Whamcloud - gitweb
d10626d81d932f939a65be6638f936fd41e8d807
[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_handler.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, 2016, 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/osd/osd_handler.c
33  *
34  * Top-level entry points into osd module
35  *
36  * Author: Nikita Danilov <nikita@clusterfs.com>
37  *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
38  */
39
40 #define DEBUG_SUBSYSTEM S_OSD
41
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #ifdef HAVE_UIDGID_HEADER
45 # include <linux/uidgid.h>
46 #endif
47
48 /* LUSTRE_VERSION_CODE */
49 #include <lustre_ver.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/types.h>
52 /* prerequisite for linux/xattr.h */
53 #include <linux/fs.h>
54 /* XATTR_{REPLACE,CREATE} */
55 #include <linux/xattr.h>
56
57 #include <ldiskfs/ldiskfs.h>
58 #include <ldiskfs/xattr.h>
59 #include <ldiskfs/ldiskfs_extents.h>
60 #undef ENTRY
61 /*
62  * struct OBD_{ALLOC,FREE}*()
63  * OBD_FAIL_CHECK
64  */
65 #include <obd_support.h>
66 /* struct ptlrpc_thread */
67 #include <lustre_net.h>
68 #include <lustre_fid.h>
69 /* process_config */
70 #include <lustre_param.h>
71
72 #include "osd_internal.h"
73 #include "osd_dynlocks.h"
74
75 /* llo_* api support */
76 #include <md_object.h>
77 #include <lustre_quota.h>
78
79 #include <lustre_linkea.h>
80
81 int ldiskfs_pdo = 1;
82 module_param(ldiskfs_pdo, int, 0644);
83 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
84
85 int ldiskfs_track_declares_assert;
86 module_param(ldiskfs_track_declares_assert, int, 0644);
87 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
88
89 /* Slab to allocate dynlocks */
90 struct kmem_cache *dynlock_cachep;
91
92 /* Slab to allocate osd_it_ea */
93 struct kmem_cache *osd_itea_cachep;
94
95 static struct lu_kmem_descr ldiskfs_caches[] = {
96         {
97                 .ckd_cache = &dynlock_cachep,
98                 .ckd_name  = "dynlock_cache",
99                 .ckd_size  = sizeof(struct dynlock_handle)
100         },
101         {
102                 .ckd_cache = &osd_itea_cachep,
103                 .ckd_name  = "osd_itea_cache",
104                 .ckd_size  = sizeof(struct osd_it_ea)
105         },
106         {
107                 .ckd_cache = NULL
108         }
109 };
110
111 static const char dot[] = ".";
112 static const char dotdot[] = "..";
113 static const char remote_obj_dir[] = "REM_OBJ_DIR";
114
115 static const struct lu_object_operations      osd_lu_obj_ops;
116 static const struct dt_object_operations      osd_obj_ops;
117 static const struct dt_object_operations      osd_obj_ea_ops;
118 static const struct dt_object_operations      osd_obj_otable_it_ops;
119 static const struct dt_index_operations       osd_index_iam_ops;
120 static const struct dt_index_operations       osd_index_ea_ops;
121
122 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
123                           const struct lu_fid *fid);
124 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
125                                                 struct osd_device *osd);
126
127 int osd_trans_declare_op2rb[] = {
128         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
129         [OSD_OT_PUNCH]          = OSD_OT_MAX,
130         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
131         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
132         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
133         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
134         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
135         [OSD_OT_WRITE]          = OSD_OT_WRITE,
136         [OSD_OT_INSERT]         = OSD_OT_DELETE,
137         [OSD_OT_DELETE]         = OSD_OT_INSERT,
138         [OSD_OT_QUOTA]          = OSD_OT_MAX,
139 };
140
141 static int osd_has_index(const struct osd_object *obj)
142 {
143         return obj->oo_dt.do_index_ops != NULL;
144 }
145
146 static int osd_object_invariant(const struct lu_object *l)
147 {
148         return osd_invariant(osd_obj(l));
149 }
150
151 /*
152  * Concurrency: doesn't matter
153  */
154
155 /*
156  * Concurrency: doesn't matter
157  */
158 static int osd_write_locked(const struct lu_env *env, struct osd_object *o)
159 {
160         struct osd_thread_info *oti = osd_oti_get(env);
161         return oti->oti_w_locks > 0 && o->oo_owner == env;
162 }
163
164 /*
165  * Concurrency: doesn't access mutable data
166  */
167 static int osd_root_get(const struct lu_env *env,
168                         struct dt_device *dev, struct lu_fid *f)
169 {
170         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
171         return 0;
172 }
173
174 /*
175  * the following set of functions are used to maintain per-thread
176  * cache of FID->ino mapping. this mechanism is needed to resolve
177  * FID to inode at dt_insert() which in turn stores ino in the
178  * directory entries to keep ldiskfs compatible with ext[34].
179  * due to locking-originated restrictions we can't lookup ino
180  * using LU cache (deadlock is possible). lookup using OI is quite
181  * expensive. so instead we maintain this cache and methods like
182  * dt_create() fill it. so in the majority of cases dt_insert() is
183  * able to find needed mapping in lockless manner.
184  */
185 static struct osd_idmap_cache *
186 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
187              const struct lu_fid *fid)
188 {
189         struct osd_thread_info  *oti   = osd_oti_get(env);
190         struct osd_idmap_cache  *idc    = oti->oti_ins_cache;
191         int i;
192         for (i = 0; i < oti->oti_ins_cache_used; i++) {
193                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
194                         continue;
195                 if (idc[i].oic_dev != osd)
196                         continue;
197
198                 return idc + i;
199         }
200
201         return NULL;
202 }
203
204 static struct osd_idmap_cache *
205 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
206             const struct lu_fid *fid)
207 {
208         struct osd_thread_info  *oti   = osd_oti_get(env);
209         struct osd_idmap_cache  *idc;
210         int i;
211
212         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
213                 i = oti->oti_ins_cache_size * 2;
214                 if (i == 0)
215                         i = OSD_INS_CACHE_SIZE;
216                 OBD_ALLOC(idc, sizeof(*idc) * i);
217                 if (idc == NULL)
218                         return ERR_PTR(-ENOMEM);
219                 if (oti->oti_ins_cache != NULL) {
220                         memcpy(idc, oti->oti_ins_cache,
221                                oti->oti_ins_cache_used * sizeof(*idc));
222                         OBD_FREE(oti->oti_ins_cache,
223                                  oti->oti_ins_cache_used * sizeof(*idc));
224                 }
225                 oti->oti_ins_cache = idc;
226                 oti->oti_ins_cache_size = i;
227         }
228
229         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
230         idc->oic_fid = *fid;
231         idc->oic_dev = osd;
232         idc->oic_lid.oii_ino = 0;
233         idc->oic_lid.oii_gen = 0;
234         idc->oic_remote = 0;
235
236         return idc;
237 }
238
239 /*
240  * lookup mapping for the given fid in the cache, initialize a
241  * new one if not found. the initialization checks whether the
242  * object is local or remote. for local objects, OI is used to
243  * learn ino/generation. the function is used when the caller
244  * has no information about the object, e.g. at dt_insert().
245  */
246 static struct osd_idmap_cache *
247 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
248                      const struct lu_fid *fid)
249 {
250         struct osd_idmap_cache *idc;
251         int rc;
252
253         idc = osd_idc_find(env, osd, fid);
254         LASSERT(!IS_ERR(idc));
255         if (idc != NULL)
256                 return idc;
257
258         /* new mapping is needed */
259         idc = osd_idc_add(env, osd, fid);
260         if (IS_ERR(idc))
261                 return idc;
262
263         /* initialize it */
264         rc = osd_remote_fid(env, osd, fid);
265         if (unlikely(rc < 0))
266                 return ERR_PTR(rc);
267
268         if (rc == 0) {
269                 /* the object is local, lookup in OI */
270                 /* XXX: probably cheaper to lookup in LU first? */
271                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
272                                    &idc->oic_lid, 0);
273                 if (unlikely(rc < 0)) {
274                         CERROR("can't lookup: rc = %d\n", rc);
275                         return ERR_PTR(rc);
276                 }
277         } else {
278                 /* the object is remote */
279                 idc->oic_remote = 1;
280         }
281
282         return idc;
283 }
284
285 /*
286  * lookup mapping for given FID and fill it from the given object.
287  * the object is lolcal by definition.
288  */
289 static int osd_idc_find_and_init(const struct lu_env *env,
290                                  struct osd_device *osd,
291                                  struct osd_object *obj)
292 {
293         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
294         struct osd_idmap_cache  *idc;
295
296         idc = osd_idc_find(env, osd, fid);
297         LASSERT(!IS_ERR(idc));
298         if (idc != NULL) {
299                 if (obj->oo_inode == NULL)
300                         return 0;
301                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
302                         LASSERT(idc->oic_lid.oii_ino == 0);
303                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
304                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
305                 }
306                 return 0;
307         }
308
309         /* new mapping is needed */
310         idc = osd_idc_add(env, osd, fid);
311         if (IS_ERR(idc))
312                 return PTR_ERR(idc);
313
314         if (obj->oo_inode != NULL) {
315                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
316                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
317         }
318         return 0;
319 }
320
321 /*
322  * OSD object methods.
323  */
324
325 /*
326  * Concurrency: no concurrent access is possible that early in object
327  * life-cycle.
328  */
329 static struct lu_object *osd_object_alloc(const struct lu_env *env,
330                                           const struct lu_object_header *hdr,
331                                           struct lu_device *d)
332 {
333         struct osd_object *mo;
334
335         OBD_ALLOC_PTR(mo);
336         if (mo != NULL) {
337                 struct lu_object *l;
338
339                 l = &mo->oo_dt.do_lu;
340                 dt_object_init(&mo->oo_dt, NULL, d);
341                 mo->oo_dt.do_ops = &osd_obj_ea_ops;
342                 l->lo_ops = &osd_lu_obj_ops;
343                 init_rwsem(&mo->oo_sem);
344                 init_rwsem(&mo->oo_ext_idx_sem);
345                 spin_lock_init(&mo->oo_guard);
346                 INIT_LIST_HEAD(&mo->oo_xattr_list);
347                 return l;
348         } else {
349                 return NULL;
350         }
351 }
352
353 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
354                 struct dentry *dentry, struct lustre_mdt_attrs *lma)
355 {
356         int rc;
357
358         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
359         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
360                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
361         if (rc > 0) {
362                 if ((void *)lma != (void *)info->oti_mdt_attrs_old)
363                         memcpy(lma, info->oti_mdt_attrs_old, sizeof(*lma));
364                 rc = 0;
365                 lustre_lma_swab(lma);
366                 /* Check LMA compatibility */
367                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
368                         CWARN("%.16s: unsupported incompat LMA feature(s) %#x "
369                               "for fid = "DFID", ino = %lu\n",
370                               LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
371                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
372                               PFID(&lma->lma_self_fid), inode->i_ino);
373                         rc = -EOPNOTSUPP;
374                 }
375         } else if (rc == 0) {
376                 rc = -ENODATA;
377         }
378
379         return rc;
380 }
381
382 /*
383  * retrieve object from backend ext fs.
384  **/
385 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
386                        struct osd_inode_id *id)
387 {
388         int rc;
389         struct inode *inode = NULL;
390
391         /* if we look for an inode withing a running
392          * transaction, then we risk to deadlock */
393         /* osd_dirent_check_repair() breaks this */
394         /*LASSERT(current->journal_info == NULL);*/
395
396         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
397         if (IS_ERR(inode)) {
398                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
399                        id->oii_ino, PTR_ERR(inode));
400         } else if (id->oii_gen != OSD_OII_NOGEN &&
401                    inode->i_generation != id->oii_gen) {
402                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
403                        "i_generation = %u\n",
404                        id->oii_ino, id->oii_gen, inode->i_generation);
405                 iput(inode);
406                 inode = ERR_PTR(-ESTALE);
407         } else if (inode->i_nlink == 0) {
408                 /* due to parallel readdir and unlink,
409                 * we can have dead inode here. */
410                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
411                 iput(inode);
412                 inode = ERR_PTR(-ESTALE);
413         } else if (is_bad_inode(inode)) {
414                 CWARN("%.16s: bad inode: ino = %u\n",
415                 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, id->oii_ino);
416                 iput(inode);
417                 inode = ERR_PTR(-ENOENT);
418         } else if ((rc = osd_attach_jinode(inode))) {
419                 iput(inode);
420                 inode = ERR_PTR(rc);
421         } else {
422                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
423                 if (id->oii_gen == OSD_OII_NOGEN)
424                         osd_id_gen(id, inode->i_ino, inode->i_generation);
425
426                 /* Do not update file c/mtime in ldiskfs.
427                  * NB: we don't have any lock to protect this because we don't
428                  * have reference on osd_object now, but contention with
429                  * another lookup + attr_set can't happen in the tiny window
430                  * between if (...) and set S_NOCMTIME. */
431                 if (!(inode->i_flags & S_NOCMTIME))
432                         inode->i_flags |= S_NOCMTIME;
433         }
434         return inode;
435 }
436
437 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
438                           handle_t *handle, struct dentry *child,
439                           struct inode *inode, struct htree_lock *hlock)
440 {
441         int rc, rc2;
442
443         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
444         if (rc == -ENOBUFS || rc == -ENOSPC) {
445                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
446                 struct inode *parent = child->d_parent->d_inode;
447                 struct lu_fid *fid = NULL;
448
449                 rc2 = osd_get_lma(info, parent, child->d_parent, lma);
450                 if (rc2 == 0) {
451                         fid = &lma->lma_self_fid;
452                 } else if (rc2 == -ENODATA) {
453                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
454                                 fid = &info->oti_fid3;
455                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
456                         } else if (!osd->od_is_ost && osd->od_index == 0) {
457                                 fid = &info->oti_fid3;
458                                 lu_igif_build(fid, parent->i_ino,
459                                               parent->i_generation);
460                         }
461                 }
462
463                 if (fid != NULL)
464                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
465                               "maximum entry limit\n",
466                               osd_name(osd), parent->i_ino, PFID(fid),
467                               rc == -ENOSPC ? "has reached" : "is approaching");
468                 else
469                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
470                               "maximum entry limit\n",
471                               osd_name(osd), parent->i_ino,
472                               rc == -ENOSPC ? "has reached" : "is approaching");
473
474                 /* ignore such error now */
475                 if (rc == -ENOBUFS)
476                         rc = 0;
477         }
478
479         return rc;
480 }
481
482
483 static struct inode *
484 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
485              struct osd_inode_id *id, struct lu_fid *fid)
486 {
487         struct lustre_mdt_attrs *lma   = &info->oti_mdt_attrs;
488         struct inode            *inode;
489         int                      rc;
490
491         inode = osd_iget(info, dev, id);
492         if (IS_ERR(inode))
493                 return inode;
494
495         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
496         if (rc == 0) {
497                 *fid = lma->lma_self_fid;
498         } else if (rc == -ENODATA) {
499                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
500                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
501                 else
502                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
503         } else {
504                 iput(inode);
505                 inode = ERR_PTR(rc);
506         }
507         return inode;
508 }
509
510 static struct inode *osd_iget_check(struct osd_thread_info *info,
511                                     struct osd_device *dev,
512                                     const struct lu_fid *fid,
513                                     struct osd_inode_id *id,
514                                     bool cached)
515 {
516         struct inode    *inode;
517         int              rc     = 0;
518         ENTRY;
519
520         /* The cached OI mapping is trustable. If we cannot locate the inode
521          * via the cached OI mapping, then return the failure to the caller
522          * directly without further OI checking. */
523
524         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
525         if (IS_ERR(inode)) {
526                 rc = PTR_ERR(inode);
527                 if (cached || (rc != -ENOENT && rc != -ESTALE)) {
528                         CDEBUG(D_INODE, "no inode: ino = %u, rc = %d\n",
529                                id->oii_ino, rc);
530
531                         GOTO(put, rc);
532                 }
533
534                 goto check_oi;
535         }
536
537         if (is_bad_inode(inode)) {
538                 rc = -ENOENT;
539                 if (cached) {
540                         CDEBUG(D_INODE, "bad inode: ino = %u\n", id->oii_ino);
541
542                         GOTO(put, rc);
543                 }
544
545                 goto check_oi;
546         }
547
548         if (id->oii_gen != OSD_OII_NOGEN &&
549             inode->i_generation != id->oii_gen) {
550                 rc = -ESTALE;
551                 if (cached) {
552                         CDEBUG(D_INODE, "unmatched inode: ino = %u, "
553                                "oii_gen = %u, i_generation = %u\n",
554                                id->oii_ino, id->oii_gen, inode->i_generation);
555
556                         GOTO(put, rc);
557                 }
558
559                 goto check_oi;
560         }
561
562         if (inode->i_nlink == 0) {
563                 rc = -ENOENT;
564                 if (cached) {
565                         CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
566
567                         GOTO(put, rc);
568                 }
569
570                 goto check_oi;
571         }
572
573         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
574
575 check_oi:
576         if (rc != 0) {
577                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
578
579                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
580                 /* XXX: There are four possible cases:
581                  *      1. rc = 0.
582                  *         Backup/restore caused the OI invalid.
583                  *      2. rc = 0.
584                  *         Someone unlinked the object but NOT removed
585                  *         the OI mapping, such as mount target device
586                  *         as ldiskfs, and modify something directly.
587                  *      3. rc = -ENOENT.
588                  *         Someone just removed the object between the
589                  *         former oi_lookup and the iget. It is normal.
590                  *      4. Other failure cases.
591                  *
592                  *      Generally, when the device is mounted, it will
593                  *      auto check whether the system is restored from
594                  *      file-level backup or not. We trust such detect
595                  *      to distinguish the 1st case from the 2nd case:
596                  *      if the OI files are consistent but may contain
597                  *      stale OI mappings because of case 2, if iget()
598                  *      returns -ENOENT or -ESTALE, then it should be
599                  *      the case 2. */
600                 if (rc != 0)
601                         /* If the OI mapping was in OI file before the
602                          * osd_iget_check(), but now, it is disappear,
603                          * then it must be removed by race. That is a
604                          * normal race case. */
605                         GOTO(put, rc);
606
607                 if ((!IS_ERR(inode) && inode->i_generation != 0 &&
608                      inode->i_generation == id->oii_gen) ||
609                     (IS_ERR(inode) && !(dev->od_scrub.os_file.sf_flags &
610                                         SF_INCONSISTENT)))
611                         rc = -ENOENT;
612                 else
613                         rc = -EREMCHG;
614         } else {
615                 if (id->oii_gen == OSD_OII_NOGEN)
616                         osd_id_gen(id, inode->i_ino, inode->i_generation);
617
618                 /* Do not update file c/mtime in ldiskfs.
619                  * NB: we don't have any lock to protect this because we don't
620                  * have reference on osd_object now, but contention with
621                  * another lookup + attr_set can't happen in the tiny window
622                  * between if (...) and set S_NOCMTIME. */
623                 if (!(inode->i_flags & S_NOCMTIME))
624                         inode->i_flags |= S_NOCMTIME;
625         }
626
627         GOTO(put, rc);
628
629 put:
630         if (rc != 0) {
631                 if (!IS_ERR(inode))
632                         iput(inode);
633
634                 inode = ERR_PTR(rc);
635         }
636
637         return inode;
638 }
639
640 /**
641  * \retval +v: new filter_fid, does not contain self-fid
642  * \retval 0:  filter_fid_old, contains self-fid
643  * \retval -v: other failure cases
644  */
645 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
646                  struct dentry *dentry, struct lu_fid *fid)
647 {
648         struct filter_fid_old   *ff     = &info->oti_ff;
649         struct ost_id           *ostid  = &info->oti_ostid;
650         int                      rc;
651
652         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
653         if (rc == sizeof(*ff)) {
654                 rc = 0;
655                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
656                 ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
657                 /* XXX: use 0 as the index for compatibility, the caller will
658                  *      handle index related issues when necessarry. */
659                 ostid_to_fid(fid, ostid, 0);
660         } else if (rc == sizeof(struct filter_fid)) {
661                 rc = 1;
662         } else if (rc >= 0) {
663                 rc = -EINVAL;
664         }
665
666         return rc;
667 }
668
669 static int osd_lma_self_repair(struct osd_thread_info *info,
670                                struct osd_device *osd, struct inode *inode,
671                                const struct lu_fid *fid, __u32 compat)
672 {
673         handle_t *jh;
674         int       rc;
675
676         LASSERT(current->journal_info == NULL);
677
678         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
679                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
680         if (IS_ERR(jh)) {
681                 rc = PTR_ERR(jh);
682                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
683                       osd_name(osd), rc);
684                 return rc;
685         }
686
687         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
688         if (rc != 0)
689                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
690                       osd_name(osd), rc);
691         ldiskfs_journal_stop(jh);
692         return rc;
693 }
694
695 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
696 {
697         struct osd_thread_info  *info   = osd_oti_get(env);
698         struct osd_device       *osd    = osd_obj2dev(obj);
699         struct lustre_mdt_attrs *lma    = &info->oti_mdt_attrs;
700         struct inode            *inode  = obj->oo_inode;
701         struct dentry           *dentry = &info->oti_obj_dentry;
702         struct lu_fid           *fid    = NULL;
703         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
704         int                      rc;
705         ENTRY;
706
707         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
708         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
709                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
710         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
711                 fid = &lma->lma_self_fid;
712                 rc = osd_get_idif(info, inode, dentry, fid);
713                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
714                         /* For the given OST-object, if it has neither LMA nor
715                          * FID in XATTR_NAME_FID, then the given FID (which is
716                          * contained in the @obj, from client RPC for locating
717                          * the OST-object) is trusted. We use it to generate
718                          * the LMA. */
719                         osd_lma_self_repair(info, osd, inode, rfid,
720                                             LMAC_FID_ON_OST);
721                         RETURN(0);
722                 }
723         }
724
725         if (rc < 0)
726                 RETURN(rc);
727
728         if (rc > 0) {
729                 rc = 0;
730                 lustre_lma_swab(lma);
731                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
732                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
733                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
734                               "fid = "DFID", ino = %lu\n", osd_name(osd),
735                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
736                               PFID(rfid), inode->i_ino);
737                         rc = -EOPNOTSUPP;
738                 } else {
739                         fid = &lma->lma_self_fid;
740                 }
741         }
742
743         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
744                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
745                         struct ost_id   *oi   = &info->oti_ostid;
746                         struct lu_fid   *fid1 = &info->oti_fid3;
747                         __u32            idx  = fid_idif_ost_idx(rfid);
748
749                         /* For old IDIF, the OST index is not part of the IDIF,
750                          * Means that different OSTs may have the same IDIFs.
751                          * Under such case, we need to make some compatible
752                          * check to make sure to trigger OI scrub properly. */
753                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
754                                 /* Given @rfid is new, LMA is old. */
755                                 fid_to_ostid(fid, oi);
756                                 ostid_to_fid(fid1, oi, idx);
757                                 if (lu_fid_eq(fid1, rfid)) {
758                                         if (osd->od_index_in_idif)
759                                                 osd_lma_self_repair(info, osd,
760                                                         inode, rfid,
761                                                         LMAC_FID_ON_OST);
762                                         RETURN(0);
763                                 }
764                         }
765                 }
766
767                 rc = -EREMCHG;
768         }
769
770         RETURN(rc);
771 }
772
773 struct osd_check_lmv_buf {
774 #ifdef HAVE_DIR_CONTEXT
775         /* please keep it as first member */
776         struct dir_context       ctx;
777 #endif
778         struct osd_thread_info  *oclb_info;
779         struct osd_device       *oclb_dev;
780         struct osd_idmap_cache  *oclb_oic;
781 };
782
783 /**
784  * It is called internally by ->readdir() to filter out the
785  * local slave object's FID of the striped directory.
786  *
787  * \retval      1 found the local slave's FID
788  * \retval      0 continue to check next item
789  * \retval      -ve for failure
790  */
791 #ifdef HAVE_FILLDIR_USE_CTX
792 static int osd_stripe_dir_filldir(struct dir_context *buf,
793 #else
794 static int osd_stripe_dir_filldir(void *buf,
795 #endif
796                                   const char *name, int namelen,
797                                   loff_t offset, __u64 ino, unsigned d_type)
798 {
799         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
800         struct osd_thread_info *oti = oclb->oclb_info;
801         struct lu_fid *fid = &oti->oti_fid3;
802         struct osd_inode_id *id = &oti->oti_id3;
803         struct osd_device *dev = oclb->oclb_dev;
804         struct osd_idmap_cache *oic = oclb->oclb_oic;
805         struct inode *inode;
806         int rc;
807
808         if (name[0] == '.')
809                 return 0;
810
811         fid_zero(fid);
812         sscanf(name + 1, SFID, RFID(fid));
813         if (!fid_is_sane(fid))
814                 return 0;
815
816         if (osd_remote_fid(oti->oti_env, dev, fid))
817                 return 0;
818
819         osd_id_gen(id, ino, OSD_OII_NOGEN);
820         inode = osd_iget(oti, dev, id);
821         if (IS_ERR(inode))
822                 return PTR_ERR(inode);
823
824         iput(inode);
825         osd_add_oi_cache(oti, dev, id, fid);
826         oic->oic_fid = *fid;
827         oic->oic_lid = *id;
828         oic->oic_dev = dev;
829         rc = osd_oii_insert(dev, oic, true);
830
831         return rc == 0 ? 1 : rc;
832 }
833
834 /* When lookup item under striped directory, we need to locate the master
835  * MDT-object of the striped directory firstly, then the client will send
836  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
837  * and the item's name. If the system is restored from MDT file level backup,
838  * then before the OI scrub completely built the OI files, the OI mappings of
839  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
840  * not a problem for the master MDT-object. Because when locate the master
841  * MDT-object, we will do name based lookup (for the striped directory itself)
842  * firstly, during such process we can setup the correct OI mapping for the
843  * master MDT-object. But it will be trouble for the slave MDT-object. Because
844  * the client will not trigger name based lookup on the MDT to locate the slave
845  * MDT-object before locating item under the striped directory, then when
846  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
847  * is invalid and does not know what the right OI mapping is, then the MDT has
848  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
849  * the OI file, related OI mapping is unknown yet, please try again later. And
850  * then client will re-try the RPC again and again until related OI mapping has
851  * been updated. That is quite inefficient.
852  *
853  * To resolve above trouble, we will handle it as the following two cases:
854  *
855  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
856  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
857  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
858  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
859  *    directly. Please check osd_fid_lookup().
860  *
861  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
862  *    Under such case, during lookup the master MDT-object, we will lookup the
863  *    slave MDT-object via readdir against the master MDT-object, because the
864  *    slave MDT-objects information are stored as sub-directories with the name
865  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
866  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
867  *    the correct OI mapping for the slave MDT-object. */
868 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
869                          struct inode *inode, struct osd_idmap_cache *oic)
870 {
871         struct lu_buf *buf = &oti->oti_big_buf;
872         struct dentry *dentry = &oti->oti_obj_dentry;
873         struct file *filp = &oti->oti_file;
874         const struct file_operations *fops;
875         struct lmv_mds_md_v1 *lmv1;
876         struct osd_check_lmv_buf oclb = {
877 #ifdef HAVE_DIR_CONTEXT
878                 .ctx.actor = osd_stripe_dir_filldir,
879 #endif
880                 .oclb_info = oti,
881                 .oclb_dev = dev,
882                 .oclb_oic = oic
883         };
884         int rc = 0;
885         ENTRY;
886
887 again:
888         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
889                              buf->lb_len);
890         if (rc == -ERANGE) {
891                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
892                 if (rc > 0) {
893                         lu_buf_realloc(buf, rc);
894                         if (buf->lb_buf == NULL)
895                                 GOTO(out, rc = -ENOMEM);
896
897                         goto again;
898                 }
899         }
900
901         if (unlikely(rc == 0 || rc == -ENODATA))
902                 GOTO(out, rc = 0);
903
904         if (rc < 0)
905                 GOTO(out, rc);
906
907         if (unlikely(buf->lb_buf == NULL)) {
908                 lu_buf_realloc(buf, rc);
909                 if (buf->lb_buf == NULL)
910                         GOTO(out, rc = -ENOMEM);
911
912                 goto again;
913         }
914
915         lmv1 = buf->lb_buf;
916         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
917                 GOTO(out, rc = 0);
918
919         fops = inode->i_fop;
920         dentry->d_inode = inode;
921         dentry->d_sb = inode->i_sb;
922         filp->f_pos = 0;
923         filp->f_path.dentry = dentry;
924         filp->f_mode = FMODE_64BITHASH;
925         filp->f_mapping = inode->i_mapping;
926         filp->f_op = fops;
927         filp->private_data = NULL;
928         set_file_inode(filp, inode);
929
930 #ifdef HAVE_DIR_CONTEXT
931         oclb.ctx.pos = filp->f_pos;
932         rc = fops->iterate(filp, &oclb.ctx);
933         filp->f_pos = oclb.ctx.pos;
934 #else
935         rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
936 #endif
937         fops->release(inode, filp);
938
939 out:
940         if (rc < 0)
941                 CDEBUG(D_LFSCK, "%.16s: fail to check LMV EA, inode = %lu/%u,"
942                        DFID": rc = %d\n",
943                        LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
944                        inode->i_ino, inode->i_generation,
945                        PFID(&oic->oic_fid), rc);
946         else
947                 rc = 0;
948
949         RETURN(rc);
950 }
951
952 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
953                           const struct lu_fid *fid,
954                           const struct lu_object_conf *conf)
955 {
956         struct osd_thread_info *info;
957         struct lu_device       *ldev   = obj->oo_dt.do_lu.lo_dev;
958         struct osd_device      *dev;
959         struct osd_idmap_cache *oic;
960         struct osd_inode_id    *id;
961         struct osd_inode_id    *tid;
962         struct inode           *inode = NULL;
963         struct osd_scrub       *scrub;
964         struct scrub_file      *sf;
965         __u32                   flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT |
966                                         SS_AUTO_FULL;
967         __u32                   saved_ino;
968         __u32                   saved_gen;
969         int                     result  = 0;
970         int                     rc1     = 0;
971         bool                    cached  = true;
972         bool                    remote  = false;
973         ENTRY;
974
975         LINVRNT(osd_invariant(obj));
976         LASSERT(obj->oo_inode == NULL);
977         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
978
979         dev = osd_dev(ldev);
980         scrub = &dev->od_scrub;
981         sf = &scrub->os_file;
982         info = osd_oti_get(env);
983         LASSERT(info);
984         oic = &info->oti_cache;
985
986         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
987                 RETURN(-ENOENT);
988
989         /* For the object is created as locking anchor, or for the object to
990          * be created on disk. No need to osd_oi_lookup() at here because FID
991          * shouldn't never be re-used, if it's really a duplicate FID from
992          * unexpected reason, we should be able to detect it later by calling
993          * do_create->osd_oi_insert(). */
994         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
995                 GOTO(out, result = 0);
996
997         /* Search order: 1. per-thread cache. */
998         if (lu_fid_eq(fid, &oic->oic_fid) &&
999             likely(oic->oic_dev == dev)) {
1000                 id = &oic->oic_lid;
1001                 goto iget;
1002         }
1003
1004         id = &info->oti_id;
1005         if (!list_empty(&scrub->os_inconsistent_items)) {
1006                 /* Search order: 2. OI scrub pending list. */
1007                 result = osd_oii_lookup(dev, fid, id);
1008                 if (result == 0)
1009                         goto iget;
1010         }
1011
1012         cached = false;
1013         /* Search order: 3. OI files. */
1014         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1015         if (result == -ENOENT) {
1016                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1017                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1018                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1019                                       sf->sf_oi_bitmap))
1020                         GOTO(out, result = 0);
1021
1022                 goto trigger;
1023         }
1024
1025         if (result != 0)
1026                 GOTO(out, result);
1027
1028 iget:
1029         inode = osd_iget_check(info, dev, fid, id, cached);
1030         if (IS_ERR(inode)) {
1031                 result = PTR_ERR(inode);
1032                 if (result == -ENOENT || result == -ESTALE)
1033                         GOTO(out, result = 0);
1034
1035                 if (result == -EREMCHG) {
1036
1037 trigger:
1038                         /* We still have chance to get the valid inode: for the
1039                          * object which is referenced by remote name entry, the
1040                          * object on the local MDT will be linked under the dir
1041                          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1042                          *
1043                          * We do not know whether the object for the given FID
1044                          * is referenced by some remote name entry or not, and
1045                          * especially for DNE II, a multiple-linked object may
1046                          * have many name entries reside on many MDTs.
1047                          *
1048                          * To simplify the operation, OSD will not distinguish
1049                          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1050                          * only happened for the RPC from other MDT during the
1051                          * OI scrub, or for the client side RPC with FID only,
1052                          * such as FID to path, or from old connected client. */
1053                         if (!remote &&
1054                             !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1055                                 rc1 = osd_lookup_in_remote_parent(info, dev,
1056                                                                   fid, id);
1057                                 if (rc1 == 0) {
1058                                         remote = true;
1059                                         cached = true;
1060                                         flags |= SS_AUTO_PARTIAL;
1061                                         flags &= ~SS_AUTO_FULL;
1062                                         goto iget;
1063                                 }
1064                         }
1065
1066                         if (thread_is_running(&scrub->os_thread)) {
1067                                 if (scrub->os_partial_scan &&
1068                                     !scrub->os_in_join) {
1069                                         goto join;
1070                                 } else {
1071                                         if (inode != NULL && !IS_ERR(inode)) {
1072                                                 LASSERT(remote);
1073
1074                                                 osd_add_oi_cache(info, dev, id,
1075                                                                  fid);
1076                                                 osd_oii_insert(dev, oic, true);
1077                                         } else {
1078                                                 result = -EINPROGRESS;
1079                                         }
1080                                 }
1081                         } else if (!dev->od_noscrub) {
1082
1083 join:
1084                                 rc1 = osd_scrub_start(dev, flags);
1085                                 LCONSOLE_WARN("%.16s: trigger OI scrub by RPC "
1086                                               "for the "DFID" with flags 0x%x,"
1087                                               " rc = %d\n", osd_name(dev),
1088                                               PFID(fid), flags, rc1);
1089                                 if (rc1 == 0 || rc1 == -EALREADY) {
1090                                         if (inode != NULL && !IS_ERR(inode)) {
1091                                                 LASSERT(remote);
1092
1093                                                 osd_add_oi_cache(info, dev, id,
1094                                                                  fid);
1095                                                 osd_oii_insert(dev, oic, true);
1096                                         } else {
1097                                                 result = -EINPROGRESS;
1098                                         }
1099                                 } else {
1100                                         result = -EREMCHG;
1101                                 }
1102                         } else {
1103                                 result = -EREMCHG;
1104                         }
1105                 }
1106
1107                 if (inode == NULL || IS_ERR(inode))
1108                         GOTO(out, result);
1109         } else if (remote) {
1110                 goto trigger;
1111         }
1112
1113         obj->oo_inode = inode;
1114         LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
1115
1116         result = osd_check_lma(env, obj);
1117         if (result == 0)
1118                 goto found;
1119
1120         tid = &info->oti_id3;
1121         LASSERT(tid != id);
1122
1123         if (result == -ENODATA) {
1124                 if (!cached)
1125                         /* The current OI mapping is from the OI file,
1126                          * since the inode has been found via
1127                          * osd_iget_check(), no need recheck OI. */
1128                         goto found;
1129
1130                 result = osd_oi_lookup(info, dev, fid, tid, OI_CHECK_FLD);
1131                 if (result == 0) {
1132                         LASSERTF(tid->oii_ino == id->oii_ino &&
1133                                  tid->oii_gen == id->oii_gen,
1134                                  "OI mapping changed(1): %u/%u => %u/%u",
1135                                  tid->oii_ino, tid->oii_gen,
1136                                  id->oii_ino, id->oii_gen);
1137
1138                         LASSERTF(tid->oii_ino == inode->i_ino &&
1139                                  tid->oii_gen == inode->i_generation,
1140                                  "locate wrong inode(1): %u/%u => %ld/%u",
1141                                  tid->oii_ino, tid->oii_gen,
1142                                  inode->i_ino, inode->i_generation);
1143
1144                         /* "result == 0" means the cached OI mapping is still in
1145                          * the OI file, so the target the inode is valid. */
1146                         goto found;
1147                 }
1148
1149                 /* "result == -ENOENT" means that the OI mappinghas been removed
1150                  * by race, the target inode belongs to other object.
1151                  *
1152                  * Others error can be returned  directly. */
1153                 if (result == -ENOENT)
1154                         result = 0;
1155         }
1156
1157         saved_ino = inode->i_ino;
1158         saved_gen = inode->i_generation;
1159         iput(inode);
1160         inode = NULL;
1161         obj->oo_inode = NULL;
1162
1163         if (result != -EREMCHG)
1164                 GOTO(out, result);
1165
1166         if (!cached)
1167                 /* The current OI mapping is from the OI file,
1168                  * since the inode has been found via
1169                  * osd_iget_check(), no need recheck OI. */
1170                 goto trigger;
1171
1172         result = osd_oi_lookup(info, dev, fid, tid, OI_CHECK_FLD);
1173         /* "result == -ENOENT" means the cached OI mapping has been removed from
1174          * the OI file by race, above target inode belongs to other object.
1175          *
1176          * Others error can be returned directly. */
1177         if (result != 0)
1178                 GOTO(out, result = (result == -ENOENT ? 0 : result));
1179
1180         LASSERTF(tid->oii_ino == id->oii_ino && tid->oii_gen == id->oii_gen,
1181                  "OI mapping changed(2): %u/%u => %u/%u",
1182                  tid->oii_ino, tid->oii_gen, id->oii_ino, id->oii_gen);
1183
1184         LASSERTF(tid->oii_ino == saved_ino && tid->oii_gen == saved_gen,
1185                  "locate wrong inode(2): %u/%u => %u/%u",
1186                  tid->oii_ino, tid->oii_gen, saved_ino, saved_gen);
1187
1188         goto trigger;
1189
1190 found:
1191         obj->oo_compat_dot_created = 1;
1192         obj->oo_compat_dotdot_created = 1;
1193
1194         if (S_ISDIR(inode->i_mode) &&
1195             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1196                 osd_check_lmv(info, dev, inode, oic);
1197
1198         result = osd_attach_jinode(inode);
1199         if (result) {
1200                 obj->oo_inode = NULL;
1201                 iput(inode);
1202                 GOTO(out, result);
1203         }
1204
1205         if (!ldiskfs_pdo)
1206                 GOTO(out, result = 0);
1207
1208         LASSERT(obj->oo_hl_head == NULL);
1209         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1210         if (obj->oo_hl_head == NULL) {
1211                 obj->oo_inode = NULL;
1212                 iput(inode);
1213                 GOTO(out, result = -ENOMEM);
1214         }
1215         GOTO(out, result = 0);
1216
1217 out:
1218         if (result != 0 && cached)
1219                 fid_zero(&oic->oic_fid);
1220
1221         LINVRNT(osd_invariant(obj));
1222         return result;
1223 }
1224
1225 /*
1226  * Concurrency: shouldn't matter.
1227  */
1228 static void osd_object_init0(struct osd_object *obj)
1229 {
1230         LASSERT(obj->oo_inode != NULL);
1231         obj->oo_dt.do_body_ops = &osd_body_ops;
1232         obj->oo_dt.do_lu.lo_header->loh_attr |=
1233                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1234 }
1235
1236 /*
1237  * Concurrency: no concurrent access is possible that early in object
1238  * life-cycle.
1239  */
1240 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1241                            const struct lu_object_conf *conf)
1242 {
1243         struct osd_object *obj = osd_obj(l);
1244         int result;
1245
1246         LINVRNT(osd_invariant(obj));
1247
1248         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1249                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1250                 l->lo_header->loh_attr |= LOHA_EXISTS;
1251                 return 0;
1252         }
1253
1254         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1255         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1256         if (result == 0 && obj->oo_inode != NULL) {
1257                 struct osd_thread_info *oti = osd_oti_get(env);
1258                 struct lustre_mdt_attrs *lma = &oti->oti_mdt_attrs;
1259
1260                 osd_object_init0(obj);
1261                 result = osd_get_lma(oti, obj->oo_inode,
1262                                      &oti->oti_obj_dentry, lma);
1263                 if (result == 0) {
1264                         /* Convert LMAI flags to lustre LMA flags
1265                          * and cache it to oo_lma_flags */
1266                         obj->oo_lma_flags =
1267                                 lma_to_lustre_flags(lma->lma_incompat);
1268                 } else if (result == -ENODATA) {
1269                         result = 0;
1270                 }
1271         }
1272
1273         LINVRNT(osd_invariant(obj));
1274         return result;
1275 }
1276
1277 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1278  * The left part is for value, binary mode. */
1279 struct osd_xattr_entry {
1280         struct list_head        oxe_list;
1281         size_t                  oxe_len;
1282         size_t                  oxe_namelen;
1283         bool                    oxe_exist;
1284         struct rcu_head         oxe_rcu;
1285         char                    oxe_buf[0];
1286 };
1287
1288 static int osd_oxc_get(struct osd_object *obj, const char *name,
1289                        struct lu_buf *buf)
1290 {
1291         struct osd_xattr_entry *tmp;
1292         struct osd_xattr_entry *oxe = NULL;
1293         size_t namelen = strlen(name);
1294         int rc;
1295         ENTRY;
1296
1297         rcu_read_lock();
1298         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1299                 if (namelen == tmp->oxe_namelen &&
1300                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1301                         oxe = tmp;
1302                         break;
1303                 }
1304         }
1305
1306         if (oxe == NULL)
1307                 GOTO(out, rc = -ENOENT);
1308
1309         if (!oxe->oxe_exist)
1310                 GOTO(out, rc = -ENODATA);
1311
1312         /* vallen */
1313         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1314         LASSERT(rc > 0);
1315
1316         if (buf->lb_buf == NULL)
1317                 GOTO(out, rc);
1318
1319         if (buf->lb_len < rc)
1320                 GOTO(out, rc = -ERANGE);
1321
1322         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1323         EXIT;
1324 out:
1325         rcu_read_unlock();
1326
1327         return rc;
1328 }
1329
1330 static void osd_oxc_free(struct rcu_head *head)
1331 {
1332         struct osd_xattr_entry *oxe;
1333
1334         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1335         OBD_FREE(oxe, oxe->oxe_len);
1336 }
1337
1338 static void osd_oxc_add(struct osd_object *obj, const char *name,
1339                         const char *buf, int buflen)
1340 {
1341         struct osd_xattr_entry *oxe;
1342         struct osd_xattr_entry *old = NULL;
1343         struct osd_xattr_entry *tmp;
1344         size_t namelen = strlen(name);
1345         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1346
1347         OBD_ALLOC(oxe, len);
1348         if (oxe == NULL)
1349                 return;
1350
1351         INIT_LIST_HEAD(&oxe->oxe_list);
1352         oxe->oxe_len = len;
1353         oxe->oxe_namelen = namelen;
1354         memcpy(oxe->oxe_buf, name, namelen);
1355         if (buflen > 0) {
1356                 LASSERT(buf != NULL);
1357                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1358                 oxe->oxe_exist = true;
1359         } else {
1360                 oxe->oxe_exist = false;
1361         }
1362
1363         /* this should be rarely called, just remove old and add new */
1364         spin_lock(&obj->oo_guard);
1365         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1366                 if (namelen == tmp->oxe_namelen &&
1367                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1368                         old = tmp;
1369                         break;
1370                 }
1371         }
1372         if (old != NULL) {
1373                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1374                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1375         } else {
1376                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1377         }
1378         spin_unlock(&obj->oo_guard);
1379 }
1380
1381 static void osd_oxc_del(struct osd_object *obj, const char *name)
1382 {
1383         struct osd_xattr_entry *oxe;
1384         size_t namelen = strlen(name);
1385
1386         spin_lock(&obj->oo_guard);
1387         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1388                 if (namelen == oxe->oxe_namelen &&
1389                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1390                         list_del_rcu(&oxe->oxe_list);
1391                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1392                         break;
1393                 }
1394         }
1395         spin_unlock(&obj->oo_guard);
1396 }
1397
1398 static void osd_oxc_fini(struct osd_object *obj)
1399 {
1400         struct osd_xattr_entry *oxe, *next;
1401
1402         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1403                 list_del(&oxe->oxe_list);
1404                 OBD_FREE(oxe, oxe->oxe_len);
1405         }
1406 }
1407
1408 /*
1409  * Concurrency: no concurrent access is possible that late in object
1410  * life-cycle.
1411  */
1412 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1413 {
1414         struct osd_object *obj = osd_obj(l);
1415
1416         LINVRNT(osd_invariant(obj));
1417
1418         osd_oxc_fini(obj);
1419         dt_object_fini(&obj->oo_dt);
1420         if (obj->oo_hl_head != NULL)
1421                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1422         OBD_FREE_PTR(obj);
1423 }
1424
1425 /*
1426  * Concurrency: no concurrent access is possible that late in object
1427  * life-cycle.
1428  */
1429 static void osd_index_fini(struct osd_object *o)
1430 {
1431         struct iam_container *bag;
1432
1433         if (o->oo_dir != NULL) {
1434                 bag = &o->oo_dir->od_container;
1435                 if (o->oo_inode != NULL) {
1436                         if (bag->ic_object == o->oo_inode)
1437                                 iam_container_fini(bag);
1438                 }
1439                 OBD_FREE_PTR(o->oo_dir);
1440                 o->oo_dir = NULL;
1441         }
1442 }
1443
1444 /*
1445  * Concurrency: no concurrent access is possible that late in object
1446  * life-cycle (for all existing callers, that is. New callers have to provide
1447  * their own locking.)
1448  */
1449 static int osd_inode_unlinked(const struct inode *inode)
1450 {
1451         return inode->i_nlink == 0;
1452 }
1453
1454 enum {
1455         OSD_TXN_OI_DELETE_CREDITS    = 20,
1456         OSD_TXN_INODE_DELETE_CREDITS = 20
1457 };
1458
1459 /*
1460  * Journal
1461  */
1462
1463 #if OSD_THANDLE_STATS
1464 /**
1465  * Set time when the handle is allocated
1466  */
1467 static void osd_th_alloced(struct osd_thandle *oth)
1468 {
1469         oth->oth_alloced = cfs_time_current();
1470 }
1471
1472 /**
1473  * Set time when the handle started
1474  */
1475 static void osd_th_started(struct osd_thandle *oth)
1476 {
1477         oth->oth_started = cfs_time_current();
1478 }
1479
1480 /**
1481  * Helper function to convert time interval to microseconds packed in
1482  * long int.
1483  */
1484 static long interval_to_usec(cfs_time_t start, cfs_time_t end)
1485 {
1486         struct timeval val;
1487
1488         cfs_duration_usec(cfs_time_sub(end, start), &val);
1489         return val.tv_sec * 1000000 + val.tv_usec;
1490 }
1491
1492 /**
1493  * Check whether the we deal with this handle for too long.
1494  */
1495 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1496                                 cfs_time_t alloced, cfs_time_t started,
1497                                 cfs_time_t closed)
1498 {
1499         cfs_time_t now = cfs_time_current();
1500
1501         LASSERT(dev != NULL);
1502
1503         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1504                             interval_to_usec(alloced, started));
1505         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1506                             interval_to_usec(started, closed));
1507         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1508                             interval_to_usec(closed, now));
1509
1510         if (cfs_time_before(cfs_time_add(alloced, cfs_time_seconds(30)), now)) {
1511                 CWARN("transaction handle %p was open for too long: "
1512                       "now "CFS_TIME_T" ,"
1513                       "alloced "CFS_TIME_T" ,"
1514                       "started "CFS_TIME_T" ,"
1515                       "closed "CFS_TIME_T"\n",
1516                       oth, now, alloced, started, closed);
1517                 libcfs_debug_dumpstack(NULL);
1518         }
1519 }
1520
1521 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1522 {                                                                       \
1523         cfs_time_t __closed = cfs_time_current();                       \
1524         cfs_time_t __alloced = oth->oth_alloced;                        \
1525         cfs_time_t __started = oth->oth_started;                        \
1526                                                                         \
1527         expr;                                                           \
1528         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1529 }
1530
1531 #else /* OSD_THANDLE_STATS */
1532
1533 #define osd_th_alloced(h)                  do {} while(0)
1534 #define osd_th_started(h)                  do {} while(0)
1535 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1536
1537 #endif /* OSD_THANDLE_STATS */
1538
1539 /*
1540  * Concurrency: doesn't access mutable data.
1541  */
1542 static int osd_param_is_not_sane(const struct osd_device *dev,
1543                                  const struct thandle *th)
1544 {
1545         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1546
1547         return oh->ot_credits > osd_transaction_size(dev);
1548 }
1549
1550 /*
1551  * Concurrency: shouldn't matter.
1552  */
1553 static void osd_trans_commit_cb(struct super_block *sb,
1554                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1555 {
1556         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1557         struct thandle     *th  = &oh->ot_super;
1558         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
1559         struct dt_txn_commit_cb *dcb, *tmp;
1560
1561         LASSERT(oh->ot_handle == NULL);
1562
1563         if (error)
1564                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1565
1566         dt_txn_hook_commit(th);
1567
1568         /* call per-transaction callbacks if any */
1569         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1570                                  dcb_linkage) {
1571                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1572                          "commit callback entry: magic=%x name='%s'\n",
1573                          dcb->dcb_magic, dcb->dcb_name);
1574                 list_del_init(&dcb->dcb_linkage);
1575                 dcb->dcb_func(NULL, th, dcb, error);
1576         }
1577
1578         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1579         lu_device_put(lud);
1580         th->th_dev = NULL;
1581
1582         lu_context_exit(&th->th_ctx);
1583         lu_context_fini(&th->th_ctx);
1584         OBD_FREE_PTR(oh);
1585 }
1586
1587 #ifndef HAVE_SB_START_WRITE
1588 # define sb_start_write(sb) do {} while (0)
1589 # define sb_end_write(sb) do {} while (0)
1590 #endif
1591
1592 static struct thandle *osd_trans_create(const struct lu_env *env,
1593                                         struct dt_device *d)
1594 {
1595         struct osd_thread_info  *oti = osd_oti_get(env);
1596         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1597         struct osd_thandle      *oh;
1598         struct thandle          *th;
1599         ENTRY;
1600
1601         /* on pending IO in this thread should left from prev. request */
1602         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1603
1604         sb_start_write(osd_sb(osd_dt_dev(d)));
1605
1606         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1607         if (oh != NULL) {
1608                 oh->ot_quota_trans = &oti->oti_quota_trans;
1609                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1610                 th = &oh->ot_super;
1611                 th->th_dev = d;
1612                 th->th_result = 0;
1613                 th->th_tags = LCT_TX_HANDLE;
1614                 oh->ot_credits = 0;
1615                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1616                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1617                 osd_th_alloced(oh);
1618
1619                 memset(oti->oti_declare_ops, 0,
1620                        sizeof(oti->oti_declare_ops));
1621                 memset(oti->oti_declare_ops_cred, 0,
1622                        sizeof(oti->oti_declare_ops_cred));
1623                 memset(oti->oti_declare_ops_used, 0,
1624                        sizeof(oti->oti_declare_ops_used));
1625         } else {
1626                 sb_end_write(osd_sb(osd_dt_dev(d)));
1627                 th = ERR_PTR(-ENOMEM);
1628         }
1629         RETURN(th);
1630 }
1631
1632 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1633 {
1634         struct osd_thread_info  *oti = osd_oti_get(env);
1635         struct osd_thandle      *oh;
1636
1637         oh = container_of0(th, struct osd_thandle, ot_super);
1638         LASSERT(oh != NULL);
1639
1640         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1641               oti->oti_declare_ops[OSD_OT_CREATE],
1642               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1643               oti->oti_declare_ops_used[OSD_OT_CREATE],
1644               oti->oti_declare_ops[OSD_OT_DESTROY],
1645               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1646               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1647         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1648               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1649               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1650               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1651               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1652               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1653               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1654         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1655               oti->oti_declare_ops[OSD_OT_WRITE],
1656               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1657               oti->oti_declare_ops_used[OSD_OT_WRITE],
1658               oti->oti_declare_ops[OSD_OT_PUNCH],
1659               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1660               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1661               oti->oti_declare_ops[OSD_OT_QUOTA],
1662               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1663               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1664         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1665               oti->oti_declare_ops[OSD_OT_INSERT],
1666               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1667               oti->oti_declare_ops_used[OSD_OT_INSERT],
1668               oti->oti_declare_ops[OSD_OT_DELETE],
1669               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1670               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1671         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1672               oti->oti_declare_ops[OSD_OT_REF_ADD],
1673               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1674               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1675               oti->oti_declare_ops[OSD_OT_REF_DEL],
1676               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1677               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1678 }
1679
1680 /*
1681  * Concurrency: shouldn't matter.
1682  */
1683 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1684                            struct thandle *th)
1685 {
1686         struct osd_thread_info *oti = osd_oti_get(env);
1687         struct osd_device  *dev = osd_dt_dev(d);
1688         handle_t           *jh;
1689         struct osd_thandle *oh;
1690         int rc;
1691
1692         ENTRY;
1693
1694         LASSERT(current->journal_info == NULL);
1695
1696         oh = container_of0(th, struct osd_thandle, ot_super);
1697         LASSERT(oh != NULL);
1698         LASSERT(oh->ot_handle == NULL);
1699
1700         rc = dt_txn_hook_start(env, d, th);
1701         if (rc != 0)
1702                 GOTO(out, rc);
1703
1704         if (unlikely(osd_param_is_not_sane(dev, th))) {
1705                 static unsigned long last_printed;
1706                 static int last_credits;
1707
1708                 /* don't make noise on a tiny testing systems
1709                  * actual credits misuse will be caught anyway */
1710                 if (last_credits != oh->ot_credits &&
1711                     time_after(jiffies, last_printed +
1712                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1713                     osd_transaction_size(dev) > 512) {
1714                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1715                               oh->ot_credits, osd_transaction_size(dev));
1716                         osd_trans_dump_creds(env, th);
1717                         libcfs_debug_dumpstack(NULL);
1718                         last_credits = oh->ot_credits;
1719                         last_printed = jiffies;
1720                 }
1721                 /* XXX Limit the credits to 'max_transaction_buffers', and
1722                  *     let the underlying filesystem to catch the error if
1723                  *     we really need so many credits.
1724                  *
1725                  *     This should be removed when we can calculate the
1726                  *     credits precisely. */
1727                 oh->ot_credits = osd_transaction_size(dev);
1728         }
1729
1730         /*
1731          * XXX temporary stuff. Some abstraction layer should
1732          * be used.
1733          */
1734         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1735         osd_th_started(oh);
1736         if (!IS_ERR(jh)) {
1737                 oh->ot_handle = jh;
1738                 LASSERT(oti->oti_txns == 0);
1739                 lu_context_init(&th->th_ctx, th->th_tags);
1740                 lu_context_enter(&th->th_ctx);
1741
1742                 lu_device_get(&d->dd_lu_dev);
1743                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1744                               "osd-tx", th);
1745                 oti->oti_txns++;
1746                 rc = 0;
1747         } else {
1748                 rc = PTR_ERR(jh);
1749         }
1750 out:
1751         RETURN(rc);
1752 }
1753
1754 static int osd_seq_exists(const struct lu_env *env,
1755                           struct osd_device *osd, u64 seq)
1756 {
1757         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1758         struct seq_server_site  *ss = osd_seq_site(osd);
1759         int                     rc;
1760         ENTRY;
1761
1762         LASSERT(ss != NULL);
1763         LASSERT(ss->ss_server_fld != NULL);
1764
1765         rc = osd_fld_lookup(env, osd, seq, range);
1766         if (rc != 0) {
1767                 if (rc != -ENOENT)
1768                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1769                                osd_name(osd), seq, rc);
1770                 RETURN(0);
1771         }
1772
1773         RETURN(ss->ss_node_id == range->lsr_index);
1774 }
1775
1776 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1777 {
1778         struct dt_txn_commit_cb *dcb;
1779         struct dt_txn_commit_cb *tmp;
1780
1781         /* call per-transaction stop callbacks if any */
1782         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1783                                  dcb_linkage) {
1784                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1785                          "commit callback entry: magic=%x name='%s'\n",
1786                          dcb->dcb_magic, dcb->dcb_name);
1787                 list_del_init(&dcb->dcb_linkage);
1788                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1789         }
1790 }
1791
1792 /*
1793  * Concurrency: shouldn't matter.
1794  */
1795 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1796                           struct thandle *th)
1797 {
1798         int                     rc = 0, remove_agents = 0;
1799         struct osd_thandle     *oh;
1800         struct osd_thread_info *oti = osd_oti_get(env);
1801         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1802         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1803         struct qsd_instance    *qsd = osd->od_quota_slave;
1804         struct lquota_trans    *qtrans;
1805         ENTRY;
1806
1807         oh = container_of0(th, struct osd_thandle, ot_super);
1808
1809         /* reset OI cache for safety */
1810         oti->oti_ins_cache_used = 0;
1811
1812         remove_agents = oh->ot_remove_agents;
1813
1814         qtrans = oh->ot_quota_trans;
1815         oh->ot_quota_trans = NULL;
1816
1817         if (oh->ot_handle != NULL) {
1818                 int rc2;
1819                 handle_t *hdl = oh->ot_handle;
1820
1821                 /*
1822                  * add commit callback
1823                  * notice we don't do this in osd_trans_start()
1824                  * as underlying transaction can change during truncate
1825                  */
1826                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1827                                          &oh->ot_jcb);
1828
1829                 LASSERT(oti->oti_txns == 1);
1830                 oti->oti_txns--;
1831
1832                 rc = dt_txn_hook_stop(env, th);
1833                 if (rc != 0)
1834                         CERROR("%s: failed in transaction hook: rc = %d\n",
1835                                osd_name(osd), rc);
1836
1837                 osd_trans_stop_cb(oh, rc);
1838                 /* hook functions might modify th_sync */
1839                 hdl->h_sync = th->th_sync;
1840
1841                 oh->ot_handle = NULL;
1842                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1843                 if (rc2 != 0)
1844                         CERROR("%s: failed to stop transaction: rc = %d\n",
1845                                osd_name(osd), rc2);
1846                 if (!rc)
1847                         rc = rc2;
1848         } else {
1849                 osd_trans_stop_cb(oh, th->th_result);
1850                 OBD_FREE_PTR(oh);
1851         }
1852
1853         /* inform the quota slave device that the transaction is stopping */
1854         qsd_op_end(env, qsd, qtrans);
1855
1856         /* as we want IO to journal and data IO be concurrent, we don't block
1857          * awaiting data IO completion in osd_do_bio(), instead we wait here
1858          * once transaction is submitted to the journal. all reqular requests
1859          * don't do direct IO (except read/write), thus this wait_event becomes
1860          * no-op for them.
1861          *
1862          * IMPORTANT: we have to wait till any IO submited by the thread is
1863          * completed otherwise iobuf may be corrupted by different request
1864          */
1865         wait_event(iobuf->dr_wait,
1866                        atomic_read(&iobuf->dr_numreqs) == 0);
1867         osd_fini_iobuf(osd, iobuf);
1868         if (!rc)
1869                 rc = iobuf->dr_error;
1870
1871         if (unlikely(remove_agents != 0))
1872                 osd_process_scheduled_agent_removals(env, osd);
1873
1874         sb_end_write(osd_sb(osd));
1875
1876         RETURN(rc);
1877 }
1878
1879 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1880 {
1881         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1882                                                ot_super);
1883
1884         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1885         LASSERT(&dcb->dcb_func != NULL);
1886         if (dcb->dcb_flags & DCB_TRANS_STOP)
1887                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1888         else
1889                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1890
1891         return 0;
1892 }
1893
1894 /*
1895  * Called just before object is freed. Releases all resources except for
1896  * object itself (that is released by osd_object_free()).
1897  *
1898  * Concurrency: no concurrent access is possible that late in object
1899  * life-cycle.
1900  */
1901 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1902 {
1903         struct osd_object *obj   = osd_obj(l);
1904         struct inode      *inode = obj->oo_inode;
1905
1906         LINVRNT(osd_invariant(obj));
1907
1908         /*
1909          * If object is unlinked remove fid->ino mapping from object index.
1910          */
1911
1912         osd_index_fini(obj);
1913         if (inode != NULL) {
1914                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1915                 qid_t                    uid = i_uid_read(inode);
1916                 qid_t                    gid = i_gid_read(inode);
1917
1918                 iput(inode);
1919                 obj->oo_inode = NULL;
1920
1921                 if (qsd != NULL) {
1922                         struct osd_thread_info  *info = osd_oti_get(env);
1923                         struct lquota_id_info   *qi = &info->oti_qi;
1924
1925                         /* Release granted quota to master if necessary */
1926                         qi->lqi_id.qid_uid = uid;
1927                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1928
1929                         qi->lqi_id.qid_uid = gid;
1930                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1931                 }
1932         }
1933 }
1934
1935 /*
1936  * Concurrency: ->loo_object_release() is called under site spin-lock.
1937  */
1938 static void osd_object_release(const struct lu_env *env,
1939                                struct lu_object *l)
1940 {
1941         struct osd_object *o = osd_obj(l);
1942         /* nobody should be releasing a non-destroyed object with nlink=0
1943          * the API allows this, but ldiskfs doesn't like and then report
1944          * this inode as deleted */
1945         if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1946                 LBUG();
1947 }
1948
1949 /*
1950  * Concurrency: shouldn't matter.
1951  */
1952 static int osd_object_print(const struct lu_env *env, void *cookie,
1953                             lu_printer_t p, const struct lu_object *l)
1954 {
1955         struct osd_object *o = osd_obj(l);
1956         struct iam_descr  *d;
1957
1958         if (o->oo_dir != NULL)
1959                 d = o->oo_dir->od_container.ic_descr;
1960         else
1961                 d = NULL;
1962         return (*p)(env, cookie,
1963                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1964                     o, o->oo_inode,
1965                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1966                     o->oo_inode ? o->oo_inode->i_generation : 0,
1967                     d ? d->id_ops->id_name : "plain");
1968 }
1969
1970 /*
1971  * Concurrency: shouldn't matter.
1972  */
1973 int osd_statfs(const struct lu_env *env, struct dt_device *d,
1974                struct obd_statfs *sfs)
1975 {
1976         struct osd_device       *osd = osd_dt_dev(d);
1977         struct super_block      *sb = osd_sb(osd);
1978         struct kstatfs          *ksfs;
1979         __u64                    reserved;
1980         int                      result = 0;
1981
1982         if (unlikely(osd->od_mnt == NULL))
1983                 return -EINPROGRESS;
1984
1985         /* osd_lproc.c call this without env, allocate ksfs for that case */
1986         if (unlikely(env == NULL)) {
1987                 OBD_ALLOC_PTR(ksfs);
1988                 if (ksfs == NULL)
1989                         return -ENOMEM;
1990         } else {
1991                 ksfs = &osd_oti_get(env)->oti_ksfs;
1992         }
1993
1994         result = sb->s_op->statfs(sb->s_root, ksfs);
1995         if (result)
1996                 goto out;
1997
1998         statfs_pack(sfs, ksfs);
1999         if (unlikely(sb->s_flags & MS_RDONLY))
2000                 sfs->os_state |= OS_STATE_READONLY;
2001         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2002                                          LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2003                 sfs->os_maxbytes = sb->s_maxbytes;
2004         else
2005                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2006
2007         /*
2008          * Reserve some space so to avoid fragmenting the filesystem too much.
2009          * Fragmentation not only impacts performance, but can also increase
2010          * metadata overhead significantly, causing grant calculation to be
2011          * wrong.
2012          *
2013          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2014          */
2015         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2016         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2017         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2018                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2019
2020         sfs->os_blocks -= reserved;
2021         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2022         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2023
2024 out:
2025         if (unlikely(env == NULL))
2026                 OBD_FREE_PTR(ksfs);
2027         return result;
2028 }
2029
2030 /**
2031  * Estimate space needed for file creations. We assume the largest filename
2032  * which is 2^64 - 1, hence a filename of 20 chars.
2033  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2034  */
2035 #ifdef __LDISKFS_DIR_REC_LEN
2036 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2037 #else
2038 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2039 #endif
2040
2041 /*
2042  * Concurrency: doesn't access mutable data.
2043  */
2044 static void osd_conf_get(const struct lu_env *env,
2045                          const struct dt_device *dev,
2046                          struct dt_device_param *param)
2047 {
2048         struct super_block *sb = osd_sb(osd_dt_dev(dev));
2049         int                ea_overhead;
2050
2051         /*
2052          * XXX should be taken from not-yet-existing fs abstraction layer.
2053          */
2054         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2055         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2056         param->ddp_symlink_max  = sb->s_blocksize;
2057         param->ddp_mount_type     = LDD_MT_LDISKFS;
2058         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2059                 param->ddp_maxbytes = sb->s_maxbytes;
2060         else
2061                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2062         /* inode are statically allocated, so per-inode space consumption
2063          * is the space consumed by the directory entry */
2064         param->ddp_inodespace     = PER_OBJ_USAGE;
2065         /* EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
2066          * = 128MB) which is unlikely to be hit in real life. Report a smaller
2067          * maximum length to not under count the actual number of extents
2068          * needed for writing a file. */
2069         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2070         /* worst-case extent insertion metadata overhead */
2071         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2072         param->ddp_mntopts      = 0;
2073         if (test_opt(sb, XATTR_USER))
2074                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2075         if (test_opt(sb, POSIX_ACL))
2076                 param->ddp_mntopts |= MNTOPT_ACL;
2077
2078         /* LOD might calculate the max stripe count based on max_ea_size,
2079          * so we need take account in the overhead as well,
2080          * xattr_header + magic + xattr_entry_head */
2081         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2082                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2083
2084 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2085         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2086                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2087                                                                 ea_overhead;
2088         else
2089 #endif
2090                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2091 }
2092
2093 /*
2094  * Concurrency: shouldn't matter.
2095  */
2096 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2097 {
2098         int rc;
2099
2100         CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2101
2102         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2103
2104         CDEBUG(D_CACHE, "synced OSD %s: rc = %d\n",
2105                LUSTRE_OSD_LDISKFS_NAME, rc);
2106
2107         return rc;
2108 }
2109
2110 /**
2111  * Start commit for OSD device.
2112  *
2113  * An implementation of dt_commit_async method for OSD device.
2114  * Asychronously starts underlayng fs sync and thereby a transaction
2115  * commit.
2116  *
2117  * \param env environment
2118  * \param d dt device
2119  *
2120  * \see dt_device_operations
2121  */
2122 static int osd_commit_async(const struct lu_env *env,
2123                             struct dt_device *d)
2124 {
2125         struct super_block *s = osd_sb(osd_dt_dev(d));
2126         ENTRY;
2127
2128         CDEBUG(D_HA, "async commit OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2129         RETURN(s->s_op->sync_fs(s, 0));
2130 }
2131
2132 /*
2133  * Concurrency: shouldn't matter.
2134  */
2135
2136 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2137 {
2138         struct super_block *sb = osd_sb(osd_dt_dev(d));
2139         struct block_device *dev = sb->s_bdev;
2140 #ifdef HAVE_DEV_SET_RDONLY
2141         struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2142         int rc = 0;
2143 #else
2144         int rc = -EOPNOTSUPP;
2145 #endif
2146         ENTRY;
2147
2148 #ifdef HAVE_DEV_SET_RDONLY
2149         CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
2150
2151         if (sb->s_op->freeze_fs) {
2152                 rc = sb->s_op->freeze_fs(sb);
2153                 if (rc)
2154                         goto out;
2155         }
2156
2157         if (jdev && (jdev != dev)) {
2158                 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2159                        (long)jdev);
2160                 dev_set_rdonly(jdev);
2161         }
2162         CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2163         dev_set_rdonly(dev);
2164
2165         if (sb->s_op->unfreeze_fs)
2166                 sb->s_op->unfreeze_fs(sb);
2167
2168 out:
2169 #endif
2170         if (rc)
2171                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2172                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2173
2174         RETURN(rc);
2175 }
2176
2177 /**
2178  * Note: we do not count into QUOTA here.
2179  * If we mount with --data_journal we may need more.
2180  */
2181 const int osd_dto_credits_noquota[DTO_NR] = {
2182         /**
2183          * Insert.
2184          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2185          * SINGLEDATA_TRANS_BLOCKS(8)
2186          * XXX Note: maybe iam need more, since iam have more level than
2187          *           EXT3 htree.
2188          */
2189         [DTO_INDEX_INSERT]  = 16,
2190         /**
2191          * Delete
2192          * just modify a single entry, probably merge few within a block
2193          */
2194         [DTO_INDEX_DELETE]  = 1,
2195         /**
2196          * Used for OI scrub
2197          */
2198         [DTO_INDEX_UPDATE]  = 16,
2199         /**
2200          * 4(inode, inode bits, groups, GDT)
2201          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2202          */
2203         [DTO_OBJECT_CREATE] = 4,
2204         /**
2205          * 4(inode, inode bits, groups, GDT)
2206          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2207          */
2208         [DTO_OBJECT_DELETE] = 4,
2209         /**
2210          * Attr set credits (inode)
2211          */
2212         [DTO_ATTR_SET_BASE] = 1,
2213         /**
2214          * Xattr set. The same as xattr of EXT3.
2215          * DATA_TRANS_BLOCKS(14)
2216          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2217          * are also counted in. Do not know why?
2218          */
2219         [DTO_XATTR_SET]     = 14,
2220         /**
2221          * credits for inode change during write.
2222          */
2223         [DTO_WRITE_BASE]    = 3,
2224         /**
2225          * credits for single block write.
2226          */
2227         [DTO_WRITE_BLOCK]   = 14,
2228         /**
2229          * Attr set credits for chown.
2230          * This is extra credits for setattr, and it is null without quota
2231          */
2232         [DTO_ATTR_SET_CHOWN] = 0
2233 };
2234
2235 static const struct dt_device_operations osd_dt_ops = {
2236         .dt_root_get       = osd_root_get,
2237         .dt_statfs         = osd_statfs,
2238         .dt_trans_create   = osd_trans_create,
2239         .dt_trans_start    = osd_trans_start,
2240         .dt_trans_stop     = osd_trans_stop,
2241         .dt_trans_cb_add   = osd_trans_cb_add,
2242         .dt_conf_get       = osd_conf_get,
2243         .dt_sync           = osd_sync,
2244         .dt_ro             = osd_ro,
2245         .dt_commit_async   = osd_commit_async,
2246 };
2247
2248 static void osd_object_read_lock(const struct lu_env *env,
2249                                  struct dt_object *dt, unsigned role)
2250 {
2251         struct osd_object *obj = osd_dt_obj(dt);
2252         struct osd_thread_info *oti = osd_oti_get(env);
2253
2254         LINVRNT(osd_invariant(obj));
2255
2256         LASSERT(obj->oo_owner != env);
2257         down_read_nested(&obj->oo_sem, role);
2258
2259         LASSERT(obj->oo_owner == NULL);
2260         oti->oti_r_locks++;
2261 }
2262
2263 static void osd_object_write_lock(const struct lu_env *env,
2264                                   struct dt_object *dt, unsigned role)
2265 {
2266         struct osd_object *obj = osd_dt_obj(dt);
2267         struct osd_thread_info *oti = osd_oti_get(env);
2268
2269         LINVRNT(osd_invariant(obj));
2270
2271         LASSERT(obj->oo_owner != env);
2272         down_write_nested(&obj->oo_sem, role);
2273
2274         LASSERT(obj->oo_owner == NULL);
2275         obj->oo_owner = env;
2276         oti->oti_w_locks++;
2277 }
2278
2279 static void osd_object_read_unlock(const struct lu_env *env,
2280                                    struct dt_object *dt)
2281 {
2282         struct osd_object *obj = osd_dt_obj(dt);
2283         struct osd_thread_info *oti = osd_oti_get(env);
2284
2285         LINVRNT(osd_invariant(obj));
2286
2287         LASSERT(oti->oti_r_locks > 0);
2288         oti->oti_r_locks--;
2289         up_read(&obj->oo_sem);
2290 }
2291
2292 static void osd_object_write_unlock(const struct lu_env *env,
2293                                     struct dt_object *dt)
2294 {
2295         struct osd_object *obj = osd_dt_obj(dt);
2296         struct osd_thread_info *oti = osd_oti_get(env);
2297
2298         LINVRNT(osd_invariant(obj));
2299
2300         LASSERT(obj->oo_owner == env);
2301         LASSERT(oti->oti_w_locks > 0);
2302         oti->oti_w_locks--;
2303         obj->oo_owner = NULL;
2304         up_write(&obj->oo_sem);
2305 }
2306
2307 static int osd_object_write_locked(const struct lu_env *env,
2308                                    struct dt_object *dt)
2309 {
2310         struct osd_object *obj = osd_dt_obj(dt);
2311
2312         LINVRNT(osd_invariant(obj));
2313
2314         return obj->oo_owner == env;
2315 }
2316
2317 static struct timespec *osd_inode_time(const struct lu_env *env,
2318                                        struct inode *inode, __u64 seconds)
2319 {
2320         struct osd_thread_info  *oti = osd_oti_get(env);
2321         struct timespec         *t   = &oti->oti_time;
2322
2323         t->tv_sec = seconds;
2324         t->tv_nsec = 0;
2325         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2326         return t;
2327 }
2328
2329 static void osd_inode_getattr(const struct lu_env *env,
2330                               struct inode *inode, struct lu_attr *attr)
2331 {
2332         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2333                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2334                            LA_FLAGS | LA_NLINK | LA_RDEV | LA_BLKSIZE |
2335                            LA_TYPE;
2336
2337         attr->la_atime   = LTIME_S(inode->i_atime);
2338         attr->la_mtime   = LTIME_S(inode->i_mtime);
2339         attr->la_ctime   = LTIME_S(inode->i_ctime);
2340         attr->la_mode    = inode->i_mode;
2341         attr->la_size    = i_size_read(inode);
2342         attr->la_blocks  = inode->i_blocks;
2343         attr->la_uid     = i_uid_read(inode);
2344         attr->la_gid     = i_gid_read(inode);
2345         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2346         attr->la_nlink   = inode->i_nlink;
2347         attr->la_rdev    = inode->i_rdev;
2348         attr->la_blksize = 1 << inode->i_blkbits;
2349         attr->la_blkbits = inode->i_blkbits;
2350 }
2351
2352 static int osd_attr_get(const struct lu_env *env,
2353                         struct dt_object *dt,
2354                         struct lu_attr *attr)
2355 {
2356         struct osd_object *obj = osd_dt_obj(dt);
2357
2358         if (unlikely(!dt_object_exists(dt)))
2359                 return -ENOENT;
2360         if (unlikely(obj->oo_destroyed))
2361                 return -ENOENT;
2362
2363         LASSERT(!dt_object_remote(dt));
2364         LINVRNT(osd_invariant(obj));
2365
2366         spin_lock(&obj->oo_guard);
2367         osd_inode_getattr(env, obj->oo_inode, attr);
2368         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2369                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2370         spin_unlock(&obj->oo_guard);
2371
2372         return 0;
2373 }
2374
2375 static int osd_declare_attr_set(const struct lu_env *env,
2376                                 struct dt_object *dt,
2377                                 const struct lu_attr *attr,
2378                                 struct thandle *handle)
2379 {
2380         struct osd_thandle     *oh;
2381         struct osd_object      *obj;
2382         struct osd_thread_info *info = osd_oti_get(env);
2383         struct lquota_id_info  *qi = &info->oti_qi;
2384         qid_t                   uid;
2385         qid_t                   gid;
2386         long long               bspace;
2387         int                     rc = 0;
2388         bool                    enforce;
2389         ENTRY;
2390
2391         LASSERT(dt != NULL);
2392         LASSERT(handle != NULL);
2393
2394         obj = osd_dt_obj(dt);
2395         LASSERT(osd_invariant(obj));
2396
2397         oh = container_of0(handle, struct osd_thandle, ot_super);
2398         LASSERT(oh->ot_handle == NULL);
2399
2400         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2401                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2402
2403         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2404                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2405
2406         if (attr == NULL || obj->oo_inode == NULL)
2407                 RETURN(rc);
2408
2409         bspace   = obj->oo_inode->i_blocks;
2410         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2411         bspace   = toqb(bspace);
2412
2413         /* Changing ownership is always preformed by super user, it should not
2414          * fail with EDQUOT.
2415          *
2416          * We still need to call the osd_declare_qid() to calculate the journal
2417          * credits for updating quota accounting files and to trigger quota
2418          * space adjustment once the operation is completed.*/
2419         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2420                 /* USERQUOTA */
2421                 uid = i_uid_read(obj->oo_inode);
2422                 qi->lqi_type = USRQUOTA;
2423                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2424                 /* inode accounting */
2425                 qi->lqi_is_blk = false;
2426
2427                 /* one more inode for the new uid ... */
2428                 qi->lqi_id.qid_uid = attr->la_uid;
2429                 qi->lqi_space      = 1;
2430                 /* Reserve credits for the new uid */
2431                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2432                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2433                         rc = 0;
2434                 if (rc)
2435                         RETURN(rc);
2436
2437                 /* and one less inode for the current uid */
2438                 qi->lqi_id.qid_uid = uid;
2439                 qi->lqi_space      = -1;
2440                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2441                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2442                         rc = 0;
2443                 if (rc)
2444                         RETURN(rc);
2445
2446                 /* block accounting */
2447                 qi->lqi_is_blk = true;
2448
2449                 /* more blocks for the new uid ... */
2450                 qi->lqi_id.qid_uid = attr->la_uid;
2451                 qi->lqi_space      = bspace;
2452                 /*
2453                  * Credits for the new uid has been reserved, re-use "obj"
2454                  * to save credit reservation.
2455                  */
2456                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2457                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2458                         rc = 0;
2459                 if (rc)
2460                         RETURN(rc);
2461
2462                 /* and finally less blocks for the current uid */
2463                 qi->lqi_id.qid_uid = uid;
2464                 qi->lqi_space      = -bspace;
2465                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2466                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2467                         rc = 0;
2468                 if (rc)
2469                         RETURN(rc);
2470
2471                 /* GROUP QUOTA */
2472                 gid = i_gid_read(obj->oo_inode);
2473                 qi->lqi_type = GRPQUOTA;
2474                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2475
2476                 /* inode accounting */
2477                 qi->lqi_is_blk = false;
2478
2479                 /* one more inode for the new gid ... */
2480                 qi->lqi_id.qid_gid = attr->la_gid;
2481                 qi->lqi_space      = 1;
2482                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2483                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2484                         rc = 0;
2485                 if (rc)
2486                         RETURN(rc);
2487
2488                 /* and one less inode for the current gid */
2489                 qi->lqi_id.qid_gid = gid;
2490                 qi->lqi_space      = -1;
2491                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2492                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2493                         rc = 0;
2494                 if (rc)
2495                         RETURN(rc);
2496
2497                 /* block accounting */
2498                 qi->lqi_is_blk = true;
2499
2500                 /* more blocks for the new gid ... */
2501                 qi->lqi_id.qid_gid = attr->la_gid;
2502                 qi->lqi_space      = bspace;
2503                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2504                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2505                         rc = 0;
2506                 if (rc)
2507                         RETURN(rc);
2508
2509                 /* and finally less blocks for the current gid */
2510                 qi->lqi_id.qid_gid = gid;
2511                 qi->lqi_space      = -bspace;
2512                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2513                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2514                         rc = 0;
2515                 if (rc)
2516                         RETURN(rc);
2517         }
2518
2519         RETURN(rc);
2520 }
2521
2522 static int osd_inode_setattr(const struct lu_env *env,
2523                              struct inode *inode, const struct lu_attr *attr)
2524 {
2525         __u64 bits = attr->la_valid;
2526
2527         /* Only allow set size for regular file */
2528         if (!S_ISREG(inode->i_mode))
2529                 bits &= ~(LA_SIZE | LA_BLOCKS);
2530
2531         if (bits == 0)
2532                 return 0;
2533
2534         if (bits & LA_ATIME)
2535                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2536         if (bits & LA_CTIME)
2537                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2538         if (bits & LA_MTIME)
2539                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2540         if (bits & LA_SIZE) {
2541                 spin_lock(&inode->i_lock);
2542                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2543                 i_size_write(inode, attr->la_size);
2544                 spin_unlock(&inode->i_lock);
2545         }
2546
2547         /* OSD should not change "i_blocks" which is used by quota.
2548          * "i_blocks" should be changed by ldiskfs only. */
2549         if (bits & LA_MODE)
2550                 inode->i_mode = (inode->i_mode & S_IFMT) |
2551                                 (attr->la_mode & ~S_IFMT);
2552         if (bits & LA_UID)
2553                 i_uid_write(inode, attr->la_uid);
2554         if (bits & LA_GID)
2555                 i_gid_write(inode, attr->la_gid);
2556         if (bits & LA_NLINK)
2557                 set_nlink(inode, attr->la_nlink);
2558         if (bits & LA_RDEV)
2559                 inode->i_rdev = attr->la_rdev;
2560
2561         if (bits & LA_FLAGS) {
2562                 /* always keep S_NOCMTIME */
2563                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2564                                  S_NOCMTIME;
2565         }
2566         return 0;
2567 }
2568
2569 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2570 {
2571         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2572             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2573                 struct iattr    iattr;
2574                 int             rc;
2575
2576                 ll_vfs_dq_init(inode);
2577                 iattr.ia_valid = 0;
2578                 if (attr->la_valid & LA_UID)
2579                         iattr.ia_valid |= ATTR_UID;
2580                 if (attr->la_valid & LA_GID)
2581                         iattr.ia_valid |= ATTR_GID;
2582                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2583                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2584
2585                 rc = ll_vfs_dq_transfer(inode, &iattr);
2586                 if (rc) {
2587                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2588                                "enforcement enabled on the ldiskfs "
2589                                "filesystem?\n", inode->i_sb->s_id, rc);
2590                         return rc;
2591                 }
2592         }
2593         return 0;
2594 }
2595
2596 static int osd_attr_set(const struct lu_env *env,
2597                         struct dt_object *dt,
2598                         const struct lu_attr *attr,
2599                         struct thandle *handle)
2600 {
2601         struct osd_object *obj = osd_dt_obj(dt);
2602         struct inode      *inode;
2603         int rc;
2604
2605         if (!dt_object_exists(dt))
2606                 return -ENOENT;
2607
2608         LASSERT(handle != NULL);
2609         LASSERT(!dt_object_remote(dt));
2610         LASSERT(osd_invariant(obj));
2611
2612         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2613
2614         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2615                 struct osd_thread_info  *oti  = osd_oti_get(env);
2616                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
2617                 struct lu_fid           *fid1 = &oti->oti_fid;
2618                 struct osd_inode_id     *id   = &oti->oti_id;
2619                 struct iam_path_descr   *ipd;
2620                 struct iam_container    *bag;
2621                 struct osd_thandle      *oh;
2622                 int                      rc;
2623
2624                 fid_cpu_to_be(fid1, fid0);
2625                 memset(id, 1, sizeof(*id));
2626                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2627                                   fid0)->oi_dir.od_container;
2628                 ipd = osd_idx_ipd_get(env, bag);
2629                 if (unlikely(ipd == NULL))
2630                         RETURN(-ENOMEM);
2631
2632                 oh = container_of0(handle, struct osd_thandle, ot_super);
2633                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2634                                 (const struct iam_rec *)id, ipd);
2635                 osd_ipd_put(env, bag, ipd);
2636                 return(rc > 0 ? 0 : rc);
2637         }
2638
2639         inode = obj->oo_inode;
2640
2641         rc = osd_quota_transfer(inode, attr);
2642         if (rc)
2643                 return rc;
2644
2645         spin_lock(&obj->oo_guard);
2646         rc = osd_inode_setattr(env, inode, attr);
2647         spin_unlock(&obj->oo_guard);
2648         if (rc != 0)
2649                 GOTO(out, rc);
2650
2651         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2652
2653         if (!(attr->la_valid & LA_FLAGS))
2654                 GOTO(out, rc);
2655
2656         /* Let's check if there are extra flags need to be set into LMA */
2657         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2658                 struct osd_thread_info *info = osd_oti_get(env);
2659                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
2660
2661                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
2662                 if (rc != 0)
2663                         GOTO(out, rc);
2664
2665                 lma->lma_incompat |=
2666                         lustre_to_lma_flags(attr->la_flags);
2667                 lustre_lma_swab(lma);
2668                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2669                                      lma, sizeof(*lma), XATTR_REPLACE);
2670                 if (rc != 0) {
2671                         struct osd_device *osd = osd_obj2dev(obj);
2672
2673                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2674                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2675                               lma->lma_incompat, rc);
2676                 } else {
2677                         obj->oo_lma_flags =
2678                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2679                 }
2680                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2681         }
2682 out:
2683         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2684
2685         return rc;
2686 }
2687
2688 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2689                                            struct osd_object *obj,
2690                                            const char *name, const int namelen)
2691 {
2692         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2693 }
2694
2695 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2696                       umode_t mode, struct dt_allocation_hint *hint,
2697                       struct thandle *th)
2698 {
2699         int result;
2700         struct osd_device  *osd = osd_obj2dev(obj);
2701         struct osd_thandle *oth;
2702         struct dt_object   *parent = NULL;
2703         struct inode       *inode;
2704
2705         LINVRNT(osd_invariant(obj));
2706         LASSERT(obj->oo_inode == NULL);
2707         LASSERT(obj->oo_hl_head == NULL);
2708
2709         if (S_ISDIR(mode) && ldiskfs_pdo) {
2710                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2711                 if (obj->oo_hl_head == NULL)
2712                         return -ENOMEM;
2713         }
2714
2715         oth = container_of(th, struct osd_thandle, ot_super);
2716         LASSERT(oth->ot_handle->h_transaction != NULL);
2717
2718         if (hint != NULL && hint->dah_parent != NULL &&
2719             !dt_object_remote(hint->dah_parent))
2720                 parent = hint->dah_parent;
2721
2722         inode = ldiskfs_create_inode(oth->ot_handle,
2723                                      parent ? osd_dt_obj(parent)->oo_inode :
2724                                               osd_sb(osd)->s_root->d_inode,
2725                                      mode);
2726         if (!IS_ERR(inode)) {
2727                 /* Do not update file c/mtime in ldiskfs. */
2728                 inode->i_flags |= S_NOCMTIME;
2729
2730                 /* For new created object, it must be consistent,
2731                  * and it is unnecessary to scrub against it. */
2732                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2733
2734                 obj->oo_inode = inode;
2735                 result = 0;
2736         } else {
2737                 if (obj->oo_hl_head != NULL) {
2738                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2739                         obj->oo_hl_head = NULL;
2740                 }
2741                 result = PTR_ERR(inode);
2742         }
2743         LINVRNT(osd_invariant(obj));
2744         return result;
2745 }
2746
2747 enum {
2748         OSD_NAME_LEN = 255
2749 };
2750
2751 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2752                      struct lu_attr *attr,
2753                      struct dt_allocation_hint *hint,
2754                      struct dt_object_format *dof,
2755                      struct thandle *th)
2756 {
2757         int result;
2758         struct osd_thandle *oth;
2759         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2760
2761         LASSERT(S_ISDIR(attr->la_mode));
2762
2763         oth = container_of(th, struct osd_thandle, ot_super);
2764         LASSERT(oth->ot_handle->h_transaction != NULL);
2765         result = osd_mkfile(info, obj, mode, hint, th);
2766
2767         return result;
2768 }
2769
2770 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2771                         struct lu_attr *attr,
2772                         struct dt_allocation_hint *hint,
2773                         struct dt_object_format *dof,
2774                         struct thandle *th)
2775 {
2776         int result;
2777         struct osd_thandle *oth;
2778         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2779
2780         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2781
2782         LASSERT(S_ISREG(attr->la_mode));
2783
2784         oth = container_of(th, struct osd_thandle, ot_super);
2785         LASSERT(oth->ot_handle->h_transaction != NULL);
2786
2787         result = osd_mkfile(info, obj, mode, hint, th);
2788         if (result == 0) {
2789                 LASSERT(obj->oo_inode != NULL);
2790                 if (feat->dif_flags & DT_IND_VARKEY)
2791                         result = iam_lvar_create(obj->oo_inode,
2792                                                  feat->dif_keysize_max,
2793                                                  feat->dif_ptrsize,
2794                                                  feat->dif_recsize_max,
2795                                                  oth->ot_handle);
2796                 else
2797                         result = iam_lfix_create(obj->oo_inode,
2798                                                  feat->dif_keysize_max,
2799                                                  feat->dif_ptrsize,
2800                                                  feat->dif_recsize_max,
2801                                                  oth->ot_handle);
2802
2803         }
2804         return result;
2805 }
2806
2807 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2808                      struct lu_attr *attr,
2809                      struct dt_allocation_hint *hint,
2810                      struct dt_object_format *dof,
2811                      struct thandle *th)
2812 {
2813         LASSERT(S_ISREG(attr->la_mode));
2814         return osd_mkfile(info, obj, (attr->la_mode &
2815                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2816 }
2817
2818 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2819                      struct lu_attr *attr,
2820                      struct dt_allocation_hint *hint,
2821                      struct dt_object_format *dof,
2822                      struct thandle *th)
2823 {
2824         LASSERT(S_ISLNK(attr->la_mode));
2825         return osd_mkfile(info, obj, (attr->la_mode &
2826                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2827 }
2828
2829 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2830                      struct lu_attr *attr,
2831                      struct dt_allocation_hint *hint,
2832                      struct dt_object_format *dof,
2833                      struct thandle *th)
2834 {
2835         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2836         int result;
2837
2838         LINVRNT(osd_invariant(obj));
2839         LASSERT(obj->oo_inode == NULL);
2840         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2841                 S_ISFIFO(mode) || S_ISSOCK(mode));
2842
2843         result = osd_mkfile(info, obj, mode, hint, th);
2844         if (result == 0) {
2845                 LASSERT(obj->oo_inode != NULL);
2846                 /*
2847                  * This inode should be marked dirty for i_rdev.  Currently
2848                  * that is done in the osd_attr_init().
2849                  */
2850                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2851                                    attr->la_rdev);
2852         }
2853         LINVRNT(osd_invariant(obj));
2854         return result;
2855 }
2856
2857 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2858                               struct lu_attr *,
2859                               struct dt_allocation_hint *hint,
2860                               struct dt_object_format *dof,
2861                               struct thandle *);
2862
2863 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2864 {
2865         osd_obj_type_f result;
2866
2867         switch (type) {
2868         case DFT_DIR:
2869                 result = osd_mkdir;
2870                 break;
2871         case DFT_REGULAR:
2872                 result = osd_mkreg;
2873                 break;
2874         case DFT_SYM:
2875                 result = osd_mksym;
2876                 break;
2877         case DFT_NODE:
2878                 result = osd_mknod;
2879                 break;
2880         case DFT_INDEX:
2881                 result = osd_mk_index;
2882                 break;
2883
2884         default:
2885                 LBUG();
2886                 break;
2887         }
2888         return result;
2889 }
2890
2891
2892 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2893                         struct dt_object *parent, struct dt_object *child,
2894                         umode_t child_mode)
2895 {
2896         LASSERT(ah);
2897
2898         ah->dah_parent = parent;
2899         ah->dah_mode = child_mode;
2900
2901         if (parent != NULL && !dt_object_remote(parent)) {
2902                 /* will help to find FID->ino at dt_insert("..") */
2903                 struct osd_object *pobj = osd_dt_obj(parent);
2904                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2905         }
2906 }
2907
2908 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2909                           struct lu_attr *attr, struct dt_object_format *dof)
2910 {
2911         struct inode   *inode = obj->oo_inode;
2912         __u64           valid = attr->la_valid;
2913         int             result;
2914
2915         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2916
2917         if (dof->dof_type != DFT_NODE)
2918                 attr->la_valid &= ~LA_RDEV;
2919         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2920                 attr->la_valid &= ~LA_ATIME;
2921         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2922                 attr->la_valid &= ~LA_CTIME;
2923         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2924                 attr->la_valid &= ~LA_MTIME;
2925
2926         result = osd_quota_transfer(inode, attr);
2927         if (result)
2928                 return;
2929
2930         if (attr->la_valid != 0) {
2931                 result = osd_inode_setattr(info->oti_env, inode, attr);
2932                 /*
2933                  * The osd_inode_setattr() should always succeed here.  The
2934                  * only error that could be returned is EDQUOT when we are
2935                  * trying to change the UID or GID of the inode. However, this
2936                  * should not happen since quota enforcement is no longer
2937                  * enabled on ldiskfs (lquota takes care of it).
2938                  */
2939                 LASSERTF(result == 0, "%d\n", result);
2940                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2941         }
2942
2943         attr->la_valid = valid;
2944 }
2945
2946 /**
2947  * Helper function for osd_object_create()
2948  *
2949  * \retval 0, on success
2950  */
2951 static int __osd_object_create(struct osd_thread_info *info,
2952                                struct osd_object *obj, struct lu_attr *attr,
2953                                struct dt_allocation_hint *hint,
2954                                struct dt_object_format *dof,
2955                                struct thandle *th)
2956 {
2957         int     result;
2958         __u32   umask;
2959
2960         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
2961
2962         /* we drop umask so that permissions we pass are not affected */
2963         umask = current->fs->umask;
2964         current->fs->umask = 0;
2965
2966         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
2967                                                   th);
2968         if (likely(obj->oo_inode != NULL)) {
2969                 LASSERT(obj->oo_inode->i_state & I_NEW);
2970
2971                 /* Unlock the inode before attr initialization to avoid
2972                  * unnecessary dqget operations. LU-6378 */
2973                 unlock_new_inode(obj->oo_inode);
2974         }
2975
2976         if (likely(result == 0)) {
2977                 osd_attr_init(info, obj, attr, dof);
2978                 osd_object_init0(obj);
2979         }
2980
2981         /* restore previous umask value */
2982         current->fs->umask = umask;
2983
2984         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
2985
2986         return result;
2987 }
2988
2989 /**
2990  * Helper function for osd_object_create()
2991  *
2992  * \retval 0, on success
2993  */
2994 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
2995                            const struct lu_fid *fid, struct thandle *th)
2996 {
2997         struct osd_thread_info *info = osd_oti_get(env);
2998         struct osd_inode_id    *id   = &info->oti_id;
2999         struct osd_device      *osd  = osd_obj2dev(obj);
3000         struct osd_thandle     *oh;
3001         int                     rc;
3002
3003         LASSERT(obj->oo_inode != NULL);
3004
3005         oh = container_of0(th, struct osd_thandle, ot_super);
3006         LASSERT(oh->ot_handle);
3007         osd_trans_exec_op(env, th, OSD_OT_INSERT);
3008
3009         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3010         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3011                            OI_CHECK_FLD, NULL);
3012         osd_trans_exec_check(env, th, OSD_OT_INSERT);
3013
3014         return rc;
3015 }
3016
3017 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3018                    u64 seq, struct lu_seq_range *range)
3019 {
3020         struct seq_server_site  *ss = osd_seq_site(osd);
3021
3022         if (fid_seq_is_idif(seq)) {
3023                 fld_range_set_ost(range);
3024                 range->lsr_index = idif_ost_idx(seq);
3025                 return 0;
3026         }
3027
3028         if (!fid_seq_in_fldb(seq)) {
3029                 fld_range_set_mdt(range);
3030                 if (ss != NULL)
3031                         /* FIXME: If ss is NULL, it suppose not get lsr_index
3032                          * at all */
3033                         range->lsr_index = ss->ss_node_id;
3034                 return 0;
3035         }
3036
3037         LASSERT(ss != NULL);
3038         fld_range_set_any(range);
3039         /* OSD will only do local fld lookup */
3040         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3041 }
3042
3043 /*
3044  * Concurrency: no external locking is necessary.
3045  */
3046 static int osd_declare_object_create(const struct lu_env *env,
3047                                      struct dt_object *dt,
3048                                      struct lu_attr *attr,
3049                                      struct dt_allocation_hint *hint,
3050                                      struct dt_object_format *dof,
3051                                      struct thandle *handle)
3052 {
3053         struct osd_thandle      *oh;
3054         int                      rc;
3055         ENTRY;
3056
3057         LASSERT(handle != NULL);
3058
3059         oh = container_of0(handle, struct osd_thandle, ot_super);
3060         LASSERT(oh->ot_handle == NULL);
3061
3062         /* EA object consumes more credits than regular object: osd_mk_index
3063          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3064          * leaf_node, could involves the block, block bitmap, groups, GDT
3065          * change for each block, so add 4 * 2 credits in that case. */
3066         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3067                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3068                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3069         /* Reuse idle OI block may cause additional one OI block
3070          * to be changed. */
3071         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3072                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3073
3074         if (!attr)
3075                 RETURN(0);
3076
3077         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid, 1, oh,
3078                                    osd_dt_obj(dt), false, NULL, false);
3079         if (rc != 0)
3080                 RETURN(rc);
3081
3082         /* will help to find FID->ino mapping at dt_insert() */
3083         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3084                                    osd_dt_obj(dt));
3085
3086         RETURN(rc);
3087 }
3088
3089 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
3090                              struct lu_attr *attr,
3091                              struct dt_allocation_hint *hint,
3092                              struct dt_object_format *dof, struct thandle *th)
3093 {
3094         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3095         struct osd_object       *obj    = osd_dt_obj(dt);
3096         struct osd_thread_info  *info   = osd_oti_get(env);
3097         int result;
3098         ENTRY;
3099
3100         if (dt_object_exists(dt))
3101                 return -EEXIST;
3102
3103         LINVRNT(osd_invariant(obj));
3104         LASSERT(!dt_object_remote(dt));
3105         LASSERT(osd_write_locked(env, obj));
3106         LASSERT(th != NULL);
3107
3108         if (unlikely(fid_is_acct(fid)))
3109                 /* Quota files can't be created from the kernel any more,
3110                  * 'tune2fs -O quota' will take care of creating them */
3111                 RETURN(-EPERM);
3112
3113         result = __osd_object_create(info, obj, attr, hint, dof, th);
3114         if (result == 0) {
3115                 result = __osd_oi_insert(env, obj, fid, th);
3116                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3117                         obj->oo_dt.do_body_ops = &osd_body_ops;
3118         }
3119         LASSERT(ergo(result == 0,
3120                 dt_object_exists(dt) && !dt_object_remote(dt)));
3121
3122         LASSERT(osd_invariant(obj));
3123         RETURN(result);
3124 }
3125
3126 /**
3127  * Called to destroy on-disk representation of the object
3128  *
3129  * Concurrency: must be locked
3130  */
3131 static int osd_declare_object_destroy(const struct lu_env *env,
3132                                       struct dt_object *dt,
3133                                       struct thandle *th)
3134 {
3135         struct osd_object  *obj = osd_dt_obj(dt);
3136         struct inode       *inode = obj->oo_inode;
3137         struct osd_thandle *oh;
3138         int                 rc;
3139         ENTRY;
3140
3141         if (inode == NULL)
3142                 RETURN(-ENOENT);
3143
3144         oh = container_of0(th, struct osd_thandle, ot_super);
3145         LASSERT(oh->ot_handle == NULL);
3146
3147         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3148                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3149         /* Recycle idle OI leaf may cause additional three OI blocks
3150          * to be changed. */
3151         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3152                 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3153                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3154         /* one less inode */
3155         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3156                                    -1, oh, obj, false, NULL, false);
3157         if (rc)
3158                 RETURN(rc);
3159         /* data to be truncated */
3160         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3161                                    0, oh, obj, true, NULL, false);
3162         if (rc)
3163                 RETURN(rc);
3164
3165         /* will help to find FID->ino when this object is being
3166          * added to PENDING/ */
3167         rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3168
3169         RETURN(rc);
3170 }
3171
3172 static int osd_object_destroy(const struct lu_env *env,
3173                               struct dt_object *dt,
3174                               struct thandle *th)
3175 {
3176         const struct lu_fid    *fid = lu_object_fid(&dt->do_lu);
3177         struct osd_object      *obj = osd_dt_obj(dt);
3178         struct inode           *inode = obj->oo_inode;
3179         struct osd_device      *osd = osd_obj2dev(obj);
3180         struct osd_thandle     *oh;
3181         int                     result;
3182         ENTRY;
3183
3184         oh = container_of0(th, struct osd_thandle, ot_super);
3185         LASSERT(oh->ot_handle);
3186         LASSERT(inode);
3187         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3188
3189         if (unlikely(fid_is_acct(fid)))
3190                 RETURN(-EPERM);
3191
3192         if (S_ISDIR(inode->i_mode)) {
3193                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3194                         inode->i_nlink == 2);
3195                 /* it will check/delete the inode from remote parent,
3196                  * how to optimize it? unlink performance impaction XXX */
3197                 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3198                 if (result != 0)
3199                         CERROR("%s: delete inode "DFID": rc = %d\n",
3200                                osd_name(osd), PFID(fid), result);
3201
3202                 spin_lock(&obj->oo_guard);
3203                 clear_nlink(inode);
3204                 spin_unlock(&obj->oo_guard);
3205                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3206         }
3207
3208         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3209
3210         ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3211
3212         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3213                 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3214                                        oh->ot_handle, OI_CHECK_FLD);
3215
3216         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3217         /* XXX: add to ext3 orphan list */
3218         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3219
3220         /* not needed in the cache anymore */
3221         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3222         obj->oo_destroyed = 1;
3223
3224         RETURN(0);
3225 }
3226
3227 /**
3228  * Put the fid into lustre_mdt_attrs, and then place the structure
3229  * inode's ea. This fid should not be altered during the life time
3230  * of the inode.
3231  *
3232  * \retval +ve, on success
3233  * \retval -ve, on error
3234  *
3235  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3236  */
3237 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3238                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
3239 {
3240         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
3241         int                      rc;
3242         ENTRY;
3243
3244         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3245                 RETURN(0);
3246
3247         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3248                 rc = -ENOMEM;
3249
3250         lustre_lma_init(lma, fid, compat, incompat);
3251         lustre_lma_swab(lma);
3252
3253         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma, sizeof(*lma),
3254                              XATTR_CREATE);
3255         /* LMA may already exist, but we need to check that all the
3256          * desired compat/incompat flags have been added. */
3257         if (unlikely(rc == -EEXIST)) {
3258                 if (compat == 0 && incompat == 0)
3259                         RETURN(0);
3260
3261                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3262                                      XATTR_NAME_LMA, info->oti_mdt_attrs_old,
3263                                      LMA_OLD_SIZE);
3264                 if (rc <= 0)
3265                         RETURN(-EINVAL);
3266
3267                 lustre_lma_swab(lma);
3268                 if (!(~lma->lma_compat & compat) &&
3269                     !(~lma->lma_incompat & incompat))
3270                         RETURN(0);
3271
3272                 lma->lma_compat |= compat;
3273                 lma->lma_incompat |= incompat;
3274                 lustre_lma_swab(lma);
3275                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3276                                      sizeof(*lma), XATTR_REPLACE);
3277         }
3278
3279         RETURN(rc);
3280 }
3281
3282 /**
3283  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3284  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3285  * To have compatilibility with 1.8 ldiskfs driver we need to have
3286  * magic number at start of fid data.
3287  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3288  * its inmemory API.
3289  */
3290 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3291                                          const struct lu_fid *fid)
3292 {
3293         if (!fid_is_namespace_visible(fid) ||
3294             OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3295                 param->edp_magic = 0;
3296                 return;
3297         }
3298
3299         param->edp_magic = LDISKFS_LUFID_MAGIC;
3300         param->edp_len =  sizeof(struct lu_fid) + 1;
3301         fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3302 }
3303
3304 /**
3305  * Try to read the fid from inode ea into dt_rec.
3306  *
3307  * \param fid object fid.
3308  *
3309  * \retval 0 on success
3310  */
3311 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3312                           __u32 ino, struct lu_fid *fid,
3313                           struct osd_inode_id *id)
3314 {
3315         struct osd_thread_info *info  = osd_oti_get(env);
3316         struct inode           *inode;
3317         ENTRY;
3318
3319         osd_id_gen(id, ino, OSD_OII_NOGEN);
3320         inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3321         if (IS_ERR(inode))
3322                 RETURN(PTR_ERR(inode));
3323
3324         iput(inode);
3325         RETURN(0);
3326 }
3327
3328 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3329                                         struct inode *dir,
3330                                         struct inode *parent_dir,
3331                                         const struct lu_fid *dot_fid,
3332                                         const struct lu_fid *dot_dot_fid,
3333                                         struct osd_thandle *oth)
3334 {
3335         struct ldiskfs_dentry_param *dot_ldp;
3336         struct ldiskfs_dentry_param *dot_dot_ldp;
3337         __u32 saved_nlink = dir->i_nlink;
3338         int rc;
3339
3340         dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3341         osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3342
3343         dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3344         dot_ldp->edp_magic = 0;
3345
3346         rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3347                                     dir, dot_ldp, dot_dot_ldp);
3348         /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3349          * the subseqent ref_add() will increase the dir->i_nlink
3350          * as 3. That is incorrect for new created directory.
3351          *
3352          * It looks like hack, because we want to make the OSD API
3353          * to be order-independent for new created directory object
3354          * between dt_insert(..) and ref_add() operations.
3355          *
3356          * Here, we only restore the in-RAM dir-inode's nlink attr,
3357          * becuase if the nlink attr is not 2, then there will be
3358          * ref_add() called following the dt_insert(..), such call
3359          * will make both the in-RAM and on-disk dir-inode's nlink
3360          * attr to be set as 2. LU-7447 */
3361         set_nlink(dir, saved_nlink);
3362         return rc;
3363 }
3364
3365 /**
3366  * Create an local agent inode for remote entry
3367  */
3368 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3369                                                   struct osd_device *osd,
3370                                                   struct osd_object *pobj,
3371                                                   const struct lu_fid *fid,
3372                                                   __u32 type,
3373                                                   struct thandle *th)
3374 {
3375         struct osd_thread_info  *info = osd_oti_get(env);
3376         struct inode            *local;
3377         struct osd_thandle      *oh;
3378         int                     rc;
3379         ENTRY;
3380
3381         LASSERT(th);
3382         oh = container_of(th, struct osd_thandle, ot_super);
3383         LASSERT(oh->ot_handle->h_transaction != NULL);
3384
3385         local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3386         if (IS_ERR(local)) {
3387                 CERROR("%s: create local error %d\n", osd_name(osd),
3388                        (int)PTR_ERR(local));
3389                 RETURN(local);
3390         }
3391
3392         /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3393          * correct gid on remote file, not agent here */
3394         local->i_gid = current_fsgid();
3395         ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3396         unlock_new_inode(local);
3397
3398         /* Set special LMA flag for local agent inode */
3399         rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3400         if (rc != 0) {
3401                 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3402                        osd_name(osd), PFID(fid), rc);
3403                 RETURN(ERR_PTR(rc));
3404         }
3405
3406         if (!S_ISDIR(type))
3407                 RETURN(local);
3408
3409         rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3410                                          lu_object_fid(&pobj->oo_dt.do_lu),
3411                                          fid, oh);
3412         if (rc != 0) {
3413                 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3414                         osd_name(osd), PFID(fid), rc);
3415                 RETURN(ERR_PTR(rc));
3416         }
3417
3418         RETURN(local);
3419 }
3420
3421 /**
3422  * when direntry is deleted, we have to take care of possible agent inode
3423  * referenced by that. unfortunately we can't do this at that point:
3424  * iget() within a running transaction leads to deadlock and we better do
3425  * not call that every delete declaration to save performance. so we put
3426  * a potention agent inode on a list and process that once the transaction
3427  * is over. Notice it's not any worse in terms of real orphans as regular
3428  * object destroy doesn't put inodes on the on-disk orphan list. this should
3429  * be addressed separately
3430  */
3431 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3432                                             struct osd_thandle *oh,
3433                                             __u32 ino)
3434 {
3435         struct osd_device      *osd = osd_dt_dev(oh->ot_super.th_dev);
3436         struct osd_obj_orphan *oor;
3437
3438         OBD_ALLOC_PTR(oor);
3439         if (oor == NULL)
3440                 return -ENOMEM;
3441
3442         oor->oor_ino = ino;
3443         oor->oor_env = (struct lu_env *)env;
3444         spin_lock(&osd->od_osfs_lock);
3445         list_add(&oor->oor_list, &osd->od_orphan_list);
3446         spin_unlock(&osd->od_osfs_lock);
3447
3448         oh->ot_remove_agents = 1;
3449
3450         return 0;
3451
3452 }
3453
3454 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3455                                                 struct osd_device *osd)
3456 {
3457         struct osd_thread_info *info = osd_oti_get(env);
3458         struct osd_obj_orphan *oor, *tmp;
3459         struct osd_inode_id id;
3460         struct list_head list;
3461         struct inode *inode;
3462         struct lu_fid fid;
3463         handle_t *jh;
3464         __u32 ino;
3465
3466         INIT_LIST_HEAD(&list);
3467
3468         spin_lock(&osd->od_osfs_lock);
3469         list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3470                 if (oor->oor_env == env) {
3471                         list_del(&oor->oor_list);
3472                         list_add(&oor->oor_list, &list);
3473                 }
3474         }
3475         spin_unlock(&osd->od_osfs_lock);
3476
3477         list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3478
3479                 ino = oor->oor_ino;
3480
3481                 list_del(&oor->oor_list);
3482                 OBD_FREE_PTR(oor);
3483
3484                 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3485                 inode = osd_iget_fid(info, osd, &id, &fid);
3486                 if (IS_ERR(inode))
3487                         continue;
3488
3489                 if (!osd_remote_fid(env, osd, &fid)) {
3490                         iput(inode);
3491                         continue;
3492                 }
3493
3494                 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3495                 clear_nlink(inode);
3496                 mark_inode_dirty(inode);
3497                 ldiskfs_journal_stop(jh);
3498                 iput(inode);
3499         }
3500
3501         return 0;
3502 }
3503
3504 /**
3505  * OSD layer object create function for interoperability mode (b11826).
3506  * This is mostly similar to osd_object_create(). Only difference being, fid is
3507  * inserted into inode ea here.
3508  *
3509  * \retval   0, on success
3510  * \retval -ve, on error
3511  */
3512 static int osd_object_ea_create(const struct lu_env *env, struct dt_object *dt,
3513                                 struct lu_attr *attr,
3514                                 struct dt_allocation_hint *hint,
3515                                 struct dt_object_format *dof,
3516                                 struct thandle *th)
3517 {
3518         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3519         struct osd_object       *obj    = osd_dt_obj(dt);
3520         struct osd_thread_info  *info   = osd_oti_get(env);
3521         int                      result, on_ost = 0;
3522
3523         ENTRY;
3524
3525         if (dt_object_exists(dt))
3526                 RETURN(-EEXIST);
3527
3528         LASSERT(osd_invariant(obj));
3529         LASSERT(!dt_object_remote(dt));
3530         LASSERT(osd_write_locked(env, obj));
3531         LASSERT(th != NULL);
3532
3533         if (unlikely(fid_is_acct(fid)))
3534                 /* Quota files can't be created from the kernel any more,
3535                  * 'tune2fs -O quota' will take care of creating them */
3536                 RETURN(-EPERM);
3537
3538         result = __osd_object_create(info, obj, attr, hint, dof, th);
3539         if (result == 0) {
3540                 if (fid_is_idif(fid) &&
3541                     !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3542                         struct lu_fid *tfid = &info->oti_fid;
3543                         struct ost_id *oi   = &info->oti_ostid;
3544
3545                         fid_to_ostid(fid, oi);
3546                         ostid_to_fid(tfid, oi, 0);
3547                         on_ost = 1;
3548                         result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3549                                                 LMAC_FID_ON_OST, 0);
3550                 } else {
3551                         on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3552                                                fid, OI_CHECK_FLD);
3553                         result = osd_ea_fid_set(info, obj->oo_inode, fid,
3554                                                 on_ost ? LMAC_FID_ON_OST : 0,
3555                                                 0);
3556                 }
3557                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3558                         obj->oo_dt.do_body_ops = &osd_body_ops;
3559         }
3560
3561         if (result == 0)
3562                 result = __osd_oi_insert(env, obj, fid, th);
3563
3564         /* a small optimization - dt_insert() isn't usually applied
3565          * to OST objects, so we don't need to cache OI mapping for
3566          * OST objects */
3567         if (result == 0 && on_ost == 0) {
3568                 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3569                 result = osd_idc_find_and_init(env, osd, obj);
3570                 LASSERT(result == 0);
3571         }
3572
3573         LASSERT(ergo(result == 0,
3574                      dt_object_exists(dt) && !dt_object_remote(dt)));
3575         LINVRNT(osd_invariant(obj));
3576         RETURN(result);
3577 }
3578
3579 static int osd_declare_object_ref_add(const struct lu_env *env,
3580                                       struct dt_object *dt,
3581                                       struct thandle *handle)
3582 {
3583         struct osd_thandle       *oh;
3584
3585         /* it's possible that object doesn't exist yet */
3586         LASSERT(handle != NULL);
3587
3588         oh = container_of0(handle, struct osd_thandle, ot_super);
3589         LASSERT(oh->ot_handle == NULL);
3590
3591         osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3592                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3593
3594         return 0;
3595 }
3596
3597 /*
3598  * Concurrency: @dt is write locked.
3599  */
3600 static int osd_object_ref_add(const struct lu_env *env,
3601                               struct dt_object *dt, struct thandle *th)
3602 {
3603         struct osd_object  *obj = osd_dt_obj(dt);
3604         struct inode       *inode = obj->oo_inode;
3605         struct osd_thandle *oh;
3606         int                 rc = 0;
3607
3608         if (!dt_object_exists(dt) || obj->oo_destroyed)
3609                 return -ENOENT;
3610
3611         LINVRNT(osd_invariant(obj));
3612         LASSERT(!dt_object_remote(dt));
3613         LASSERT(osd_write_locked(env, obj));
3614         LASSERT(th != NULL);
3615
3616         oh = container_of0(th, struct osd_thandle, ot_super);
3617         LASSERT(oh->ot_handle != NULL);
3618
3619         osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3620
3621         CDEBUG(D_INODE, DFID" increase nlink %d\n",
3622                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3623         /*
3624          * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3625          * (65000) subdirectories by storing "1" in i_nlink if the link count
3626          * would otherwise overflow. Directory tranversal tools understand
3627          * that (st_nlink == 1) indicates that the filesystem dose not track
3628          * hard links count on the directory, and will not abort subdirectory
3629          * scanning early once (st_nlink - 2) subdirs have been found.
3630          *
3631          * This also has to properly handle the case of inodes with nlink == 0
3632          * in case they are being linked into the PENDING directory
3633          */
3634         spin_lock(&obj->oo_guard);
3635         if (unlikely(inode->i_nlink == 0))
3636            &nbs