Whamcloud - gitweb
b49f9b13f3343991ca6e5da3982977c7b90a4f89
[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 <uapi/linux/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 #define PFID_STRIPE_IDX_BITS    16
82 #define PFID_STRIPE_COUNT_MASK  ((1 << PFID_STRIPE_IDX_BITS) - 1)
83
84 int ldiskfs_pdo = 1;
85 module_param(ldiskfs_pdo, int, 0644);
86 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
87
88 int ldiskfs_track_declares_assert;
89 module_param(ldiskfs_track_declares_assert, int, 0644);
90 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
91
92 /* Slab to allocate dynlocks */
93 struct kmem_cache *dynlock_cachep;
94
95 /* Slab to allocate osd_it_ea */
96 struct kmem_cache *osd_itea_cachep;
97
98 static struct lu_kmem_descr ldiskfs_caches[] = {
99         {
100                 .ckd_cache = &dynlock_cachep,
101                 .ckd_name  = "dynlock_cache",
102                 .ckd_size  = sizeof(struct dynlock_handle)
103         },
104         {
105                 .ckd_cache = &osd_itea_cachep,
106                 .ckd_name  = "osd_itea_cache",
107                 .ckd_size  = sizeof(struct osd_it_ea)
108         },
109         {
110                 .ckd_cache = NULL
111         }
112 };
113
114 static const char dot[] = ".";
115 static const char dotdot[] = "..";
116
117 static const struct lu_object_operations      osd_lu_obj_ops;
118 static const struct dt_object_operations      osd_obj_ops;
119 static const struct dt_object_operations      osd_obj_otable_it_ops;
120 static const struct dt_index_operations       osd_index_iam_ops;
121 static const struct dt_index_operations       osd_index_ea_ops;
122
123 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
124                           const struct lu_fid *fid);
125 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
126                                                 struct osd_device *osd);
127
128 int osd_trans_declare_op2rb[] = {
129         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
130         [OSD_OT_PUNCH]          = OSD_OT_MAX,
131         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
132         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
133         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
134         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
135         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
136         [OSD_OT_WRITE]          = OSD_OT_WRITE,
137         [OSD_OT_INSERT]         = OSD_OT_DELETE,
138         [OSD_OT_DELETE]         = OSD_OT_INSERT,
139         [OSD_OT_QUOTA]          = OSD_OT_MAX,
140 };
141
142 static int osd_has_index(const struct osd_object *obj)
143 {
144         return obj->oo_dt.do_index_ops != NULL;
145 }
146
147 static int osd_object_invariant(const struct lu_object *l)
148 {
149         return osd_invariant(osd_obj(l));
150 }
151
152 /*
153  * Concurrency: doesn't matter
154  */
155 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
156 {
157         struct osd_thread_info *oti = osd_oti_get(env);
158         return oti->oti_w_locks > 0 && o->oo_owner == env;
159 }
160
161 /*
162  * Concurrency: doesn't access mutable data
163  */
164 static int osd_root_get(const struct lu_env *env,
165                         struct dt_device *dev, struct lu_fid *f)
166 {
167         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
168         return 0;
169 }
170
171 /*
172  * the following set of functions are used to maintain per-thread
173  * cache of FID->ino mapping. this mechanism is needed to resolve
174  * FID to inode at dt_insert() which in turn stores ino in the
175  * directory entries to keep ldiskfs compatible with ext[34].
176  * due to locking-originated restrictions we can't lookup ino
177  * using LU cache (deadlock is possible). lookup using OI is quite
178  * expensive. so instead we maintain this cache and methods like
179  * dt_create() fill it. so in the majority of cases dt_insert() is
180  * able to find needed mapping in lockless manner.
181  */
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184              const struct lu_fid *fid)
185 {
186         struct osd_thread_info  *oti   = osd_oti_get(env);
187         struct osd_idmap_cache  *idc    = oti->oti_ins_cache;
188         int i;
189         for (i = 0; i < oti->oti_ins_cache_used; i++) {
190                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
191                         continue;
192                 if (idc[i].oic_dev != osd)
193                         continue;
194
195                 return idc + i;
196         }
197
198         return NULL;
199 }
200
201 static struct osd_idmap_cache *
202 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
203             const struct lu_fid *fid)
204 {
205         struct osd_thread_info  *oti   = osd_oti_get(env);
206         struct osd_idmap_cache  *idc;
207         int i;
208
209         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
210                 i = oti->oti_ins_cache_size * 2;
211                 if (i == 0)
212                         i = OSD_INS_CACHE_SIZE;
213                 OBD_ALLOC(idc, sizeof(*idc) * i);
214                 if (idc == NULL)
215                         return ERR_PTR(-ENOMEM);
216                 if (oti->oti_ins_cache != NULL) {
217                         memcpy(idc, oti->oti_ins_cache,
218                                oti->oti_ins_cache_used * sizeof(*idc));
219                         OBD_FREE(oti->oti_ins_cache,
220                                  oti->oti_ins_cache_used * sizeof(*idc));
221                 }
222                 oti->oti_ins_cache = idc;
223                 oti->oti_ins_cache_size = i;
224         }
225
226         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
227         idc->oic_fid = *fid;
228         idc->oic_dev = osd;
229         idc->oic_lid.oii_ino = 0;
230         idc->oic_lid.oii_gen = 0;
231         idc->oic_remote = 0;
232
233         return idc;
234 }
235
236 /*
237  * lookup mapping for the given fid in the cache, initialize a
238  * new one if not found. the initialization checks whether the
239  * object is local or remote. for local objects, OI is used to
240  * learn ino/generation. the function is used when the caller
241  * has no information about the object, e.g. at dt_insert().
242  */
243 static struct osd_idmap_cache *
244 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
245                      const struct lu_fid *fid)
246 {
247         struct osd_idmap_cache *idc;
248         int rc;
249
250         idc = osd_idc_find(env, osd, fid);
251         LASSERT(!IS_ERR(idc));
252         if (idc != NULL)
253                 return idc;
254
255         /* new mapping is needed */
256         idc = osd_idc_add(env, osd, fid);
257         if (IS_ERR(idc))
258                 return idc;
259
260         /* initialize it */
261         rc = osd_remote_fid(env, osd, fid);
262         if (unlikely(rc < 0))
263                 return ERR_PTR(rc);
264
265         if (rc == 0) {
266                 /* the object is local, lookup in OI */
267                 /* XXX: probably cheaper to lookup in LU first? */
268                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
269                                    &idc->oic_lid, 0);
270                 if (unlikely(rc < 0)) {
271                         CERROR("can't lookup: rc = %d\n", rc);
272                         return ERR_PTR(rc);
273                 }
274         } else {
275                 /* the object is remote */
276                 idc->oic_remote = 1;
277         }
278
279         return idc;
280 }
281
282 /*
283  * lookup mapping for given FID and fill it from the given object.
284  * the object is lolcal by definition.
285  */
286 static int osd_idc_find_and_init(const struct lu_env *env,
287                                  struct osd_device *osd,
288                                  struct osd_object *obj)
289 {
290         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
291         struct osd_idmap_cache  *idc;
292
293         idc = osd_idc_find(env, osd, fid);
294         LASSERT(!IS_ERR(idc));
295         if (idc != NULL) {
296                 if (obj->oo_inode == NULL)
297                         return 0;
298                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
299                         LASSERT(idc->oic_lid.oii_ino == 0);
300                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
301                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
302                 }
303                 return 0;
304         }
305
306         /* new mapping is needed */
307         idc = osd_idc_add(env, osd, fid);
308         if (IS_ERR(idc))
309                 return PTR_ERR(idc);
310
311         if (obj->oo_inode != NULL) {
312                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
313                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
314         }
315         return 0;
316 }
317
318 /*
319  * OSD object methods.
320  */
321
322 /*
323  * Concurrency: no concurrent access is possible that early in object
324  * life-cycle.
325  */
326 static struct lu_object *osd_object_alloc(const struct lu_env *env,
327                                           const struct lu_object_header *hdr,
328                                           struct lu_device *d)
329 {
330         struct osd_object *mo;
331
332         OBD_ALLOC_PTR(mo);
333         if (mo != NULL) {
334                 struct lu_object *l;
335
336                 l = &mo->oo_dt.do_lu;
337                 dt_object_init(&mo->oo_dt, NULL, d);
338                 mo->oo_dt.do_ops = &osd_obj_ops;
339                 l->lo_ops = &osd_lu_obj_ops;
340                 init_rwsem(&mo->oo_sem);
341                 init_rwsem(&mo->oo_ext_idx_sem);
342                 spin_lock_init(&mo->oo_guard);
343                 INIT_LIST_HEAD(&mo->oo_xattr_list);
344                 return l;
345         } else {
346                 return NULL;
347         }
348 }
349
350 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
351                 struct dentry *dentry, struct lustre_ost_attrs *loa)
352 {
353         int rc;
354
355         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
356                              (void *)loa, sizeof(*loa));
357         if (rc > 0) {
358                 struct lustre_mdt_attrs *lma = &loa->loa_lma;
359
360                 if (rc < sizeof(*lma))
361                         return -EINVAL;
362
363                 rc = 0;
364                 lustre_loa_swab(loa, true);
365                 /* Check LMA compatibility */
366                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
367                         CWARN("%s: unsupported incompat LMA feature(s) %#x "
368                               "for fid = "DFID", ino = %lu\n",
369                               osd_ino2name(inode),
370                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
371                               PFID(&lma->lma_self_fid), inode->i_ino);
372                         rc = -EOPNOTSUPP;
373                 }
374         } else if (rc == 0) {
375                 rc = -ENODATA;
376         }
377
378         return rc;
379 }
380
381 /*
382  * retrieve object from backend ext fs.
383  **/
384 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
385                        struct osd_inode_id *id)
386 {
387         int rc;
388         struct inode *inode = NULL;
389
390         /* if we look for an inode withing a running
391          * transaction, then we risk to deadlock */
392         /* osd_dirent_check_repair() breaks this */
393         /*LASSERT(current->journal_info == NULL);*/
394
395         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
396         if (IS_ERR(inode)) {
397                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
398                        id->oii_ino, PTR_ERR(inode));
399         } else if (id->oii_gen != OSD_OII_NOGEN &&
400                    inode->i_generation != id->oii_gen) {
401                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
402                        "i_generation = %u\n",
403                        id->oii_ino, id->oii_gen, inode->i_generation);
404                 iput(inode);
405                 inode = ERR_PTR(-ESTALE);
406         } else if (inode->i_nlink == 0) {
407                 /* due to parallel readdir and unlink,
408                 * we can have dead inode here. */
409                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
410                 iput(inode);
411                 inode = ERR_PTR(-ESTALE);
412         } else if (is_bad_inode(inode)) {
413                 CWARN("%s: bad inode: ino = %u\n",
414                 osd_dev2name(dev), id->oii_ino);
415                 iput(inode);
416                 inode = ERR_PTR(-ENOENT);
417         } else if ((rc = osd_attach_jinode(inode))) {
418                 iput(inode);
419                 inode = ERR_PTR(rc);
420         } else {
421                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
422                 if (id->oii_gen == OSD_OII_NOGEN)
423                         osd_id_gen(id, inode->i_ino, inode->i_generation);
424
425                 /* Do not update file c/mtime in ldiskfs.
426                  * NB: we don't have any lock to protect this because we don't
427                  * have reference on osd_object now, but contention with
428                  * another lookup + attr_set can't happen in the tiny window
429                  * between if (...) and set S_NOCMTIME. */
430                 if (!(inode->i_flags & S_NOCMTIME))
431                         inode->i_flags |= S_NOCMTIME;
432         }
433         return inode;
434 }
435
436 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
437                           handle_t *handle, struct dentry *child,
438                           struct inode *inode, struct htree_lock *hlock)
439 {
440         int rc, rc2;
441
442         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
443         if (rc == -ENOBUFS || rc == -ENOSPC) {
444                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
445                 struct inode *parent = child->d_parent->d_inode;
446                 struct lu_fid *fid = NULL;
447
448                 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
449                 if (!rc2) {
450                         fid = &loa->loa_lma.lma_self_fid;
451                 } else if (rc2 == -ENODATA) {
452                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
453                                 fid = &info->oti_fid3;
454                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
455                         } else if (!osd->od_is_ost && osd->od_index == 0) {
456                                 fid = &info->oti_fid3;
457                                 lu_igif_build(fid, parent->i_ino,
458                                               parent->i_generation);
459                         }
460                 }
461
462                 if (fid != NULL)
463                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
464                               "maximum entry limit\n",
465                               osd_name(osd), parent->i_ino, PFID(fid),
466                               rc == -ENOSPC ? "has reached" : "is approaching");
467                 else
468                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
469                               "maximum entry limit\n",
470                               osd_name(osd), parent->i_ino,
471                               rc == -ENOSPC ? "has reached" : "is approaching");
472
473                 /* ignore such error now */
474                 if (rc == -ENOBUFS)
475                         rc = 0;
476         }
477
478         return rc;
479 }
480
481
482 static struct inode *
483 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
484              struct osd_inode_id *id, struct lu_fid *fid)
485 {
486         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
487         struct inode *inode;
488         int rc;
489
490         inode = osd_iget(info, dev, id);
491         if (IS_ERR(inode))
492                 return inode;
493
494         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
495         if (!rc) {
496                 *fid = loa->loa_lma.lma_self_fid;
497         } else if (rc == -ENODATA) {
498                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
499                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
500                 else
501                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
502         } else {
503                 iput(inode);
504                 inode = ERR_PTR(rc);
505         }
506         return inode;
507 }
508
509 static struct inode *osd_iget_check(struct osd_thread_info *info,
510                                     struct osd_device *dev,
511                                     const struct lu_fid *fid,
512                                     struct osd_inode_id *id,
513                                     bool trusted)
514 {
515         struct inode *inode;
516         int rc = 0;
517         ENTRY;
518
519         /* The cached OI mapping is trustable. If we cannot locate the inode
520          * via the cached OI mapping, then return the failure to the caller
521          * directly without further OI checking. */
522
523 again:
524         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
525         if (IS_ERR(inode)) {
526                 rc = PTR_ERR(inode);
527                 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
528                         goto check_oi;
529
530                 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
531                        PFID(fid), id->oii_ino, rc);
532                 GOTO(put, rc);
533         }
534
535         if (is_bad_inode(inode)) {
536                 rc = -ENOENT;
537                 if (!trusted)
538                         goto check_oi;
539
540                 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
541                        PFID(fid), id->oii_ino);
542                 GOTO(put, rc);
543         }
544
545         if (id->oii_gen != OSD_OII_NOGEN &&
546             inode->i_generation != id->oii_gen) {
547                 rc = -ESTALE;
548                 if (!trusted)
549                         goto check_oi;
550
551                 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
552                        "oii_gen = %u, i_generation = %u\n", PFID(fid),
553                        id->oii_ino, id->oii_gen, inode->i_generation);
554                 GOTO(put, rc);
555         }
556
557         if (inode->i_nlink == 0) {
558                 rc = -ENOENT;
559                 if (!trusted)
560                         goto check_oi;
561
562                 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
563                        PFID(fid), id->oii_ino);
564                 GOTO(put, rc);
565         }
566
567         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
568
569 check_oi:
570         if (rc != 0) {
571                 __u32 saved_ino = id->oii_ino;
572                 __u32 saved_gen = id->oii_gen;
573
574                 LASSERT(!trusted);
575                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
576
577                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
578                 /* XXX: There are four possible cases:
579                  *      1. rc = 0.
580                  *         Backup/restore caused the OI invalid.
581                  *      2. rc = 0.
582                  *         Someone unlinked the object but NOT removed
583                  *         the OI mapping, such as mount target device
584                  *         as ldiskfs, and modify something directly.
585                  *      3. rc = -ENOENT.
586                  *         Someone just removed the object between the
587                  *         former oi_lookup and the iget. It is normal.
588                  *      4. Other failure cases.
589                  *
590                  *      Generally, when the device is mounted, it will
591                  *      auto check whether the system is restored from
592                  *      file-level backup or not. We trust such detect
593                  *      to distinguish the 1st case from the 2nd case:
594                  *      if the OI files are consistent but may contain
595                  *      stale OI mappings because of case 2, if iget()
596                  *      returns -ENOENT or -ESTALE, then it should be
597                  *      the case 2. */
598                 if (rc != 0)
599                         /* If the OI mapping was in OI file before the
600                          * osd_iget_check(), but now, it is disappear,
601                          * then it must be removed by race. That is a
602                          * normal race case. */
603                         GOTO(put, rc);
604
605                 /* It is the OI scrub updated the OI mapping by race.
606                  * The new OI mapping must be valid. */
607                 if (saved_ino != id->oii_ino || saved_gen != id->oii_gen) {
608                         trusted = true;
609                         goto again;
610                 }
611
612                 if (IS_ERR(inode)) {
613                         if (dev->od_scrub.os_file.sf_flags & SF_INCONSISTENT)
614                                 /* It still can be the case 2, but we cannot
615                                  * distinguish it from the case 1. So return
616                                  * -EREMCHG to block current operation until
617                                  *  OI scrub rebuilt the OI mappings. */
618                                 rc = -EREMCHG;
619                         else
620                                 rc = -ENOENT;
621
622                         GOTO(put, rc);
623                 }
624
625                 if (inode->i_generation == id->oii_gen)
626                         rc = -ENOENT;
627                 else
628                         rc = -EREMCHG;
629         } else {
630                 if (id->oii_gen == OSD_OII_NOGEN)
631                         osd_id_gen(id, inode->i_ino, inode->i_generation);
632
633                 /* Do not update file c/mtime in ldiskfs.
634                  * NB: we don't have any lock to protect this because we don't
635                  * have reference on osd_object now, but contention with
636                  * another lookup + attr_set can't happen in the tiny window
637                  * between if (...) and set S_NOCMTIME. */
638                 if (!(inode->i_flags & S_NOCMTIME))
639                         inode->i_flags |= S_NOCMTIME;
640         }
641
642         GOTO(put, rc);
643
644 put:
645         if (rc != 0) {
646                 if (!IS_ERR(inode))
647                         iput(inode);
648
649                 inode = ERR_PTR(rc);
650         }
651
652         return inode;
653 }
654
655 /**
656  * \retval +v: new filter_fid, does not contain self-fid
657  * \retval 0:  filter_fid_old, contains self-fid
658  * \retval -v: other failure cases
659  */
660 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
661                  struct dentry *dentry, struct lu_fid *fid)
662 {
663         struct filter_fid_old   *ff     = &info->oti_ff;
664         struct ost_id           *ostid  = &info->oti_ostid;
665         int                      rc;
666
667         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
668         if (rc == sizeof(*ff)) {
669                 rc = 0;
670                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
671                 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
672                 /*
673                  * XXX: use 0 as the index for compatibility, the caller will
674                  *      handle index related issues when necessary.
675                  */
676                 if (!rc)
677                         ostid_to_fid(fid, ostid, 0);
678         } else if (rc == sizeof(struct filter_fid)) {
679                 rc = 1;
680         } else if (rc >= 0) {
681                 rc = -EINVAL;
682         }
683
684         return rc;
685 }
686
687 static int osd_lma_self_repair(struct osd_thread_info *info,
688                                struct osd_device *osd, struct inode *inode,
689                                const struct lu_fid *fid, __u32 compat)
690 {
691         handle_t *jh;
692         int       rc;
693
694         LASSERT(current->journal_info == NULL);
695
696         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
697                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
698         if (IS_ERR(jh)) {
699                 rc = PTR_ERR(jh);
700                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
701                       osd_name(osd), rc);
702                 return rc;
703         }
704
705         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
706         if (rc != 0)
707                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
708                       osd_name(osd), rc);
709         ldiskfs_journal_stop(jh);
710         return rc;
711 }
712
713 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
714 {
715         struct osd_thread_info  *info   = osd_oti_get(env);
716         struct osd_device       *osd    = osd_obj2dev(obj);
717         struct lustre_ost_attrs *loa    = &info->oti_ost_attrs;
718         struct lustre_mdt_attrs *lma    = &loa->loa_lma;
719         struct inode            *inode  = obj->oo_inode;
720         struct dentry           *dentry = &info->oti_obj_dentry;
721         struct lu_fid           *fid    = NULL;
722         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
723         int                      rc;
724         ENTRY;
725
726         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
727                              (void *)loa, sizeof(*loa));
728         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
729                 fid = &lma->lma_self_fid;
730                 rc = osd_get_idif(info, inode, dentry, fid);
731                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
732                         /* For the given OST-object, if it has neither LMA nor
733                          * FID in XATTR_NAME_FID, then the given FID (which is
734                          * contained in the @obj, from client RPC for locating
735                          * the OST-object) is trusted. We use it to generate
736                          * the LMA. */
737                         osd_lma_self_repair(info, osd, inode, rfid,
738                                             LMAC_FID_ON_OST);
739                         RETURN(0);
740                 }
741         }
742
743         if (rc < 0)
744                 RETURN(rc);
745
746         if (rc > 0) {
747                 rc = 0;
748                 lustre_lma_swab(lma);
749                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
750                              (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
751                               S_ISREG(inode->i_mode)))) {
752                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
753                               "fid = "DFID", ino = %lu\n", osd_name(osd),
754                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
755                               PFID(rfid), inode->i_ino);
756                         rc = -EOPNOTSUPP;
757                 } else {
758                         fid = &lma->lma_self_fid;
759                 }
760         }
761
762         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
763                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
764                         struct ost_id   *oi   = &info->oti_ostid;
765                         struct lu_fid   *fid1 = &info->oti_fid3;
766                         __u32            idx  = fid_idif_ost_idx(rfid);
767
768                         /* For old IDIF, the OST index is not part of the IDIF,
769                          * Means that different OSTs may have the same IDIFs.
770                          * Under such case, we need to make some compatible
771                          * check to make sure to trigger OI scrub properly. */
772                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
773                                 /* Given @rfid is new, LMA is old. */
774                                 fid_to_ostid(fid, oi);
775                                 ostid_to_fid(fid1, oi, idx);
776                                 if (lu_fid_eq(fid1, rfid)) {
777                                         if (osd->od_index_in_idif)
778                                                 osd_lma_self_repair(info, osd,
779                                                         inode, rfid,
780                                                         LMAC_FID_ON_OST);
781                                         RETURN(0);
782                                 }
783                         }
784                 }
785
786                 rc = -EREMCHG;
787         }
788
789         RETURN(rc);
790 }
791
792 struct osd_check_lmv_buf {
793 #ifdef HAVE_DIR_CONTEXT
794         /* please keep it as first member */
795         struct dir_context       ctx;
796 #endif
797         struct osd_thread_info  *oclb_info;
798         struct osd_device       *oclb_dev;
799         struct osd_idmap_cache  *oclb_oic;
800 };
801
802 /**
803  * It is called internally by ->readdir() to filter out the
804  * local slave object's FID of the striped directory.
805  *
806  * \retval      1 found the local slave's FID
807  * \retval      0 continue to check next item
808  * \retval      -ve for failure
809  */
810 #ifdef HAVE_FILLDIR_USE_CTX
811 static int osd_stripe_dir_filldir(struct dir_context *buf,
812 #else
813 static int osd_stripe_dir_filldir(void *buf,
814 #endif
815                                   const char *name, int namelen,
816                                   loff_t offset, __u64 ino, unsigned d_type)
817 {
818         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
819         struct osd_thread_info *oti = oclb->oclb_info;
820         struct lu_fid *fid = &oti->oti_fid3;
821         struct osd_inode_id *id = &oti->oti_id3;
822         struct osd_device *dev = oclb->oclb_dev;
823         struct osd_idmap_cache *oic = oclb->oclb_oic;
824         struct inode *inode;
825         int rc;
826
827         if (name[0] == '.')
828                 return 0;
829
830         fid_zero(fid);
831         sscanf(name + 1, SFID, RFID(fid));
832         if (!fid_is_sane(fid))
833                 return 0;
834
835         if (osd_remote_fid(oti->oti_env, dev, fid))
836                 return 0;
837
838         osd_id_gen(id, ino, OSD_OII_NOGEN);
839         inode = osd_iget(oti, dev, id);
840         if (IS_ERR(inode))
841                 return PTR_ERR(inode);
842
843         iput(inode);
844         osd_add_oi_cache(oti, dev, id, fid);
845         oic->oic_fid = *fid;
846         oic->oic_lid = *id;
847         oic->oic_dev = dev;
848         rc = osd_oii_insert(dev, oic, true);
849
850         return rc == 0 ? 1 : rc;
851 }
852
853 /* When lookup item under striped directory, we need to locate the master
854  * MDT-object of the striped directory firstly, then the client will send
855  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
856  * and the item's name. If the system is restored from MDT file level backup,
857  * then before the OI scrub completely built the OI files, the OI mappings of
858  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
859  * not a problem for the master MDT-object. Because when locate the master
860  * MDT-object, we will do name based lookup (for the striped directory itself)
861  * firstly, during such process we can setup the correct OI mapping for the
862  * master MDT-object. But it will be trouble for the slave MDT-object. Because
863  * the client will not trigger name based lookup on the MDT to locate the slave
864  * MDT-object before locating item under the striped directory, then when
865  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
866  * is invalid and does not know what the right OI mapping is, then the MDT has
867  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
868  * the OI file, related OI mapping is unknown yet, please try again later. And
869  * then client will re-try the RPC again and again until related OI mapping has
870  * been updated. That is quite inefficient.
871  *
872  * To resolve above trouble, we will handle it as the following two cases:
873  *
874  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
875  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
876  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
877  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
878  *    directly. Please check osd_fid_lookup().
879  *
880  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
881  *    Under such case, during lookup the master MDT-object, we will lookup the
882  *    slave MDT-object via readdir against the master MDT-object, because the
883  *    slave MDT-objects information are stored as sub-directories with the name
884  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
885  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
886  *    the correct OI mapping for the slave MDT-object. */
887 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
888                          struct inode *inode, struct osd_idmap_cache *oic)
889 {
890         struct lu_buf *buf = &oti->oti_big_buf;
891         struct dentry *dentry = &oti->oti_obj_dentry;
892         struct file *filp = &oti->oti_file;
893         const struct file_operations *fops;
894         struct lmv_mds_md_v1 *lmv1;
895         struct osd_check_lmv_buf oclb = {
896 #ifdef HAVE_DIR_CONTEXT
897                 .ctx.actor = osd_stripe_dir_filldir,
898 #endif
899                 .oclb_info = oti,
900                 .oclb_dev = dev,
901                 .oclb_oic = oic
902         };
903         int rc = 0;
904         ENTRY;
905
906 again:
907         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
908                              buf->lb_len);
909         if (rc == -ERANGE) {
910                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
911                 if (rc > 0) {
912                         lu_buf_realloc(buf, rc);
913                         if (buf->lb_buf == NULL)
914                                 GOTO(out, rc = -ENOMEM);
915
916                         goto again;
917                 }
918         }
919
920         if (unlikely(rc == 0 || rc == -ENODATA))
921                 GOTO(out, rc = 0);
922
923         if (rc < 0)
924                 GOTO(out, rc);
925
926         if (unlikely(buf->lb_buf == NULL)) {
927                 lu_buf_realloc(buf, rc);
928                 if (buf->lb_buf == NULL)
929                         GOTO(out, rc = -ENOMEM);
930
931                 goto again;
932         }
933
934         lmv1 = buf->lb_buf;
935         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
936                 GOTO(out, rc = 0);
937
938         fops = inode->i_fop;
939         dentry->d_inode = inode;
940         dentry->d_sb = inode->i_sb;
941         filp->f_pos = 0;
942         filp->f_path.dentry = dentry;
943         filp->f_mode = FMODE_64BITHASH;
944         filp->f_mapping = inode->i_mapping;
945         filp->f_op = fops;
946         filp->private_data = NULL;
947         set_file_inode(filp, inode);
948
949 #ifdef HAVE_DIR_CONTEXT
950         oclb.ctx.pos = filp->f_pos;
951         rc = fops->iterate(filp, &oclb.ctx);
952         filp->f_pos = oclb.ctx.pos;
953 #else
954         rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
955 #endif
956         fops->release(inode, filp);
957
958 out:
959         if (rc < 0)
960                 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
961                        DFID": rc = %d\n", osd_ino2name(inode),
962                        inode->i_ino, inode->i_generation,
963                        PFID(&oic->oic_fid), rc);
964         else
965                 rc = 0;
966
967         RETURN(rc);
968 }
969
970 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
971                           const struct lu_fid *fid,
972                           const struct lu_object_conf *conf)
973 {
974         struct osd_thread_info *info;
975         struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
976         struct osd_device *dev;
977         struct osd_idmap_cache *oic;
978         struct osd_inode_id *id;
979         struct inode *inode = NULL;
980         struct osd_scrub *scrub;
981         struct scrub_file *sf;
982         __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
983         __u32 saved_ino;
984         __u32 saved_gen;
985         int result = 0;
986         int rc1 = 0;
987         bool remote = false;
988         bool trusted = true;
989         bool updated = false;
990         ENTRY;
991
992         LINVRNT(osd_invariant(obj));
993         LASSERT(obj->oo_inode == NULL);
994         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
995
996         dev = osd_dev(ldev);
997         scrub = &dev->od_scrub;
998         sf = &scrub->os_file;
999         info = osd_oti_get(env);
1000         LASSERT(info);
1001         oic = &info->oti_cache;
1002
1003         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1004                 RETURN(-ENOENT);
1005
1006         /* For the object is created as locking anchor, or for the object to
1007          * be created on disk. No need to osd_oi_lookup() at here because FID
1008          * shouldn't never be re-used, if it's really a duplicate FID from
1009          * unexpected reason, we should be able to detect it later by calling
1010          * do_create->osd_oi_insert(). */
1011         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
1012                 GOTO(out, result = 0);
1013
1014         /* Search order: 1. per-thread cache. */
1015         if (lu_fid_eq(fid, &oic->oic_fid) &&
1016             likely(oic->oic_dev == dev)) {
1017                 id = &oic->oic_lid;
1018                 goto iget;
1019         }
1020
1021         id = &info->oti_id;
1022         if (!list_empty(&scrub->os_inconsistent_items)) {
1023                 /* Search order: 2. OI scrub pending list. */
1024                 result = osd_oii_lookup(dev, fid, id);
1025                 if (result == 0)
1026                         goto iget;
1027         }
1028
1029         /* The OI mapping in the OI file can be updated by the OI scrub
1030          * when we locate the inode via FID. So it may be not trustable. */
1031         trusted = false;
1032
1033         /* Search order: 3. OI files. */
1034         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1035         if (result == -ENOENT) {
1036                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1037                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1038                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1039                                       sf->sf_oi_bitmap))
1040                         GOTO(out, result = 0);
1041
1042                 goto trigger;
1043         }
1044
1045         if (result != 0)
1046                 GOTO(out, result);
1047
1048 iget:
1049         inode = osd_iget_check(info, dev, fid, id, trusted);
1050         if (IS_ERR(inode)) {
1051                 result = PTR_ERR(inode);
1052                 if (result == -ENOENT || result == -ESTALE)
1053                         GOTO(out, result = 0);
1054
1055                 if (result == -EREMCHG) {
1056
1057 trigger:
1058                         /* We still have chance to get the valid inode: for the
1059                          * object which is referenced by remote name entry, the
1060                          * object on the local MDT will be linked under the dir
1061                          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1062                          *
1063                          * We do not know whether the object for the given FID
1064                          * is referenced by some remote name entry or not, and
1065                          * especially for DNE II, a multiple-linked object may
1066                          * have many name entries reside on many MDTs.
1067                          *
1068                          * To simplify the operation, OSD will not distinguish
1069                          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1070                          * only happened for the RPC from other MDT during the
1071                          * OI scrub, or for the client side RPC with FID only,
1072                          * such as FID to path, or from old connected client. */
1073                         if (!remote &&
1074                             !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1075                                 rc1 = osd_lookup_in_remote_parent(info, dev,
1076                                                                   fid, id);
1077                                 if (rc1 == 0) {
1078                                         remote = true;
1079                                         trusted = true;
1080                                         flags |= SS_AUTO_PARTIAL;
1081                                         flags &= ~SS_AUTO_FULL;
1082                                         goto iget;
1083                                 }
1084                         }
1085
1086                         if (thread_is_running(&scrub->os_thread)) {
1087                                 if (scrub->os_partial_scan &&
1088                                     !scrub->os_in_join) {
1089                                         goto join;
1090                                 } else {
1091                                         if (inode != NULL && !IS_ERR(inode)) {
1092                                                 LASSERT(remote);
1093
1094                                                 osd_add_oi_cache(info, dev, id,
1095                                                                  fid);
1096                                                 osd_oii_insert(dev, oic, true);
1097                                         } else {
1098                                                 result = -EINPROGRESS;
1099                                         }
1100                                 }
1101                         } else if (!dev->od_noscrub) {
1102
1103 join:
1104                                 rc1 = osd_scrub_start(dev, flags);
1105                                 LCONSOLE_WARN("%s: trigger OI scrub by RPC "
1106                                               "for the "DFID" with flags 0x%x,"
1107                                               " rc = %d\n", osd_name(dev),
1108                                               PFID(fid), flags, rc1);
1109                                 if (rc1 == 0 || rc1 == -EALREADY) {
1110                                         if (inode != NULL && !IS_ERR(inode)) {
1111                                                 LASSERT(remote);
1112
1113                                                 osd_add_oi_cache(info, dev, id,
1114                                                                  fid);
1115                                                 osd_oii_insert(dev, oic, true);
1116                                         } else {
1117                                                 result = -EINPROGRESS;
1118                                         }
1119                                 } else {
1120                                         result = -EREMCHG;
1121                                 }
1122                         } else {
1123                                 result = -EREMCHG;
1124                         }
1125                 }
1126
1127                 if (inode == NULL || IS_ERR(inode))
1128                         GOTO(out, result);
1129         } else if (remote) {
1130                 goto trigger;
1131         }
1132
1133         obj->oo_inode = inode;
1134         LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
1135
1136         result = osd_check_lma(env, obj);
1137         if (result == 0)
1138                 goto found;
1139
1140         LASSERTF(id->oii_ino == inode->i_ino &&
1141                  id->oii_gen == inode->i_generation,
1142                  "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1143                  PFID(fid), id->oii_ino, id->oii_gen,
1144                  inode->i_ino, inode->i_generation);
1145
1146         saved_ino = inode->i_ino;
1147         saved_gen = inode->i_generation;
1148
1149         if (unlikely(result == -ENODATA)) {
1150                 /* If the OI scrub updated the OI mapping by race, it
1151                  * must be valid. Trust the inode that has no LMA EA. */
1152                 if (updated)
1153                         goto found;
1154
1155                 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1156                 if (result == 0) {
1157                         /* The OI mapping is still there, the inode is still
1158                          * valid. It is just becaues the inode has no LMA EA. */
1159                         if (saved_ino == id->oii_ino &&
1160                             saved_gen == id->oii_gen)
1161                                 goto found;
1162
1163                         /* It is the OI scrub updated the OI mapping by race.
1164                          * The new OI mapping must be valid. */
1165                         iput(inode);
1166                         inode = NULL;
1167                         obj->oo_inode = NULL;
1168                         trusted = true;
1169                         updated = true;
1170                         goto iget;
1171                 }
1172
1173                 /* "result == -ENOENT" means that the OI mappinghas been
1174                  * removed by race, so the inode belongs to other object.
1175                  *
1176                  * Others error can be returned  directly. */
1177                 if (result == -ENOENT) {
1178                         LASSERT(trusted);
1179
1180                         result = 0;
1181                 }
1182         }
1183
1184         iput(inode);
1185         inode = NULL;
1186         obj->oo_inode = NULL;
1187
1188         if (result != -EREMCHG)
1189                 GOTO(out, result);
1190
1191         LASSERT(!updated);
1192
1193         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1194         /* "result == -ENOENT" means the cached OI mapping has been removed
1195          * from the OI file by race, above inode belongs to other object. */
1196         if (result == -ENOENT) {
1197                 LASSERT(trusted);
1198
1199                 GOTO(out, result = 0);
1200         }
1201
1202         if (result != 0)
1203                 GOTO(out, result);
1204
1205         if (saved_ino == id->oii_ino && saved_gen == id->oii_gen)
1206                 goto trigger;
1207
1208         /* It is the OI scrub updated the OI mapping by race.
1209          * The new OI mapping must be valid. */
1210         trusted = true;
1211         updated = true;
1212         goto iget;
1213
1214 found:
1215         obj->oo_compat_dot_created = 1;
1216         obj->oo_compat_dotdot_created = 1;
1217
1218         if (S_ISDIR(inode->i_mode) &&
1219             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1220                 osd_check_lmv(info, dev, inode, oic);
1221
1222         result = osd_attach_jinode(inode);
1223         if (result) {
1224                 obj->oo_inode = NULL;
1225                 iput(inode);
1226                 GOTO(out, result);
1227         }
1228
1229         if (!ldiskfs_pdo)
1230                 GOTO(out, result = 0);
1231
1232         LASSERT(obj->oo_hl_head == NULL);
1233         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1234         if (obj->oo_hl_head == NULL) {
1235                 obj->oo_inode = NULL;
1236                 iput(inode);
1237                 GOTO(out, result = -ENOMEM);
1238         }
1239         GOTO(out, result = 0);
1240
1241 out:
1242         if (result != 0 && trusted)
1243                 fid_zero(&oic->oic_fid);
1244
1245         LINVRNT(osd_invariant(obj));
1246         return result;
1247 }
1248
1249 /*
1250  * Concurrency: shouldn't matter.
1251  */
1252 static void osd_object_init0(struct osd_object *obj)
1253 {
1254         LASSERT(obj->oo_inode != NULL);
1255         obj->oo_dt.do_body_ops = &osd_body_ops;
1256         obj->oo_dt.do_lu.lo_header->loh_attr |=
1257                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1258 }
1259
1260 /*
1261  * Concurrency: no concurrent access is possible that early in object
1262  * life-cycle.
1263  */
1264 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1265                            const struct lu_object_conf *conf)
1266 {
1267         struct osd_object *obj = osd_obj(l);
1268         int result;
1269
1270         LINVRNT(osd_invariant(obj));
1271
1272         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1273                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1274                 l->lo_header->loh_attr |= LOHA_EXISTS;
1275                 return 0;
1276         }
1277
1278         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1279         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1280         if (result == 0 && obj->oo_inode != NULL) {
1281                 struct osd_thread_info *oti = osd_oti_get(env);
1282                 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1283
1284                 osd_object_init0(obj);
1285                 result = osd_get_lma(oti, obj->oo_inode,
1286                                      &oti->oti_obj_dentry, loa);
1287                 if (!result) {
1288                         /* Convert LMAI flags to lustre LMA flags
1289                          * and cache it to oo_lma_flags */
1290                         obj->oo_lma_flags =
1291                                 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1292                 } else if (result == -ENODATA) {
1293                         result = 0;
1294                 }
1295         }
1296
1297         LINVRNT(osd_invariant(obj));
1298         return result;
1299 }
1300
1301 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1302  * The left part is for value, binary mode. */
1303 struct osd_xattr_entry {
1304         struct list_head        oxe_list;
1305         size_t                  oxe_len;
1306         size_t                  oxe_namelen;
1307         bool                    oxe_exist;
1308         struct rcu_head         oxe_rcu;
1309         char                    oxe_buf[0];
1310 };
1311
1312 static int osd_oxc_get(struct osd_object *obj, const char *name,
1313                        struct lu_buf *buf)
1314 {
1315         struct osd_xattr_entry *tmp;
1316         struct osd_xattr_entry *oxe = NULL;
1317         size_t namelen = strlen(name);
1318         int rc;
1319         ENTRY;
1320
1321         rcu_read_lock();
1322         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1323                 if (namelen == tmp->oxe_namelen &&
1324                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1325                         oxe = tmp;
1326                         break;
1327                 }
1328         }
1329
1330         if (oxe == NULL)
1331                 GOTO(out, rc = -ENOENT);
1332
1333         if (!oxe->oxe_exist)
1334                 GOTO(out, rc = -ENODATA);
1335
1336         /* vallen */
1337         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1338         LASSERT(rc > 0);
1339
1340         if (buf->lb_buf == NULL)
1341                 GOTO(out, rc);
1342
1343         if (buf->lb_len < rc)
1344                 GOTO(out, rc = -ERANGE);
1345
1346         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1347         EXIT;
1348 out:
1349         rcu_read_unlock();
1350
1351         return rc;
1352 }
1353
1354 static void osd_oxc_free(struct rcu_head *head)
1355 {
1356         struct osd_xattr_entry *oxe;
1357
1358         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1359         OBD_FREE(oxe, oxe->oxe_len);
1360 }
1361
1362 static void osd_oxc_add(struct osd_object *obj, const char *name,
1363                         const char *buf, int buflen)
1364 {
1365         struct osd_xattr_entry *oxe;
1366         struct osd_xattr_entry *old = NULL;
1367         struct osd_xattr_entry *tmp;
1368         size_t namelen = strlen(name);
1369         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1370
1371         OBD_ALLOC(oxe, len);
1372         if (oxe == NULL)
1373                 return;
1374
1375         INIT_LIST_HEAD(&oxe->oxe_list);
1376         oxe->oxe_len = len;
1377         oxe->oxe_namelen = namelen;
1378         memcpy(oxe->oxe_buf, name, namelen);
1379         if (buflen > 0) {
1380                 LASSERT(buf != NULL);
1381                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1382                 oxe->oxe_exist = true;
1383         } else {
1384                 oxe->oxe_exist = false;
1385         }
1386
1387         /* this should be rarely called, just remove old and add new */
1388         spin_lock(&obj->oo_guard);
1389         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1390                 if (namelen == tmp->oxe_namelen &&
1391                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1392                         old = tmp;
1393                         break;
1394                 }
1395         }
1396         if (old != NULL) {
1397                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1398                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1399         } else {
1400                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1401         }
1402         spin_unlock(&obj->oo_guard);
1403 }
1404
1405 static void osd_oxc_del(struct osd_object *obj, const char *name)
1406 {
1407         struct osd_xattr_entry *oxe;
1408         size_t namelen = strlen(name);
1409
1410         spin_lock(&obj->oo_guard);
1411         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1412                 if (namelen == oxe->oxe_namelen &&
1413                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1414                         list_del_rcu(&oxe->oxe_list);
1415                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1416                         break;
1417                 }
1418         }
1419         spin_unlock(&obj->oo_guard);
1420 }
1421
1422 static void osd_oxc_fini(struct osd_object *obj)
1423 {
1424         struct osd_xattr_entry *oxe, *next;
1425
1426         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1427                 list_del(&oxe->oxe_list);
1428                 OBD_FREE(oxe, oxe->oxe_len);
1429         }
1430 }
1431
1432 /*
1433  * Concurrency: no concurrent access is possible that late in object
1434  * life-cycle.
1435  */
1436 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1437 {
1438         struct osd_object *obj = osd_obj(l);
1439
1440         LINVRNT(osd_invariant(obj));
1441
1442         osd_oxc_fini(obj);
1443         dt_object_fini(&obj->oo_dt);
1444         if (obj->oo_hl_head != NULL)
1445                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1446         OBD_FREE_PTR(obj);
1447 }
1448
1449 /*
1450  * Concurrency: no concurrent access is possible that late in object
1451  * life-cycle.
1452  */
1453 static void osd_index_fini(struct osd_object *o)
1454 {
1455         struct iam_container *bag;
1456
1457         if (o->oo_dir != NULL) {
1458                 bag = &o->oo_dir->od_container;
1459                 if (o->oo_inode != NULL) {
1460                         if (bag->ic_object == o->oo_inode)
1461                                 iam_container_fini(bag);
1462                 }
1463                 OBD_FREE_PTR(o->oo_dir);
1464                 o->oo_dir = NULL;
1465         }
1466 }
1467
1468 /*
1469  * Concurrency: no concurrent access is possible that late in object
1470  * life-cycle (for all existing callers, that is. New callers have to provide
1471  * their own locking.)
1472  */
1473 static int osd_inode_unlinked(const struct inode *inode)
1474 {
1475         return inode->i_nlink == 0;
1476 }
1477
1478 enum {
1479         OSD_TXN_OI_DELETE_CREDITS    = 20,
1480         OSD_TXN_INODE_DELETE_CREDITS = 20
1481 };
1482
1483 /*
1484  * Journal
1485  */
1486
1487 #if OSD_THANDLE_STATS
1488 /**
1489  * Set time when the handle is allocated
1490  */
1491 static void osd_th_alloced(struct osd_thandle *oth)
1492 {
1493         oth->oth_alloced = ktime_get();
1494 }
1495
1496 /**
1497  * Set time when the handle started
1498  */
1499 static void osd_th_started(struct osd_thandle *oth)
1500 {
1501         oth->oth_started = ktime_get();
1502 }
1503
1504 /**
1505  * Check whether the we deal with this handle for too long.
1506  */
1507 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1508                                 ktime_t alloced, ktime_t started,
1509                                 ktime_t closed)
1510 {
1511         ktime_t now = ktime_get();
1512
1513         LASSERT(dev != NULL);
1514
1515         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1516                             ktime_us_delta(started, alloced));
1517         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1518                             ktime_us_delta(closed, started));
1519         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1520                             ktime_us_delta(now, closed));
1521
1522         if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1523                 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1524                       oth, now, alloced, started, closed);
1525                 libcfs_debug_dumpstack(NULL);
1526         }
1527 }
1528
1529 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1530 {                                                                       \
1531         ktime_t __closed = ktime_get();                                 \
1532         ktime_t __alloced = oth->oth_alloced;                           \
1533         ktime_t __started = oth->oth_started;                           \
1534                                                                         \
1535         expr;                                                           \
1536         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1537 }
1538
1539 #else /* OSD_THANDLE_STATS */
1540
1541 #define osd_th_alloced(h)                  do {} while(0)
1542 #define osd_th_started(h)                  do {} while(0)
1543 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1544
1545 #endif /* OSD_THANDLE_STATS */
1546
1547 /*
1548  * Concurrency: doesn't access mutable data.
1549  */
1550 static int osd_param_is_not_sane(const struct osd_device *dev,
1551                                  const struct thandle *th)
1552 {
1553         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1554
1555         return oh->ot_credits > osd_transaction_size(dev);
1556 }
1557
1558 /*
1559  * Concurrency: shouldn't matter.
1560  */
1561 static void osd_trans_commit_cb(struct super_block *sb,
1562                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1563 {
1564         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1565         struct thandle     *th  = &oh->ot_super;
1566         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
1567         struct dt_txn_commit_cb *dcb, *tmp;
1568
1569         LASSERT(oh->ot_handle == NULL);
1570
1571         if (error)
1572                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1573
1574         dt_txn_hook_commit(th);
1575
1576         /* call per-transaction callbacks if any */
1577         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1578                                  dcb_linkage) {
1579                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1580                          "commit callback entry: magic=%x name='%s'\n",
1581                          dcb->dcb_magic, dcb->dcb_name);
1582                 list_del_init(&dcb->dcb_linkage);
1583                 dcb->dcb_func(NULL, th, dcb, error);
1584         }
1585
1586         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1587         lu_device_put(lud);
1588         th->th_dev = NULL;
1589
1590         lu_context_exit(&th->th_ctx);
1591         lu_context_fini(&th->th_ctx);
1592         OBD_FREE_PTR(oh);
1593 }
1594
1595 #ifndef HAVE_SB_START_WRITE
1596 # define sb_start_write(sb) do {} while (0)
1597 # define sb_end_write(sb) do {} while (0)
1598 #endif
1599
1600 static struct thandle *osd_trans_create(const struct lu_env *env,
1601                                         struct dt_device *d)
1602 {
1603         struct osd_thread_info  *oti = osd_oti_get(env);
1604         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1605         struct osd_thandle      *oh;
1606         struct thandle          *th;
1607         ENTRY;
1608
1609         if (d->dd_rdonly) {
1610                 CERROR("%s: someone try to start transaction under "
1611                        "readonly mode, should be disabled.\n",
1612                        osd_name(osd_dt_dev(d)));
1613                 dump_stack();
1614                 RETURN(ERR_PTR(-EROFS));
1615         }
1616
1617         /* on pending IO in this thread should left from prev. request */
1618         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1619
1620         sb_start_write(osd_sb(osd_dt_dev(d)));
1621
1622         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1623         if (oh != NULL) {
1624                 oh->ot_quota_trans = &oti->oti_quota_trans;
1625                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1626                 th = &oh->ot_super;
1627                 th->th_dev = d;
1628                 th->th_result = 0;
1629                 th->th_tags = LCT_TX_HANDLE;
1630                 oh->ot_credits = 0;
1631                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1632                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1633                 osd_th_alloced(oh);
1634
1635                 memset(oti->oti_declare_ops, 0,
1636                        sizeof(oti->oti_declare_ops));
1637                 memset(oti->oti_declare_ops_cred, 0,
1638                        sizeof(oti->oti_declare_ops_cred));
1639                 memset(oti->oti_declare_ops_used, 0,
1640                        sizeof(oti->oti_declare_ops_used));
1641         } else {
1642                 sb_end_write(osd_sb(osd_dt_dev(d)));
1643                 th = ERR_PTR(-ENOMEM);
1644         }
1645         RETURN(th);
1646 }
1647
1648 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1649 {
1650         struct osd_thread_info  *oti = osd_oti_get(env);
1651         struct osd_thandle      *oh;
1652
1653         oh = container_of0(th, struct osd_thandle, ot_super);
1654         LASSERT(oh != NULL);
1655
1656         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1657               oti->oti_declare_ops[OSD_OT_CREATE],
1658               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1659               oti->oti_declare_ops_used[OSD_OT_CREATE],
1660               oti->oti_declare_ops[OSD_OT_DESTROY],
1661               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1662               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1663         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1664               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1665               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1666               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1667               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1668               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1669               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1670         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1671               oti->oti_declare_ops[OSD_OT_WRITE],
1672               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1673               oti->oti_declare_ops_used[OSD_OT_WRITE],
1674               oti->oti_declare_ops[OSD_OT_PUNCH],
1675               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1676               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1677               oti->oti_declare_ops[OSD_OT_QUOTA],
1678               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1679               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1680         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1681               oti->oti_declare_ops[OSD_OT_INSERT],
1682               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1683               oti->oti_declare_ops_used[OSD_OT_INSERT],
1684               oti->oti_declare_ops[OSD_OT_DELETE],
1685               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1686               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1687         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1688               oti->oti_declare_ops[OSD_OT_REF_ADD],
1689               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1690               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1691               oti->oti_declare_ops[OSD_OT_REF_DEL],
1692               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1693               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1694 }
1695
1696 /*
1697  * Concurrency: shouldn't matter.
1698  */
1699 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1700                            struct thandle *th)
1701 {
1702         struct osd_thread_info *oti = osd_oti_get(env);
1703         struct osd_device  *dev = osd_dt_dev(d);
1704         handle_t           *jh;
1705         struct osd_thandle *oh;
1706         int rc;
1707
1708         ENTRY;
1709
1710         LASSERT(current->journal_info == NULL);
1711
1712         oh = container_of0(th, struct osd_thandle, ot_super);
1713         LASSERT(oh != NULL);
1714         LASSERT(oh->ot_handle == NULL);
1715
1716         rc = dt_txn_hook_start(env, d, th);
1717         if (rc != 0)
1718                 GOTO(out, rc);
1719
1720         if (unlikely(osd_param_is_not_sane(dev, th))) {
1721                 static unsigned long last_printed;
1722                 static int last_credits;
1723
1724                 /* don't make noise on a tiny testing systems
1725                  * actual credits misuse will be caught anyway */
1726                 if (last_credits != oh->ot_credits &&
1727                     time_after(jiffies, last_printed +
1728                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1729                     osd_transaction_size(dev) > 512) {
1730                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1731                               oh->ot_credits, osd_transaction_size(dev));
1732                         osd_trans_dump_creds(env, th);
1733                         libcfs_debug_dumpstack(NULL);
1734                         last_credits = oh->ot_credits;
1735                         last_printed = jiffies;
1736                 }
1737                 /* XXX Limit the credits to 'max_transaction_buffers', and
1738                  *     let the underlying filesystem to catch the error if
1739                  *     we really need so many credits.
1740                  *
1741                  *     This should be removed when we can calculate the
1742                  *     credits precisely. */
1743                 oh->ot_credits = osd_transaction_size(dev);
1744         } else if (ldiskfs_track_declares_assert != 0) {
1745                 /* reserve few credits to prevent an assertion in JBD
1746                  * our debugging mechanism will be able to detected
1747                  * overuse. this can help to debug single-update
1748                  * transactions */
1749                 oh->ot_credits += 10;
1750                 if (unlikely(osd_param_is_not_sane(dev, th)))
1751                         oh->ot_credits = osd_transaction_size(dev);
1752         }
1753
1754         /*
1755          * XXX temporary stuff. Some abstraction layer should
1756          * be used.
1757          */
1758         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1759         osd_th_started(oh);
1760         if (!IS_ERR(jh)) {
1761                 oh->ot_handle = jh;
1762                 LASSERT(oti->oti_txns == 0);
1763                 lu_context_init(&th->th_ctx, th->th_tags);
1764                 lu_context_enter(&th->th_ctx);
1765
1766                 lu_device_get(&d->dd_lu_dev);
1767                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1768                               "osd-tx", th);
1769                 oti->oti_txns++;
1770                 rc = 0;
1771         } else {
1772                 rc = PTR_ERR(jh);
1773         }
1774 out:
1775         RETURN(rc);
1776 }
1777
1778 static int osd_seq_exists(const struct lu_env *env,
1779                           struct osd_device *osd, u64 seq)
1780 {
1781         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1782         struct seq_server_site  *ss = osd_seq_site(osd);
1783         int                     rc;
1784         ENTRY;
1785
1786         LASSERT(ss != NULL);
1787         LASSERT(ss->ss_server_fld != NULL);
1788
1789         rc = osd_fld_lookup(env, osd, seq, range);
1790         if (rc != 0) {
1791                 if (rc != -ENOENT)
1792                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1793                                osd_name(osd), seq, rc);
1794                 RETURN(0);
1795         }
1796
1797         RETURN(ss->ss_node_id == range->lsr_index);
1798 }
1799
1800 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1801 {
1802         struct dt_txn_commit_cb *dcb;
1803         struct dt_txn_commit_cb *tmp;
1804
1805         /* call per-transaction stop callbacks if any */
1806         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1807                                  dcb_linkage) {
1808                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1809                          "commit callback entry: magic=%x name='%s'\n",
1810                          dcb->dcb_magic, dcb->dcb_name);
1811                 list_del_init(&dcb->dcb_linkage);
1812                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1813         }
1814 }
1815
1816 /*
1817  * Concurrency: shouldn't matter.
1818  */
1819 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1820                           struct thandle *th)
1821 {
1822         int                     rc = 0, remove_agents = 0;
1823         struct osd_thandle     *oh;
1824         struct osd_thread_info *oti = osd_oti_get(env);
1825         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1826         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1827         struct qsd_instance    *qsd = osd->od_quota_slave;
1828         struct lquota_trans    *qtrans;
1829         ENTRY;
1830
1831         oh = container_of0(th, struct osd_thandle, ot_super);
1832
1833         /* reset OI cache for safety */
1834         oti->oti_ins_cache_used = 0;
1835
1836         remove_agents = oh->ot_remove_agents;
1837
1838         qtrans = oh->ot_quota_trans;
1839         oh->ot_quota_trans = NULL;
1840
1841         if (oh->ot_handle != NULL) {
1842                 int rc2;
1843                 handle_t *hdl = oh->ot_handle;
1844
1845                 /*
1846                  * add commit callback
1847                  * notice we don't do this in osd_trans_start()
1848                  * as underlying transaction can change during truncate
1849                  */
1850                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1851                                          &oh->ot_jcb);
1852
1853                 LASSERT(oti->oti_txns == 1);
1854                 oti->oti_txns--;
1855
1856                 rc = dt_txn_hook_stop(env, th);
1857                 if (rc != 0)
1858                         CERROR("%s: failed in transaction hook: rc = %d\n",
1859                                osd_name(osd), rc);
1860
1861                 osd_trans_stop_cb(oh, rc);
1862                 /* hook functions might modify th_sync */
1863                 hdl->h_sync = th->th_sync;
1864
1865                 oh->ot_handle = NULL;
1866                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1867                 if (rc2 != 0)
1868                         CERROR("%s: failed to stop transaction: rc = %d\n",
1869                                osd_name(osd), rc2);
1870                 if (!rc)
1871                         rc = rc2;
1872         } else {
1873                 osd_trans_stop_cb(oh, th->th_result);
1874                 OBD_FREE_PTR(oh);
1875         }
1876
1877         /* inform the quota slave device that the transaction is stopping */
1878         qsd_op_end(env, qsd, qtrans);
1879
1880         /* as we want IO to journal and data IO be concurrent, we don't block
1881          * awaiting data IO completion in osd_do_bio(), instead we wait here
1882          * once transaction is submitted to the journal. all reqular requests
1883          * don't do direct IO (except read/write), thus this wait_event becomes
1884          * no-op for them.
1885          *
1886          * IMPORTANT: we have to wait till any IO submited by the thread is
1887          * completed otherwise iobuf may be corrupted by different request
1888          */
1889         wait_event(iobuf->dr_wait,
1890                        atomic_read(&iobuf->dr_numreqs) == 0);
1891         osd_fini_iobuf(osd, iobuf);
1892         if (!rc)
1893                 rc = iobuf->dr_error;
1894
1895         if (unlikely(remove_agents != 0))
1896                 osd_process_scheduled_agent_removals(env, osd);
1897
1898         sb_end_write(osd_sb(osd));
1899
1900         RETURN(rc);
1901 }
1902
1903 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1904 {
1905         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1906                                                ot_super);
1907
1908         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1909         LASSERT(&dcb->dcb_func != NULL);
1910         if (dcb->dcb_flags & DCB_TRANS_STOP)
1911                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1912         else
1913                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1914
1915         return 0;
1916 }
1917
1918 /*
1919  * Called just before object is freed. Releases all resources except for
1920  * object itself (that is released by osd_object_free()).
1921  *
1922  * Concurrency: no concurrent access is possible that late in object
1923  * life-cycle.
1924  */
1925 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1926 {
1927         struct osd_object *obj   = osd_obj(l);
1928         struct inode      *inode = obj->oo_inode;
1929
1930         LINVRNT(osd_invariant(obj));
1931
1932         /*
1933          * If object is unlinked remove fid->ino mapping from object index.
1934          */
1935
1936         osd_index_fini(obj);
1937         if (inode != NULL) {
1938                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1939                 qid_t                    uid = i_uid_read(inode);
1940                 qid_t                    gid = i_gid_read(inode);
1941
1942                 iput(inode);
1943                 obj->oo_inode = NULL;
1944
1945                 if (qsd != NULL) {
1946                         struct osd_thread_info  *info = osd_oti_get(env);
1947                         struct lquota_id_info   *qi = &info->oti_qi;
1948
1949                         /* Release granted quota to master if necessary */
1950                         qi->lqi_id.qid_uid = uid;
1951                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1952
1953                         qi->lqi_id.qid_uid = gid;
1954                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1955
1956                         qi->lqi_id.qid_uid = i_projid_read(inode);
1957                         qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1958                 }
1959         }
1960 }
1961
1962 /*
1963  * Concurrency: ->loo_object_release() is called under site spin-lock.
1964  */
1965 static void osd_object_release(const struct lu_env *env,
1966                                struct lu_object *l)
1967 {
1968         struct osd_object *o = osd_obj(l);
1969         /* nobody should be releasing a non-destroyed object with nlink=0
1970          * the API allows this, but ldiskfs doesn't like and then report
1971          * this inode as deleted */
1972         if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1973                 LBUG();
1974 }
1975
1976 /*
1977  * Concurrency: shouldn't matter.
1978  */
1979 static int osd_object_print(const struct lu_env *env, void *cookie,
1980                             lu_printer_t p, const struct lu_object *l)
1981 {
1982         struct osd_object *o = osd_obj(l);
1983         struct iam_descr  *d;
1984
1985         if (o->oo_dir != NULL)
1986                 d = o->oo_dir->od_container.ic_descr;
1987         else
1988                 d = NULL;
1989         return (*p)(env, cookie,
1990                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1991                     o, o->oo_inode,
1992                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1993                     o->oo_inode ? o->oo_inode->i_generation : 0,
1994                     d ? d->id_ops->id_name : "plain");
1995 }
1996
1997 /*
1998  * Concurrency: shouldn't matter.
1999  */
2000 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2001                struct obd_statfs *sfs)
2002 {
2003         struct osd_device       *osd = osd_dt_dev(d);
2004         struct super_block      *sb = osd_sb(osd);
2005         struct kstatfs          *ksfs;
2006         __u64                    reserved;
2007         int                      result = 0;
2008
2009         if (unlikely(osd->od_mnt == NULL))
2010                 return -EINPROGRESS;
2011
2012         /* osd_lproc.c call this without env, allocate ksfs for that case */
2013         if (unlikely(env == NULL)) {
2014                 OBD_ALLOC_PTR(ksfs);
2015                 if (ksfs == NULL)
2016                         return -ENOMEM;
2017         } else {
2018                 ksfs = &osd_oti_get(env)->oti_ksfs;
2019         }
2020
2021         result = sb->s_op->statfs(sb->s_root, ksfs);
2022         if (result)
2023                 goto out;
2024
2025         statfs_pack(sfs, ksfs);
2026         if (unlikely(sb->s_flags & MS_RDONLY))
2027                 sfs->os_state |= OS_STATE_READONLY;
2028         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2029                                          LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2030                 sfs->os_maxbytes = sb->s_maxbytes;
2031         else
2032                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2033
2034         /*
2035          * Reserve some space so to avoid fragmenting the filesystem too much.
2036          * Fragmentation not only impacts performance, but can also increase
2037          * metadata overhead significantly, causing grant calculation to be
2038          * wrong.
2039          *
2040          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2041          */
2042         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2043         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2044         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2045                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2046
2047         sfs->os_blocks -= reserved;
2048         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2049         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2050
2051 out:
2052         if (unlikely(env == NULL))
2053                 OBD_FREE_PTR(ksfs);
2054         return result;
2055 }
2056
2057 /**
2058  * Estimate space needed for file creations. We assume the largest filename
2059  * which is 2^64 - 1, hence a filename of 20 chars.
2060  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2061  */
2062 #ifdef __LDISKFS_DIR_REC_LEN
2063 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2064 #else
2065 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2066 #endif
2067
2068 /*
2069  * Concurrency: doesn't access mutable data.
2070  */
2071 static void osd_conf_get(const struct lu_env *env,
2072                          const struct dt_device *dev,
2073                          struct dt_device_param *param)
2074 {
2075         struct super_block *sb = osd_sb(osd_dt_dev(dev));
2076         int                ea_overhead;
2077
2078         /*
2079          * XXX should be taken from not-yet-existing fs abstraction layer.
2080          */
2081         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2082         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2083         param->ddp_symlink_max  = sb->s_blocksize;
2084         param->ddp_mount_type     = LDD_MT_LDISKFS;
2085         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2086                 param->ddp_maxbytes = sb->s_maxbytes;
2087         else
2088                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2089         /* inode are statically allocated, so per-inode space consumption
2090          * is the space consumed by the directory entry */
2091         param->ddp_inodespace     = PER_OBJ_USAGE;
2092         /* EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
2093          * = 128MB) which is unlikely to be hit in real life. Report a smaller
2094          * maximum length to not under count the actual number of extents
2095          * needed for writing a file. */
2096         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2097         /* worst-case extent insertion metadata overhead */
2098         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2099         param->ddp_mntopts      = 0;
2100         if (test_opt(sb, XATTR_USER))
2101                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2102         if (test_opt(sb, POSIX_ACL))
2103                 param->ddp_mntopts |= MNTOPT_ACL;
2104
2105         /* LOD might calculate the max stripe count based on max_ea_size,
2106          * so we need take account in the overhead as well,
2107          * xattr_header + magic + xattr_entry_head */
2108         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2109                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2110
2111 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2112         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2113                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2114                                                                 ea_overhead;
2115         else
2116 #endif
2117                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2118 }
2119
2120 /*
2121  * Concurrency: shouldn't matter.
2122  */
2123 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2124 {
2125         int rc;
2126
2127         CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2128
2129         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2130
2131         CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2132                rc);
2133
2134         return rc;
2135 }
2136
2137 /**
2138  * Start commit for OSD device.
2139  *
2140  * An implementation of dt_commit_async method for OSD device.
2141  * Asychronously starts underlayng fs sync and thereby a transaction
2142  * commit.
2143  *
2144  * \param env environment
2145  * \param d dt device
2146  *
2147  * \see dt_device_operations
2148  */
2149 static int osd_commit_async(const struct lu_env *env,
2150                             struct dt_device *d)
2151 {
2152         struct super_block *s = osd_sb(osd_dt_dev(d));
2153         ENTRY;
2154
2155         CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2156         RETURN(s->s_op->sync_fs(s, 0));
2157 }
2158
2159 /*
2160  * Concurrency: shouldn't matter.
2161  */
2162
2163 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2164 {
2165         struct super_block *sb = osd_sb(osd_dt_dev(d));
2166         struct block_device *dev = sb->s_bdev;
2167 #ifdef HAVE_DEV_SET_RDONLY
2168         struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2169         int rc = 0;
2170 #else
2171         int rc = -EOPNOTSUPP;
2172 #endif
2173         ENTRY;
2174
2175 #ifdef HAVE_DEV_SET_RDONLY
2176         CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
2177
2178         if (sb->s_op->freeze_fs) {
2179                 rc = sb->s_op->freeze_fs(sb);
2180                 if (rc)
2181                         goto out;
2182         }
2183
2184         if (jdev && (jdev != dev)) {
2185                 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2186                        (long)jdev);
2187                 dev_set_rdonly(jdev);
2188         }
2189         CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2190         dev_set_rdonly(dev);
2191
2192         if (sb->s_op->unfreeze_fs)
2193                 sb->s_op->unfreeze_fs(sb);
2194
2195 out:
2196 #endif
2197         if (rc)
2198                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2199                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2200
2201         RETURN(rc);
2202 }
2203
2204 /**
2205  * Note: we do not count into QUOTA here.
2206  * If we mount with --data_journal we may need more.
2207  */
2208 const int osd_dto_credits_noquota[DTO_NR] = {
2209         /**
2210          * Insert.
2211          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2212          * SINGLEDATA_TRANS_BLOCKS(8)
2213          * XXX Note: maybe iam need more, since iam have more level than
2214          *           EXT3 htree.
2215          */
2216         [DTO_INDEX_INSERT]  = 16,
2217         /**
2218          * Delete
2219          * just modify a single entry, probably merge few within a block
2220          */
2221         [DTO_INDEX_DELETE]  = 1,
2222         /**
2223          * Used for OI scrub
2224          */
2225         [DTO_INDEX_UPDATE]  = 16,
2226         /**
2227          * 4(inode, inode bits, groups, GDT)
2228          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2229          */
2230         [DTO_OBJECT_CREATE] = 4,
2231         /**
2232          * 4(inode, inode bits, groups, GDT)
2233          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2234          */
2235         [DTO_OBJECT_DELETE] = 4,
2236         /**
2237          * Attr set credits (inode)
2238          */
2239         [DTO_ATTR_SET_BASE] = 1,
2240         /**
2241          * Xattr set. The same as xattr of EXT3.
2242          * DATA_TRANS_BLOCKS(14)
2243          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2244          * are also counted in. Do not know why?
2245          */
2246         [DTO_XATTR_SET]     = 14,
2247         /**
2248          * credits for inode change during write.
2249          */
2250         [DTO_WRITE_BASE]    = 3,
2251         /**
2252          * credits for single block write.
2253          */
2254         [DTO_WRITE_BLOCK]   = 14,
2255         /**
2256          * Attr set credits for chown.
2257          * This is extra credits for setattr, and it is null without quota
2258          */
2259         [DTO_ATTR_SET_CHOWN] = 0
2260 };
2261
2262 static const struct dt_device_operations osd_dt_ops = {
2263         .dt_root_get       = osd_root_get,
2264         .dt_statfs         = osd_statfs,
2265         .dt_trans_create   = osd_trans_create,
2266         .dt_trans_start    = osd_trans_start,
2267         .dt_trans_stop     = osd_trans_stop,
2268         .dt_trans_cb_add   = osd_trans_cb_add,
2269         .dt_conf_get       = osd_conf_get,
2270         .dt_sync           = osd_sync,
2271         .dt_ro             = osd_ro,
2272         .dt_commit_async   = osd_commit_async,
2273 };
2274
2275 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2276                           unsigned role)
2277 {
2278         struct osd_object *obj = osd_dt_obj(dt);
2279         struct osd_thread_info *oti = osd_oti_get(env);
2280
2281         LINVRNT(osd_invariant(obj));
2282
2283         LASSERT(obj->oo_owner != env);
2284         down_read_nested(&obj->oo_sem, role);
2285
2286         LASSERT(obj->oo_owner == NULL);
2287         oti->oti_r_locks++;
2288 }
2289
2290 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2291                            unsigned role)
2292 {
2293         struct osd_object *obj = osd_dt_obj(dt);
2294         struct osd_thread_info *oti = osd_oti_get(env);
2295
2296         LINVRNT(osd_invariant(obj));
2297
2298         LASSERT(obj->oo_owner != env);
2299         down_write_nested(&obj->oo_sem, role);
2300
2301         LASSERT(obj->oo_owner == NULL);
2302         obj->oo_owner = env;
2303         oti->oti_w_locks++;
2304 }
2305
2306 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2307 {
2308         struct osd_object *obj = osd_dt_obj(dt);
2309         struct osd_thread_info *oti = osd_oti_get(env);
2310
2311         LINVRNT(osd_invariant(obj));
2312
2313         LASSERT(oti->oti_r_locks > 0);
2314         oti->oti_r_locks--;
2315         up_read(&obj->oo_sem);
2316 }
2317
2318 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2319 {
2320         struct osd_object *obj = osd_dt_obj(dt);
2321         struct osd_thread_info *oti = osd_oti_get(env);
2322
2323         LINVRNT(osd_invariant(obj));
2324
2325         LASSERT(obj->oo_owner == env);
2326         LASSERT(oti->oti_w_locks > 0);
2327         oti->oti_w_locks--;
2328         obj->oo_owner = NULL;
2329         up_write(&obj->oo_sem);
2330 }
2331
2332 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2333 {
2334         struct osd_object *obj = osd_dt_obj(dt);
2335
2336         LINVRNT(osd_invariant(obj));
2337
2338         return obj->oo_owner == env;
2339 }
2340
2341 static struct timespec *osd_inode_time(const struct lu_env *env,
2342                                        struct inode *inode, __u64 seconds)
2343 {
2344         struct osd_thread_info  *oti = osd_oti_get(env);
2345         struct timespec         *t   = &oti->oti_time;
2346
2347         t->tv_sec = seconds;
2348         t->tv_nsec = 0;
2349         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2350         return t;
2351 }
2352
2353 static void osd_inode_getattr(const struct lu_env *env,
2354                               struct inode *inode, struct lu_attr *attr)
2355 {
2356         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2357                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2358                            LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2359                            LA_BLKSIZE | LA_TYPE;
2360
2361         attr->la_atime   = LTIME_S(inode->i_atime);
2362         attr->la_mtime   = LTIME_S(inode->i_mtime);
2363         attr->la_ctime   = LTIME_S(inode->i_ctime);
2364         attr->la_mode    = inode->i_mode;
2365         attr->la_size    = i_size_read(inode);
2366         attr->la_blocks  = inode->i_blocks;
2367         attr->la_uid     = i_uid_read(inode);
2368         attr->la_gid     = i_gid_read(inode);
2369         attr->la_projid  = i_projid_read(inode);
2370         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2371         attr->la_nlink   = inode->i_nlink;
2372         attr->la_rdev    = inode->i_rdev;
2373         attr->la_blksize = 1 << inode->i_blkbits;
2374         attr->la_blkbits = inode->i_blkbits;
2375         /*
2376          * Ext4 did not transfer inherit flags from raw inode
2377          * to inode flags, and ext4 internally test raw inode
2378          * @i_flags directly. Instead of patching ext4, we do it here.
2379          */
2380         if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2381                 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2382 }
2383
2384 static int osd_attr_get(const struct lu_env *env,
2385                         struct dt_object *dt,
2386                         struct lu_attr *attr)
2387 {
2388         struct osd_object *obj = osd_dt_obj(dt);
2389
2390         if (unlikely(!dt_object_exists(dt)))
2391                 return -ENOENT;
2392         if (unlikely(obj->oo_destroyed))
2393                 return -ENOENT;
2394
2395         LASSERT(!dt_object_remote(dt));
2396         LINVRNT(osd_invariant(obj));
2397
2398         spin_lock(&obj->oo_guard);
2399         osd_inode_getattr(env, obj->oo_inode, attr);
2400         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2401                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2402         spin_unlock(&obj->oo_guard);
2403
2404         return 0;
2405 }
2406
2407 static int osd_declare_attr_qid(const struct lu_env *env,
2408                                 struct osd_object *obj,
2409                                 struct osd_thandle *oh, long long bspace,
2410                                 qid_t old_id, qid_t new_id, bool enforce,
2411                                 unsigned type)
2412 {
2413         int rc;
2414         struct osd_thread_info *info = osd_oti_get(env);
2415         struct lquota_id_info  *qi = &info->oti_qi;
2416
2417         qi->lqi_type = type;
2418         /* inode accounting */
2419         qi->lqi_is_blk = false;
2420
2421         /* one more inode for the new id ... */
2422         qi->lqi_id.qid_uid = new_id;
2423         qi->lqi_space      = 1;
2424         /* Reserve credits for the new id */
2425         rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2426         if (rc == -EDQUOT || rc == -EINPROGRESS)
2427                 rc = 0;
2428         if (rc)
2429                 RETURN(rc);
2430
2431         /* and one less inode for the current id */
2432         qi->lqi_id.qid_uid = old_id;
2433         qi->lqi_space      = -1;
2434         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2435         if (rc == -EDQUOT || rc == -EINPROGRESS)
2436                 rc = 0;
2437         if (rc)
2438                 RETURN(rc);
2439
2440         /* block accounting */
2441         qi->lqi_is_blk = true;
2442
2443         /* more blocks for the new id ... */
2444         qi->lqi_id.qid_uid = new_id;
2445         qi->lqi_space      = bspace;
2446         /*
2447          * Credits for the new uid has been reserved, re-use "obj"
2448          * to save credit reservation.
2449          */
2450         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2451         if (rc == -EDQUOT || rc == -EINPROGRESS)
2452                 rc = 0;
2453         if (rc)
2454                 RETURN(rc);
2455
2456         /* and finally less blocks for the current uid */
2457         qi->lqi_id.qid_uid = old_id;
2458         qi->lqi_space      = -bspace;
2459         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2460         if (rc == -EDQUOT || rc == -EINPROGRESS)
2461                 rc = 0;
2462
2463         RETURN(rc);
2464 }
2465
2466 static int osd_declare_attr_set(const struct lu_env *env,
2467                                 struct dt_object *dt,
2468                                 const struct lu_attr *attr,
2469                                 struct thandle *handle)
2470 {
2471         struct osd_thandle     *oh;
2472         struct osd_object      *obj;
2473         qid_t                   uid;
2474         qid_t                   gid;
2475         long long               bspace;
2476         int                     rc = 0;
2477         bool                    enforce;
2478         ENTRY;
2479
2480         LASSERT(dt != NULL);
2481         LASSERT(handle != NULL);
2482
2483         obj = osd_dt_obj(dt);
2484         LASSERT(osd_invariant(obj));
2485
2486         oh = container_of0(handle, struct osd_thandle, ot_super);
2487         LASSERT(oh->ot_handle == NULL);
2488
2489         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2490                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2491
2492         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2493                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2494
2495         if (attr == NULL || obj->oo_inode == NULL)
2496                 RETURN(rc);
2497
2498         bspace   = obj->oo_inode->i_blocks;
2499         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2500         bspace   = toqb(bspace);
2501
2502         /* Changing ownership is always preformed by super user, it should not
2503          * fail with EDQUOT.
2504          *
2505          * We still need to call the osd_declare_qid() to calculate the journal
2506          * credits for updating quota accounting files and to trigger quota
2507          * space adjustment once the operation is completed.*/
2508         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2509                 /* USERQUOTA */
2510                 uid = i_uid_read(obj->oo_inode);
2511                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2512                 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2513                                           attr->la_uid, enforce, USRQUOTA);
2514                 if (rc)
2515                         RETURN(rc);
2516
2517                 gid = i_gid_read(obj->oo_inode);
2518                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2519                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2520                                           i_gid_read(obj->oo_inode),
2521                                           attr->la_gid, enforce, GRPQUOTA);
2522                 if (rc)
2523                         RETURN(rc);
2524
2525         }
2526 #ifdef HAVE_PROJECT_QUOTA
2527         if (attr->la_valid & LA_PROJID) {
2528                 __u32 projid = i_projid_read(obj->oo_inode);
2529                 enforce = (attr->la_valid & LA_PROJID) &&
2530                                         (attr->la_projid != projid);
2531                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2532                                           (qid_t)projid, (qid_t)attr->la_projid,
2533                                           enforce, PRJQUOTA);
2534                 if (rc)
2535                         RETURN(rc);
2536         }
2537 #endif
2538         RETURN(rc);
2539 }
2540
2541 static int osd_inode_setattr(const struct lu_env *env,
2542                              struct inode *inode, const struct lu_attr *attr)
2543 {
2544         __u64 bits = attr->la_valid;
2545
2546         /* Only allow set size for regular file */
2547         if (!S_ISREG(inode->i_mode))
2548                 bits &= ~(LA_SIZE | LA_BLOCKS);
2549
2550         if (bits == 0)
2551                 return 0;
2552
2553         if (bits & LA_ATIME)
2554                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2555         if (bits & LA_CTIME)
2556                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2557         if (bits & LA_MTIME)
2558                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2559         if (bits & LA_SIZE) {
2560                 spin_lock(&inode->i_lock);
2561                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2562                 i_size_write(inode, attr->la_size);
2563                 spin_unlock(&inode->i_lock);
2564         }
2565
2566         /* OSD should not change "i_blocks" which is used by quota.
2567          * "i_blocks" should be changed by ldiskfs only. */
2568         if (bits & LA_MODE)
2569                 inode->i_mode = (inode->i_mode & S_IFMT) |
2570                                 (attr->la_mode & ~S_IFMT);
2571         if (bits & LA_UID)
2572                 i_uid_write(inode, attr->la_uid);
2573         if (bits & LA_GID)
2574                 i_gid_write(inode, attr->la_gid);
2575         if (bits & LA_PROJID)
2576                 i_projid_write(inode, attr->la_projid);
2577         if (bits & LA_NLINK)
2578                 set_nlink(inode, attr->la_nlink);
2579         if (bits & LA_RDEV)
2580                 inode->i_rdev = attr->la_rdev;
2581
2582         if (bits & LA_FLAGS) {
2583                 /* always keep S_NOCMTIME */
2584                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2585                                  S_NOCMTIME;
2586                 /*
2587                  * Ext4 did not transfer inherit flags from
2588                  * @inode->i_flags to raw inode i_flags when writing
2589                  * flags, we do it explictly here.
2590                  */
2591                 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2592                         LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2593         }
2594         return 0;
2595 }
2596
2597 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2598 {
2599         int rc;
2600
2601         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2602             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2603                 struct iattr    iattr;
2604
2605                 ll_vfs_dq_init(inode);
2606                 iattr.ia_valid = 0;
2607                 if (attr->la_valid & LA_UID)
2608                         iattr.ia_valid |= ATTR_UID;
2609                 if (attr->la_valid & LA_GID)
2610                         iattr.ia_valid |= ATTR_GID;
2611                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2612                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2613
2614                 rc = ll_vfs_dq_transfer(inode, &iattr);
2615                 if (rc) {
2616                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2617                                "enforcement enabled on the ldiskfs "
2618                                "filesystem?\n", inode->i_sb->s_id, rc);
2619                         return rc;
2620                 }
2621         }
2622
2623 #ifdef HAVE_PROJECT_QUOTA
2624         /* Handle project id transfer here properly */
2625         if (attr->la_valid & LA_PROJID &&
2626             attr->la_projid != i_projid_read(inode)) {
2627                 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2628                 if (rc) {
2629                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2630                                "enforcement enabled on the ldiskfs "
2631                                "filesystem?\n", inode->i_sb->s_id, rc);
2632                         return rc;
2633                 }
2634         }
2635 #endif
2636         return 0;
2637 }
2638
2639 static int osd_attr_set(const struct lu_env *env,
2640                         struct dt_object *dt,
2641                         const struct lu_attr *attr,
2642                         struct thandle *handle)
2643 {
2644         struct osd_object *obj = osd_dt_obj(dt);
2645         struct inode      *inode;
2646         int rc;
2647
2648         if (!dt_object_exists(dt))
2649                 return -ENOENT;
2650
2651         LASSERT(handle != NULL);
2652         LASSERT(!dt_object_remote(dt));
2653         LASSERT(osd_invariant(obj));
2654
2655         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2656
2657         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2658                 struct osd_thread_info  *oti  = osd_oti_get(env);
2659                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
2660                 struct lu_fid           *fid1 = &oti->oti_fid;
2661                 struct osd_inode_id     *id   = &oti->oti_id;
2662                 struct iam_path_descr   *ipd;
2663                 struct iam_container    *bag;
2664                 struct osd_thandle      *oh;
2665                 int                      rc;
2666
2667                 fid_cpu_to_be(fid1, fid0);
2668                 memset(id, 1, sizeof(*id));
2669                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2670                                   fid0)->oi_dir.od_container;
2671                 ipd = osd_idx_ipd_get(env, bag);
2672                 if (unlikely(ipd == NULL))
2673                         RETURN(-ENOMEM);
2674
2675                 oh = container_of0(handle, struct osd_thandle, ot_super);
2676                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2677                                 (const struct iam_rec *)id, ipd);
2678                 osd_ipd_put(env, bag, ipd);
2679                 return(rc > 0 ? 0 : rc);
2680         }
2681
2682         inode = obj->oo_inode;
2683
2684         rc = osd_quota_transfer(inode, attr);
2685         if (rc)
2686                 return rc;
2687
2688         spin_lock(&obj->oo_guard);
2689         rc = osd_inode_setattr(env, inode, attr);
2690         spin_unlock(&obj->oo_guard);
2691         if (rc != 0)
2692                 GOTO(out, rc);
2693
2694         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2695
2696         if (!(attr->la_valid & LA_FLAGS))
2697                 GOTO(out, rc);
2698
2699         /* Let's check if there are extra flags need to be set into LMA */
2700         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2701                 struct osd_thread_info *info = osd_oti_get(env);
2702                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2703
2704                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2705                                  &info->oti_ost_attrs);
2706                 if (rc)
2707                         GOTO(out, rc);
2708
2709                 lma->lma_incompat |=
2710                         lustre_to_lma_flags(attr->la_flags);
2711                 lustre_lma_swab(lma);
2712                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2713                                      lma, sizeof(*lma), XATTR_REPLACE);
2714                 if (rc != 0) {
2715                         struct osd_device *osd = osd_obj2dev(obj);
2716
2717                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2718                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2719                               lma->lma_incompat, rc);
2720                 } else {
2721                         obj->oo_lma_flags =
2722                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2723                 }
2724                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2725         }
2726 out:
2727         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2728
2729         return rc;
2730 }
2731
2732 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2733                                            struct osd_object *obj,
2734                                            const char *name, const int namelen)
2735 {
2736         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2737 }
2738
2739 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2740                       umode_t mode, struct dt_allocation_hint *hint,
2741                       struct thandle *th)
2742 {
2743         int result;
2744         struct osd_device  *osd = osd_obj2dev(obj);
2745         struct osd_thandle *oth;
2746         struct dt_object   *parent = NULL;
2747         struct inode       *inode;
2748
2749         LINVRNT(osd_invariant(obj));
2750         LASSERT(obj->oo_inode == NULL);
2751         LASSERT(obj->oo_hl_head == NULL);
2752
2753         if (S_ISDIR(mode) && ldiskfs_pdo) {
2754                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2755                 if (obj->oo_hl_head == NULL)
2756                         return -ENOMEM;
2757         }
2758
2759         oth = container_of(th, struct osd_thandle, ot_super);
2760         LASSERT(oth->ot_handle->h_transaction != NULL);
2761
2762         if (hint != NULL && hint->dah_parent != NULL &&
2763             !dt_object_remote(hint->dah_parent))
2764                 parent = hint->dah_parent;
2765
2766         inode = ldiskfs_create_inode(oth->ot_handle,
2767                                      parent ? osd_dt_obj(parent)->oo_inode :
2768                                               osd_sb(osd)->s_root->d_inode,
2769                                      mode);
2770         if (!IS_ERR(inode)) {
2771                 /* Do not update file c/mtime in ldiskfs. */
2772                 inode->i_flags |= S_NOCMTIME;
2773
2774                 /* For new created object, it must be consistent,
2775                  * and it is unnecessary to scrub against it. */
2776                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2777
2778                 obj->oo_inode = inode;
2779                 result = 0;
2780         } else {
2781                 if (obj->oo_hl_head != NULL) {
2782                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2783                         obj->oo_hl_head = NULL;
2784                 }
2785                 result = PTR_ERR(inode);
2786         }
2787         LINVRNT(osd_invariant(obj));
2788         return result;
2789 }
2790
2791 enum {
2792         OSD_NAME_LEN = 255
2793 };
2794
2795 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2796                      struct lu_attr *attr,
2797                      struct dt_allocation_hint *hint,
2798                      struct dt_object_format *dof,
2799                      struct thandle *th)
2800 {
2801         int result;
2802         struct osd_thandle *oth;
2803         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2804
2805         LASSERT(S_ISDIR(attr->la_mode));
2806
2807         oth = container_of(th, struct osd_thandle, ot_super);
2808         LASSERT(oth->ot_handle->h_transaction != NULL);
2809         result = osd_mkfile(info, obj, mode, hint, th);
2810
2811         return result;
2812 }
2813
2814 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2815                         struct lu_attr *attr,
2816                         struct dt_allocation_hint *hint,
2817                         struct dt_object_format *dof,
2818                         struct thandle *th)
2819 {
2820         int result;
2821         struct osd_thandle *oth;
2822         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2823
2824         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2825
2826         LASSERT(S_ISREG(attr->la_mode));
2827
2828         oth = container_of(th, struct osd_thandle, ot_super);
2829         LASSERT(oth->ot_handle->h_transaction != NULL);
2830
2831         result = osd_mkfile(info, obj, mode, hint, th);
2832         if (result == 0) {
2833                 LASSERT(obj->oo_inode != NULL);
2834                 if (feat->dif_flags & DT_IND_VARKEY)
2835                         result = iam_lvar_create(obj->oo_inode,
2836                                                  feat->dif_keysize_max,
2837                                                  feat->dif_ptrsize,
2838                                                  feat->dif_recsize_max,
2839                                                  oth->ot_handle);
2840                 else
2841                         result = iam_lfix_create(obj->oo_inode,
2842                                                  feat->dif_keysize_max,
2843                                                  feat->dif_ptrsize,
2844                                                  feat->dif_recsize_max,
2845                                                  oth->ot_handle);
2846
2847         }
2848         return result;
2849 }
2850
2851 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2852                      struct lu_attr *attr,
2853                      struct dt_allocation_hint *hint,
2854                      struct dt_object_format *dof,
2855                      struct thandle *th)
2856 {
2857         LASSERT(S_ISREG(attr->la_mode));
2858         return osd_mkfile(info, obj, (attr->la_mode &
2859                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2860 }
2861
2862 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2863                      struct lu_attr *attr,
2864                      struct dt_allocation_hint *hint,
2865                      struct dt_object_format *dof,
2866                      struct thandle *th)
2867 {
2868         LASSERT(S_ISLNK(attr->la_mode));
2869         return osd_mkfile(info, obj, (attr->la_mode &
2870                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2871 }
2872
2873 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2874                      struct lu_attr *attr,
2875                      struct dt_allocation_hint *hint,
2876                      struct dt_object_format *dof,
2877                      struct thandle *th)
2878 {
2879         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2880         int result;
2881
2882         LINVRNT(osd_invariant(obj));
2883         LASSERT(obj->oo_inode == NULL);
2884         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2885                 S_ISFIFO(mode) || S_ISSOCK(mode));
2886
2887         result = osd_mkfile(info, obj, mode, hint, th);
2888         if (result == 0) {
2889                 LASSERT(obj->oo_inode != NULL);
2890                 /*
2891                  * This inode should be marked dirty for i_rdev.  Currently
2892                  * that is done in the osd_attr_init().
2893                  */
2894                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2895                                    attr->la_rdev);
2896         }
2897         LINVRNT(osd_invariant(obj));
2898         return result;
2899 }
2900
2901 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2902                               struct lu_attr *,
2903                               struct dt_allocation_hint *hint,
2904                               struct dt_object_format *dof,
2905                               struct thandle *);
2906
2907 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2908 {
2909         osd_obj_type_f result;
2910
2911         switch (type) {
2912         case DFT_DIR:
2913                 result = osd_mkdir;
2914                 break;
2915         case DFT_REGULAR:
2916                 result = osd_mkreg;
2917                 break;
2918         case DFT_SYM:
2919                 result = osd_mksym;
2920                 break;
2921         case DFT_NODE:
2922                 result = osd_mknod;
2923                 break;
2924         case DFT_INDEX:
2925                 result = osd_mk_index;
2926                 break;
2927
2928         default:
2929                 LBUG();
2930                 break;
2931         }
2932         return result;
2933 }
2934
2935
2936 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2937                         struct dt_object *parent, struct dt_object *child,
2938                         umode_t child_mode)
2939 {
2940         LASSERT(ah);
2941
2942         ah->dah_parent = parent;
2943         ah->dah_mode = child_mode;
2944
2945         if (parent != NULL && !dt_object_remote(parent)) {
2946                 /* will help to find FID->ino at dt_insert("..") */
2947                 struct osd_object *pobj = osd_dt_obj(parent);
2948                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2949         }
2950 }
2951
2952 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2953                           struct lu_attr *attr, struct dt_object_format *dof)
2954 {
2955         struct inode   *inode = obj->oo_inode;
2956         __u64           valid = attr->la_valid;
2957         int             result;
2958
2959         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2960
2961         if (dof->dof_type != DFT_NODE)
2962                 attr->la_valid &= ~LA_RDEV;
2963         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2964                 attr->la_valid &= ~LA_ATIME;
2965         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2966                 attr->la_valid &= ~LA_CTIME;
2967         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2968                 attr->la_valid &= ~LA_MTIME;
2969
2970         result = osd_quota_transfer(inode, attr);
2971         if (result)
2972                 return;
2973
2974         if (attr->la_valid != 0) {
2975                 result = osd_inode_setattr(info->oti_env, inode, attr);
2976                 /*
2977                  * The osd_inode_setattr() should always succeed here.  The
2978                  * only error that could be returned is EDQUOT when we are
2979                  * trying to change the UID or GID of the inode. However, this
2980                  * should not happen since quota enforcement is no longer
2981                  * enabled on ldiskfs (lquota takes care of it).
2982                  */
2983                 LASSERTF(result == 0, "%d\n", result);
2984                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2985         }
2986
2987         attr->la_valid = valid;
2988 }
2989
2990 /**
2991  * Helper function for osd_create()
2992  *
2993  * \retval 0, on success
2994  */
2995 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
2996                         struct lu_attr *attr, struct dt_allocation_hint *hint,
2997                         struct dt_object_format *dof, struct thandle *th)
2998 {
2999         int     result;
3000         __u32   umask;
3001
3002         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3003
3004         /* we drop umask so that permissions we pass are not affected */
3005         umask = current->fs->umask;
3006         current->fs->umask = 0;
3007
3008         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3009                                                   th);
3010         if (likely(obj->oo_inode != NULL)) {
3011                 LASSERT(obj->oo_inode->i_state & I_NEW);
3012
3013                 /* Unlock the inode before attr initialization to avoid
3014                  * unnecessary dqget operations. LU-6378 */
3015                 unlock_new_inode(obj->oo_inode);
3016         }
3017
3018         if (likely(result == 0)) {
3019                 osd_attr_init(info, obj, attr, dof);
3020                 osd_object_init0(obj);
3021         }
3022
3023         /* restore previous umask value */
3024         current->fs->umask = umask;
3025
3026         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3027
3028         return result;
3029 }
3030
3031 /**
3032  * Helper function for osd_create()
3033  *
3034  * \retval 0, on success
3035  */
3036 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3037                            const struct lu_fid *fid, struct thandle *th)
3038 {
3039         struct osd_thread_info *info = osd_oti_get(env);
3040         struct osd_inode_id    *id   = &info->oti_id;
3041         struct osd_device      *osd  = osd_obj2dev(obj);
3042         struct osd_thandle     *oh;
3043         int                     rc;
3044
3045         LASSERT(obj->oo_inode != NULL);
3046
3047         oh = container_of0(th, struct osd_thandle, ot_super);
3048         LASSERT(oh->ot_handle);
3049         osd_trans_exec_op(env, th, OSD_OT_INSERT);
3050
3051         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3052         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3053                            OI_CHECK_FLD, NULL);
3054         osd_trans_exec_check(env, th, OSD_OT_INSERT);
3055
3056         return rc;
3057 }
3058
3059 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3060                    u64 seq, struct lu_seq_range *range)
3061 {
3062         struct seq_server_site  *ss = osd_seq_site(osd);
3063
3064         if (fid_seq_is_idif(seq)) {
3065                 fld_range_set_ost(range);
3066                 range->lsr_index = idif_ost_idx(seq);
3067                 return 0;
3068         }
3069
3070         if (!fid_seq_in_fldb(seq)) {
3071                 fld_range_set_mdt(range);
3072                 if (ss != NULL)
3073                         /* FIXME: If ss is NULL, it suppose not get lsr_index
3074                          * at all */
3075                         range->lsr_index = ss->ss_node_id;
3076                 return 0;
3077         }
3078
3079         LASSERT(ss != NULL);
3080         fld_range_set_any(range);
3081         /* OSD will only do local fld lookup */
3082         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3083 }
3084
3085 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3086                               struct lu_attr *attr,
3087                               struct dt_allocation_hint *hint,
3088                               struct dt_object_format *dof,
3089                               struct thandle *handle)
3090 {
3091         struct osd_thandle *oh;
3092         int rc;
3093         ENTRY;
3094
3095         LASSERT(handle != NULL);
3096
3097         oh = container_of0(handle, struct osd_thandle, ot_super);
3098         LASSERT(oh->ot_handle == NULL);
3099
3100         /* EA object consumes more credits than regular object: osd_mk_index
3101          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3102          * leaf_node, could involves the block, block bitmap, groups, GDT
3103          * change for each block, so add 4 * 2 credits in that case. */
3104         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3105                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3106                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3107         /* Reuse idle OI block may cause additional one OI block
3108          * to be changed. */
3109         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3110                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3111
3112         if (!attr)
3113                 RETURN(0);
3114
3115         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3116                                    attr->la_projid, 1, oh, osd_dt_obj(dt),
3117                                    NULL, OSD_QID_INODE);
3118         if (rc != 0)
3119                 RETURN(rc);
3120
3121         /* will help to find FID->ino mapping at dt_insert() */
3122         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3123                                    osd_dt_obj(dt));
3124
3125         RETURN(rc);
3126 }
3127
3128 /**
3129  * Called to destroy on-disk representation of the object
3130  *
3131  * Concurrency: must be locked
3132  */
3133 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3134                                struct thandle *th)
3135 {
3136         struct osd_object  *obj = osd_dt_obj(dt);
3137         struct inode       *inode = obj->oo_inode;
3138         struct osd_thandle *oh;
3139         int                 rc;
3140         ENTRY;
3141
3142         if (inode == NULL)
3143                 RETURN(-ENOENT);
3144
3145         oh = container_of0(th, struct osd_thandle, ot_super);
3146         LASSERT(oh->ot_handle == NULL);
3147
3148         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3149                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3150         /* Recycle idle OI leaf may cause additional three OI blocks
3151          * to be changed. */
3152         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3153                 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3154                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3155         /* one less inode */
3156         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3157                                    i_projid_read(inode), -1, oh, obj, NULL,
3158                                    OSD_QID_INODE);
3159         if (rc)
3160                 RETURN(rc);
3161         /* data to be truncated */
3162         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3163                                    i_projid_read(inode), 0, oh, obj, NULL,
3164                                    OSD_QID_BLK);
3165         if (rc)
3166                 RETURN(rc);
3167
3168         /* will help to find FID->ino when this object is being
3169          * added to PENDING/ */
3170         rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3171
3172         RETURN(rc);
3173 }
3174
3175 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3176                        struct thandle *th)
3177 {
3178         const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3179         struct osd_object *obj = osd_dt_obj(dt);
3180         struct inode *inode = obj->oo_inode;
3181         struct osd_device *osd = osd_obj2dev(obj);
3182         struct osd_thandle *oh;
3183         int result;
3184         ENTRY;
3185
3186         oh = container_of0(th, struct osd_thandle, ot_super);
3187         LASSERT(oh->ot_handle);
3188         LASSERT(inode);
3189         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3190
3191         if (unlikely(fid_is_acct(fid)))
3192                 RETURN(-EPERM);
3193
3194         if (S_ISDIR(inode->i_mode)) {
3195                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3196                         inode->i_nlink == 2);
3197                 /* it will check/delete the inode from remote parent,
3198                  * how to optimize it? unlink performance impaction XXX */
3199                 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3200                 if (result != 0)
3201                         CERROR("%s: delete inode "DFID": rc = %d\n",
3202                                osd_name(osd), PFID(fid), result);
3203
3204                 spin_lock(&obj->oo_guard);
3205                 clear_nlink(inode);
3206                 spin_unlock(&obj->oo_guard);
3207                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3208         }
3209
3210         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3211
3212         ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3213
3214         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3215                 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3216                                        oh->ot_handle, OI_CHECK_FLD);
3217
3218         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3219         /* XXX: add to ext3 orphan list */
3220         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3221
3222         /* not needed in the cache anymore */
3223         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3224         obj->oo_destroyed = 1;
3225
3226         RETURN(0);
3227 }
3228
3229 /**
3230  * Put the fid into lustre_mdt_attrs, and then place the structure
3231  * inode's ea. This fid should not be altered during the life time
3232  * of the inode.
3233  *
3234  * \retval +ve, on success
3235  * \retval -ve, on error
3236  *
3237  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3238  */
3239 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3240                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
3241 {
3242         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3243         struct lustre_mdt_attrs *lma = &loa->loa_lma;
3244         int rc;
3245         ENTRY;
3246
3247         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3248                 RETURN(0);
3249
3250         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3251                 rc = -ENOMEM;
3252
3253         lustre_loa_init(loa, fid, compat, incompat);
3254         lustre_loa_swab(loa, false);
3255
3256         /* For the OST device with 256 bytes inode size by default,
3257          * the PFID EA will be stored together with LMA EA to avoid
3258          * performance trouble. Otherwise the PFID EA can be stored
3259          * independently. LU-8998 */
3260         if ((compat & LMAC_FID_ON_OST) &&
3261             LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3262                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3263                                      sizeof(*loa), XATTR_CREATE);
3264         else
3265                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3266                                      sizeof(*lma), XATTR_CREATE);
3267         /* LMA may already exist, but we need to check that all the
3268          * desired compat/incompat flags have been added. */
3269         if (unlikely(rc == -EEXIST)) {
3270                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3271                                      XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3272                 if (rc < 0)
3273                         RETURN(rc);
3274
3275                 if (rc < sizeof(*lma))
3276                         RETURN(-EINVAL);
3277
3278                 lustre_loa_swab(loa, true);
3279                 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3280                     ((compat == 0 && incompat == 0) ||
3281                      (!(~lma->lma_compat & compat) &&
3282                       !(~lma->lma_incompat & incompat))))
3283                         RETURN(0);
3284
3285                 lma->lma_self_fid = *fid;
3286                 lma->lma_compat |= compat;
3287                 lma->lma_incompat |= incompat;
3288                 if (rc == sizeof(*lma)) {
3289                         lustre_lma_swab(lma);
3290                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3291                                              sizeof(*lma), XATTR_REPLACE);
3292                 } else {
3293                         lustre_loa_swab(loa, false);
3294                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3295                                              sizeof(*loa), XATTR_REPLACE);
3296                 }
3297         }
3298
3299         RETURN(rc);
3300 }
3301
3302 /**
3303  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3304  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3305  * To have compatilibility with 1.8 ldiskfs driver we need to have
3306  * magic number at start of fid data.
3307  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3308  * its inmemory API.
3309  */
3310 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3311                                          const struct lu_fid *fid)
3312 {
3313         if (!fid_is_namespace_visible(fid) ||
3314             OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3315                 param->edp_magic = 0;
3316                 return;
3317         }
3318
3319         param->edp_magic = LDISKFS_LUFID_MAGIC;
3320         param->edp_len =  sizeof(struct lu_fid) + 1;
3321         fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3322 }
3323
3324 /**
3325  * Try to read the fid from inode ea into dt_rec.
3326  *
3327  * \param fid object fid.
3328  *
3329  * \retval 0 on success
3330  */
3331 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3332                           __u32 ino, struct lu_fid *fid,
3333                           struct osd_inode_id *id)
3334 {
3335         struct osd_thread_info *info  = osd_oti_get(env);
3336         struct inode           *inode;
3337         ENTRY;
3338
3339         osd_id_gen(id, ino, OSD_OII_NOGEN);
3340         inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3341         if (IS_ERR(inode))
3342                 RETURN(PTR_ERR(inode));
3343
3344         iput(inode);
3345         RETURN(0);
3346 }
3347
3348 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3349                                         struct inode *dir,
3350                                         struct inode *parent_dir,
3351                                         const struct lu_fid *dot_fid,
3352                                         const struct lu_fid *dot_dot_fid,
3353                                         struct osd_thandle *oth)
3354 {
3355         struct ldiskfs_dentry_param *dot_ldp;
3356         struct ldiskfs_dentry_param *dot_dot_ldp;
3357         __u32 saved_nlink = dir->i_nlink;
3358         int rc;
3359
3360         dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3361         osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3362
3363         dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3364         dot_ldp->edp_magic = 0;
3365
3366         rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3367                                     dir, dot_ldp, dot_dot_ldp);
3368         /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3369          * the subseqent ref_add() will increase the dir->i_nlink
3370          * as 3. That is incorrect for new created directory.
3371          *
3372          * It looks like hack, because we want to make the OSD API
3373          * to be order-independent for new created directory object
3374          * between dt_insert(..) and ref_add() operations.
3375          *
3376          * Here, we only restore the in-RAM dir-inode's nlink attr,
3377          * becuase if the nlink attr is not 2, then there will be
3378          * ref_add() called following the dt_insert(..), such call
3379          * will make both the in-RAM and on-disk dir-inode's nlink
3380          * attr to be set as 2. LU-7447 */
3381         set_nlink(dir, saved_nlink);
3382         return rc;
3383 }
3384
3385 /**
3386  * Create an local agent inode for remote entry
3387  */
3388 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3389                                                   struct osd_device *osd,
3390                                                   struct osd_object *pobj,
3391                                                   const struct lu_fid *fid,
3392                                                   __u32 type,
3393                                                   struct thandle *th)
3394 {
3395         struct osd_thread_info  *info = osd_oti_get(env);
3396         struct inode            *local;
3397         struct osd_thandle      *oh;
3398         int                     rc;
3399         ENTRY;
3400
3401         LASSERT(th);
3402         oh = container_of(th, struct osd_thandle, ot_super);
3403         LASSERT(oh->ot_handle->h_transaction != NULL);
3404
3405         local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3406         if (IS_ERR(local)) {
3407                 CERROR("%s: create local error %d\n", osd_name(osd),
3408                        (int)PTR_ERR(local));
3409                 RETURN(local);
3410         }
3411
3412         /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3413          * correct gid on remote file, not agent here */
3414         local->i_gid = current_fsgid();
3415         ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3416         unlock_new_inode(local);
3417
3418         /* Agent inode should not have project ID */
3419 #ifdef  HAVE_PROJECT_QUOTA
3420         if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3421                 rc = __ldiskfs_ioctl_setproject(local, 0);
3422                 if (rc) {
3423                         CERROR("%s: quota transfer failed: rc = %d. Is project "
3424                                "quota enforcement enabled on the ldiskfs "
3425                                "filesystem?\n", local->i_sb->s_id, rc);
3426                         RETURN(ERR_PTR(rc));
3427                 }
3428         }
3429 #endif
3430         /* Set special LMA flag for local agent inode */
3431         rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3432         if (rc != 0) {
3433                 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3434                        osd_name(osd), PFID(fid), rc);
3435                 RETURN(ERR_PTR(rc));
3436         }
3437
3438         if (!S_ISDIR(type))
3439                 RETURN(local);
3440
3441         rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3442                                          lu_object_fid(&pobj->oo_dt.do_lu),
3443                                          fid, oh);
3444         if (rc != 0) {
3445                 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3446                         osd_name(osd), PFID(fid), rc);
3447                 RETURN(ERR_PTR(rc));
3448         }
3449
3450         RETURN(local);
3451 }
3452
3453 /**
3454  * when direntry is deleted, we have to take care of possible agent inode
3455  * referenced by that. unfortunately we can't do this at that point:
3456  * iget() within a running transaction leads to deadlock and we better do
3457  * not call that every delete declaration to save performance. so we put
3458  * a potention agent inode on a list and process that once the transaction
3459  * is over. Notice it's not any worse in terms of real orphans as regular
3460  * object destroy doesn't put inodes on the on-disk orphan list. this should
3461  * be addressed separately
3462  */
3463 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3464                                             struct osd_thandle *oh,
3465                                             __u32 ino)
3466 {
3467         struct osd_device      *osd = osd_dt_dev(oh->ot_super.th_dev);
3468         struct osd_obj_orphan *oor;
3469
3470         OBD_ALLOC_PTR(oor);
3471         if (oor == NULL)
3472                 return -ENOMEM;
3473
3474         oor->oor_ino = ino;
3475         oor->oor_env = (struct lu_env *)env;
3476         spin_lock(&osd->od_osfs_lock);
3477         list_add(&oor->oor_list, &osd->od_orphan_list);
3478         spin_unlock(&osd->od_osfs_lock);
3479
3480         oh->ot_remove_agents = 1;
3481
3482         return 0;
3483
3484 }
3485
3486 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3487                                                 struct osd_device *osd)
3488 {
3489         struct osd_thread_info *info = osd_oti_get(env);
3490         struct osd_obj_orphan *oor, *tmp;
3491         struct osd_inode_id id;
3492         struct list_head list;
3493         struct inode *inode;
3494         struct lu_fid fid;
3495         handle_t *jh;
3496         __u32 ino;
3497
3498         INIT_LIST_HEAD(&list);
3499
3500         spin_lock(&osd->od_osfs_lock);
3501         list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3502                 if (oor->oor_env == env) {
3503                         list_del(&oor->oor_list);
3504                         list_add(&oor->oor_list, &list);
3505                 }
3506         }
3507         spin_unlock(&osd->od_osfs_lock);
3508
3509         list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3510
3511                 ino = oor->oor_ino;
3512
3513                 list_del(&oor->oor_list);
3514                 OBD_FREE_PTR(oor);
3515
3516                 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3517                 inode = osd_iget_fid(info, osd, &id, &fid);
3518                 if (IS_ERR(inode))
3519                         continue;
3520
3521                 if (!osd_remote_fid(env, osd, &fid)) {
3522                         iput(inode);
3523                         continue;
3524                 }
3525
3526                 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3527                 clear_nlink(inode);
3528                 mark_inode_dirty(inode);
3529                 ldiskfs_journal_stop(jh);
3530                 iput(inode);
3531         }
3532
3533         return 0;
3534 }
3535
3536 /**
3537  * OSD layer object create function for OST objects (b=11826).
3538  *
3539  * The FID is inserted into inode xattr here.
3540  *
3541  * \retval   0, on success
3542  * \retval -ve, on error
3543  */
3544 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3545                       struct lu_attr *attr, struct dt_allocation_hint *hint,
3546                       struct dt_object_format *dof, struct thandle *th)
3547 {
3548         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3549         struct osd_object       *obj    = osd_dt_obj(dt);
3550         struct osd_thread_info  *info   = osd_oti_get(env);
3551         int                      result, on_ost = 0;
3552
3553         ENTRY;
3554
3555         if (dt_object_exists(dt))
3556                 RETURN(-EEXIST);
3557
3558         LINVRNT(osd_invariant(obj));
3559         LASSERT(!dt_object_remote(dt));
3560         LASSERT(osd_is_write_locked(env, obj));
3561         LASSERT(th != NULL);
3562
3563         if (unlikely(fid_is_acct(fid)))
3564                 /* Quota files can't be created from the kernel any more,
3565                  * 'tune2fs -O quota' will take care of creating them */
3566                 RETURN(-EPERM);
3567
3568         result = __osd_create(info, obj, attr, hint, dof, th);
3569         if (result == 0) {
3570                 if (fid_is_idif(fid) &&
3571                     !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3572                         struct lu_fid *tfid = &info->oti_fid;
3573                         struct ost_id *oi   = &info->oti_ostid;
3574
3575                         fid_to_ostid(fid, oi);
3576                         ostid_to_fid(tfid, oi, 0);
3577                         on_ost = 1;
3578                         result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3579                                                 LMAC_FID_ON_OST, 0);
3580                 } else {
3581                         on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3582                                                fid, OI_CHECK_FLD);
3583                         result = osd_ea_fid_set(info, obj->oo_inode, fid,
3584                                                 on_ost ? LMAC_FID_ON_OST : 0,
3585                                                 0);
3586                 }
3587                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3588                         obj->oo_dt.do_body_ops = &osd_body_ops;
3589         }
3590
3591         if (result == 0)
3592                 result = __osd_oi_insert(env, obj, fid, th);
3593
3594         /* a small optimization - dt_insert() isn't usually applied
3595          * to OST objects, so we don't need to cache OI mapping for
3596          * OST objects */
3597         if (result == 0 && on_ost == 0) {
3598                 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3599                 result = osd_idc_find_and_init(env, osd, obj);
3600                 LASSERT(result == 0);
3601         }
3602
3603         LASSERT(ergo(result == 0,
3604                      dt_object_exists(dt) && !dt_object_remote(dt)));
3605         LINVRNT(osd_invariant(obj));
3606         RETURN(result);
3607 }
3608
3609 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3610                                struct thandle *handle)
3611 {
3612         struct osd_thandle *oh;
3613
3614         /* it's possible that object doesn't exist yet */
3615         LASSERT(handle != NULL);
3616
3617         oh = container_of0(handle, struct osd_thandle, ot_super);
3618         LASSERT(oh->ot_handle == NULL);
3619
3620         osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3621                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3622
3623         return 0;
3624 }
3625
3626 /*
3627  * Concurrency: @dt is write locked.
3628  */
3629 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3630                        struct thandle *th)
3631 {
3632         struct osd_object  *obj = osd_dt_obj(dt);
3633         struct inode       *inode = obj->oo_inode;
3634         struct osd_thandle *oh;
3635         int                 rc = 0;
3636
3637         if (!dt_object_exists(dt) || obj->oo_destroyed)
3638                 return -ENOENT;
3639
3640         LINVRNT(osd_invariant(obj));
3641         LASSERT(!dt_object_remote(dt));
3642         LASSERT(osd