Whamcloud - gitweb
LU-1403 ucred: ucred cleanup
[fs/lustre-release.git] / lustre / obdclass / lu_object.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
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
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdclass/lu_object.c
37  *
38  * Lustre Object.
39  * These are the only exported functions, they provide some generic
40  * infrastructure for managing object devices
41  *
42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
43  */
44
45 #define DEBUG_SUBSYSTEM S_CLASS
46
47 #include <libcfs/libcfs.h>
48
49 #ifdef __KERNEL__
50 # include <linux/module.h>
51 #endif
52
53 /* hash_long() */
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>
60 #include <libcfs/list.h>
61 /* lu_time_global_{init,fini}() */
62 #include <lu_time.h>
63
64 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
65
66 /**
67  * Decrease reference counter on object. If last reference is freed, return
68  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
69  * case, free object immediately.
70  */
71 void lu_object_put(const struct lu_env *env, struct lu_object *o)
72 {
73         struct lu_site_bkt_data *bkt;
74         struct lu_object_header *top;
75         struct lu_site          *site;
76         struct lu_object        *orig;
77         cfs_hash_bd_t            bd;
78         const struct lu_fid     *fid;
79
80         top  = o->lo_header;
81         site = o->lo_dev->ld_site;
82         orig = o;
83
84         /*
85          * till we have full fids-on-OST implemented anonymous objects
86          * are possible in OSP. such an object isn't listed in the site
87          * so we should not remove it from the site.
88          */
89         fid = lu_object_fid(o);
90         if (fid_is_zero(fid)) {
91                 LASSERT(top->loh_hash.next == NULL
92                         && top->loh_hash.pprev == NULL);
93                 LASSERT(cfs_list_empty(&top->loh_lru));
94                 if (!cfs_atomic_dec_and_test(&top->loh_ref))
95                         return;
96                 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
97                         if (o->lo_ops->loo_object_release != NULL)
98                                 o->lo_ops->loo_object_release(env, o);
99                 }
100                 lu_object_free(env, orig);
101                 return;
102         }
103
104         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
105         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
106
107         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
108                 if (lu_object_is_dying(top)) {
109
110                         /*
111                          * somebody may be waiting for this, currently only
112                          * used for cl_object, see cl_object_put_last().
113                          */
114                         cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
115                 }
116                 return;
117         }
118
119         LASSERT(bkt->lsb_busy > 0);
120         bkt->lsb_busy--;
121         /*
122          * When last reference is released, iterate over object
123          * layers, and notify them that object is no longer busy.
124          */
125         cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
126                 if (o->lo_ops->loo_object_release != NULL)
127                         o->lo_ops->loo_object_release(env, o);
128         }
129
130         if (!lu_object_is_dying(top)) {
131                 LASSERT(cfs_list_empty(&top->loh_lru));
132                 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
133                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
134                 return;
135         }
136
137         /*
138          * If object is dying (will not be cached), removed it
139          * from hash table and LRU.
140          *
141          * This is done with hash table and LRU lists locked. As the only
142          * way to acquire first reference to previously unreferenced
143          * object is through hash-table lookup (lu_object_find()),
144          * or LRU scanning (lu_site_purge()), that are done under hash-table
145          * and LRU lock, no race with concurrent object lookup is possible
146          * and we can safely destroy object below.
147          */
148         cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
149         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
150         /*
151          * Object was already removed from hash and lru above, can
152          * kill it.
153          */
154         lu_object_free(env, orig);
155 }
156 EXPORT_SYMBOL(lu_object_put);
157
158 /**
159  * Put object and don't keep in cache. This is temporary solution for
160  * multi-site objects when its layering is not constant.
161  */
162 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
163 {
164         set_bit(LU_OBJECT_HEARD_BANSHEE,
165                     &o->lo_header->loh_flags);
166         return lu_object_put(env, o);
167 }
168 EXPORT_SYMBOL(lu_object_put_nocache);
169
170 /**
171  * Allocate new object.
172  *
173  * This follows object creation protocol, described in the comment within
174  * struct lu_device_operations definition.
175  */
176 static struct lu_object *lu_object_alloc(const struct lu_env *env,
177                                          struct lu_device *dev,
178                                          const struct lu_fid *f,
179                                          const struct lu_object_conf *conf)
180 {
181         struct lu_object *scan;
182         struct lu_object *top;
183         cfs_list_t *layers;
184         int clean;
185         int result;
186         ENTRY;
187
188         /*
189          * Create top-level object slice. This will also create
190          * lu_object_header.
191          */
192         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
193         if (top == NULL)
194                 RETURN(ERR_PTR(-ENOMEM));
195         /*
196          * This is the only place where object fid is assigned. It's constant
197          * after this point.
198          */
199         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
200         top->lo_header->loh_fid = *f;
201         layers = &top->lo_header->loh_layers;
202         do {
203                 /*
204                  * Call ->loo_object_init() repeatedly, until no more new
205                  * object slices are created.
206                  */
207                 clean = 1;
208                 cfs_list_for_each_entry(scan, layers, lo_linkage) {
209                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
210                                 continue;
211                         clean = 0;
212                         scan->lo_header = top->lo_header;
213                         result = scan->lo_ops->loo_object_init(env, scan, conf);
214                         if (result != 0) {
215                                 lu_object_free(env, top);
216                                 RETURN(ERR_PTR(result));
217                         }
218                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
219                 }
220         } while (!clean);
221
222         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
223                 if (scan->lo_ops->loo_object_start != NULL) {
224                         result = scan->lo_ops->loo_object_start(env, scan);
225                         if (result != 0) {
226                                 lu_object_free(env, top);
227                                 RETURN(ERR_PTR(result));
228                         }
229                 }
230         }
231
232         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
233         RETURN(top);
234 }
235
236 /**
237  * Free an object.
238  */
239 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
240 {
241         struct lu_site_bkt_data *bkt;
242         struct lu_site          *site;
243         struct lu_object        *scan;
244         cfs_list_t              *layers;
245         cfs_list_t               splice;
246
247         site   = o->lo_dev->ld_site;
248         layers = &o->lo_header->loh_layers;
249         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
250         /*
251          * First call ->loo_object_delete() method to release all resources.
252          */
253         cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
254                 if (scan->lo_ops->loo_object_delete != NULL)
255                         scan->lo_ops->loo_object_delete(env, scan);
256         }
257
258         /*
259          * Then, splice object layers into stand-alone list, and call
260          * ->loo_object_free() on all layers to free memory. Splice is
261          * necessary, because lu_object_header is freed together with the
262          * top-level slice.
263          */
264         CFS_INIT_LIST_HEAD(&splice);
265         cfs_list_splice_init(layers, &splice);
266         while (!cfs_list_empty(&splice)) {
267                 /*
268                  * Free layers in bottom-to-top order, so that object header
269                  * lives as long as possible and ->loo_object_free() methods
270                  * can look at its contents.
271                  */
272                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
273                 cfs_list_del_init(&o->lo_linkage);
274                 LASSERT(o->lo_ops->loo_object_free != NULL);
275                 o->lo_ops->loo_object_free(env, o);
276         }
277
278         if (cfs_waitq_active(&bkt->lsb_marche_funebre))
279                 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
280 }
281
282 /**
283  * Free \a nr objects from the cold end of the site LRU list.
284  */
285 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
286 {
287         struct lu_object_header *h;
288         struct lu_object_header *temp;
289         struct lu_site_bkt_data *bkt;
290         cfs_hash_bd_t            bd;
291         cfs_hash_bd_t            bd2;
292         cfs_list_t               dispose;
293         int                      did_sth;
294         int                      start;
295         int                      count;
296         int                      bnr;
297         int                      i;
298
299         CFS_INIT_LIST_HEAD(&dispose);
300         /*
301          * Under LRU list lock, scan LRU list and move unreferenced objects to
302          * the dispose list, removing them from LRU and hash table.
303          */
304         start = s->ls_purge_start;
305         bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
306  again:
307         did_sth = 0;
308         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
309                 if (i < start)
310                         continue;
311                 count = bnr;
312                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
313                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
314
315                 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
316                         LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
317
318                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
319                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
320
321                         cfs_hash_bd_del_locked(s->ls_obj_hash,
322                                                &bd2, &h->loh_hash);
323                         cfs_list_move(&h->loh_lru, &dispose);
324                         if (did_sth == 0)
325                                 did_sth = 1;
326
327                         if (nr != ~0 && --nr == 0)
328                                 break;
329
330                         if (count > 0 && --count == 0)
331                                 break;
332
333                 }
334                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
335                 cfs_cond_resched();
336                 /*
337                  * Free everything on the dispose list. This is safe against
338                  * races due to the reasons described in lu_object_put().
339                  */
340                 while (!cfs_list_empty(&dispose)) {
341                         h = container_of0(dispose.next,
342                                           struct lu_object_header, loh_lru);
343                         cfs_list_del_init(&h->loh_lru);
344                         lu_object_free(env, lu_object_top(h));
345                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
346                 }
347
348                 if (nr == 0)
349                         break;
350         }
351
352         if (nr != 0 && did_sth && start != 0) {
353                 start = 0; /* restart from the first bucket */
354                 goto again;
355         }
356         /* race on s->ls_purge_start, but nobody cares */
357         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
358
359         return nr;
360 }
361 EXPORT_SYMBOL(lu_site_purge);
362
363 /*
364  * Object printing.
365  *
366  * Code below has to jump through certain loops to output object description
367  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
368  * composes object description from strings that are parts of _lines_ of
369  * output (i.e., strings that are not terminated by newline). This doesn't fit
370  * very well into libcfs_debug_msg() interface that assumes that each message
371  * supplied to it is a self-contained output line.
372  *
373  * To work around this, strings are collected in a temporary buffer
374  * (implemented as a value of lu_cdebug_key key), until terminating newline
375  * character is detected.
376  *
377  */
378
379 enum {
380         /**
381          * Maximal line size.
382          *
383          * XXX overflow is not handled correctly.
384          */
385         LU_CDEBUG_LINE = 512
386 };
387
388 struct lu_cdebug_data {
389         /**
390          * Temporary buffer.
391          */
392         char lck_area[LU_CDEBUG_LINE];
393 };
394
395 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
396 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
397
398 /**
399  * Key, holding temporary buffer. This key is registered very early by
400  * lu_global_init().
401  */
402 struct lu_context_key lu_global_key = {
403         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
404                     LCT_MG_THREAD | LCT_CL_THREAD,
405         .lct_init = lu_global_key_init,
406         .lct_fini = lu_global_key_fini
407 };
408
409 /**
410  * Printer function emitting messages through libcfs_debug_msg().
411  */
412 int lu_cdebug_printer(const struct lu_env *env,
413                       void *cookie, const char *format, ...)
414 {
415         struct libcfs_debug_msg_data *msgdata = cookie;
416         struct lu_cdebug_data        *key;
417         int used;
418         int complete;
419         va_list args;
420
421         va_start(args, format);
422
423         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
424         LASSERT(key != NULL);
425
426         used = strlen(key->lck_area);
427         complete = format[strlen(format) - 1] == '\n';
428         /*
429          * Append new chunk to the buffer.
430          */
431         vsnprintf(key->lck_area + used,
432                   ARRAY_SIZE(key->lck_area) - used, format, args);
433         if (complete) {
434                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
435                         libcfs_debug_msg(msgdata, "%s", key->lck_area);
436                 key->lck_area[0] = 0;
437         }
438         va_end(args);
439         return 0;
440 }
441 EXPORT_SYMBOL(lu_cdebug_printer);
442
443 /**
444  * Print object header.
445  */
446 void lu_object_header_print(const struct lu_env *env, void *cookie,
447                             lu_printer_t printer,
448                             const struct lu_object_header *hdr)
449 {
450         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
451                    hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
452                    PFID(&hdr->loh_fid),
453                    cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
454                    cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
455                    "" : " lru",
456                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
457 }
458 EXPORT_SYMBOL(lu_object_header_print);
459
460 /**
461  * Print human readable representation of the \a o to the \a printer.
462  */
463 void lu_object_print(const struct lu_env *env, void *cookie,
464                      lu_printer_t printer, const struct lu_object *o)
465 {
466         static const char ruler[] = "........................................";
467         struct lu_object_header *top;
468         int depth;
469
470         top = o->lo_header;
471         lu_object_header_print(env, cookie, printer, top);
472         (*printer)(env, cookie, "{ \n");
473         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
474                 depth = o->lo_depth + 4;
475
476                 /*
477                  * print `.' \a depth times followed by type name and address
478                  */
479                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
480                            o->lo_dev->ld_type->ldt_name, o);
481                 if (o->lo_ops->loo_object_print != NULL)
482                         o->lo_ops->loo_object_print(env, cookie, printer, o);
483                 (*printer)(env, cookie, "\n");
484         }
485         (*printer)(env, cookie, "} header@%p\n", top);
486 }
487 EXPORT_SYMBOL(lu_object_print);
488
489 /**
490  * Check object consistency.
491  */
492 int lu_object_invariant(const struct lu_object *o)
493 {
494         struct lu_object_header *top;
495
496         top = o->lo_header;
497         cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
498                 if (o->lo_ops->loo_object_invariant != NULL &&
499                     !o->lo_ops->loo_object_invariant(o))
500                         return 0;
501         }
502         return 1;
503 }
504 EXPORT_SYMBOL(lu_object_invariant);
505
506 static struct lu_object *htable_lookup(struct lu_site *s,
507                                        cfs_hash_bd_t *bd,
508                                        const struct lu_fid *f,
509                                        cfs_waitlink_t *waiter,
510                                        __u64 *version)
511 {
512         struct lu_site_bkt_data *bkt;
513         struct lu_object_header *h;
514         cfs_hlist_node_t        *hnode;
515         __u64  ver = cfs_hash_bd_version_get(bd);
516
517         if (*version == ver)
518                 return NULL;
519
520         *version = ver;
521         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
522         /* cfs_hash_bd_lookup_intent is a somehow "internal" function
523          * of cfs_hash, but we don't want refcount on object right now */
524         hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
525         if (hnode == NULL) {
526                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
527                 return NULL;
528         }
529
530         h = container_of0(hnode, struct lu_object_header, loh_hash);
531         if (likely(!lu_object_is_dying(h))) {
532                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
533                 cfs_list_del_init(&h->loh_lru);
534                 return lu_object_top(h);
535         }
536
537         /*
538          * Lookup found an object being destroyed this object cannot be
539          * returned (to assure that references to dying objects are eventually
540          * drained), and moreover, lookup has to wait until object is freed.
541          */
542         cfs_atomic_dec(&h->loh_ref);
543
544         cfs_waitlink_init(waiter);
545         cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
546         cfs_set_current_state(CFS_TASK_UNINT);
547         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
548         return ERR_PTR(-EAGAIN);
549 }
550
551 /**
552  * Search cache for an object with the fid \a f. If such object is found,
553  * return it. Otherwise, create new object, insert it into cache and return
554  * it. In any case, additional reference is acquired on the returned object.
555  */
556 struct lu_object *lu_object_find(const struct lu_env *env,
557                                  struct lu_device *dev, const struct lu_fid *f,
558                                  const struct lu_object_conf *conf)
559 {
560         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
561 }
562 EXPORT_SYMBOL(lu_object_find);
563
564 static struct lu_object *lu_object_new(const struct lu_env *env,
565                                        struct lu_device *dev,
566                                        const struct lu_fid *f,
567                                        const struct lu_object_conf *conf)
568 {
569         struct lu_object        *o;
570         cfs_hash_t              *hs;
571         cfs_hash_bd_t            bd;
572         struct lu_site_bkt_data *bkt;
573
574         o = lu_object_alloc(env, dev, f, conf);
575         if (unlikely(IS_ERR(o)))
576                 return o;
577
578         hs = dev->ld_site->ls_obj_hash;
579         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
580         bkt = cfs_hash_bd_extra_get(hs, &bd);
581         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
582         bkt->lsb_busy++;
583         cfs_hash_bd_unlock(hs, &bd, 1);
584         return o;
585 }
586
587 /**
588  * Core logic of lu_object_find*() functions.
589  */
590 static struct lu_object *lu_object_find_try(const struct lu_env *env,
591                                             struct lu_device *dev,
592                                             const struct lu_fid *f,
593                                             const struct lu_object_conf *conf,
594                                             cfs_waitlink_t *waiter)
595 {
596         struct lu_object      *o;
597         struct lu_object      *shadow;
598         struct lu_site        *s;
599         cfs_hash_t            *hs;
600         cfs_hash_bd_t          bd;
601         __u64                  version = 0;
602
603         /*
604          * This uses standard index maintenance protocol:
605          *
606          *     - search index under lock, and return object if found;
607          *     - otherwise, unlock index, allocate new object;
608          *     - lock index and search again;
609          *     - if nothing is found (usual case), insert newly created
610          *       object into index;
611          *     - otherwise (race: other thread inserted object), free
612          *       object just allocated.
613          *     - unlock index;
614          *     - return object.
615          *
616          * For "LOC_F_NEW" case, we are sure the object is new established.
617          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
618          * just alloc and insert directly.
619          *
620          * If dying object is found during index search, add @waiter to the
621          * site wait-queue and return ERR_PTR(-EAGAIN).
622          */
623         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
624                 return lu_object_new(env, dev, f, conf);
625
626         s  = dev->ld_site;
627         hs = s->ls_obj_hash;
628         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
629         o = htable_lookup(s, &bd, f, waiter, &version);
630         cfs_hash_bd_unlock(hs, &bd, 1);
631         if (o != NULL)
632                 return o;
633
634         /*
635          * Allocate new object. This may result in rather complicated
636          * operations, including fld queries, inode loading, etc.
637          */
638         o = lu_object_alloc(env, dev, f, conf);
639         if (unlikely(IS_ERR(o)))
640                 return o;
641
642         LASSERT(lu_fid_eq(lu_object_fid(o), f));
643
644         cfs_hash_bd_lock(hs, &bd, 1);
645
646         shadow = htable_lookup(s, &bd, f, waiter, &version);
647         if (likely(shadow == NULL)) {
648                 struct lu_site_bkt_data *bkt;
649
650                 bkt = cfs_hash_bd_extra_get(hs, &bd);
651                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
652                 bkt->lsb_busy++;
653                 cfs_hash_bd_unlock(hs, &bd, 1);
654                 return o;
655         }
656
657         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
658         cfs_hash_bd_unlock(hs, &bd, 1);
659         lu_object_free(env, o);
660         return shadow;
661 }
662
663 /**
664  * Much like lu_object_find(), but top level device of object is specifically
665  * \a dev rather than top level device of the site. This interface allows
666  * objects of different "stacking" to be created within the same site.
667  */
668 struct lu_object *lu_object_find_at(const struct lu_env *env,
669                                     struct lu_device *dev,
670                                     const struct lu_fid *f,
671                                     const struct lu_object_conf *conf)
672 {
673         struct lu_site_bkt_data *bkt;
674         struct lu_object        *obj;
675         cfs_waitlink_t           wait;
676
677         while (1) {
678                 obj = lu_object_find_try(env, dev, f, conf, &wait);
679                 if (obj != ERR_PTR(-EAGAIN))
680                         return obj;
681                 /*
682                  * lu_object_find_try() already added waiter into the
683                  * wait queue.
684                  */
685                 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
686                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
687                 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
688         }
689 }
690 EXPORT_SYMBOL(lu_object_find_at);
691
692 /**
693  * Find object with given fid, and return its slice belonging to given device.
694  */
695 struct lu_object *lu_object_find_slice(const struct lu_env *env,
696                                        struct lu_device *dev,
697                                        const struct lu_fid *f,
698                                        const struct lu_object_conf *conf)
699 {
700         struct lu_object *top;
701         struct lu_object *obj;
702
703         top = lu_object_find(env, dev, f, conf);
704         if (!IS_ERR(top)) {
705                 obj = lu_object_locate(top->lo_header, dev->ld_type);
706                 if (obj == NULL)
707                         lu_object_put(env, top);
708         } else
709                 obj = top;
710         return obj;
711 }
712 EXPORT_SYMBOL(lu_object_find_slice);
713
714 /**
715  * Global list of all device types.
716  */
717 static CFS_LIST_HEAD(lu_device_types);
718
719 int lu_device_type_init(struct lu_device_type *ldt)
720 {
721         int result;
722
723         CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
724         result = ldt->ldt_ops->ldto_init(ldt);
725         if (result == 0)
726                 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
727         return result;
728 }
729 EXPORT_SYMBOL(lu_device_type_init);
730
731 void lu_device_type_fini(struct lu_device_type *ldt)
732 {
733         cfs_list_del_init(&ldt->ldt_linkage);
734         ldt->ldt_ops->ldto_fini(ldt);
735 }
736 EXPORT_SYMBOL(lu_device_type_fini);
737
738 void lu_types_stop(void)
739 {
740         struct lu_device_type *ldt;
741
742         cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
743                 if (ldt->ldt_device_nr == 0)
744                         ldt->ldt_ops->ldto_stop(ldt);
745         }
746 }
747 EXPORT_SYMBOL(lu_types_stop);
748
749 /**
750  * Global list of all sites on this node
751  */
752 static CFS_LIST_HEAD(lu_sites);
753 static DEFINE_MUTEX(lu_sites_guard);
754
755 /**
756  * Global environment used by site shrinker.
757  */
758 static struct lu_env lu_shrink_env;
759
760 struct lu_site_print_arg {
761         struct lu_env   *lsp_env;
762         void            *lsp_cookie;
763         lu_printer_t     lsp_printer;
764 };
765
766 static int
767 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
768                   cfs_hlist_node_t *hnode, void *data)
769 {
770         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
771         struct lu_object_header  *h;
772
773         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
774         if (!cfs_list_empty(&h->loh_layers)) {
775                 const struct lu_object *o;
776
777                 o = lu_object_top(h);
778                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
779                                 arg->lsp_printer, o);
780         } else {
781                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
782                                        arg->lsp_printer, h);
783         }
784         return 0;
785 }
786
787 /**
788  * Print all objects in \a s.
789  */
790 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
791                    lu_printer_t printer)
792 {
793         struct lu_site_print_arg arg = {
794                 .lsp_env     = (struct lu_env *)env,
795                 .lsp_cookie  = cookie,
796                 .lsp_printer = printer,
797         };
798
799         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
800 }
801 EXPORT_SYMBOL(lu_site_print);
802
803 enum {
804         LU_CACHE_PERCENT_MAX     = 50,
805         LU_CACHE_PERCENT_DEFAULT = 20
806 };
807
808 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
809 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
810                 "Percentage of memory to be used as lu_object cache");
811
812 /**
813  * Return desired hash table order.
814  */
815 static int lu_htable_order(void)
816 {
817         unsigned long cache_size;
818         int bits;
819
820         /*
821          * Calculate hash table size, assuming that we want reasonable
822          * performance when 20% of total memory is occupied by cache of
823          * lu_objects.
824          *
825          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
826          */
827         cache_size = cfs_num_physpages;
828
829 #if BITS_PER_LONG == 32
830         /* limit hashtable size for lowmem systems to low RAM */
831         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
832                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
833 #endif
834
835         /* clear off unreasonable cache setting. */
836         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
837                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
838                       " the range of (0, %u]. Will use default value: %u.\n",
839                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
840                       LU_CACHE_PERCENT_DEFAULT);
841
842                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
843         }
844         cache_size = cache_size / 100 * lu_cache_percent *
845                 (CFS_PAGE_SIZE / 1024);
846
847         for (bits = 1; (1 << bits) < cache_size; ++bits) {
848                 ;
849         }
850         return bits;
851 }
852
853 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
854                                 const void *key, unsigned mask)
855 {
856         struct lu_fid  *fid = (struct lu_fid *)key;
857         __u32           hash;
858
859         hash = fid_flatten32(fid);
860         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
861         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
862
863         /* give me another random factor */
864         hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
865
866         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
867         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
868
869         return hash & mask;
870 }
871
872 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
873 {
874         return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
875 }
876
877 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
878 {
879         struct lu_object_header *h;
880
881         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
882         return &h->loh_fid;
883 }
884
885 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
886 {
887         struct lu_object_header *h;
888
889         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
890         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
891 }
892
893 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
894 {
895         struct lu_object_header *h;
896
897         h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
898         if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
899                 struct lu_site_bkt_data *bkt;
900                 cfs_hash_bd_t            bd;
901
902                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
903                 bkt = cfs_hash_bd_extra_get(hs, &bd);
904                 bkt->lsb_busy++;
905         }
906 }
907
908 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
909 {
910         LBUG(); /* we should never called it */
911 }
912
913 cfs_hash_ops_t lu_site_hash_ops = {
914         .hs_hash        = lu_obj_hop_hash,
915         .hs_key         = lu_obj_hop_key,
916         .hs_keycmp      = lu_obj_hop_keycmp,
917         .hs_object      = lu_obj_hop_object,
918         .hs_get         = lu_obj_hop_get,
919         .hs_put_locked  = lu_obj_hop_put_locked,
920 };
921
922 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
923 {
924         spin_lock(&s->ls_ld_lock);
925         if (cfs_list_empty(&d->ld_linkage))
926                 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
927         spin_unlock(&s->ls_ld_lock);
928 }
929 EXPORT_SYMBOL(lu_dev_add_linkage);
930
931 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
932 {
933         spin_lock(&s->ls_ld_lock);
934         cfs_list_del_init(&d->ld_linkage);
935         spin_unlock(&s->ls_ld_lock);
936 }
937 EXPORT_SYMBOL(lu_dev_del_linkage);
938
939 /**
940  * Initialize site \a s, with \a d as the top level device.
941  */
942 #define LU_SITE_BITS_MIN    12
943 #define LU_SITE_BITS_MAX    24
944 /**
945  * total 256 buckets, we don't want too many buckets because:
946  * - consume too much memory
947  * - avoid unbalanced LRU list
948  */
949 #define LU_SITE_BKT_BITS    8
950
951 int lu_site_init(struct lu_site *s, struct lu_device *top)
952 {
953         struct lu_site_bkt_data *bkt;
954         cfs_hash_bd_t bd;
955         char name[16];
956         int bits;
957         int i;
958         ENTRY;
959
960         memset(s, 0, sizeof *s);
961         bits = lu_htable_order();
962         snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
963         for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
964              bits >= LU_SITE_BITS_MIN; bits--) {
965                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
966                                                  bits - LU_SITE_BKT_BITS,
967                                                  sizeof(*bkt), 0, 0,
968                                                  &lu_site_hash_ops,
969                                                  CFS_HASH_SPIN_BKTLOCK |
970                                                  CFS_HASH_NO_ITEMREF |
971                                                  CFS_HASH_DEPTH |
972                                                  CFS_HASH_ASSERT_EMPTY);
973                 if (s->ls_obj_hash != NULL)
974                         break;
975         }
976
977         if (s->ls_obj_hash == NULL) {
978                 CERROR("failed to create lu_site hash with bits: %d\n", bits);
979                 return -ENOMEM;
980         }
981
982         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
983                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
984                 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
985                 cfs_waitq_init(&bkt->lsb_marche_funebre);
986         }
987
988         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
989         if (s->ls_stats == NULL) {
990                 cfs_hash_putref(s->ls_obj_hash);
991                 s->ls_obj_hash = NULL;
992                 return -ENOMEM;
993         }
994
995         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
996                              0, "created", "created");
997         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
998                              0, "cache_hit", "cache_hit");
999         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1000                              0, "cache_miss", "cache_miss");
1001         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1002                              0, "cache_race", "cache_race");
1003         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1004                              0, "cache_death_race", "cache_death_race");
1005         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1006                              0, "lru_purged", "lru_purged");
1007
1008         CFS_INIT_LIST_HEAD(&s->ls_linkage);
1009         s->ls_top_dev = top;
1010         top->ld_site = s;
1011         lu_device_get(top);
1012         lu_ref_add(&top->ld_reference, "site-top", s);
1013
1014         CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1015         spin_lock_init(&s->ls_ld_lock);
1016
1017         lu_dev_add_linkage(s, top);
1018
1019         RETURN(0);
1020 }
1021 EXPORT_SYMBOL(lu_site_init);
1022
1023 /**
1024  * Finalize \a s and release its resources.
1025  */
1026 void lu_site_fini(struct lu_site *s)
1027 {
1028         mutex_lock(&lu_sites_guard);
1029         cfs_list_del_init(&s->ls_linkage);
1030         mutex_unlock(&lu_sites_guard);
1031
1032         if (s->ls_obj_hash != NULL) {
1033                 cfs_hash_putref(s->ls_obj_hash);
1034                 s->ls_obj_hash = NULL;
1035         }
1036
1037         if (s->ls_top_dev != NULL) {
1038                 s->ls_top_dev->ld_site = NULL;
1039                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1040                 lu_device_put(s->ls_top_dev);
1041                 s->ls_top_dev = NULL;
1042         }
1043
1044         if (s->ls_stats != NULL)
1045                 lprocfs_free_stats(&s->ls_stats);
1046 }
1047 EXPORT_SYMBOL(lu_site_fini);
1048
1049 /**
1050  * Called when initialization of stack for this site is completed.
1051  */
1052 int lu_site_init_finish(struct lu_site *s)
1053 {
1054         int result;
1055         mutex_lock(&lu_sites_guard);
1056         result = lu_context_refill(&lu_shrink_env.le_ctx);
1057         if (result == 0)
1058                 cfs_list_add(&s->ls_linkage, &lu_sites);
1059         mutex_unlock(&lu_sites_guard);
1060         return result;
1061 }
1062 EXPORT_SYMBOL(lu_site_init_finish);
1063
1064 /**
1065  * Acquire additional reference on device \a d
1066  */
1067 void lu_device_get(struct lu_device *d)
1068 {
1069         cfs_atomic_inc(&d->ld_ref);
1070 }
1071 EXPORT_SYMBOL(lu_device_get);
1072
1073 /**
1074  * Release reference on device \a d.
1075  */
1076 void lu_device_put(struct lu_device *d)
1077 {
1078         LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1079         cfs_atomic_dec(&d->ld_ref);
1080 }
1081 EXPORT_SYMBOL(lu_device_put);
1082
1083 /**
1084  * Initialize device \a d of type \a t.
1085  */
1086 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1087 {
1088         if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1089                 t->ldt_ops->ldto_start(t);
1090         memset(d, 0, sizeof *d);
1091         cfs_atomic_set(&d->ld_ref, 0);
1092         d->ld_type = t;
1093         lu_ref_init(&d->ld_reference);
1094         CFS_INIT_LIST_HEAD(&d->ld_linkage);
1095         return 0;
1096 }
1097 EXPORT_SYMBOL(lu_device_init);
1098
1099 /**
1100  * Finalize device \a d.
1101  */
1102 void lu_device_fini(struct lu_device *d)
1103 {
1104         struct lu_device_type *t;
1105
1106         t = d->ld_type;
1107         if (d->ld_obd != NULL) {
1108                 d->ld_obd->obd_lu_dev = NULL;
1109                 d->ld_obd = NULL;
1110         }
1111
1112         lu_ref_fini(&d->ld_reference);
1113         LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1114                  "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1115         LASSERT(t->ldt_device_nr > 0);
1116         if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1117                 t->ldt_ops->ldto_stop(t);
1118 }
1119 EXPORT_SYMBOL(lu_device_fini);
1120
1121 /**
1122  * Initialize object \a o that is part of compound object \a h and was created
1123  * by device \a d.
1124  */
1125 int lu_object_init(struct lu_object *o,
1126                    struct lu_object_header *h, struct lu_device *d)
1127 {
1128         memset(o, 0, sizeof *o);
1129         o->lo_header = h;
1130         o->lo_dev    = d;
1131         lu_device_get(d);
1132         o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1133         CFS_INIT_LIST_HEAD(&o->lo_linkage);
1134         return 0;
1135 }
1136 EXPORT_SYMBOL(lu_object_init);
1137
1138 /**
1139  * Finalize object and release its resources.
1140  */
1141 void lu_object_fini(struct lu_object *o)
1142 {
1143         struct lu_device *dev = o->lo_dev;
1144
1145         LASSERT(cfs_list_empty(&o->lo_linkage));
1146
1147         if (dev != NULL) {
1148                 lu_ref_del_at(&dev->ld_reference,
1149                               o->lo_dev_ref , "lu_object", o);
1150                 lu_device_put(dev);
1151                 o->lo_dev = NULL;
1152         }
1153 }
1154 EXPORT_SYMBOL(lu_object_fini);
1155
1156 /**
1157  * Add object \a o as first layer of compound object \a h
1158  *
1159  * This is typically called by the ->ldo_object_alloc() method of top-level
1160  * device.
1161  */
1162 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1163 {
1164         cfs_list_move(&o->lo_linkage, &h->loh_layers);
1165 }
1166 EXPORT_SYMBOL(lu_object_add_top);
1167
1168 /**
1169  * Add object \a o as a layer of compound object, going after \a before.
1170  *
1171  * This is typically called by the ->ldo_object_alloc() method of \a
1172  * before->lo_dev.
1173  */
1174 void lu_object_add(struct lu_object *before, struct lu_object *o)
1175 {
1176         cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1177 }
1178 EXPORT_SYMBOL(lu_object_add);
1179
1180 /**
1181  * Initialize compound object.
1182  */
1183 int lu_object_header_init(struct lu_object_header *h)
1184 {
1185         memset(h, 0, sizeof *h);
1186         cfs_atomic_set(&h->loh_ref, 1);
1187         CFS_INIT_HLIST_NODE(&h->loh_hash);
1188         CFS_INIT_LIST_HEAD(&h->loh_lru);
1189         CFS_INIT_LIST_HEAD(&h->loh_layers);
1190         lu_ref_init(&h->loh_reference);
1191         return 0;
1192 }
1193 EXPORT_SYMBOL(lu_object_header_init);
1194
1195 /**
1196  * Finalize compound object.
1197  */
1198 void lu_object_header_fini(struct lu_object_header *h)
1199 {
1200         LASSERT(cfs_list_empty(&h->loh_layers));
1201         LASSERT(cfs_list_empty(&h->loh_lru));
1202         LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1203         lu_ref_fini(&h->loh_reference);
1204 }
1205 EXPORT_SYMBOL(lu_object_header_fini);
1206
1207 /**
1208  * Given a compound object, find its slice, corresponding to the device type
1209  * \a dtype.
1210  */
1211 struct lu_object *lu_object_locate(struct lu_object_header *h,
1212                                    const struct lu_device_type *dtype)
1213 {
1214         struct lu_object *o;
1215
1216         cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1217                 if (o->lo_dev->ld_type == dtype)
1218                         return o;
1219         }
1220         return NULL;
1221 }
1222 EXPORT_SYMBOL(lu_object_locate);
1223
1224
1225
1226 /**
1227  * Finalize and free devices in the device stack.
1228  *
1229  * Finalize device stack by purging object cache, and calling
1230  * lu_device_type_operations::ldto_device_fini() and
1231  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1232  */
1233 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1234 {
1235         struct lu_site   *site = top->ld_site;
1236         struct lu_device *scan;
1237         struct lu_device *next;
1238
1239         lu_site_purge(env, site, ~0);
1240         for (scan = top; scan != NULL; scan = next) {
1241                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1242                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1243                 lu_device_put(scan);
1244         }
1245
1246         /* purge again. */
1247         lu_site_purge(env, site, ~0);
1248
1249         for (scan = top; scan != NULL; scan = next) {
1250                 const struct lu_device_type *ldt = scan->ld_type;
1251                 struct obd_type             *type;
1252
1253                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1254                 type = ldt->ldt_obd_type;
1255                 if (type != NULL) {
1256                         type->typ_refcnt--;
1257                         class_put_type(type);
1258                 }
1259         }
1260 }
1261 EXPORT_SYMBOL(lu_stack_fini);
1262
1263 enum {
1264         /**
1265          * Maximal number of tld slots.
1266          */
1267         LU_CONTEXT_KEY_NR = 40
1268 };
1269
1270 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1271
1272 static DEFINE_SPINLOCK(lu_keys_guard);
1273
1274 /**
1275  * Global counter incremented whenever key is registered, unregistered,
1276  * revived or quiesced. This is used to void unnecessary calls to
1277  * lu_context_refill(). No locking is provided, as initialization and shutdown
1278  * are supposed to be externally serialized.
1279  */
1280 static unsigned key_set_version = 0;
1281
1282 /**
1283  * Register new key.
1284  */
1285 int lu_context_key_register(struct lu_context_key *key)
1286 {
1287         int result;
1288         int i;
1289
1290         LASSERT(key->lct_init != NULL);
1291         LASSERT(key->lct_fini != NULL);
1292         LASSERT(key->lct_tags != 0);
1293         LASSERT(key->lct_owner != NULL);
1294
1295         result = -ENFILE;
1296         spin_lock(&lu_keys_guard);
1297         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1298                 if (lu_keys[i] == NULL) {
1299                         key->lct_index = i;
1300                         cfs_atomic_set(&key->lct_used, 1);
1301                         lu_keys[i] = key;
1302                         lu_ref_init(&key->lct_reference);
1303                         result = 0;
1304                         ++key_set_version;
1305                         break;
1306                 }
1307         }
1308         spin_unlock(&lu_keys_guard);
1309         return result;
1310 }
1311 EXPORT_SYMBOL(lu_context_key_register);
1312
1313 static void key_fini(struct lu_context *ctx, int index)
1314 {
1315         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1316                 struct lu_context_key *key;
1317
1318                 key = lu_keys[index];
1319                 LASSERT(key != NULL);
1320                 LASSERT(key->lct_fini != NULL);
1321                 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1322
1323                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1324                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1325                 cfs_atomic_dec(&key->lct_used);
1326
1327                 LASSERT(key->lct_owner != NULL);
1328                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1329                         LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1330                         cfs_module_put(key->lct_owner);
1331                 }
1332                 ctx->lc_value[index] = NULL;
1333         }
1334 }
1335
1336 /**
1337  * Deregister key.
1338  */
1339 void lu_context_key_degister(struct lu_context_key *key)
1340 {
1341         LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1342         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1343
1344         lu_context_key_quiesce(key);
1345
1346         ++key_set_version;
1347         spin_lock(&lu_keys_guard);
1348         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1349         if (lu_keys[key->lct_index]) {
1350                 lu_keys[key->lct_index] = NULL;
1351                 lu_ref_fini(&key->lct_reference);
1352         }
1353         spin_unlock(&lu_keys_guard);
1354
1355         LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1356                  "key has instances: %d\n",
1357                  cfs_atomic_read(&key->lct_used));
1358 }
1359 EXPORT_SYMBOL(lu_context_key_degister);
1360
1361 /**
1362  * Register a number of keys. This has to be called after all keys have been
1363  * initialized by a call to LU_CONTEXT_KEY_INIT().
1364  */
1365 int lu_context_key_register_many(struct lu_context_key *k, ...)
1366 {
1367         struct lu_context_key *key = k;
1368         va_list args;
1369         int result;
1370
1371         va_start(args, k);
1372         do {
1373                 result = lu_context_key_register(key);
1374                 if (result)
1375                         break;
1376                 key = va_arg(args, struct lu_context_key *);
1377         } while (key != NULL);
1378         va_end(args);
1379
1380         if (result != 0) {
1381                 va_start(args, k);
1382                 while (k != key) {
1383                         lu_context_key_degister(k);
1384                         k = va_arg(args, struct lu_context_key *);
1385                 }
1386                 va_end(args);
1387         }
1388
1389         return result;
1390 }
1391 EXPORT_SYMBOL(lu_context_key_register_many);
1392
1393 /**
1394  * De-register a number of keys. This is a dual to
1395  * lu_context_key_register_many().
1396  */
1397 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1398 {
1399         va_list args;
1400
1401         va_start(args, k);
1402         do {
1403                 lu_context_key_degister(k);
1404                 k = va_arg(args, struct lu_context_key*);
1405         } while (k != NULL);
1406         va_end(args);
1407 }
1408 EXPORT_SYMBOL(lu_context_key_degister_many);
1409
1410 /**
1411  * Revive a number of keys.
1412  */
1413 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1414 {
1415         va_list args;
1416
1417         va_start(args, k);
1418         do {
1419                 lu_context_key_revive(k);
1420                 k = va_arg(args, struct lu_context_key*);
1421         } while (k != NULL);
1422         va_end(args);
1423 }
1424 EXPORT_SYMBOL(lu_context_key_revive_many);
1425
1426 /**
1427  * Quiescent a number of keys.
1428  */
1429 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1430 {
1431         va_list args;
1432
1433         va_start(args, k);
1434         do {
1435                 lu_context_key_quiesce(k);
1436                 k = va_arg(args, struct lu_context_key*);
1437         } while (k != NULL);
1438         va_end(args);
1439 }
1440 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1441
1442 /**
1443  * Return value associated with key \a key in context \a ctx.
1444  */
1445 void *lu_context_key_get(const struct lu_context *ctx,
1446                          const struct lu_context_key *key)
1447 {
1448         LINVRNT(ctx->lc_state == LCS_ENTERED);
1449         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1450         LASSERT(lu_keys[key->lct_index] == key);
1451         return ctx->lc_value[key->lct_index];
1452 }
1453 EXPORT_SYMBOL(lu_context_key_get);
1454
1455 /**
1456  * List of remembered contexts. XXX document me.
1457  */
1458 static CFS_LIST_HEAD(lu_context_remembered);
1459
1460 /**
1461  * Destroy \a key in all remembered contexts. This is used to destroy key
1462  * values in "shared" contexts (like service threads), when a module owning
1463  * the key is about to be unloaded.
1464  */
1465 void lu_context_key_quiesce(struct lu_context_key *key)
1466 {
1467         struct lu_context *ctx;
1468         extern unsigned cl_env_cache_purge(unsigned nr);
1469
1470         if (!(key->lct_tags & LCT_QUIESCENT)) {
1471                 /*
1472                  * XXX layering violation.
1473                  */
1474                 cl_env_cache_purge(~0);
1475                 key->lct_tags |= LCT_QUIESCENT;
1476                 /*
1477                  * XXX memory barrier has to go here.
1478                  */
1479                 spin_lock(&lu_keys_guard);
1480                 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1481                                         lc_remember)
1482                         key_fini(ctx, key->lct_index);
1483                 spin_unlock(&lu_keys_guard);
1484                 ++key_set_version;
1485         }
1486 }
1487 EXPORT_SYMBOL(lu_context_key_quiesce);
1488
1489 void lu_context_key_revive(struct lu_context_key *key)
1490 {
1491         key->lct_tags &= ~LCT_QUIESCENT;
1492         ++key_set_version;
1493 }
1494 EXPORT_SYMBOL(lu_context_key_revive);
1495
1496 static void keys_fini(struct lu_context *ctx)
1497 {
1498         int     i;
1499
1500         if (ctx->lc_value == NULL)
1501                 return;
1502
1503         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1504                 key_fini(ctx, i);
1505
1506         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1507         ctx->lc_value = NULL;
1508 }
1509
1510 static int keys_fill(struct lu_context *ctx)
1511 {
1512         int i;
1513
1514         LINVRNT(ctx->lc_value != NULL);
1515         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1516                 struct lu_context_key *key;
1517
1518                 key = lu_keys[i];
1519                 if (ctx->lc_value[i] == NULL && key != NULL &&
1520                     (key->lct_tags & ctx->lc_tags) &&
1521                     /*
1522                      * Don't create values for a LCT_QUIESCENT key, as this
1523                      * will pin module owning a key.
1524                      */
1525                     !(key->lct_tags & LCT_QUIESCENT)) {
1526                         void *value;
1527
1528                         LINVRNT(key->lct_init != NULL);
1529                         LINVRNT(key->lct_index == i);
1530
1531                         value = key->lct_init(ctx, key);
1532                         if (unlikely(IS_ERR(value)))
1533                                 return PTR_ERR(value);
1534
1535                         LASSERT(key->lct_owner != NULL);
1536                         if (!(ctx->lc_tags & LCT_NOREF))
1537                                 cfs_try_module_get(key->lct_owner);
1538                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1539                         cfs_atomic_inc(&key->lct_used);
1540                         /*
1541                          * This is the only place in the code, where an
1542                          * element of ctx->lc_value[] array is set to non-NULL
1543                          * value.
1544                          */
1545                         ctx->lc_value[i] = value;
1546                         if (key->lct_exit != NULL)
1547                                 ctx->lc_tags |= LCT_HAS_EXIT;
1548                 }
1549                 ctx->lc_version = key_set_version;
1550         }
1551         return 0;
1552 }
1553
1554 static int keys_init(struct lu_context *ctx)
1555 {
1556         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1557         if (likely(ctx->lc_value != NULL))
1558                 return keys_fill(ctx);
1559
1560         return -ENOMEM;
1561 }
1562
1563 /**
1564  * Initialize context data-structure. Create values for all keys.
1565  */
1566 int lu_context_init(struct lu_context *ctx, __u32 tags)
1567 {
1568         int     rc;
1569
1570         memset(ctx, 0, sizeof *ctx);
1571         ctx->lc_state = LCS_INITIALIZED;
1572         ctx->lc_tags = tags;
1573         if (tags & LCT_REMEMBER) {
1574                 spin_lock(&lu_keys_guard);
1575                 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1576                 spin_unlock(&lu_keys_guard);
1577         } else {
1578                 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1579         }
1580
1581         rc = keys_init(ctx);
1582         if (rc != 0)
1583                 lu_context_fini(ctx);
1584
1585         return rc;
1586 }
1587 EXPORT_SYMBOL(lu_context_init);
1588
1589 /**
1590  * Finalize context data-structure. Destroy key values.
1591  */
1592 void lu_context_fini(struct lu_context *ctx)
1593 {
1594         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1595         ctx->lc_state = LCS_FINALIZED;
1596
1597         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1598                 LASSERT(cfs_list_empty(&ctx->lc_remember));
1599                 keys_fini(ctx);
1600
1601         } else { /* could race with key degister */
1602                 spin_lock(&lu_keys_guard);
1603                 keys_fini(ctx);
1604                 cfs_list_del_init(&ctx->lc_remember);
1605                 spin_unlock(&lu_keys_guard);
1606         }
1607 }
1608 EXPORT_SYMBOL(lu_context_fini);
1609
1610 /**
1611  * Called before entering context.
1612  */
1613 void lu_context_enter(struct lu_context *ctx)
1614 {
1615         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1616         ctx->lc_state = LCS_ENTERED;
1617 }
1618 EXPORT_SYMBOL(lu_context_enter);
1619
1620 /**
1621  * Called after exiting from \a ctx
1622  */
1623 void lu_context_exit(struct lu_context *ctx)
1624 {
1625         int i;
1626
1627         LINVRNT(ctx->lc_state == LCS_ENTERED);
1628         ctx->lc_state = LCS_LEFT;
1629         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1630                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1631                         if (ctx->lc_value[i] != NULL) {
1632                                 struct lu_context_key *key;
1633
1634                                 key = lu_keys[i];
1635                                 LASSERT(key != NULL);
1636                                 if (key->lct_exit != NULL)
1637                                         key->lct_exit(ctx,
1638                                                       key, ctx->lc_value[i]);
1639                         }
1640                 }
1641         }
1642 }
1643 EXPORT_SYMBOL(lu_context_exit);
1644
1645 /**
1646  * Allocate for context all missing keys that were registered after context
1647  * creation. key_set_version is only changed in rare cases when modules
1648  * are loaded and removed.
1649  */
1650 int lu_context_refill(struct lu_context *ctx)
1651 {
1652         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1653 }
1654 EXPORT_SYMBOL(lu_context_refill);
1655
1656 /**
1657  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1658  * obd being added. Currently, this is only used on client side, specifically
1659  * for echo device client, for other stack (like ptlrpc threads), context are
1660  * predefined when the lu_device type are registered, during the module probe
1661  * phase.
1662  */
1663 __u32 lu_context_tags_default = 0;
1664 __u32 lu_session_tags_default = 0;
1665
1666 void lu_context_tags_update(__u32 tags)
1667 {
1668         spin_lock(&lu_keys_guard);
1669         lu_context_tags_default |= tags;
1670         key_set_version++;
1671         spin_unlock(&lu_keys_guard);
1672 }
1673 EXPORT_SYMBOL(lu_context_tags_update);
1674
1675 void lu_context_tags_clear(__u32 tags)
1676 {
1677         spin_lock(&lu_keys_guard);
1678         lu_context_tags_default &= ~tags;
1679         key_set_version++;
1680         spin_unlock(&lu_keys_guard);
1681 }
1682 EXPORT_SYMBOL(lu_context_tags_clear);
1683
1684 void lu_session_tags_update(__u32 tags)
1685 {
1686         spin_lock(&lu_keys_guard);
1687         lu_session_tags_default |= tags;
1688         key_set_version++;
1689         spin_unlock(&lu_keys_guard);
1690 }
1691 EXPORT_SYMBOL(lu_session_tags_update);
1692
1693 void lu_session_tags_clear(__u32 tags)
1694 {
1695         spin_lock(&lu_keys_guard);
1696         lu_session_tags_default &= ~tags;
1697         key_set_version++;
1698         spin_unlock(&lu_keys_guard);
1699 }
1700 EXPORT_SYMBOL(lu_session_tags_clear);
1701
1702 int lu_env_init(struct lu_env *env, __u32 tags)
1703 {
1704         int result;
1705
1706         env->le_ses = NULL;
1707         result = lu_context_init(&env->le_ctx, tags);
1708         if (likely(result == 0))
1709                 lu_context_enter(&env->le_ctx);
1710         return result;
1711 }
1712 EXPORT_SYMBOL(lu_env_init);
1713
1714 void lu_env_fini(struct lu_env *env)
1715 {
1716         lu_context_exit(&env->le_ctx);
1717         lu_context_fini(&env->le_ctx);
1718         env->le_ses = NULL;
1719 }
1720 EXPORT_SYMBOL(lu_env_fini);
1721
1722 int lu_env_refill(struct lu_env *env)
1723 {
1724         int result;
1725
1726         result = lu_context_refill(&env->le_ctx);
1727         if (result == 0 && env->le_ses != NULL)
1728                 result = lu_context_refill(env->le_ses);
1729         return result;
1730 }
1731 EXPORT_SYMBOL(lu_env_refill);
1732
1733 /**
1734  * Currently, this API will only be used by echo client.
1735  * Because echo client and normal lustre client will share
1736  * same cl_env cache. So echo client needs to refresh
1737  * the env context after it get one from the cache, especially
1738  * when normal client and echo client co-exist in the same client.
1739  */
1740 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1741                           __u32 stags)
1742 {
1743         int    result;
1744
1745         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1746                 env->le_ctx.lc_version = 0;
1747                 env->le_ctx.lc_tags |= ctags;
1748         }
1749
1750         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1751                 env->le_ses->lc_version = 0;
1752                 env->le_ses->lc_tags |= stags;
1753         }
1754
1755         result = lu_env_refill(env);
1756
1757         return result;
1758 }
1759 EXPORT_SYMBOL(lu_env_refill_by_tags);
1760
1761 static struct cfs_shrinker *lu_site_shrinker = NULL;
1762
1763 typedef struct lu_site_stats{
1764         unsigned        lss_populated;
1765         unsigned        lss_max_search;
1766         unsigned        lss_total;
1767         unsigned        lss_busy;
1768 } lu_site_stats_t;
1769
1770 static void lu_site_stats_get(cfs_hash_t *hs,
1771                               lu_site_stats_t *stats, int populated)
1772 {
1773         cfs_hash_bd_t bd;
1774         int           i;
1775
1776         cfs_hash_for_each_bucket(hs, &bd, i) {
1777                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1778                 cfs_hlist_head_t        *hhead;
1779
1780                 cfs_hash_bd_lock(hs, &bd, 1);
1781                 stats->lss_busy  += bkt->lsb_busy;
1782                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1783                 stats->lss_max_search = max((int)stats->lss_max_search,
1784                                             cfs_hash_bd_depmax_get(&bd));
1785                 if (!populated) {
1786                         cfs_hash_bd_unlock(hs, &bd, 1);
1787                         continue;
1788                 }
1789
1790                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1791                         if (!cfs_hlist_empty(hhead))
1792                                 stats->lss_populated++;
1793                 }
1794                 cfs_hash_bd_unlock(hs, &bd, 1);
1795         }
1796 }
1797
1798 #ifdef __KERNEL__
1799
1800 /*
1801  * There exists a potential lock inversion deadlock scenario when using
1802  * Lustre on top of ZFS. This occurs between one of ZFS's
1803  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1804  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1805  * while thread B will take the ht_lock and sleep on the lu_sites_guard
1806  * lock. Obviously neither thread will wake and drop their respective hold
1807  * on their lock.
1808  *
1809  * To prevent this from happening we must ensure the lu_sites_guard lock is
1810  * not taken while down this code path. ZFS reliably does not set the
1811  * __GFP_FS bit in its code paths, so this can be used to determine if it
1812  * is safe to take the lu_sites_guard lock.
1813  *
1814  * Ideally we should accurately return the remaining number of cached
1815  * objects without taking the  lu_sites_guard lock, but this is not
1816  * possible in the current implementation.
1817  */
1818 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1819 {
1820         lu_site_stats_t stats;
1821         struct lu_site *s;
1822         struct lu_site *tmp;
1823         int cached = 0;
1824         int remain = shrink_param(sc, nr_to_scan);
1825         CFS_LIST_HEAD(splice);
1826
1827         if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1828                 if (remain != 0)
1829                         return -1;
1830                 else
1831                         /* We must not take the lu_sites_guard lock when
1832                          * __GFP_FS is *not* set because of the deadlock
1833                          * possibility detailed above. Additionally,
1834                          * since we cannot determine the number of
1835                          * objects in the cache without taking this
1836                          * lock, we're in a particularly tough spot. As
1837                          * a result, we'll just lie and say our cache is
1838                          * empty. This _should_ be ok, as we can't
1839                          * reclaim objects when __GFP_FS is *not* set
1840                          * anyways.
1841                          */
1842                         return 0;
1843         }
1844
1845         CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1846
1847         mutex_lock(&lu_sites_guard);
1848         cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1849                 if (shrink_param(sc, nr_to_scan) != 0) {
1850                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1851                         /*
1852                          * Move just shrunk site to the tail of site list to
1853                          * assure shrinking fairness.
1854                          */
1855                         cfs_list_move_tail(&s->ls_linkage, &splice);
1856                 }
1857
1858                 memset(&stats, 0, sizeof(stats));
1859                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1860                 cached += stats.lss_total - stats.lss_busy;
1861                 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1862                         break;
1863         }
1864         cfs_list_splice(&splice, lu_sites.prev);
1865         mutex_unlock(&lu_sites_guard);
1866
1867         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1868         if (shrink_param(sc, nr_to_scan) == 0)
1869                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1870         return cached;
1871 }
1872
1873 /*
1874  * Debugging stuff.
1875  */
1876
1877 /**
1878  * Environment to be used in debugger, contains all tags.
1879  */
1880 struct lu_env lu_debugging_env;
1881
1882 /**
1883  * Debugging printer function using printk().
1884  */
1885 int lu_printk_printer(const struct lu_env *env,
1886                       void *unused, const char *format, ...)
1887 {
1888         va_list args;
1889
1890         va_start(args, format);
1891         vprintk(format, args);
1892         va_end(args);
1893         return 0;
1894 }
1895
1896 void lu_debugging_setup(void)
1897 {
1898         lu_env_init(&lu_debugging_env, ~0);
1899 }
1900
1901 void lu_context_keys_dump(void)
1902 {
1903         int i;
1904
1905         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1906                 struct lu_context_key *key;
1907
1908                 key = lu_keys[i];
1909                 if (key != NULL) {
1910                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1911                                i, key, key->lct_tags,
1912                                key->lct_init, key->lct_fini, key->lct_exit,
1913                                key->lct_index, cfs_atomic_read(&key->lct_used),
1914                                key->lct_owner ? key->lct_owner->name : "",
1915                                key->lct_owner);
1916                         lu_ref_print(&key->lct_reference);
1917                 }
1918         }
1919 }
1920 EXPORT_SYMBOL(lu_context_keys_dump);
1921 #else  /* !__KERNEL__ */
1922 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1923 {
1924         return 0;
1925 }
1926 #endif /* __KERNEL__ */
1927
1928 int  cl_global_init(void);
1929 void cl_global_fini(void);
1930 int  lu_ref_global_init(void);
1931 void lu_ref_global_fini(void);
1932
1933 int dt_global_init(void);
1934 void dt_global_fini(void);
1935
1936 int llo_global_init(void);
1937 void llo_global_fini(void);
1938
1939 /* context key constructor/destructor: lu_ucred_key_init, lu_ucred_key_fini */
1940 LU_KEY_INIT_FINI(lu_ucred, struct lu_ucred);
1941
1942 static struct lu_context_key lu_ucred_key = {
1943         .lct_tags = LCT_SESSION,
1944         .lct_init = lu_ucred_key_init,
1945         .lct_fini = lu_ucred_key_fini
1946 };
1947
1948 /**
1949  * Get ucred key if session exists and ucred key is allocated on it.
1950  * Return NULL otherwise.
1951  */
1952 struct lu_ucred *lu_ucred(const struct lu_env *env)
1953 {
1954         if (!env->le_ses)
1955                 return NULL;
1956         return lu_context_key_get(env->le_ses, &lu_ucred_key);
1957 }
1958 EXPORT_SYMBOL(lu_ucred);
1959
1960 /**
1961  * Get ucred key and check if it is properly initialized.
1962  * Return NULL otherwise.
1963  */
1964 struct lu_ucred *lu_ucred_check(const struct lu_env *env)
1965 {
1966         struct lu_ucred *uc = lu_ucred(env);
1967         if (uc && uc->uc_valid != UCRED_OLD && uc->uc_valid != UCRED_NEW)
1968                 return NULL;
1969         return uc;
1970 }
1971 EXPORT_SYMBOL(lu_ucred_check);
1972
1973 /**
1974  * Get ucred key, which must exist and must be properly initialized.
1975  * Assert otherwise.
1976  */
1977 struct lu_ucred *lu_ucred_assert(const struct lu_env *env)
1978 {
1979         struct lu_ucred *uc = lu_ucred_check(env);
1980         LASSERT(uc != NULL);
1981         return uc;
1982 }
1983 EXPORT_SYMBOL(lu_ucred_assert);
1984
1985 /**
1986  * Initialization of global lu_* data.
1987  */
1988 int lu_global_init(void)
1989 {
1990         int result;
1991
1992         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1993
1994         result = lu_ref_global_init();
1995         if (result != 0)
1996                 return result;
1997
1998         LU_CONTEXT_KEY_INIT(&lu_global_key);
1999         result = lu_context_key_register(&lu_global_key);
2000         if (result != 0)
2001                 return result;
2002
2003         LU_CONTEXT_KEY_INIT(&lu_ucred_key);
2004         result = lu_context_key_register(&lu_ucred_key);
2005         if (result != 0)
2006                 return result;
2007
2008         /*
2009          * At this level, we don't know what tags are needed, so allocate them
2010          * conservatively. This should not be too bad, because this
2011          * environment is global.
2012          */
2013         mutex_lock(&lu_sites_guard);
2014         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2015         mutex_unlock(&lu_sites_guard);
2016         if (result != 0)
2017                 return result;
2018
2019         /*
2020          * seeks estimation: 3 seeks to read a record from oi, one to read
2021          * inode, one for ea. Unfortunately setting this high value results in
2022          * lu_object/inode cache consuming all the memory.
2023          */
2024         lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
2025         if (lu_site_shrinker == NULL)
2026                 return -ENOMEM;
2027
2028         result = lu_time_global_init();
2029         if (result)
2030                 GOTO(out, result);
2031
2032 #ifdef __KERNEL__
2033         result = dt_global_init();
2034         if (result)
2035                 GOTO(out, result);
2036
2037         result = llo_global_init();
2038         if (result)
2039                 GOTO(out, result);
2040 #endif
2041         result = cl_global_init();
2042 out:
2043
2044         return result;
2045 }
2046
2047 /**
2048  * Dual to lu_global_init().
2049  */
2050 void lu_global_fini(void)
2051 {
2052         cl_global_fini();
2053 #ifdef __KERNEL__
2054         llo_global_fini();
2055         dt_global_fini();
2056 #endif
2057         lu_time_global_fini();
2058         if (lu_site_shrinker != NULL) {
2059                 cfs_remove_shrinker(lu_site_shrinker);
2060                 lu_site_shrinker = NULL;
2061         }
2062
2063         lu_context_key_degister(&lu_global_key);
2064         lu_context_key_degister(&lu_ucred_key);
2065
2066         /*
2067          * Tear shrinker environment down _after_ de-registering
2068          * lu_global_key, because the latter has a value in the former.
2069          */
2070         mutex_lock(&lu_sites_guard);
2071         lu_env_fini(&lu_shrink_env);
2072         mutex_unlock(&lu_sites_guard);
2073
2074         lu_ref_global_fini();
2075 }
2076
2077 struct lu_buf LU_BUF_NULL = {
2078         .lb_buf = NULL,
2079         .lb_len = 0
2080 };
2081 EXPORT_SYMBOL(LU_BUF_NULL);
2082
2083 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2084 {
2085 #ifdef LPROCFS
2086         struct lprocfs_counter ret;
2087
2088         lprocfs_stats_collect(stats, idx, &ret);
2089         return (__u32)ret.lc_count;
2090 #else
2091         return 0;
2092 #endif
2093 }
2094
2095 /**
2096  * Output site statistical counters into a buffer. Suitable for
2097  * lprocfs_rd_*()-style functions.
2098  */
2099 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2100 {
2101         lu_site_stats_t stats;
2102
2103         memset(&stats, 0, sizeof(stats));
2104         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2105
2106         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2107                         stats.lss_busy,
2108                         stats.lss_total,
2109                         stats.lss_populated,
2110                         CFS_HASH_NHLIST(s->ls_obj_hash),
2111                         stats.lss_max_search,
2112                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
2113                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2114                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2115                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2116                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2117                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2118 }
2119 EXPORT_SYMBOL(lu_site_stats_print);
2120
2121 const char *lu_time_names[LU_TIME_NR] = {
2122         [LU_TIME_FIND_LOOKUP] = "find_lookup",
2123         [LU_TIME_FIND_ALLOC]  = "find_alloc",
2124         [LU_TIME_FIND_INSERT] = "find_insert"
2125 };
2126 EXPORT_SYMBOL(lu_time_names);
2127
2128 /**
2129  * Helper function to initialize a number of kmem slab caches at once.
2130  */
2131 int lu_kmem_init(struct lu_kmem_descr *caches)
2132 {
2133         int result;
2134         struct lu_kmem_descr *iter = caches;
2135
2136         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2137                 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2138                                                         iter->ckd_size,
2139                                                         0, 0);
2140                 if (*iter->ckd_cache == NULL) {
2141                         result = -ENOMEM;
2142                         /* free all previously allocated caches */
2143                         lu_kmem_fini(caches);
2144                         break;
2145                 }
2146         }
2147         return result;
2148 }
2149 EXPORT_SYMBOL(lu_kmem_init);
2150
2151 /**
2152  * Helper function to finalize a number of kmem slab cached at once. Dual to
2153  * lu_kmem_init().
2154  */
2155 void lu_kmem_fini(struct lu_kmem_descr *caches)
2156 {
2157         int rc;
2158
2159         for (; caches->ckd_cache != NULL; ++caches) {
2160                 if (*caches->ckd_cache != NULL) {
2161                         rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2162                         LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2163                                  caches->ckd_name);
2164                         *caches->ckd_cache = NULL;
2165                 }
2166         }
2167 }
2168 EXPORT_SYMBOL(lu_kmem_fini);
2169
2170 /**
2171  * Temporary solution to be able to assign fid in ->do_create()
2172  * till we have fully-functional OST fids
2173  */
2174 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2175                           const struct lu_fid *fid)
2176 {
2177         struct lu_site          *s = o->lo_dev->ld_site;
2178         struct lu_fid           *old = &o->lo_header->loh_fid;
2179         struct lu_site_bkt_data *bkt;
2180         struct lu_object        *shadow;
2181         cfs_waitlink_t           waiter;
2182         cfs_hash_t              *hs;
2183         cfs_hash_bd_t            bd;
2184         __u64                    version = 0;
2185
2186         LASSERT(fid_is_zero(old));
2187
2188         hs = s->ls_obj_hash;
2189         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2190         shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2191         /* supposed to be unique */
2192         LASSERT(shadow == NULL);
2193         *old = *fid;
2194         bkt = cfs_hash_bd_extra_get(hs, &bd);
2195         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2196         bkt->lsb_busy++;
2197         cfs_hash_bd_unlock(hs, &bd, 1);
2198 }
2199 EXPORT_SYMBOL(lu_object_assign_fid);
2200
2201 /**
2202  * allocates object with 0 (non-assiged) fid
2203  * XXX: temporary solution to be able to assign fid in ->do_create()
2204  *      till we have fully-functional OST fids
2205  */
2206 struct lu_object *lu_object_anon(const struct lu_env *env,
2207                                  struct lu_device *dev,
2208                                  const struct lu_object_conf *conf)
2209 {
2210         struct lu_fid     fid;
2211         struct lu_object *o;
2212
2213         fid_zero(&fid);
2214         o = lu_object_alloc(env, dev, &fid, conf);
2215
2216         return o;
2217 }
2218 EXPORT_SYMBOL(lu_object_anon);