Whamcloud - gitweb
3ddade986edb98e0cded6f7c74ec5a1b86940e69
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/obdclass/lu_object.c
33  *
34  * Lustre Object.
35  * These are the only exported functions, they provide some generic
36  * infrastructure for managing object devices
37  *
38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
39  */
40
41 #define DEBUG_SUBSYSTEM S_CLASS
42
43 #include <linux/module.h>
44 #include <linux/list.h>
45 #ifdef HAVE_PROCESSOR_H
46 #include <linux/processor.h>
47 #else
48 #include <libcfs/linux/processor.h>
49 #endif
50
51 #include <libcfs/libcfs.h>
52 #include <libcfs/libcfs_hash.h> /* hash_long() */
53 #include <libcfs/linux/linux-mem.h>
54 #include <obd_class.h>
55 #include <obd_support.h>
56 #include <lustre_disk.h>
57 #include <lustre_fid.h>
58 #include <lu_object.h>
59 #include <lu_ref.h>
60
61 struct lu_site_bkt_data {
62         /**
63          * LRU list, updated on each access to object. Protected by
64          * bucket lock of lu_site::ls_obj_hash.
65          *
66          * "Cold" end of LRU is lu_site::ls_lru.next. Accessed object are
67          * moved to the lu_site::ls_lru.prev (this is due to the non-existence
68          * of list_for_each_entry_safe_reverse()).
69          */
70         struct list_head                lsb_lru;
71         /**
72          * Wait-queue signaled when an object in this site is ultimately
73          * destroyed (lu_object_free()). It is used by lu_object_find() to
74          * wait before re-trying when object in the process of destruction is
75          * found in the hash table.
76          *
77          * \see htable_lookup().
78          */
79         wait_queue_head_t               lsb_marche_funebre;
80 };
81
82 enum {
83         LU_CACHE_PERCENT_MAX     = 50,
84         LU_CACHE_PERCENT_DEFAULT = 20
85 };
86
87 #define LU_CACHE_NR_MAX_ADJUST          512
88 #define LU_CACHE_NR_UNLIMITED           -1
89 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
90 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
91 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
92 #define LU_CACHE_NR_ZFS_LIMIT           10240
93
94 #define LU_SITE_BITS_MIN    12
95 #define LU_SITE_BITS_MAX    24
96 #define LU_SITE_BITS_MAX_CL 19
97 /**
98  * total 256 buckets, we don't want too many buckets because:
99  * - consume too much memory
100  * - avoid unbalanced LRU list
101  */
102 #define LU_SITE_BKT_BITS    8
103
104
105 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
106 module_param(lu_cache_percent, int, 0644);
107 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
108
109 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
110 module_param(lu_cache_nr, long, 0644);
111 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
112
113 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
114 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
115
116 wait_queue_head_t *
117 lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
118 {
119         struct cfs_hash_bd bd;
120         struct lu_site_bkt_data *bkt;
121
122         cfs_hash_bd_get(site->ls_obj_hash, fid, &bd);
123         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
124         return &bkt->lsb_marche_funebre;
125 }
126 EXPORT_SYMBOL(lu_site_wq_from_fid);
127
128 /**
129  * Decrease reference counter on object. If last reference is freed, return
130  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
131  * case, free object immediately.
132  */
133 void lu_object_put(const struct lu_env *env, struct lu_object *o)
134 {
135         struct lu_site_bkt_data *bkt;
136         struct lu_object_header *top;
137         struct lu_site *site;
138         struct lu_object *orig;
139         struct cfs_hash_bd bd;
140         const struct lu_fid *fid;
141
142         top  = o->lo_header;
143         site = o->lo_dev->ld_site;
144         orig = o;
145
146         /*
147          * till we have full fids-on-OST implemented anonymous objects
148          * are possible in OSP. such an object isn't listed in the site
149          * so we should not remove it from the site.
150          */
151         fid = lu_object_fid(o);
152         if (fid_is_zero(fid)) {
153                 LASSERT(top->loh_hash.next == NULL
154                         && top->loh_hash.pprev == NULL);
155                 LASSERT(list_empty(&top->loh_lru));
156                 if (!atomic_dec_and_test(&top->loh_ref))
157                         return;
158                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
159                         if (o->lo_ops->loo_object_release != NULL)
160                                 o->lo_ops->loo_object_release(env, o);
161                 }
162                 lu_object_free(env, orig);
163                 return;
164         }
165
166         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
167         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
168
169         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
170                 if (lu_object_is_dying(top)) {
171                         /*
172                          * somebody may be waiting for this, currently only
173                          * used for cl_object, see cl_object_put_last().
174                          */
175                         wake_up_all(&bkt->lsb_marche_funebre);
176                 }
177                 return;
178         }
179
180         /*
181          * When last reference is released, iterate over object
182          * layers, and notify them that object is no longer busy.
183          */
184         list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
185                 if (o->lo_ops->loo_object_release != NULL)
186                         o->lo_ops->loo_object_release(env, o);
187         }
188
189         if (!lu_object_is_dying(top) &&
190             (lu_object_exists(orig) || lu_object_is_cl(orig))) {
191                 LASSERT(list_empty(&top->loh_lru));
192                 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
193                 percpu_counter_inc(&site->ls_lru_len_counter);
194                 CDEBUG(D_INODE, "Add %p/%p to site lru. hash: %p, bkt: %p\n",
195                        orig, top, site->ls_obj_hash, bkt);
196                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
197                 return;
198         }
199
200         /*
201          * If object is dying (will not be cached) then remove it
202          * from hash table and LRU.
203          *
204          * This is done with hash table and LRU lists locked. As the only
205          * way to acquire first reference to previously unreferenced
206          * object is through hash-table lookup (lu_object_find()),
207          * or LRU scanning (lu_site_purge()), that are done under hash-table
208          * and LRU lock, no race with concurrent object lookup is possible
209          * and we can safely destroy object below.
210          */
211         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
212                 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
213         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
214         /*
215          * Object was already removed from hash and lru above, can
216          * kill it.
217          */
218         lu_object_free(env, orig);
219 }
220 EXPORT_SYMBOL(lu_object_put);
221
222 /**
223  * Put object and don't keep in cache. This is temporary solution for
224  * multi-site objects when its layering is not constant.
225  */
226 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
227 {
228         set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
229         return lu_object_put(env, o);
230 }
231 EXPORT_SYMBOL(lu_object_put_nocache);
232
233 /**
234  * Kill the object and take it out of LRU cache.
235  * Currently used by client code for layout change.
236  */
237 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
238 {
239         struct lu_object_header *top;
240
241         top = o->lo_header;
242         set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
243         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
244                 struct lu_site *site = o->lo_dev->ld_site;
245                 struct cfs_hash *obj_hash = site->ls_obj_hash;
246                 struct cfs_hash_bd bd;
247
248                 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
249                 if (!list_empty(&top->loh_lru)) {
250                         struct lu_site_bkt_data *bkt;
251
252                         list_del_init(&top->loh_lru);
253                         bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
254                         percpu_counter_dec(&site->ls_lru_len_counter);
255                 }
256                 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
257                 cfs_hash_bd_unlock(obj_hash, &bd, 1);
258         }
259 }
260 EXPORT_SYMBOL(lu_object_unhash);
261
262 /**
263  * Allocate new object.
264  *
265  * This follows object creation protocol, described in the comment within
266  * struct lu_device_operations definition.
267  */
268 static struct lu_object *lu_object_alloc(const struct lu_env *env,
269                                          struct lu_device *dev,
270                                          const struct lu_fid *f,
271                                          const struct lu_object_conf *conf)
272 {
273         struct lu_object *scan;
274         struct lu_object *top;
275         struct list_head *layers;
276         unsigned int init_mask = 0;
277         unsigned int init_flag;
278         int clean;
279         int result;
280         ENTRY;
281
282         /*
283          * Create top-level object slice. This will also create
284          * lu_object_header.
285          */
286         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
287         if (top == NULL)
288                 RETURN(ERR_PTR(-ENOMEM));
289         if (IS_ERR(top))
290                 RETURN(top);
291         /*
292          * This is the only place where object fid is assigned. It's constant
293          * after this point.
294          */
295         top->lo_header->loh_fid = *f;
296         layers = &top->lo_header->loh_layers;
297
298         do {
299                 /*
300                  * Call ->loo_object_init() repeatedly, until no more new
301                  * object slices are created.
302                  */
303                 clean = 1;
304                 init_flag = 1;
305                 list_for_each_entry(scan, layers, lo_linkage) {
306                         if (init_mask & init_flag)
307                                 goto next;
308                         clean = 0;
309                         scan->lo_header = top->lo_header;
310                         result = scan->lo_ops->loo_object_init(env, scan, conf);
311                         if (result != 0) {
312                                 lu_object_free(env, top);
313                                 RETURN(ERR_PTR(result));
314                         }
315                         init_mask |= init_flag;
316 next:
317                         init_flag <<= 1;
318                 }
319         } while (!clean);
320
321         list_for_each_entry_reverse(scan, layers, lo_linkage) {
322                 if (scan->lo_ops->loo_object_start != NULL) {
323                         result = scan->lo_ops->loo_object_start(env, scan);
324                         if (result != 0) {
325                                 lu_object_free(env, top);
326                                 RETURN(ERR_PTR(result));
327                         }
328                 }
329         }
330
331         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
332         RETURN(top);
333 }
334
335 /**
336  * Free an object.
337  */
338 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
339 {
340         wait_queue_head_t *wq;
341         struct lu_site          *site;
342         struct lu_object        *scan;
343         struct list_head        *layers;
344         struct list_head         splice;
345
346         site = o->lo_dev->ld_site;
347         layers = &o->lo_header->loh_layers;
348         wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
349         /*
350          * First call ->loo_object_delete() method to release all resources.
351          */
352         list_for_each_entry_reverse(scan, layers, lo_linkage) {
353                 if (scan->lo_ops->loo_object_delete != NULL)
354                         scan->lo_ops->loo_object_delete(env, scan);
355         }
356
357         /*
358          * Then, splice object layers into stand-alone list, and call
359          * ->loo_object_free() on all layers to free memory. Splice is
360          * necessary, because lu_object_header is freed together with the
361          * top-level slice.
362          */
363         INIT_LIST_HEAD(&splice);
364         list_splice_init(layers, &splice);
365         while (!list_empty(&splice)) {
366                 /*
367                  * Free layers in bottom-to-top order, so that object header
368                  * lives as long as possible and ->loo_object_free() methods
369                  * can look at its contents.
370                  */
371                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
372                 list_del_init(&o->lo_linkage);
373                 LASSERT(o->lo_ops->loo_object_free != NULL);
374                 o->lo_ops->loo_object_free(env, o);
375         }
376
377         if (waitqueue_active(wq))
378                 wake_up_all(wq);
379 }
380
381 /**
382  * Free \a nr objects from the cold end of the site LRU list.
383  * if canblock is 0, then don't block awaiting for another
384  * instance of lu_site_purge() to complete
385  */
386 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
387                           int nr, int canblock)
388 {
389         struct lu_object_header *h;
390         struct lu_object_header *temp;
391         struct lu_site_bkt_data *bkt;
392         struct cfs_hash_bd            bd;
393         struct cfs_hash_bd            bd2;
394         struct list_head         dispose;
395         int                      did_sth;
396         unsigned int             start = 0;
397         int                      count;
398         int                      bnr;
399         unsigned int             i;
400
401         if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
402                 RETURN(0);
403
404         INIT_LIST_HEAD(&dispose);
405         /*
406          * Under LRU list lock, scan LRU list and move unreferenced objects to
407          * the dispose list, removing them from LRU and hash table.
408          */
409         if (nr != ~0)
410                 start = s->ls_purge_start;
411         bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
412  again:
413         /*
414          * It doesn't make any sense to make purge threads parallel, that can
415          * only bring troubles to us. See LU-5331.
416          */
417         if (canblock != 0)
418                 mutex_lock(&s->ls_purge_mutex);
419         else if (mutex_trylock(&s->ls_purge_mutex) == 0)
420                 goto out;
421
422         did_sth = 0;
423         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
424                 if (i < start)
425                         continue;
426                 count = bnr;
427                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
428                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
429
430                 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
431                         LASSERT(atomic_read(&h->loh_ref) == 0);
432
433                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
434                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
435
436                         cfs_hash_bd_del_locked(s->ls_obj_hash,
437                                                &bd2, &h->loh_hash);
438                         list_move(&h->loh_lru, &dispose);
439                         percpu_counter_dec(&s->ls_lru_len_counter);
440                         if (did_sth == 0)
441                                 did_sth = 1;
442
443                         if (nr != ~0 && --nr == 0)
444                                 break;
445
446                         if (count > 0 && --count == 0)
447                                 break;
448
449                 }
450                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
451                 cond_resched();
452                 /*
453                  * Free everything on the dispose list. This is safe against
454                  * races due to the reasons described in lu_object_put().
455                  */
456                 while (!list_empty(&dispose)) {
457                         h = container_of0(dispose.next,
458                                           struct lu_object_header, loh_lru);
459                         list_del_init(&h->loh_lru);
460                         lu_object_free(env, lu_object_top(h));
461                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
462                 }
463
464                 if (nr == 0)
465                         break;
466         }
467         mutex_unlock(&s->ls_purge_mutex);
468
469         if (nr != 0 && did_sth && start != 0) {
470                 start = 0; /* restart from the first bucket */
471                 goto again;
472         }
473         /* race on s->ls_purge_start, but nobody cares */
474         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
475
476 out:
477         return nr;
478 }
479 EXPORT_SYMBOL(lu_site_purge_objects);
480
481 /*
482  * Object printing.
483  *
484  * Code below has to jump through certain loops to output object description
485  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
486  * composes object description from strings that are parts of _lines_ of
487  * output (i.e., strings that are not terminated by newline). This doesn't fit
488  * very well into libcfs_debug_msg() interface that assumes that each message
489  * supplied to it is a self-contained output line.
490  *
491  * To work around this, strings are collected in a temporary buffer
492  * (implemented as a value of lu_cdebug_key key), until terminating newline
493  * character is detected.
494  *
495  */
496
497 enum {
498         /**
499          * Maximal line size.
500          *
501          * XXX overflow is not handled correctly.
502          */
503         LU_CDEBUG_LINE = 512
504 };
505
506 struct lu_cdebug_data {
507         /**
508          * Temporary buffer.
509          */
510         char lck_area[LU_CDEBUG_LINE];
511 };
512
513 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
514 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
515
516 /**
517  * Key, holding temporary buffer. This key is registered very early by
518  * lu_global_init().
519  */
520 static struct lu_context_key lu_global_key = {
521         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
522                     LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
523         .lct_init = lu_global_key_init,
524         .lct_fini = lu_global_key_fini
525 };
526
527 /**
528  * Printer function emitting messages through libcfs_debug_msg().
529  */
530 int lu_cdebug_printer(const struct lu_env *env,
531                       void *cookie, const char *format, ...)
532 {
533         struct libcfs_debug_msg_data *msgdata = cookie;
534         struct lu_cdebug_data        *key;
535         int used;
536         int complete;
537         va_list args;
538
539         va_start(args, format);
540
541         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
542         LASSERT(key != NULL);
543
544         used = strlen(key->lck_area);
545         complete = format[strlen(format) - 1] == '\n';
546         /*
547          * Append new chunk to the buffer.
548          */
549         vsnprintf(key->lck_area + used,
550                   ARRAY_SIZE(key->lck_area) - used, format, args);
551         if (complete) {
552                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
553                         libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
554                 key->lck_area[0] = 0;
555         }
556         va_end(args);
557         return 0;
558 }
559 EXPORT_SYMBOL(lu_cdebug_printer);
560
561 /**
562  * Print object header.
563  */
564 void lu_object_header_print(const struct lu_env *env, void *cookie,
565                             lu_printer_t printer,
566                             const struct lu_object_header *hdr)
567 {
568         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
569                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
570                    PFID(&hdr->loh_fid),
571                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
572                    list_empty((struct list_head *)&hdr->loh_lru) ? \
573                    "" : " lru",
574                    hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
575 }
576 EXPORT_SYMBOL(lu_object_header_print);
577
578 /**
579  * Print human readable representation of the \a o to the \a printer.
580  */
581 void lu_object_print(const struct lu_env *env, void *cookie,
582                      lu_printer_t printer, const struct lu_object *o)
583 {
584         static const char ruler[] = "........................................";
585         struct lu_object_header *top;
586         int depth = 4;
587
588         top = o->lo_header;
589         lu_object_header_print(env, cookie, printer, top);
590         (*printer)(env, cookie, "{\n");
591
592         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
593                 /*
594                  * print `.' \a depth times followed by type name and address
595                  */
596                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
597                            o->lo_dev->ld_type->ldt_name, o);
598
599                 if (o->lo_ops->loo_object_print != NULL)
600                         (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
601
602                 (*printer)(env, cookie, "\n");
603         }
604
605         (*printer)(env, cookie, "} header@%p\n", top);
606 }
607 EXPORT_SYMBOL(lu_object_print);
608
609 /**
610  * Check object consistency.
611  */
612 int lu_object_invariant(const struct lu_object *o)
613 {
614         struct lu_object_header *top;
615
616         top = o->lo_header;
617         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
618                 if (o->lo_ops->loo_object_invariant != NULL &&
619                     !o->lo_ops->loo_object_invariant(o))
620                         return 0;
621         }
622         return 1;
623 }
624
625 static struct lu_object *htable_lookup(struct lu_site *s,
626                                        struct cfs_hash_bd *bd,
627                                        const struct lu_fid *f,
628                                        __u64 *version)
629 {
630         struct lu_site_bkt_data *bkt;
631         struct lu_object_header *h;
632         struct hlist_node *hnode;
633         __u64 ver = cfs_hash_bd_version_get(bd);
634
635         if (*version == ver)
636                 return ERR_PTR(-ENOENT);
637
638         *version = ver;
639         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
640         /* cfs_hash_bd_peek_locked is a somehow "internal" function
641          * of cfs_hash, it doesn't add refcount on object. */
642         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
643         if (!hnode) {
644                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
645                 return ERR_PTR(-ENOENT);
646         }
647
648         h = container_of0(hnode, struct lu_object_header, loh_hash);
649         cfs_hash_get(s->ls_obj_hash, hnode);
650         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
651         if (!list_empty(&h->loh_lru)) {
652                 list_del_init(&h->loh_lru);
653                 percpu_counter_dec(&s->ls_lru_len_counter);
654         }
655         return lu_object_top(h);
656 }
657
658 /**
659  * Search cache for an object with the fid \a f. If such object is found,
660  * return it. Otherwise, create new object, insert it into cache and return
661  * it. In any case, additional reference is acquired on the returned object.
662  */
663 struct lu_object *lu_object_find(const struct lu_env *env,
664                                  struct lu_device *dev, const struct lu_fid *f,
665                                  const struct lu_object_conf *conf)
666 {
667         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
668 }
669 EXPORT_SYMBOL(lu_object_find);
670
671 /*
672  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
673  * the calculation for the number of objects to reclaim is not covered by
674  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
675  * This ensures that many concurrent threads will not accidentally purge
676  * the entire cache.
677  */
678 static void lu_object_limit(const struct lu_env *env,
679                             struct lu_device *dev)
680 {
681         __u64 size, nr;
682
683         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
684                 return;
685
686         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
687         nr = (__u64)lu_cache_nr;
688         if (size <= nr)
689                 return;
690
691         lu_site_purge_objects(env, dev->ld_site,
692                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
693 }
694
695 /**
696  * Core logic of lu_object_find*() functions.
697  *
698  * Much like lu_object_find(), but top level device of object is specifically
699  * \a dev rather than top level device of the site. This interface allows
700  * objects of different "stacking" to be created within the same site.
701  */
702 struct lu_object *lu_object_find_at(const struct lu_env *env,
703                                     struct lu_device *dev,
704                                     const struct lu_fid *f,
705                                     const struct lu_object_conf *conf)
706 {
707         struct lu_object *o;
708         struct lu_object *shadow;
709         struct lu_site *s;
710         struct cfs_hash *hs;
711         struct cfs_hash_bd bd;
712         __u64 version = 0;
713
714         /*
715          * This uses standard index maintenance protocol:
716          *
717          *     - search index under lock, and return object if found;
718          *     - otherwise, unlock index, allocate new object;
719          *     - lock index and search again;
720          *     - if nothing is found (usual case), insert newly created
721          *       object into index;
722          *     - otherwise (race: other thread inserted object), free
723          *       object just allocated.
724          *     - unlock index;
725          *     - return object.
726          *
727          * For "LOC_F_NEW" case, we are sure the object is new established.
728          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
729          * just alloc and insert directly.
730          *
731          */
732         s  = dev->ld_site;
733         hs = s->ls_obj_hash;
734         cfs_hash_bd_get(hs, f, &bd);
735         if (!(conf && conf->loc_flags & LOC_F_NEW)) {
736                 cfs_hash_bd_lock(hs, &bd, 1);
737                 o = htable_lookup(s, &bd, f, &version);
738                 cfs_hash_bd_unlock(hs, &bd, 1);
739
740                 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
741                         return o;
742         }
743         /*
744          * Allocate new object. This may result in rather complicated
745          * operations, including fld queries, inode loading, etc.
746          */
747         o = lu_object_alloc(env, dev, f, conf);
748         if (IS_ERR(o))
749                 return o;
750
751         LASSERT(lu_fid_eq(lu_object_fid(o), f));
752
753         cfs_hash_bd_lock(hs, &bd, 1);
754
755         if (conf && conf->loc_flags & LOC_F_NEW)
756                 shadow = ERR_PTR(-ENOENT);
757         else
758                 shadow = htable_lookup(s, &bd, f, &version);
759         if (likely(PTR_ERR(shadow) == -ENOENT)) {
760                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
761                 cfs_hash_bd_unlock(hs, &bd, 1);
762
763                 lu_object_limit(env, dev);
764
765                 return o;
766         }
767
768         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
769         cfs_hash_bd_unlock(hs, &bd, 1);
770         lu_object_free(env, o);
771         return shadow;
772 }
773 EXPORT_SYMBOL(lu_object_find_at);
774
775 /**
776  * Find object with given fid, and return its slice belonging to given device.
777  */
778 struct lu_object *lu_object_find_slice(const struct lu_env *env,
779                                        struct lu_device *dev,
780                                        const struct lu_fid *f,
781                                        const struct lu_object_conf *conf)
782 {
783         struct lu_object *top;
784         struct lu_object *obj;
785
786         top = lu_object_find(env, dev, f, conf);
787         if (IS_ERR(top))
788                 return top;
789
790         obj = lu_object_locate(top->lo_header, dev->ld_type);
791         if (unlikely(obj == NULL)) {
792                 lu_object_put(env, top);
793                 obj = ERR_PTR(-ENOENT);
794         }
795
796         return obj;
797 }
798 EXPORT_SYMBOL(lu_object_find_slice);
799
800 int lu_device_type_init(struct lu_device_type *ldt)
801 {
802         int result = 0;
803
804         atomic_set(&ldt->ldt_device_nr, 0);
805         if (ldt->ldt_ops->ldto_init)
806                 result = ldt->ldt_ops->ldto_init(ldt);
807
808         return result;
809 }
810 EXPORT_SYMBOL(lu_device_type_init);
811
812 void lu_device_type_fini(struct lu_device_type *ldt)
813 {
814         if (ldt->ldt_ops->ldto_fini)
815                 ldt->ldt_ops->ldto_fini(ldt);
816 }
817 EXPORT_SYMBOL(lu_device_type_fini);
818
819 /**
820  * Global list of all sites on this node
821  */
822 static LIST_HEAD(lu_sites);
823 static DECLARE_RWSEM(lu_sites_guard);
824
825 /**
826  * Global environment used by site shrinker.
827  */
828 static struct lu_env lu_shrink_env;
829
830 struct lu_site_print_arg {
831         struct lu_env   *lsp_env;
832         void            *lsp_cookie;
833         lu_printer_t     lsp_printer;
834 };
835
836 static int
837 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
838                   struct hlist_node *hnode, void *data)
839 {
840         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
841         struct lu_object_header  *h;
842
843         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
844         if (!list_empty(&h->loh_layers)) {
845                 const struct lu_object *o;
846
847                 o = lu_object_top(h);
848                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
849                                 arg->lsp_printer, o);
850         } else {
851                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
852                                        arg->lsp_printer, h);
853         }
854         return 0;
855 }
856
857 /**
858  * Print all objects in \a s.
859  */
860 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
861                    lu_printer_t printer)
862 {
863         struct lu_site_print_arg arg = {
864                 .lsp_env     = (struct lu_env *)env,
865                 .lsp_cookie  = cookie,
866                 .lsp_printer = printer,
867         };
868
869         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
870 }
871 EXPORT_SYMBOL(lu_site_print);
872
873 /**
874  * Return desired hash table order.
875  */
876 static unsigned long lu_htable_order(struct lu_device *top)
877 {
878         unsigned long cache_size;
879         unsigned long bits;
880         unsigned long bits_max = LU_SITE_BITS_MAX;
881
882         /*
883          * For ZFS based OSDs the cache should be disabled by default.  This
884          * allows the ZFS ARC maximum flexibility in determining what buffers
885          * to cache.  If Lustre has objects or buffer which it wants to ensure
886          * always stay cached it must maintain a hold on them.
887          */
888         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
889                 lu_cache_percent = 1;
890                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
891                 return LU_SITE_BITS_MIN;
892         }
893
894         if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
895                 bits_max = LU_SITE_BITS_MAX_CL;
896
897         /*
898          * Calculate hash table size, assuming that we want reasonable
899          * performance when 20% of total memory is occupied by cache of
900          * lu_objects.
901          *
902          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
903          */
904         cache_size = totalram_pages;
905
906 #if BITS_PER_LONG == 32
907         /* limit hashtable size for lowmem systems to low RAM */
908         if (cache_size > 1 << (30 - PAGE_SHIFT))
909                 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
910 #endif
911
912         /* clear off unreasonable cache setting. */
913         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
914                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
915                       " the range of (0, %u]. Will use default value: %u.\n",
916                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
917                       LU_CACHE_PERCENT_DEFAULT);
918
919                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
920         }
921         cache_size = cache_size / 100 * lu_cache_percent *
922                 (PAGE_SIZE / 1024);
923
924         for (bits = 1; (1 << bits) < cache_size; ++bits) {
925                 ;
926         }
927
928         return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
929 }
930
931 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
932                                 const void *key, unsigned mask)
933 {
934         struct lu_fid  *fid = (struct lu_fid *)key;
935         __u32           hash;
936
937         hash = fid_flatten32(fid);
938         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
939         hash = hash_long(hash, hs->hs_bkt_bits);
940
941         /* give me another random factor */
942         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
943
944         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
945         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
946
947         return hash & mask;
948 }
949
950 static void *lu_obj_hop_object(struct hlist_node *hnode)
951 {
952         return hlist_entry(hnode, struct lu_object_header, loh_hash);
953 }
954
955 static void *lu_obj_hop_key(struct hlist_node *hnode)
956 {
957         struct lu_object_header *h;
958
959         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
960         return &h->loh_fid;
961 }
962
963 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
964 {
965         struct lu_object_header *h;
966
967         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
968         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
969 }
970
971 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
972 {
973         struct lu_object_header *h;
974
975         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
976         atomic_inc(&h->loh_ref);
977 }
978
979 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
980 {
981         LBUG(); /* we should never called it */
982 }
983
984 static struct cfs_hash_ops lu_site_hash_ops = {
985         .hs_hash        = lu_obj_hop_hash,
986         .hs_key         = lu_obj_hop_key,
987         .hs_keycmp      = lu_obj_hop_keycmp,
988         .hs_object      = lu_obj_hop_object,
989         .hs_get         = lu_obj_hop_get,
990         .hs_put_locked  = lu_obj_hop_put_locked,
991 };
992
993 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
994 {
995         spin_lock(&s->ls_ld_lock);
996         if (list_empty(&d->ld_linkage))
997                 list_add(&d->ld_linkage, &s->ls_ld_linkage);
998         spin_unlock(&s->ls_ld_lock);
999 }
1000 EXPORT_SYMBOL(lu_dev_add_linkage);
1001
1002 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1003 {
1004         spin_lock(&s->ls_ld_lock);
1005         list_del_init(&d->ld_linkage);
1006         spin_unlock(&s->ls_ld_lock);
1007 }
1008 EXPORT_SYMBOL(lu_dev_del_linkage);
1009
1010 /**
1011   * Initialize site \a s, with \a d as the top level device.
1012   */
1013 int lu_site_init(struct lu_site *s, struct lu_device *top)
1014 {
1015         struct lu_site_bkt_data *bkt;
1016         struct cfs_hash_bd bd;
1017         char name[16];
1018         unsigned long bits;
1019         unsigned int i;
1020         int rc;
1021         ENTRY;
1022
1023         memset(s, 0, sizeof *s);
1024         mutex_init(&s->ls_purge_mutex);
1025
1026 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1027         rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1028 #else
1029         rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1030 #endif
1031         if (rc)
1032                 return -ENOMEM;
1033
1034         snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1035         for (bits = lu_htable_order(top);
1036              bits >= LU_SITE_BITS_MIN; bits--) {
1037                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1038                                                  bits - LU_SITE_BKT_BITS,
1039                                                  sizeof(*bkt), 0, 0,
1040                                                  &lu_site_hash_ops,
1041                                                  CFS_HASH_SPIN_BKTLOCK |
1042                                                  CFS_HASH_NO_ITEMREF |
1043                                                  CFS_HASH_DEPTH |
1044                                                  CFS_HASH_ASSERT_EMPTY |
1045                                                  CFS_HASH_COUNTER);
1046                 if (s->ls_obj_hash != NULL)
1047                         break;
1048         }
1049
1050         if (s->ls_obj_hash == NULL) {
1051                 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1052                 return -ENOMEM;
1053         }
1054
1055         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1056                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1057                 INIT_LIST_HEAD(&bkt->lsb_lru);
1058                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1059         }
1060
1061         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1062         if (s->ls_stats == NULL) {
1063                 cfs_hash_putref(s->ls_obj_hash);
1064                 s->ls_obj_hash = NULL;
1065                 return -ENOMEM;
1066         }
1067
1068         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1069                              0, "created", "created");
1070         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1071                              0, "cache_hit", "cache_hit");
1072         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1073                              0, "cache_miss", "cache_miss");
1074         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1075                              0, "cache_race", "cache_race");
1076         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1077                              0, "cache_death_race", "cache_death_race");
1078         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1079                              0, "lru_purged", "lru_purged");
1080
1081         INIT_LIST_HEAD(&s->ls_linkage);
1082         s->ls_top_dev = top;
1083         top->ld_site = s;
1084         lu_device_get(top);
1085         lu_ref_add(&top->ld_reference, "site-top", s);
1086
1087         INIT_LIST_HEAD(&s->ls_ld_linkage);
1088         spin_lock_init(&s->ls_ld_lock);
1089
1090         lu_dev_add_linkage(s, top);
1091
1092         RETURN(0);
1093 }
1094 EXPORT_SYMBOL(lu_site_init);
1095
1096 /**
1097  * Finalize \a s and release its resources.
1098  */
1099 void lu_site_fini(struct lu_site *s)
1100 {
1101         down_write(&lu_sites_guard);
1102         list_del_init(&s->ls_linkage);
1103         up_write(&lu_sites_guard);
1104
1105         percpu_counter_destroy(&s->ls_lru_len_counter);
1106
1107         if (s->ls_obj_hash != NULL) {
1108                 cfs_hash_putref(s->ls_obj_hash);
1109                 s->ls_obj_hash = NULL;
1110         }
1111
1112         if (s->ls_top_dev != NULL) {
1113                 s->ls_top_dev->ld_site = NULL;
1114                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1115                 lu_device_put(s->ls_top_dev);
1116                 s->ls_top_dev = NULL;
1117         }
1118
1119         if (s->ls_stats != NULL)
1120                 lprocfs_free_stats(&s->ls_stats);
1121 }
1122 EXPORT_SYMBOL(lu_site_fini);
1123
1124 /**
1125  * Called when initialization of stack for this site is completed.
1126  */
1127 int lu_site_init_finish(struct lu_site *s)
1128 {
1129         int result;
1130         down_write(&lu_sites_guard);
1131         result = lu_context_refill(&lu_shrink_env.le_ctx);
1132         if (result == 0)
1133                 list_add(&s->ls_linkage, &lu_sites);
1134         up_write(&lu_sites_guard);
1135         return result;
1136 }
1137 EXPORT_SYMBOL(lu_site_init_finish);
1138
1139 /**
1140  * Acquire additional reference on device \a d
1141  */
1142 void lu_device_get(struct lu_device *d)
1143 {
1144         atomic_inc(&d->ld_ref);
1145 }
1146 EXPORT_SYMBOL(lu_device_get);
1147
1148 /**
1149  * Release reference on device \a d.
1150  */
1151 void lu_device_put(struct lu_device *d)
1152 {
1153         LASSERT(atomic_read(&d->ld_ref) > 0);
1154         atomic_dec(&d->ld_ref);
1155 }
1156 EXPORT_SYMBOL(lu_device_put);
1157
1158 /**
1159  * Initialize device \a d of type \a t.
1160  */
1161 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1162 {
1163         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1164             t->ldt_ops->ldto_start != NULL)
1165                 t->ldt_ops->ldto_start(t);
1166
1167         memset(d, 0, sizeof *d);
1168         d->ld_type = t;
1169         lu_ref_init(&d->ld_reference);
1170         INIT_LIST_HEAD(&d->ld_linkage);
1171
1172         return 0;
1173 }
1174 EXPORT_SYMBOL(lu_device_init);
1175
1176 /**
1177  * Finalize device \a d.
1178  */
1179 void lu_device_fini(struct lu_device *d)
1180 {
1181         struct lu_device_type *t = d->ld_type;
1182
1183         if (d->ld_obd != NULL) {
1184                 d->ld_obd->obd_lu_dev = NULL;
1185                 d->ld_obd = NULL;
1186         }
1187
1188         lu_ref_fini(&d->ld_reference);
1189         LASSERTF(atomic_read(&d->ld_ref) == 0,
1190                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1191         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1192
1193         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1194             t->ldt_ops->ldto_stop != NULL)
1195                 t->ldt_ops->ldto_stop(t);
1196 }
1197 EXPORT_SYMBOL(lu_device_fini);
1198
1199 /**
1200  * Initialize object \a o that is part of compound object \a h and was created
1201  * by device \a d.
1202  */
1203 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1204                    struct lu_device *d)
1205 {
1206         memset(o, 0, sizeof(*o));
1207         o->lo_header = h;
1208         o->lo_dev = d;
1209         lu_device_get(d);
1210         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1211         INIT_LIST_HEAD(&o->lo_linkage);
1212
1213         return 0;
1214 }
1215 EXPORT_SYMBOL(lu_object_init);
1216
1217 /**
1218  * Finalize object and release its resources.
1219  */
1220 void lu_object_fini(struct lu_object *o)
1221 {
1222         struct lu_device *dev = o->lo_dev;
1223
1224         LASSERT(list_empty(&o->lo_linkage));
1225
1226         if (dev != NULL) {
1227                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1228                               "lu_object", o);
1229                 lu_device_put(dev);
1230                 o->lo_dev = NULL;
1231         }
1232 }
1233 EXPORT_SYMBOL(lu_object_fini);
1234
1235 /**
1236  * Add object \a o as first layer of compound object \a h
1237  *
1238  * This is typically called by the ->ldo_object_alloc() method of top-level
1239  * device.
1240  */
1241 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1242 {
1243         list_move(&o->lo_linkage, &h->loh_layers);
1244 }
1245 EXPORT_SYMBOL(lu_object_add_top);
1246
1247 /**
1248  * Add object \a o as a layer of compound object, going after \a before.
1249  *
1250  * This is typically called by the ->ldo_object_alloc() method of \a
1251  * before->lo_dev.
1252  */
1253 void lu_object_add(struct lu_object *before, struct lu_object *o)
1254 {
1255         list_move(&o->lo_linkage, &before->lo_linkage);
1256 }
1257 EXPORT_SYMBOL(lu_object_add);
1258
1259 /**
1260  * Initialize compound object.
1261  */
1262 int lu_object_header_init(struct lu_object_header *h)
1263 {
1264         memset(h, 0, sizeof *h);
1265         atomic_set(&h->loh_ref, 1);
1266         INIT_HLIST_NODE(&h->loh_hash);
1267         INIT_LIST_HEAD(&h->loh_lru);
1268         INIT_LIST_HEAD(&h->loh_layers);
1269         lu_ref_init(&h->loh_reference);
1270         return 0;
1271 }
1272 EXPORT_SYMBOL(lu_object_header_init);
1273
1274 /**
1275  * Finalize compound object.
1276  */
1277 void lu_object_header_fini(struct lu_object_header *h)
1278 {
1279         LASSERT(list_empty(&h->loh_layers));
1280         LASSERT(list_empty(&h->loh_lru));
1281         LASSERT(hlist_unhashed(&h->loh_hash));
1282         lu_ref_fini(&h->loh_reference);
1283 }
1284 EXPORT_SYMBOL(lu_object_header_fini);
1285
1286 /**
1287  * Given a compound object, find its slice, corresponding to the device type
1288  * \a dtype.
1289  */
1290 struct lu_object *lu_object_locate(struct lu_object_header *h,
1291                                    const struct lu_device_type *dtype)
1292 {
1293         struct lu_object *o;
1294
1295         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1296                 if (o->lo_dev->ld_type == dtype)
1297                         return o;
1298         }
1299         return NULL;
1300 }
1301 EXPORT_SYMBOL(lu_object_locate);
1302
1303 /**
1304  * Finalize and free devices in the device stack.
1305  *
1306  * Finalize device stack by purging object cache, and calling
1307  * lu_device_type_operations::ldto_device_fini() and
1308  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1309  */
1310 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1311 {
1312         struct lu_site   *site = top->ld_site;
1313         struct lu_device *scan;
1314         struct lu_device *next;
1315
1316         lu_site_purge(env, site, ~0);
1317         for (scan = top; scan != NULL; scan = next) {
1318                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1319                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1320                 lu_device_put(scan);
1321         }
1322
1323         /* purge again. */
1324         lu_site_purge(env, site, ~0);
1325
1326         for (scan = top; scan != NULL; scan = next) {
1327                 const struct lu_device_type *ldt = scan->ld_type;
1328                 struct obd_type             *type;
1329
1330                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1331                 type = ldt->ldt_obd_type;
1332                 if (type != NULL) {
1333                         type->typ_refcnt--;
1334                         class_put_type(type);
1335                 }
1336         }
1337 }
1338
1339 enum {
1340         /**
1341          * Maximal number of tld slots.
1342          */
1343         LU_CONTEXT_KEY_NR = 40
1344 };
1345
1346 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1347
1348 DEFINE_RWLOCK(lu_keys_guard);
1349 static DECLARE_RWSEM(lu_key_initing);
1350
1351 /**
1352  * Global counter incremented whenever key is registered, unregistered,
1353  * revived or quiesced. This is used to void unnecessary calls to
1354  * lu_context_refill(). No locking is provided, as initialization and shutdown
1355  * are supposed to be externally serialized.
1356  */
1357 static atomic_t key_set_version = ATOMIC_INIT(0);
1358
1359 /**
1360  * Register new key.
1361  */
1362 int lu_context_key_register(struct lu_context_key *key)
1363 {
1364         int result;
1365         unsigned int i;
1366
1367         LASSERT(key->lct_init != NULL);
1368         LASSERT(key->lct_fini != NULL);
1369         LASSERT(key->lct_tags != 0);
1370         LASSERT(key->lct_owner != NULL);
1371
1372         result = -ENFILE;
1373         write_lock(&lu_keys_guard);
1374         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1375                 if (lu_keys[i] == NULL) {
1376                         key->lct_index = i;
1377                         atomic_set(&key->lct_used, 1);
1378                         lu_keys[i] = key;
1379                         lu_ref_init(&key->lct_reference);
1380                         result = 0;
1381                         atomic_inc(&key_set_version);
1382                         break;
1383                 }
1384         }
1385         write_unlock(&lu_keys_guard);
1386         return result;
1387 }
1388 EXPORT_SYMBOL(lu_context_key_register);
1389
1390 static void key_fini(struct lu_context *ctx, int index)
1391 {
1392         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1393                 struct lu_context_key *key;
1394
1395                 key = lu_keys[index];
1396                 LASSERT(key != NULL);
1397                 LASSERT(key->lct_fini != NULL);
1398                 LASSERT(atomic_read(&key->lct_used) > 1);
1399
1400                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1401                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1402                 if (atomic_dec_and_test(&key->lct_used))
1403                         wake_up_var(&key->lct_used);
1404
1405                 LASSERT(key->lct_owner != NULL);
1406                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1407                         LINVRNT(module_refcount(key->lct_owner) > 0);
1408                         module_put(key->lct_owner);
1409                 }
1410                 ctx->lc_value[index] = NULL;
1411         }
1412 }
1413
1414 /**
1415  * Deregister key.
1416  */
1417 void lu_context_key_degister(struct lu_context_key *key)
1418 {
1419         LASSERT(atomic_read(&key->lct_used) >= 1);
1420         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1421
1422         lu_context_key_quiesce(key);
1423
1424         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1425
1426         /**
1427          * Wait until all transient contexts referencing this key have
1428          * run lu_context_key::lct_fini() method.
1429          */
1430         atomic_dec(&key->lct_used);
1431         wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1432
1433         write_lock(&lu_keys_guard);
1434         if (lu_keys[key->lct_index]) {
1435                 lu_keys[key->lct_index] = NULL;
1436                 lu_ref_fini(&key->lct_reference);
1437         }
1438         write_unlock(&lu_keys_guard);
1439
1440         LASSERTF(atomic_read(&key->lct_used) == 0,
1441                  "key has instances: %d\n",
1442                  atomic_read(&key->lct_used));
1443 }
1444 EXPORT_SYMBOL(lu_context_key_degister);
1445
1446 /**
1447  * Register a number of keys. This has to be called after all keys have been
1448  * initialized by a call to LU_CONTEXT_KEY_INIT().
1449  */
1450 int lu_context_key_register_many(struct lu_context_key *k, ...)
1451 {
1452         struct lu_context_key *key = k;
1453         va_list args;
1454         int result;
1455
1456         va_start(args, k);
1457         do {
1458                 result = lu_context_key_register(key);
1459                 if (result)
1460                         break;
1461                 key = va_arg(args, struct lu_context_key *);
1462         } while (key != NULL);
1463         va_end(args);
1464
1465         if (result != 0) {
1466                 va_start(args, k);
1467                 while (k != key) {
1468                         lu_context_key_degister(k);
1469                         k = va_arg(args, struct lu_context_key *);
1470                 }
1471                 va_end(args);
1472         }
1473
1474         return result;
1475 }
1476 EXPORT_SYMBOL(lu_context_key_register_many);
1477
1478 /**
1479  * De-register a number of keys. This is a dual to
1480  * lu_context_key_register_many().
1481  */
1482 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1483 {
1484         va_list args;
1485
1486         va_start(args, k);
1487         do {
1488                 lu_context_key_degister(k);
1489                 k = va_arg(args, struct lu_context_key*);
1490         } while (k != NULL);
1491         va_end(args);
1492 }
1493 EXPORT_SYMBOL(lu_context_key_degister_many);
1494
1495 /**
1496  * Revive a number of keys.
1497  */
1498 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1499 {
1500         va_list args;
1501
1502         va_start(args, k);
1503         do {
1504                 lu_context_key_revive(k);
1505                 k = va_arg(args, struct lu_context_key*);
1506         } while (k != NULL);
1507         va_end(args);
1508 }
1509 EXPORT_SYMBOL(lu_context_key_revive_many);
1510
1511 /**
1512  * Quiescent a number of keys.
1513  */
1514 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1515 {
1516         va_list args;
1517
1518         va_start(args, k);
1519         do {
1520                 lu_context_key_quiesce(k);
1521                 k = va_arg(args, struct lu_context_key*);
1522         } while (k != NULL);
1523         va_end(args);
1524 }
1525 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1526
1527 /**
1528  * Return value associated with key \a key in context \a ctx.
1529  */
1530 void *lu_context_key_get(const struct lu_context *ctx,
1531                          const struct lu_context_key *key)
1532 {
1533         LINVRNT(ctx->lc_state == LCS_ENTERED);
1534         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1535         LASSERT(lu_keys[key->lct_index] == key);
1536         return ctx->lc_value[key->lct_index];
1537 }
1538 EXPORT_SYMBOL(lu_context_key_get);
1539
1540 /**
1541  * List of remembered contexts. XXX document me.
1542  */
1543 static LIST_HEAD(lu_context_remembered);
1544
1545 /**
1546  * Destroy \a key in all remembered contexts. This is used to destroy key
1547  * values in "shared" contexts (like service threads), when a module owning
1548  * the key is about to be unloaded.
1549  */
1550 void lu_context_key_quiesce(struct lu_context_key *key)
1551 {
1552         struct lu_context *ctx;
1553
1554         if (!(key->lct_tags & LCT_QUIESCENT)) {
1555                 /*
1556                  * The write-lock on lu_key_initing will ensure that any
1557                  * keys_fill() which didn't see LCT_QUIESCENT will have
1558                  * finished before we call key_fini().
1559                  */
1560                 down_write(&lu_key_initing);
1561                 key->lct_tags |= LCT_QUIESCENT;
1562                 up_write(&lu_key_initing);
1563
1564                 write_lock(&lu_keys_guard);
1565                 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1566                         spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1567                         key_fini(ctx, key->lct_index);
1568                 }
1569
1570                 write_unlock(&lu_keys_guard);
1571         }
1572 }
1573
1574 void lu_context_key_revive(struct lu_context_key *key)
1575 {
1576         key->lct_tags &= ~LCT_QUIESCENT;
1577         atomic_inc(&key_set_version);
1578 }
1579
1580 static void keys_fini(struct lu_context *ctx)
1581 {
1582         unsigned int i;
1583
1584         if (ctx->lc_value == NULL)
1585                 return;
1586
1587         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1588                 key_fini(ctx, i);
1589
1590         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1591         ctx->lc_value = NULL;
1592 }
1593
1594 static int keys_fill(struct lu_context *ctx)
1595 {
1596         unsigned int i;
1597         int rc = 0;
1598
1599         /*
1600          * A serialisation with lu_context_key_quiesce() is needed, to
1601          * ensure we see LCT_QUIESCENT and don't allocate a new value
1602          * after it freed one.  The rwsem provides this.  As down_read()
1603          * does optimistic spinning while the writer is active, this is
1604          * unlikely to ever sleep.
1605          */
1606         down_read(&lu_key_initing);
1607         ctx->lc_version = atomic_read(&key_set_version);
1608
1609         LINVRNT(ctx->lc_value);
1610         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1611                 struct lu_context_key *key;
1612
1613                 key = lu_keys[i];
1614                 if (!ctx->lc_value[i] && key &&
1615                     (key->lct_tags & ctx->lc_tags) &&
1616                     /*
1617                      * Don't create values for a LCT_QUIESCENT key, as this
1618                      * will pin module owning a key.
1619                      */
1620                     !(key->lct_tags & LCT_QUIESCENT)) {
1621                         void *value;
1622
1623                         LINVRNT(key->lct_init != NULL);
1624                         LINVRNT(key->lct_index == i);
1625
1626                         LASSERT(key->lct_owner != NULL);
1627                         if (!(ctx->lc_tags & LCT_NOREF) &&
1628                             try_module_get(key->lct_owner) == 0) {
1629                                 /* module is unloading, skip this key */
1630                                 continue;
1631                         }
1632
1633                         value = key->lct_init(ctx, key);
1634                         if (unlikely(IS_ERR(value))) {
1635                                 rc = PTR_ERR(value);
1636                                 break;
1637                         }
1638
1639                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1640                         atomic_inc(&key->lct_used);
1641                         /*
1642                          * This is the only place in the code, where an
1643                          * element of ctx->lc_value[] array is set to non-NULL
1644                          * value.
1645                          */
1646                         ctx->lc_value[i] = value;
1647                         if (key->lct_exit != NULL)
1648                                 ctx->lc_tags |= LCT_HAS_EXIT;
1649                 }
1650         }
1651
1652         up_read(&lu_key_initing);
1653         return rc;
1654 }
1655
1656 static int keys_init(struct lu_context *ctx)
1657 {
1658         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1659         if (likely(ctx->lc_value != NULL))
1660                 return keys_fill(ctx);
1661
1662         return -ENOMEM;
1663 }
1664
1665 /**
1666  * Initialize context data-structure. Create values for all keys.
1667  */
1668 int lu_context_init(struct lu_context *ctx, __u32 tags)
1669 {
1670         int     rc;
1671
1672         memset(ctx, 0, sizeof *ctx);
1673         ctx->lc_state = LCS_INITIALIZED;
1674         ctx->lc_tags = tags;
1675         if (tags & LCT_REMEMBER) {
1676                 write_lock(&lu_keys_guard);
1677                 list_add(&ctx->lc_remember, &lu_context_remembered);
1678                 write_unlock(&lu_keys_guard);
1679         } else {
1680                 INIT_LIST_HEAD(&ctx->lc_remember);
1681         }
1682
1683         rc = keys_init(ctx);
1684         if (rc != 0)
1685                 lu_context_fini(ctx);
1686
1687         return rc;
1688 }
1689 EXPORT_SYMBOL(lu_context_init);
1690
1691 /**
1692  * Finalize context data-structure. Destroy key values.
1693  */
1694 void lu_context_fini(struct lu_context *ctx)
1695 {
1696         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1697         ctx->lc_state = LCS_FINALIZED;
1698
1699         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1700                 LASSERT(list_empty(&ctx->lc_remember));
1701                 keys_fini(ctx);
1702
1703         } else { /* could race with key degister */
1704                 write_lock(&lu_keys_guard);
1705                 keys_fini(ctx);
1706                 list_del_init(&ctx->lc_remember);
1707                 write_unlock(&lu_keys_guard);
1708         }
1709 }
1710 EXPORT_SYMBOL(lu_context_fini);
1711
1712 /**
1713  * Called before entering context.
1714  */
1715 void lu_context_enter(struct lu_context *ctx)
1716 {
1717         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1718         ctx->lc_state = LCS_ENTERED;
1719 }
1720 EXPORT_SYMBOL(lu_context_enter);
1721
1722 /**
1723  * Called after exiting from \a ctx
1724  */
1725 void lu_context_exit(struct lu_context *ctx)
1726 {
1727         unsigned int i;
1728
1729         LINVRNT(ctx->lc_state == LCS_ENTERED);
1730         /*
1731          * Disable preempt to ensure we get a warning if
1732          * any lct_exit ever tries to sleep.  That would hurt
1733          * lu_context_key_quiesce() which spins waiting for us.
1734          * This also ensure we aren't preempted while the state
1735          * is LCS_LEAVING, as that too would cause problems for
1736          * lu_context_key_quiesce().
1737          */
1738         preempt_disable();
1739         /*
1740          * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1741          * or we see LCT_QUIESCENT
1742          */
1743         smp_store_mb(ctx->lc_state, LCS_LEAVING);
1744         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1745                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1746                         struct lu_context_key *key;
1747
1748                         key = lu_keys[i];
1749                         if (ctx->lc_value[i] &&
1750                             !(key->lct_tags & LCT_QUIESCENT) &&
1751                             key->lct_exit)
1752                                 key->lct_exit(ctx, key, ctx->lc_value[i]);
1753                 }
1754         }
1755
1756         smp_store_release(&ctx->lc_state, LCS_LEFT);
1757         preempt_enable();
1758 }
1759 EXPORT_SYMBOL(lu_context_exit);
1760
1761 /**
1762  * Allocate for context all missing keys that were registered after context
1763  * creation. key_set_version is only changed in rare cases when modules
1764  * are loaded and removed.
1765  */
1766 int lu_context_refill(struct lu_context *ctx)
1767 {
1768         if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1769                 return 0;
1770
1771         return keys_fill(ctx);
1772 }
1773
1774 /**
1775  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1776  * obd being added. Currently, this is only used on client side, specifically
1777  * for echo device client, for other stack (like ptlrpc threads), context are
1778  * predefined when the lu_device type are registered, during the module probe
1779  * phase.
1780  */
1781 u32 lu_context_tags_default;
1782 u32 lu_session_tags_default;
1783
1784 #ifdef HAVE_SERVER_SUPPORT
1785 void lu_context_tags_update(__u32 tags)
1786 {
1787         write_lock(&lu_keys_guard);
1788         lu_context_tags_default |= tags;
1789         atomic_inc(&key_set_version);
1790         write_unlock(&lu_keys_guard);
1791 }
1792 EXPORT_SYMBOL(lu_context_tags_update);
1793
1794 void lu_context_tags_clear(__u32 tags)
1795 {
1796         write_lock(&lu_keys_guard);
1797         lu_context_tags_default &= ~tags;
1798         atomic_inc(&key_set_version);
1799         write_unlock(&lu_keys_guard);
1800 }
1801 EXPORT_SYMBOL(lu_context_tags_clear);
1802
1803 void lu_session_tags_update(__u32 tags)
1804 {
1805         write_lock(&lu_keys_guard);
1806         lu_session_tags_default |= tags;
1807         atomic_inc(&key_set_version);
1808         write_unlock(&lu_keys_guard);
1809 }
1810 EXPORT_SYMBOL(lu_session_tags_update);
1811
1812 void lu_session_tags_clear(__u32 tags)
1813 {
1814         write_lock(&lu_keys_guard);
1815         lu_session_tags_default &= ~tags;
1816         atomic_inc(&key_set_version);
1817         write_unlock(&lu_keys_guard);
1818 }
1819 EXPORT_SYMBOL(lu_session_tags_clear);
1820 #endif /* HAVE_SERVER_SUPPORT */
1821
1822 int lu_env_init(struct lu_env *env, __u32 tags)
1823 {
1824         int result;
1825
1826         env->le_ses = NULL;
1827         result = lu_context_init(&env->le_ctx, tags);
1828         if (likely(result == 0))
1829                 lu_context_enter(&env->le_ctx);
1830         return result;
1831 }
1832 EXPORT_SYMBOL(lu_env_init);
1833
1834 void lu_env_fini(struct lu_env *env)
1835 {
1836         lu_context_exit(&env->le_ctx);
1837         lu_context_fini(&env->le_ctx);
1838         env->le_ses = NULL;
1839 }
1840 EXPORT_SYMBOL(lu_env_fini);
1841
1842 int lu_env_refill(struct lu_env *env)
1843 {
1844         int result;
1845
1846         result = lu_context_refill(&env->le_ctx);
1847         if (result == 0 && env->le_ses != NULL)
1848                 result = lu_context_refill(env->le_ses);
1849         return result;
1850 }
1851 EXPORT_SYMBOL(lu_env_refill);
1852
1853 /**
1854  * Currently, this API will only be used by echo client.
1855  * Because echo client and normal lustre client will share
1856  * same cl_env cache. So echo client needs to refresh
1857  * the env context after it get one from the cache, especially
1858  * when normal client and echo client co-exist in the same client.
1859  */
1860 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1861                           __u32 stags)
1862 {
1863         int    result;
1864
1865         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1866                 env->le_ctx.lc_version = 0;
1867                 env->le_ctx.lc_tags |= ctags;
1868         }
1869
1870         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1871                 env->le_ses->lc_version = 0;
1872                 env->le_ses->lc_tags |= stags;
1873         }
1874
1875         result = lu_env_refill(env);
1876
1877         return result;
1878 }
1879 EXPORT_SYMBOL(lu_env_refill_by_tags);
1880
1881
1882 struct lu_env_item {
1883         struct task_struct *lei_task;   /* rhashtable key */
1884         struct rhash_head lei_linkage;
1885         struct lu_env *lei_env;
1886 };
1887
1888 static const struct rhashtable_params lu_env_rhash_params = {
1889         .key_len     = sizeof(struct task_struct *),
1890         .key_offset  = offsetof(struct lu_env_item, lei_task),
1891         .head_offset = offsetof(struct lu_env_item, lei_linkage),
1892     };
1893
1894 struct rhashtable lu_env_rhash;
1895
1896 struct lu_env_percpu {
1897         struct task_struct *lep_task;
1898         struct lu_env *lep_env ____cacheline_aligned_in_smp;
1899 };
1900
1901 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
1902
1903 int lu_env_add(struct lu_env *env)
1904 {
1905         struct lu_env_item *lei, *old;
1906
1907         LASSERT(env);
1908
1909         OBD_ALLOC_PTR(lei);
1910         if (!lei)
1911                 return -ENOMEM;
1912
1913         lei->lei_task = current;
1914         lei->lei_env = env;
1915
1916         old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
1917                                                 &lei->lei_linkage,
1918                                                 lu_env_rhash_params);
1919         LASSERT(!old);
1920
1921         return 0;
1922 }
1923 EXPORT_SYMBOL(lu_env_add);
1924
1925 void lu_env_remove(struct lu_env *env)
1926 {
1927         struct lu_env_item *lei;
1928         const void *task = current;
1929         int i;
1930
1931         for_each_possible_cpu(i) {
1932                 if (lu_env_percpu[i].lep_env == env) {
1933                         LASSERT(lu_env_percpu[i].lep_task == task);
1934                         lu_env_percpu[i].lep_task = NULL;
1935                         lu_env_percpu[i].lep_env = NULL;
1936                 }
1937         }
1938
1939         rcu_read_lock();
1940         lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
1941                                      lu_env_rhash_params);
1942         if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
1943                                           lu_env_rhash_params) == 0)
1944                 OBD_FREE_PTR(lei);
1945         rcu_read_unlock();
1946 }
1947 EXPORT_SYMBOL(lu_env_remove);
1948
1949 struct lu_env *lu_env_find(void)
1950 {
1951         struct lu_env *env = NULL;
1952         struct lu_env_item *lei;
1953         const void *task = current;
1954         int i = get_cpu();
1955
1956         if (lu_env_percpu[i].lep_task == current) {
1957                 env = lu_env_percpu[i].lep_env;
1958                 put_cpu();
1959                 LASSERT(env);
1960                 return env;
1961         }
1962
1963         lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
1964                                      lu_env_rhash_params);
1965         if (lei) {
1966                 env = lei->lei_env;
1967                 lu_env_percpu[i].lep_task = current;
1968                 lu_env_percpu[i].lep_env = env;
1969         }
1970         put_cpu();
1971
1972         return env;
1973 }
1974 EXPORT_SYMBOL(lu_env_find);
1975
1976 static struct shrinker *lu_site_shrinker;
1977
1978 typedef struct lu_site_stats{
1979         unsigned        lss_populated;
1980         unsigned        lss_max_search;
1981         unsigned        lss_total;
1982         unsigned        lss_busy;
1983 } lu_site_stats_t;
1984
1985 static void lu_site_stats_get(const struct lu_site *s,
1986                               lu_site_stats_t *stats, int populated)
1987 {
1988         struct cfs_hash *hs = s->ls_obj_hash;
1989         struct cfs_hash_bd bd;
1990         unsigned int i;
1991         /*
1992          * percpu_counter_sum_positive() won't accept a const pointer
1993          * as it does modify the struct by taking a spinlock
1994          */
1995         struct lu_site *s2 = (struct lu_site *)s;
1996
1997         stats->lss_busy += cfs_hash_size_get(hs) -
1998                 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
1999         cfs_hash_for_each_bucket(hs, &bd, i) {
2000                 struct hlist_head *hhead;
2001
2002                 cfs_hash_bd_lock(hs, &bd, 1);
2003                 stats->lss_total += cfs_hash_bd_count_get(&bd);
2004                 stats->lss_max_search = max((int)stats->lss_max_search,
2005                                             cfs_hash_bd_depmax_get(&bd));
2006                 if (!populated) {
2007                         cfs_hash_bd_unlock(hs, &bd, 1);
2008                         continue;
2009                 }
2010
2011                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
2012                         if (!hlist_empty(hhead))
2013                                 stats->lss_populated++;
2014                 }
2015                 cfs_hash_bd_unlock(hs, &bd, 1);
2016         }
2017 }
2018
2019
2020 /*
2021  * lu_cache_shrink_count() returns an approximate number of cached objects
2022  * that can be freed by shrink_slab(). A counter, which tracks the
2023  * number of items in the site's lru, is maintained in a percpu_counter
2024  * for each site. The percpu values are incremented and decremented as
2025  * objects are added or removed from the lru. The percpu values are summed
2026  * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2027  * summed value at any given time may not accurately reflect the current
2028  * lru length. But this value is sufficiently accurate for the needs of
2029  * a shrinker.
2030  *
2031  * Using a per cpu counter is a compromise solution to concurrent access:
2032  * lu_object_put() can update the counter without locking the site and
2033  * lu_cache_shrink_count can sum the counters without locking each
2034  * ls_obj_hash bucket.
2035  */
2036 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2037                                            struct shrink_control *sc)
2038 {
2039         struct lu_site *s;
2040         struct lu_site *tmp;
2041         unsigned long cached = 0;
2042
2043         if (!(sc->gfp_mask & __GFP_FS))
2044                 return 0;
2045
2046         down_read(&lu_sites_guard);
2047         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2048                 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2049         up_read(&lu_sites_guard);
2050
2051         cached = (cached / 100) * sysctl_vfs_cache_pressure;
2052         CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2053                cached, sysctl_vfs_cache_pressure);
2054
2055         return cached;
2056 }
2057
2058 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2059                                           struct shrink_control *sc)
2060 {
2061         struct lu_site *s;
2062         struct lu_site *tmp;
2063         unsigned long remain = sc->nr_to_scan;
2064         LIST_HEAD(splice);
2065
2066         if (!(sc->gfp_mask & __GFP_FS))
2067                 /* We must not take the lu_sites_guard lock when
2068                  * __GFP_FS is *not* set because of the deadlock
2069                  * possibility detailed above. Additionally,
2070                  * since we cannot determine the number of
2071                  * objects in the cache without taking this
2072                  * lock, we're in a particularly tough spot. As
2073                  * a result, we'll just lie and say our cache is
2074                  * empty. This _should_ be ok, as we can't
2075                  * reclaim objects when __GFP_FS is *not* set
2076                  * anyways.
2077                  */
2078                 return SHRINK_STOP;
2079
2080         down_write(&lu_sites_guard);
2081         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2082                 remain = lu_site_purge(&lu_shrink_env, s, remain);
2083                 /*
2084                  * Move just shrunk site to the tail of site list to
2085                  * assure shrinking fairness.
2086                  */
2087                 list_move_tail(&s->ls_linkage, &splice);
2088         }
2089         list_splice(&splice, lu_sites.prev);
2090         up_write(&lu_sites_guard);
2091
2092         return sc->nr_to_scan - remain;
2093 }
2094
2095 #ifndef HAVE_SHRINKER_COUNT
2096 /*
2097  * There exists a potential lock inversion deadlock scenario when using
2098  * Lustre on top of ZFS. This occurs between one of ZFS's
2099  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2100  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2101  * while thread B will take the ht_lock and sleep on the lu_sites_guard
2102  * lock. Obviously neither thread will wake and drop their respective hold
2103  * on their lock.
2104  *
2105  * To prevent this from happening we must ensure the lu_sites_guard lock is
2106  * not taken while down this code path. ZFS reliably does not set the
2107  * __GFP_FS bit in its code paths, so this can be used to determine if it
2108  * is safe to take the lu_sites_guard lock.
2109  *
2110  * Ideally we should accurately return the remaining number of cached
2111  * objects without taking the lu_sites_guard lock, but this is not
2112  * possible in the current implementation.
2113  */
2114 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2115 {
2116         int cached = 0;
2117         struct shrink_control scv = {
2118                  .nr_to_scan = shrink_param(sc, nr_to_scan),
2119                  .gfp_mask   = shrink_param(sc, gfp_mask)
2120         };
2121 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2122         struct shrinker* shrinker = NULL;
2123 #endif
2124
2125
2126         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2127
2128         if (scv.nr_to_scan != 0)
2129                 lu_cache_shrink_scan(shrinker, &scv);
2130
2131         cached = lu_cache_shrink_count(shrinker, &scv);
2132         return cached;
2133 }
2134
2135 #endif /* HAVE_SHRINKER_COUNT */
2136
2137
2138 /*
2139  * Debugging stuff.
2140  */
2141
2142 /**
2143  * Environment to be used in debugger, contains all tags.
2144  */
2145 static struct lu_env lu_debugging_env;
2146
2147 /**
2148  * Debugging printer function using printk().
2149  */
2150 int lu_printk_printer(const struct lu_env *env,
2151                       void *unused, const char *format, ...)
2152 {
2153         va_list args;
2154
2155         va_start(args, format);
2156         vprintk(format, args);
2157         va_end(args);
2158         return 0;
2159 }
2160
2161 int lu_debugging_setup(void)
2162 {
2163         return lu_env_init(&lu_debugging_env, ~0);
2164 }
2165
2166 void lu_context_keys_dump(void)
2167 {
2168         unsigned int i;
2169
2170         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2171                 struct lu_context_key *key;
2172
2173                 key = lu_keys[i];
2174                 if (key != NULL) {
2175                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2176                                i, key, key->lct_tags,
2177                                key->lct_init, key->lct_fini, key->lct_exit,
2178                                key->lct_index, atomic_read(&key->lct_used),
2179                                key->lct_owner ? key->lct_owner->name : "",
2180                                key->lct_owner);
2181                         lu_ref_print(&key->lct_reference);
2182                 }
2183         }
2184 }
2185
2186 /**
2187  * Initialization of global lu_* data.
2188  */
2189 int lu_global_init(void)
2190 {
2191         int result;
2192         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2193                          lu_cache_shrink_count, lu_cache_shrink_scan);
2194
2195         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2196
2197         result = lu_ref_global_init();
2198         if (result != 0)
2199                 return result;
2200
2201         LU_CONTEXT_KEY_INIT(&lu_global_key);
2202         result = lu_context_key_register(&lu_global_key);
2203         if (result != 0)
2204                 return result;
2205
2206         /*
2207          * At this level, we don't know what tags are needed, so allocate them
2208          * conservatively. This should not be too bad, because this
2209          * environment is global.
2210          */
2211         down_write(&lu_sites_guard);
2212         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2213         up_write(&lu_sites_guard);
2214         if (result != 0)
2215                 return result;
2216
2217         /*
2218          * seeks estimation: 3 seeks to read a record from oi, one to read
2219          * inode, one for ea. Unfortunately setting this high value results in
2220          * lu_object/inode cache consuming all the memory.
2221          */
2222         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2223         if (lu_site_shrinker == NULL)
2224                 return -ENOMEM;
2225
2226         result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2227
2228         return result;
2229 }
2230
2231 /**
2232  * Dual to lu_global_init().
2233  */
2234 void lu_global_fini(void)
2235 {
2236         if (lu_site_shrinker != NULL) {
2237                 remove_shrinker(lu_site_shrinker);
2238                 lu_site_shrinker = NULL;
2239         }
2240
2241         lu_context_key_degister(&lu_global_key);
2242
2243         /*
2244          * Tear shrinker environment down _after_ de-registering
2245          * lu_global_key, because the latter has a value in the former.
2246          */
2247         down_write(&lu_sites_guard);
2248         lu_env_fini(&lu_shrink_env);
2249         up_write(&lu_sites_guard);
2250
2251         rhashtable_destroy(&lu_env_rhash);
2252
2253         lu_ref_global_fini();
2254 }
2255
2256 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2257 {
2258 #ifdef CONFIG_PROC_FS
2259         struct lprocfs_counter ret;
2260
2261         lprocfs_stats_collect(stats, idx, &ret);
2262         return (__u32)ret.lc_count;
2263 #else
2264         return 0;
2265 #endif
2266 }
2267
2268 /**
2269  * Output site statistical counters into a buffer. Suitable for
2270  * lprocfs_rd_*()-style functions.
2271  */
2272 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2273 {
2274         lu_site_stats_t stats;
2275
2276         memset(&stats, 0, sizeof(stats));
2277         lu_site_stats_get(s, &stats, 1);
2278
2279         seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2280                    stats.lss_busy,
2281                    stats.lss_total,
2282                    stats.lss_populated,
2283                    CFS_HASH_NHLIST(s->ls_obj_hash),
2284                    stats.lss_max_search,
2285                    ls_stats_read(s->ls_stats, LU_SS_CREATED),
2286                    ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2287                    ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2288                    ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2289                    ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2290                    ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2291         return 0;
2292 }
2293 EXPORT_SYMBOL(lu_site_stats_seq_print);
2294
2295 /**
2296  * Helper function to initialize a number of kmem slab caches at once.
2297  */
2298 int lu_kmem_init(struct lu_kmem_descr *caches)
2299 {
2300         int result;
2301         struct lu_kmem_descr *iter = caches;
2302
2303         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2304                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2305                                                      iter->ckd_size,
2306                                                      0, 0, NULL);
2307                 if (*iter->ckd_cache == NULL) {
2308                         result = -ENOMEM;
2309                         /* free all previously allocated caches */
2310                         lu_kmem_fini(caches);
2311                         break;
2312                 }
2313         }
2314         return result;
2315 }
2316 EXPORT_SYMBOL(lu_kmem_init);
2317
2318 /**
2319  * Helper function to finalize a number of kmem slab cached at once. Dual to
2320  * lu_kmem_init().
2321  */
2322 void lu_kmem_fini(struct lu_kmem_descr *caches)
2323 {
2324         for (; caches->ckd_cache != NULL; ++caches) {
2325                 if (*caches->ckd_cache != NULL) {
2326                         kmem_cache_destroy(*caches->ckd_cache);
2327                         *caches->ckd_cache = NULL;
2328                 }
2329         }
2330 }
2331 EXPORT_SYMBOL(lu_kmem_fini);
2332
2333 /**
2334  * Temporary solution to be able to assign fid in ->do_create()
2335  * till we have fully-functional OST fids
2336  */
2337 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2338                           const struct lu_fid *fid)
2339 {
2340         struct lu_site          *s = o->lo_dev->ld_site;
2341         struct lu_fid           *old = &o->lo_header->loh_fid;
2342         struct cfs_hash         *hs;
2343         struct cfs_hash_bd       bd;
2344
2345         LASSERT(fid_is_zero(old));
2346
2347         /* supposed to be unique */
2348         hs = s->ls_obj_hash;
2349         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2350 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2351         {
2352                 __u64 version = 0;
2353                 struct lu_object *shadow;
2354
2355                 shadow = htable_lookup(s, &bd, fid, &version);
2356                 /* supposed to be unique */
2357                 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2358         }
2359 #endif
2360         *old = *fid;
2361         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2362         cfs_hash_bd_unlock(hs, &bd, 1);
2363 }
2364 EXPORT_SYMBOL(lu_object_assign_fid);
2365
2366 /**
2367  * allocates object with 0 (non-assiged) fid
2368  * XXX: temporary solution to be able to assign fid in ->do_create()
2369  *      till we have fully-functional OST fids
2370  */
2371 struct lu_object *lu_object_anon(const struct lu_env *env,
2372                                  struct lu_device *dev,
2373                                  const struct lu_object_conf *conf)
2374 {
2375         struct lu_fid     fid;
2376         struct lu_object *o;
2377
2378         fid_zero(&fid);
2379         o = lu_object_alloc(env, dev, &fid, conf);
2380
2381         return o;
2382 }
2383 EXPORT_SYMBOL(lu_object_anon);
2384
2385 struct lu_buf LU_BUF_NULL = {
2386         .lb_buf = NULL,
2387         .lb_len = 0
2388 };
2389 EXPORT_SYMBOL(LU_BUF_NULL);
2390
2391 void lu_buf_free(struct lu_buf *buf)
2392 {
2393         LASSERT(buf);
2394         if (buf->lb_buf) {
2395                 LASSERT(buf->lb_len > 0);
2396                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2397                 buf->lb_buf = NULL;
2398                 buf->lb_len = 0;
2399         }
2400 }
2401 EXPORT_SYMBOL(lu_buf_free);
2402
2403 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2404 {
2405         LASSERT(buf);
2406         LASSERT(buf->lb_buf == NULL);
2407         LASSERT(buf->lb_len == 0);
2408         OBD_ALLOC_LARGE(buf->lb_buf, size);
2409         if (likely(buf->lb_buf))
2410                 buf->lb_len = size;
2411 }
2412 EXPORT_SYMBOL(lu_buf_alloc);
2413
2414 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2415 {
2416         lu_buf_free(buf);
2417         lu_buf_alloc(buf, size);
2418 }
2419 EXPORT_SYMBOL(lu_buf_realloc);
2420
2421 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2422 {
2423         if (buf->lb_buf == NULL && buf->lb_len == 0)
2424                 lu_buf_alloc(buf, len);
2425
2426         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2427                 lu_buf_realloc(buf, len);
2428
2429         return buf;
2430 }
2431 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2432
2433 /**
2434  * Increase the size of the \a buf.
2435  * preserves old data in buffer
2436  * old buffer remains unchanged on error
2437  * \retval 0 or -ENOMEM
2438  */
2439 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2440 {
2441         char *ptr;
2442
2443         if (len <= buf->lb_len)
2444                 return 0;
2445
2446         OBD_ALLOC_LARGE(ptr, len);
2447         if (ptr == NULL)
2448                 return -ENOMEM;
2449
2450         /* Free the old buf */
2451         if (buf->lb_buf != NULL) {
2452                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2453                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2454         }
2455
2456         buf->lb_buf = ptr;
2457         buf->lb_len = len;
2458         return 0;
2459 }
2460 EXPORT_SYMBOL(lu_buf_check_and_grow);