
1988 * a result, we'll just lie and say our cache is
1989 * empty. This _should_ be ok, as we can't
1990 * reclaim objects when __GFP_FS is *not* set
1995 mutex_lock(&lu_sites_guard);
1996 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1997 remain = lu_site_purge(&lu_shrink_env, s, remain);
1999 * Move just shrunk site to the tail of site list to
2000 * assure shrinking fairness.
2002 list_move_tail(&s->ls_linkage, &splice);
2004 list_splice(&splice, lu_sites.prev);
2005 mutex_unlock(&lu_sites_guard);
2007 return sc->nr_to_scan - remain;
2010 #ifndef HAVE_SHRINKER_COUNT
2012 * There exists a potential lock inversion deadlock scenario when using
2013 * Lustre on top of ZFS. This occurs between one of ZFS's
2014 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2015 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2016 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2017 * lock. Obviously neither thread will wake and drop their respective hold
2020 * To prevent this from happening we must ensure the lu_sites_guard lock is
2021 * not taken while down this code path. ZFS reliably does not set the
2022 * __GFP_FS bit in its code paths, so this can be used to determine if it
2023 * is safe to take the lu_sites_guard lock.
2025 * Ideally we should accurately return the remaining number of cached
2026 * objects without taking the lu_sites_guard lock, but this is not
2027 * possible in the current implementation.
2029 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2032 struct shrink_control scv = {
2033 .nr_to_scan = shrink_param(sc, nr_to_scan),
2034 .gfp_mask = shrink_param(sc, gfp_mask)
2036 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2037 struct shrinker* shrinker = NULL;
2041 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2043 lu_cache_shrink_scan(shrinker, &scv);
2045 cached = lu_cache_shrink_count(shrinker, &scv);
2046 if (scv.nr_to_scan == 0)
2047 CDEBUG(D_INODE, "%d objects cached\n", cached);
2051 #endif /* HAVE_SHRINKER_COUNT */
2059 * Environment to be used in debugger, contains all tags.
2061 struct lu_env lu_debugging_env;
2064 * Debugging printer function using printk().
2066 int lu_printk_printer(const struct lu_env *env,
2067 void *unused, const char *format, ...)
2071 va_start(args, format);
2072 vprintk(format, args);
2077 int lu_debugging_setup(void)
2079 return lu_env_init(&lu_debugging_env, ~0);
2082 void lu_context_keys_dump(void)
2086 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2087 struct lu_context_key *key;
2091 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2092 i, key, key->lct_tags,
2093 key->lct_init, key->lct_fini, key->lct_exit,
2094 key->lct_index, atomic_read(&key->lct_used),
2095 key->lct_owner ? key->lct_owner->name : "",
2097 lu_ref_print(&key->lct_reference);
2101 EXPORT_SYMBOL(lu_context_keys_dump);
2102 #endif /* __KERNEL__ */
2105 * Initialization of global lu_* data.
2107 int lu_global_init(void)
2110 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2111 lu_cache_shrink_count, lu_cache_shrink_scan);
2113 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2115 result = lu_ref_global_init();
2119 LU_CONTEXT_KEY_INIT(&lu_global_key);
2120 result = lu_context_key_register(&lu_global_key);
2125 * At this level, we don't know what tags are needed, so allocate them
2126 * conservatively. This should not be too bad, because this
2127 * environment is global.
2129 mutex_lock(&lu_sites_guard);
2130 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2131 mutex_unlock(&lu_sites_guard);
2136 * seeks estimation: 3 seeks to read a record from oi, one to read
2137 * inode, one for ea. Unfortunately setting this high value results in
2138 * lu_object/inode cache consuming all the memory.
2140 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2141 if (lu_site_shrinker == NULL)
2148 * Dual to lu_global_init().
2150 void lu_global_fini(void)
2152 if (lu_site_shrinker != NULL) {
2153 remove_shrinker(lu_site_shrinker);
2154 lu_site_shrinker = NULL;
2157 lu_context_key_degister(&lu_global_key);
2160 * Tear shrinker environment down _after_ de-registering
2161 * lu_global_key, because the latter has a value in the former.
2163 mutex_lock(&lu_sites_guard);
2164 lu_env_fini(&lu_shrink_env);
2165 mutex_unlock(&lu_sites_guard);
2167 lu_ref_global_fini();
2170 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2173 struct lprocfs_counter ret;
2175 lprocfs_stats_collect(stats, idx, &ret);
2176 return (__u32)ret.lc_count;
2183 * Output site statistical counters into a buffer. Suitable for
2184 * lprocfs_rd_*()-style functions.
2186 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2188 lu_site_stats_t stats;
2190 memset(&stats, 0, sizeof(stats));
2191 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2193 return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2196 stats.lss_populated,
2197 CFS_HASH_NHLIST(s->ls_obj_hash),
2198 stats.lss_max_search,
2199 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2200 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2201 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2202 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2203 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2204 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2206 EXPORT_SYMBOL(lu_site_stats_seq_print);
2208 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2210 lu_site_stats_t stats;
2212 memset(&stats, 0, sizeof(stats));
2213 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2215 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2218 stats.lss_populated,
2219 CFS_HASH_NHLIST(s->ls_obj_hash),
2220 stats.lss_max_search,
2221 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2222 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2223 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2224 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2225 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2226 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2228 EXPORT_SYMBOL(lu_site_stats_print);
2231 * Helper function to initialize a number of kmem slab caches at once.
2233 int lu_kmem_init(struct lu_kmem_descr *caches)
2236 struct lu_kmem_descr *iter = caches;
2238 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2239 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2242 if (*iter->ckd_cache == NULL) {
2244 /* free all previously allocated caches */
2245 lu_kmem_fini(caches);
2251 EXPORT_SYMBOL(lu_kmem_init);
2254 * Helper function to finalize a number of kmem slab cached at once. Dual to
2257 void lu_kmem_fini(struct lu_kmem_descr *caches)
2259 for (; caches->ckd_cache != NULL; ++caches) {
2260 if (*caches->ckd_cache != NULL) {
2261 kmem_cache_destroy(*caches->ckd_cache);
2262 *caches->ckd_cache = NULL;
2266 EXPORT_SYMBOL(lu_kmem_fini);
2269 * Temporary solution to be able to assign fid in ->do_create()
2270 * till we have fully-functional OST fids
2272 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2273 const struct lu_fid *fid)
2275 struct lu_site *s = o->lo_dev->ld_site;
2276 struct lu_fid *old = &o->lo_header->loh_fid;
2277 struct lu_site_bkt_data *bkt;
2278 struct lu_object *shadow;
2279 wait_queue_t waiter;
2284 LASSERT(fid_is_zero(old));
2286 hs = s->ls_obj_hash;
2287 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2288 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2289 /* supposed to be unique */
2290 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2292 bkt = cfs_hash_bd_extra_get(hs, &bd);
2293 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2295 cfs_hash_bd_unlock(hs, &bd, 1);
2297 EXPORT_SYMBOL(lu_object_assign_fid);
2300 * allocates object with 0 (non-assiged) fid
2301 * XXX: temporary solution to be able to assign fid in ->do_create()
2302 * till we have fully-functional OST fids
2304 struct lu_object *lu_object_anon(const struct lu_env *env,
2305 struct lu_device *dev,
2306 const struct lu_object_conf *conf)
2309 struct lu_object *o;
2312 o = lu_object_alloc(env, dev, &fid, conf);
2316 EXPORT_SYMBOL(lu_object_anon);
2318 struct lu_buf LU_BUF_NULL = {
2322 EXPORT_SYMBOL(LU_BUF_NULL);
2324 void lu_buf_free(struct lu_buf *buf)
2328 LASSERT(buf->lb_len > 0);
2329 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2334 EXPORT_SYMBOL(lu_buf_free);
2336 void lu_buf_alloc(struct lu_buf *buf, int size)
2339 LASSERT(buf->lb_buf == NULL);
2340 LASSERT(buf->lb_len == 0);
2341 OBD_ALLOC_LARGE(buf->lb_buf, size);
2342 if (likely(buf->lb_buf))
2345 EXPORT_SYMBOL(lu_buf_alloc);
2347 void lu_buf_realloc(struct lu_buf *buf, int size)
2350 lu_buf_alloc(buf, size);
2352 EXPORT_SYMBOL(lu_buf_realloc);
2354 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, int len)
2356 if (buf->lb_buf == NULL && buf->lb_len == 0)
2357 lu_buf_alloc(buf, len);
2359 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2360 lu_buf_realloc(buf, len);
2364 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2367 * Increase the size of the \a buf.
2368 * preserves old data in buffer
2369 * old buffer remains unchanged on error
2370 * \retval 0 or -ENOMEM
2372 int lu_buf_check_and_grow(struct lu_buf *buf, int len)
2376 if (len <= buf->lb_len)
2379 OBD_ALLOC_LARGE(ptr, len);
2383 /* Free the old buf */
2384 if (buf->lb_buf != NULL) {
2385 memcpy(ptr, buf->lb_buf, buf->lb_len);
2386 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2393 EXPORT_SYMBOL(lu_buf_check_and_grow);