Whamcloud - gitweb
1a82aee6973c56f78f03db3397b37538fb70a274
[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 /* prerequisite for linux/xattr.h */
49 #include <linux/types.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/fs.h>
52 /* XATTR_{REPLACE,CREATE} */
53 #include <linux/xattr.h>
54
55 #include <ldiskfs/ldiskfs.h>
56 #include <ldiskfs/xattr.h>
57 #include <ldiskfs/ldiskfs_extents.h>
58 #undef ENTRY
59 /*
60  * struct OBD_{ALLOC,FREE}*()
61  * OBD_FAIL_CHECK
62  */
63 #include <obd_support.h>
64 /* struct ptlrpc_thread */
65 #include <lustre_net.h>
66 #include <lustre_fid.h>
67 /* process_config */
68 #include <uapi/linux/lustre/lustre_param.h>
69
70 #include "osd_internal.h"
71 #include "osd_dynlocks.h"
72
73 /* llo_* api support */
74 #include <md_object.h>
75 #include <lustre_quota.h>
76
77 #include <lustre_linkea.h>
78
79 #define PFID_STRIPE_IDX_BITS    16
80 #define PFID_STRIPE_COUNT_MASK  ((1 << PFID_STRIPE_IDX_BITS) - 1)
81
82 int ldiskfs_pdo = 1;
83 module_param(ldiskfs_pdo, int, 0644);
84 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
85
86 int ldiskfs_track_declares_assert;
87 module_param(ldiskfs_track_declares_assert, int, 0644);
88 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
89
90 /* Slab to allocate dynlocks */
91 struct kmem_cache *dynlock_cachep;
92
93 /* Slab to allocate osd_it_ea */
94 struct kmem_cache *osd_itea_cachep;
95
96 static struct lu_kmem_descr ldiskfs_caches[] = {
97         {
98                 .ckd_cache = &dynlock_cachep,
99                 .ckd_name  = "dynlock_cache",
100                 .ckd_size  = sizeof(struct dynlock_handle)
101         },
102         {
103                 .ckd_cache = &osd_itea_cachep,
104                 .ckd_name  = "osd_itea_cache",
105                 .ckd_size  = sizeof(struct osd_it_ea)
106         },
107         {
108                 .ckd_cache = NULL
109         }
110 };
111
112 static const char dot[] = ".";
113 static const char dotdot[] = "..";
114
115 static const struct lu_object_operations      osd_lu_obj_ops;
116 static const struct dt_object_operations      osd_obj_ops;
117 static const struct dt_object_operations      osd_obj_otable_it_ops;
118 static const struct dt_index_operations       osd_index_iam_ops;
119 static const struct dt_index_operations       osd_index_ea_ops;
120
121 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
122                           const struct lu_fid *fid);
123 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
124                                                 struct osd_device *osd);
125
126 int osd_trans_declare_op2rb[] = {
127         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
128         [OSD_OT_PUNCH]          = OSD_OT_MAX,
129         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
130         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
131         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
132         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
133         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
134         [OSD_OT_WRITE]          = OSD_OT_WRITE,
135         [OSD_OT_INSERT]         = OSD_OT_DELETE,
136         [OSD_OT_DELETE]         = OSD_OT_INSERT,
137         [OSD_OT_QUOTA]          = OSD_OT_MAX,
138 };
139
140 static int osd_has_index(const struct osd_object *obj)
141 {
142         return obj->oo_dt.do_index_ops != NULL;
143 }
144
145 static int osd_object_invariant(const struct lu_object *l)
146 {
147         return osd_invariant(osd_obj(l));
148 }
149
150 /*
151  * Concurrency: doesn't matter
152  */
153 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
154 {
155         struct osd_thread_info *oti = osd_oti_get(env);
156         return oti->oti_w_locks > 0 && o->oo_owner == env;
157 }
158
159 /*
160  * Concurrency: doesn't access mutable data
161  */
162 static int osd_root_get(const struct lu_env *env,
163                         struct dt_device *dev, struct lu_fid *f)
164 {
165         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
166         return 0;
167 }
168
169 /*
170  * the following set of functions are used to maintain per-thread
171  * cache of FID->ino mapping. this mechanism is needed to resolve
172  * FID to inode at dt_insert() which in turn stores ino in the
173  * directory entries to keep ldiskfs compatible with ext[34].
174  * due to locking-originated restrictions we can't lookup ino
175  * using LU cache (deadlock is possible). lookup using OI is quite
176  * expensive. so instead we maintain this cache and methods like
177  * dt_create() fill it. so in the majority of cases dt_insert() is
178  * able to find needed mapping in lockless manner.
179  */
180 static struct osd_idmap_cache *
181 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
182              const struct lu_fid *fid)
183 {
184         struct osd_thread_info  *oti   = osd_oti_get(env);
185         struct osd_idmap_cache  *idc    = oti->oti_ins_cache;
186         int i;
187         for (i = 0; i < oti->oti_ins_cache_used; i++) {
188                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
189                         continue;
190                 if (idc[i].oic_dev != osd)
191                         continue;
192
193                 return idc + i;
194         }
195
196         return NULL;
197 }
198
199 static struct osd_idmap_cache *
200 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
201             const struct lu_fid *fid)
202 {
203         struct osd_thread_info  *oti   = osd_oti_get(env);
204         struct osd_idmap_cache  *idc;
205         int i;
206
207         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
208                 i = oti->oti_ins_cache_size * 2;
209                 if (i == 0)
210                         i = OSD_INS_CACHE_SIZE;
211                 OBD_ALLOC(idc, sizeof(*idc) * i);
212                 if (idc == NULL)
213                         return ERR_PTR(-ENOMEM);
214                 if (oti->oti_ins_cache != NULL) {
215                         memcpy(idc, oti->oti_ins_cache,
216                                oti->oti_ins_cache_used * sizeof(*idc));
217                         OBD_FREE(oti->oti_ins_cache,
218                                  oti->oti_ins_cache_used * sizeof(*idc));
219                 }
220                 oti->oti_ins_cache = idc;
221                 oti->oti_ins_cache_size = i;
222         }
223
224         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
225         idc->oic_fid = *fid;
226         idc->oic_dev = osd;
227         idc->oic_lid.oii_ino = 0;
228         idc->oic_lid.oii_gen = 0;
229         idc->oic_remote = 0;
230
231         return idc;
232 }
233
234 /*
235  * lookup mapping for the given fid in the cache, initialize a
236  * new one if not found. the initialization checks whether the
237  * object is local or remote. for local objects, OI is used to
238  * learn ino/generation. the function is used when the caller
239  * has no information about the object, e.g. at dt_insert().
240  */
241 static struct osd_idmap_cache *
242 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
243                      const struct lu_fid *fid)
244 {
245         struct osd_idmap_cache *idc;
246         int rc;
247
248         idc = osd_idc_find(env, osd, fid);
249         LASSERT(!IS_ERR(idc));
250         if (idc != NULL)
251                 return idc;
252
253         /* new mapping is needed */
254         idc = osd_idc_add(env, osd, fid);
255         if (IS_ERR(idc))
256                 return idc;
257
258         /* initialize it */
259         rc = osd_remote_fid(env, osd, fid);
260         if (unlikely(rc < 0))
261                 return ERR_PTR(rc);
262
263         if (rc == 0) {
264                 /* the object is local, lookup in OI */
265                 /* XXX: probably cheaper to lookup in LU first? */
266                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
267                                    &idc->oic_lid, 0);
268                 if (unlikely(rc < 0)) {
269                         CERROR("can't lookup: rc = %d\n", rc);
270                         return ERR_PTR(rc);
271                 }
272         } else {
273                 /* the object is remote */
274                 idc->oic_remote = 1;
275         }
276
277         return idc;
278 }
279
280 /*
281  * lookup mapping for given FID and fill it from the given object.
282  * the object is lolcal by definition.
283  */
284 static int osd_idc_find_and_init(const struct lu_env *env,
285                                  struct osd_device *osd,
286                                  struct osd_object *obj)
287 {
288         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
289         struct osd_idmap_cache  *idc;
290
291         idc = osd_idc_find(env, osd, fid);
292         LASSERT(!IS_ERR(idc));
293         if (idc != NULL) {
294                 if (obj->oo_inode == NULL)
295                         return 0;
296                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
297                         LASSERT(idc->oic_lid.oii_ino == 0);
298                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
299                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
300                 }
301                 return 0;
302         }
303
304         /* new mapping is needed */
305         idc = osd_idc_add(env, osd, fid);
306         if (IS_ERR(idc))
307                 return PTR_ERR(idc);
308
309         if (obj->oo_inode != NULL) {
310                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
311                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
312         }
313         return 0;
314 }
315
316 /*
317  * OSD object methods.
318  */
319
320 /*
321  * Concurrency: no concurrent access is possible that early in object
322  * life-cycle.
323  */
324 static struct lu_object *osd_object_alloc(const struct lu_env *env,
325                                           const struct lu_object_header *hdr,
326                                           struct lu_device *d)
327 {
328         struct osd_object *mo;
329
330         OBD_ALLOC_PTR(mo);
331         if (mo != NULL) {
332                 struct lu_object *l;
333
334                 l = &mo->oo_dt.do_lu;
335                 dt_object_init(&mo->oo_dt, NULL, d);
336                 mo->oo_dt.do_ops = &osd_obj_ops;
337                 l->lo_ops = &osd_lu_obj_ops;
338                 init_rwsem(&mo->oo_sem);
339                 init_rwsem(&mo->oo_ext_idx_sem);
340                 spin_lock_init(&mo->oo_guard);
341                 INIT_LIST_HEAD(&mo->oo_xattr_list);
342                 return l;
343         } else {
344                 return NULL;
345         }
346 }
347
348 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
349                 struct dentry *dentry, struct lustre_ost_attrs *loa)
350 {
351         int rc;
352
353         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
354                              (void *)loa, sizeof(*loa));
355         if (rc > 0) {
356                 struct lustre_mdt_attrs *lma = &loa->loa_lma;
357
358                 if (rc < sizeof(*lma))
359                         return -EINVAL;
360
361                 rc = 0;
362                 lustre_loa_swab(loa, true);
363                 /* Check LMA compatibility */
364                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
365                         CWARN("%s: unsupported incompat LMA feature(s) %#x "
366                               "for fid = "DFID", ino = %lu\n",
367                               osd_ino2name(inode),
368                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
369                               PFID(&lma->lma_self_fid), inode->i_ino);
370                         rc = -EOPNOTSUPP;
371                 }
372         } else if (rc == 0) {
373                 rc = -ENODATA;
374         }
375
376         return rc;
377 }
378
379 /*
380  * retrieve object from backend ext fs.
381  **/
382 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
383                        struct osd_inode_id *id)
384 {
385         int rc;
386         struct inode *inode = NULL;
387
388         /* if we look for an inode withing a running
389          * transaction, then we risk to deadlock */
390         /* osd_dirent_check_repair() breaks this */
391         /*LASSERT(current->journal_info == NULL);*/
392
393         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
394         if (IS_ERR(inode)) {
395                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
396                        id->oii_ino, PTR_ERR(inode));
397         } else if (id->oii_gen != OSD_OII_NOGEN &&
398                    inode->i_generation != id->oii_gen) {
399                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
400                        "i_generation = %u\n",
401                        id->oii_ino, id->oii_gen, inode->i_generation);
402                 iput(inode);
403                 inode = ERR_PTR(-ESTALE);
404         } else if (inode->i_nlink == 0) {
405                 /* due to parallel readdir and unlink,
406                 * we can have dead inode here. */
407                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
408                 iput(inode);
409                 inode = ERR_PTR(-ESTALE);
410         } else if (is_bad_inode(inode)) {
411                 CWARN("%s: bad inode: ino = %u\n",
412                 osd_dev2name(dev), id->oii_ino);
413                 iput(inode);
414                 inode = ERR_PTR(-ENOENT);
415         } else if ((rc = osd_attach_jinode(inode))) {
416                 iput(inode);
417                 inode = ERR_PTR(rc);
418         } else {
419                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
420                 if (id->oii_gen == OSD_OII_NOGEN)
421                         osd_id_gen(id, inode->i_ino, inode->i_generation);
422
423                 /* Do not update file c/mtime in ldiskfs.
424                  * NB: we don't have any lock to protect this because we don't
425                  * have reference on osd_object now, but contention with
426                  * another lookup + attr_set can't happen in the tiny window
427                  * between if (...) and set S_NOCMTIME. */
428                 if (!(inode->i_flags & S_NOCMTIME))
429                         inode->i_flags |= S_NOCMTIME;
430         }
431         return inode;
432 }
433
434 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
435                           handle_t *handle, struct dentry *child,
436                           struct inode *inode, struct htree_lock *hlock)
437 {
438         int rc, rc2;
439
440         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
441         if (rc == -ENOBUFS || rc == -ENOSPC) {
442                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
443                 struct inode *parent = child->d_parent->d_inode;
444                 struct lu_fid *fid = NULL;
445
446                 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
447                 if (!rc2) {
448                         fid = &loa->loa_lma.lma_self_fid;
449                 } else if (rc2 == -ENODATA) {
450                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
451                                 fid = &info->oti_fid3;
452                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
453                         } else if (!osd->od_is_ost && osd->od_index == 0) {
454                                 fid = &info->oti_fid3;
455                                 lu_igif_build(fid, parent->i_ino,
456                                               parent->i_generation);
457                         }
458                 }
459
460                 if (fid != NULL)
461                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
462                               "maximum entry limit\n",
463                               osd_name(osd), parent->i_ino, PFID(fid),
464                               rc == -ENOSPC ? "has reached" : "is approaching");
465                 else
466                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
467                               "maximum entry limit\n",
468                               osd_name(osd), parent->i_ino,
469                               rc == -ENOSPC ? "has reached" : "is approaching");
470
471                 /* ignore such error now */
472                 if (rc == -ENOBUFS)
473                         rc = 0;
474         }
475
476         return rc;
477 }
478
479
480 static struct inode *
481 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
482              struct osd_inode_id *id, struct lu_fid *fid)
483 {
484         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
485         struct inode *inode;
486         int rc;
487
488         inode = osd_iget(info, dev, id);
489         if (IS_ERR(inode))
490                 return inode;
491
492         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
493         if (!rc) {
494                 *fid = loa->loa_lma.lma_self_fid;
495         } else if (rc == -ENODATA) {
496                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
497                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
498                 else
499                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
500         } else {
501                 iput(inode);
502                 inode = ERR_PTR(rc);
503         }
504         return inode;
505 }
506
507 static struct inode *osd_iget_check(struct osd_thread_info *info,
508                                     struct osd_device *dev,
509                                     const struct lu_fid *fid,
510                                     struct osd_inode_id *id,
511                                     bool trusted)
512 {
513         struct inode *inode;
514         int rc = 0;
515         ENTRY;
516
517         /* The cached OI mapping is trustable. If we cannot locate the inode
518          * via the cached OI mapping, then return the failure to the caller
519          * directly without further OI checking. */
520
521 again:
522         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
523         if (IS_ERR(inode)) {
524                 rc = PTR_ERR(inode);
525                 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
526                         goto check_oi;
527
528                 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
529                        PFID(fid), id->oii_ino, rc);
530                 GOTO(put, rc);
531         }
532
533         if (is_bad_inode(inode)) {
534                 rc = -ENOENT;
535                 if (!trusted)
536                         goto check_oi;
537
538                 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
539                        PFID(fid), id->oii_ino);
540                 GOTO(put, rc);
541         }
542
543         if (id->oii_gen != OSD_OII_NOGEN &&
544             inode->i_generation != id->oii_gen) {
545                 rc = -ESTALE;
546                 if (!trusted)
547                         goto check_oi;
548
549                 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
550                        "oii_gen = %u, i_generation = %u\n", PFID(fid),
551                        id->oii_ino, id->oii_gen, inode->i_generation);
552                 GOTO(put, rc);
553         }
554
555         if (inode->i_nlink == 0) {
556                 rc = -ENOENT;
557                 if (!trusted)
558                         goto check_oi;
559
560                 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
561                        PFID(fid), id->oii_ino);
562                 GOTO(put, rc);
563         }
564
565         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
566
567 check_oi:
568         if (rc != 0) {
569                 __u32 saved_ino = id->oii_ino;
570                 __u32 saved_gen = id->oii_gen;
571
572                 LASSERT(!trusted);
573                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
574
575                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
576                 /* XXX: There are four possible cases:
577                  *      1. rc = 0.
578                  *         Backup/restore caused the OI invalid.
579                  *      2. rc = 0.
580                  *         Someone unlinked the object but NOT removed
581                  *         the OI mapping, such as mount target device
582                  *         as ldiskfs, and modify something directly.
583                  *      3. rc = -ENOENT.
584                  *         Someone just removed the object between the
585                  *         former oi_lookup and the iget. It is normal.
586                  *      4. Other failure cases.
587                  *
588                  *      Generally, when the device is mounted, it will
589                  *      auto check whether the system is restored from
590                  *      file-level backup or not. We trust such detect
591                  *      to distinguish the 1st case from the 2nd case:
592                  *      if the OI files are consistent but may contain
593                  *      stale OI mappings because of case 2, if iget()
594                  *      returns -ENOENT or -ESTALE, then it should be
595                  *      the case 2. */
596                 if (rc != 0)
597                         /* If the OI mapping was in OI file before the
598                          * osd_iget_check(), but now, it is disappear,
599                          * then it must be removed by race. That is a
600                          * normal race case. */
601                         GOTO(put, rc);
602
603                 /* It is the OI scrub updated the OI mapping by race.
604                  * The new OI mapping must be valid. */
605                 if (saved_ino != id->oii_ino ||
606                     (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
607                         if (!IS_ERR(inode))
608                                 iput(inode);
609
610                         trusted = true;
611                         goto again;
612                 }
613
614                 if (IS_ERR(inode)) {
615                         if (dev->od_scrub.os_file.sf_flags & SF_INCONSISTENT)
616                                 /* It still can be the case 2, but we cannot
617                                  * distinguish it from the case 1. So return
618                                  * -EREMCHG to block current operation until
619                                  *  OI scrub rebuilt the OI mappings. */
620                                 rc = -EREMCHG;
621                         else
622                                 rc = -ENOENT;
623
624                         GOTO(put, rc);
625                 }
626
627                 if (inode->i_generation == id->oii_gen)
628                         rc = -ENOENT;
629                 else
630                         rc = -EREMCHG;
631         } else {
632                 if (id->oii_gen == OSD_OII_NOGEN)
633                         osd_id_gen(id, inode->i_ino, inode->i_generation);
634
635                 /* Do not update file c/mtime in ldiskfs.
636                  * NB: we don't have any lock to protect this because we don't
637                  * have reference on osd_object now, but contention with
638                  * another lookup + attr_set can't happen in the tiny window
639                  * between if (...) and set S_NOCMTIME. */
640                 if (!(inode->i_flags & S_NOCMTIME))
641                         inode->i_flags |= S_NOCMTIME;
642         }
643
644         GOTO(put, rc);
645
646 put:
647         if (rc != 0) {
648                 if (!IS_ERR(inode))
649                         iput(inode);
650
651                 inode = ERR_PTR(rc);
652         }
653
654         return inode;
655 }
656
657 /**
658  * \retval +v: new filter_fid, does not contain self-fid
659  * \retval 0:  filter_fid_old, contains self-fid
660  * \retval -v: other failure cases
661  */
662 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
663                  struct dentry *dentry, struct lu_fid *fid)
664 {
665         struct filter_fid_old   *ff     = &info->oti_ff;
666         struct ost_id           *ostid  = &info->oti_ostid;
667         int                      rc;
668
669         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
670         if (rc == sizeof(*ff)) {
671                 rc = 0;
672                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
673                 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
674                 /*
675                  * XXX: use 0 as the index for compatibility, the caller will
676                  *      handle index related issues when necessary.
677                  */
678                 if (!rc)
679                         ostid_to_fid(fid, ostid, 0);
680         } else if (rc == sizeof(struct filter_fid)) {
681                 rc = 1;
682         } else if (rc >= 0) {
683                 rc = -EINVAL;
684         }
685
686         return rc;
687 }
688
689 static int osd_lma_self_repair(struct osd_thread_info *info,
690                                struct osd_device *osd, struct inode *inode,
691                                const struct lu_fid *fid, __u32 compat)
692 {
693         handle_t *jh;
694         int       rc;
695
696         LASSERT(current->journal_info == NULL);
697
698         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
699                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
700         if (IS_ERR(jh)) {
701                 rc = PTR_ERR(jh);
702                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
703                       osd_name(osd), rc);
704                 return rc;
705         }
706
707         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
708         if (rc != 0)
709                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
710                       osd_name(osd), rc);
711         ldiskfs_journal_stop(jh);
712         return rc;
713 }
714
715 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
716 {
717         struct osd_thread_info  *info   = osd_oti_get(env);
718         struct osd_device       *osd    = osd_obj2dev(obj);
719         struct lustre_ost_attrs *loa    = &info->oti_ost_attrs;
720         struct lustre_mdt_attrs *lma    = &loa->loa_lma;
721         struct inode            *inode  = obj->oo_inode;
722         struct dentry           *dentry = &info->oti_obj_dentry;
723         struct lu_fid           *fid    = NULL;
724         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
725         int                      rc;
726         ENTRY;
727
728         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
729                              (void *)loa, sizeof(*loa));
730         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
731                 fid = &lma->lma_self_fid;
732                 rc = osd_get_idif(info, inode, dentry, fid);
733                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
734                         /* For the given OST-object, if it has neither LMA nor
735                          * FID in XATTR_NAME_FID, then the given FID (which is
736                          * contained in the @obj, from client RPC for locating
737                          * the OST-object) is trusted. We use it to generate
738                          * the LMA. */
739                         osd_lma_self_repair(info, osd, inode, rfid,
740                                             LMAC_FID_ON_OST);
741                         RETURN(0);
742                 }
743         }
744
745         if (rc < 0)
746                 RETURN(rc);
747
748         if (rc > 0) {
749                 rc = 0;
750                 lustre_lma_swab(lma);
751                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
752                              (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
753                               S_ISREG(inode->i_mode)))) {
754                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
755                               "fid = "DFID", ino = %lu\n", osd_name(osd),
756                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
757                               PFID(rfid), inode->i_ino);
758                         rc = -EOPNOTSUPP;
759                 } else {
760                         fid = &lma->lma_self_fid;
761                 }
762         }
763
764         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
765                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
766                         struct ost_id   *oi   = &info->oti_ostid;
767                         struct lu_fid   *fid1 = &info->oti_fid3;
768                         __u32            idx  = fid_idif_ost_idx(rfid);
769
770                         /* For old IDIF, the OST index is not part of the IDIF,
771                          * Means that different OSTs may have the same IDIFs.
772                          * Under such case, we need to make some compatible
773                          * check to make sure to trigger OI scrub properly. */
774                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
775                                 /* Given @rfid is new, LMA is old. */
776                                 fid_to_ostid(fid, oi);
777                                 ostid_to_fid(fid1, oi, idx);
778                                 if (lu_fid_eq(fid1, rfid)) {
779                                         if (osd->od_index_in_idif)
780                                                 osd_lma_self_repair(info, osd,
781                                                         inode, rfid,
782                                                         LMAC_FID_ON_OST);
783                                         RETURN(0);
784                                 }
785                         }
786                 }
787
788                 rc = -EREMCHG;
789         }
790
791         RETURN(rc);
792 }
793
794 struct osd_check_lmv_buf {
795 #ifdef HAVE_DIR_CONTEXT
796         /* please keep it as first member */
797         struct dir_context       ctx;
798 #endif
799         struct osd_thread_info  *oclb_info;
800         struct osd_device       *oclb_dev;
801         struct osd_idmap_cache  *oclb_oic;
802 };
803
804 /**
805  * It is called internally by ->readdir() to filter out the
806  * local slave object's FID of the striped directory.
807  *
808  * \retval      1 found the local slave's FID
809  * \retval      0 continue to check next item
810  * \retval      -ve for failure
811  */
812 #ifdef HAVE_FILLDIR_USE_CTX
813 static int osd_stripe_dir_filldir(struct dir_context *buf,
814 #else
815 static int osd_stripe_dir_filldir(void *buf,
816 #endif
817                                   const char *name, int namelen,
818                                   loff_t offset, __u64 ino, unsigned d_type)
819 {
820         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
821         struct osd_thread_info *oti = oclb->oclb_info;
822         struct lu_fid *fid = &oti->oti_fid3;
823         struct osd_inode_id *id = &oti->oti_id3;
824         struct osd_device *dev = oclb->oclb_dev;
825         struct osd_idmap_cache *oic = oclb->oclb_oic;
826         struct inode *inode;
827         int rc;
828
829         if (name[0] == '.')
830                 return 0;
831
832         fid_zero(fid);
833         sscanf(name + 1, SFID, RFID(fid));
834         if (!fid_is_sane(fid))
835                 return 0;
836
837         if (osd_remote_fid(oti->oti_env, dev, fid))
838                 return 0;
839
840         osd_id_gen(id, ino, OSD_OII_NOGEN);
841         inode = osd_iget(oti, dev, id);
842         if (IS_ERR(inode))
843                 return PTR_ERR(inode);
844
845         iput(inode);
846         osd_add_oi_cache(oti, dev, id, fid);
847         oic->oic_fid = *fid;
848         oic->oic_lid = *id;
849         oic->oic_dev = dev;
850         rc = osd_oii_insert(dev, oic, true);
851
852         return rc == 0 ? 1 : rc;
853 }
854
855 /* When lookup item under striped directory, we need to locate the master
856  * MDT-object of the striped directory firstly, then the client will send
857  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
858  * and the item's name. If the system is restored from MDT file level backup,
859  * then before the OI scrub completely built the OI files, the OI mappings of
860  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
861  * not a problem for the master MDT-object. Because when locate the master
862  * MDT-object, we will do name based lookup (for the striped directory itself)
863  * firstly, during such process we can setup the correct OI mapping for the
864  * master MDT-object. But it will be trouble for the slave MDT-object. Because
865  * the client will not trigger name based lookup on the MDT to locate the slave
866  * MDT-object before locating item under the striped directory, then when
867  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
868  * is invalid and does not know what the right OI mapping is, then the MDT has
869  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
870  * the OI file, related OI mapping is unknown yet, please try again later. And
871  * then client will re-try the RPC again and again until related OI mapping has
872  * been updated. That is quite inefficient.
873  *
874  * To resolve above trouble, we will handle it as the following two cases:
875  *
876  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
877  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
878  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
879  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
880  *    directly. Please check osd_fid_lookup().
881  *
882  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
883  *    Under such case, during lookup the master MDT-object, we will lookup the
884  *    slave MDT-object via readdir against the master MDT-object, because the
885  *    slave MDT-objects information are stored as sub-directories with the name
886  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
887  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
888  *    the correct OI mapping for the slave MDT-object. */
889 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
890                          struct inode *inode, struct osd_idmap_cache *oic)
891 {
892         struct lu_buf *buf = &oti->oti_big_buf;
893         struct dentry *dentry = &oti->oti_obj_dentry;
894         struct file *filp = &oti->oti_file;
895         const struct file_operations *fops;
896         struct lmv_mds_md_v1 *lmv1;
897         struct osd_check_lmv_buf oclb = {
898 #ifdef HAVE_DIR_CONTEXT
899                 .ctx.actor = osd_stripe_dir_filldir,
900 #endif
901                 .oclb_info = oti,
902                 .oclb_dev = dev,
903                 .oclb_oic = oic
904         };
905         int rc = 0;
906         ENTRY;
907
908 again:
909         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
910                              buf->lb_len);
911         if (rc == -ERANGE) {
912                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
913                 if (rc > 0) {
914                         lu_buf_realloc(buf, rc);
915                         if (buf->lb_buf == NULL)
916                                 GOTO(out, rc = -ENOMEM);
917
918                         goto again;
919                 }
920         }
921
922         if (unlikely(rc == 0 || rc == -ENODATA))
923                 GOTO(out, rc = 0);
924
925         if (rc < 0)
926                 GOTO(out, rc);
927
928         if (unlikely(buf->lb_buf == NULL)) {
929                 lu_buf_realloc(buf, rc);
930                 if (buf->lb_buf == NULL)
931                         GOTO(out, rc = -ENOMEM);
932
933                 goto again;
934         }
935
936         lmv1 = buf->lb_buf;
937         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
938                 GOTO(out, rc = 0);
939
940         fops = inode->i_fop;
941         dentry->d_inode = inode;
942         dentry->d_sb = inode->i_sb;
943         filp->f_pos = 0;
944         filp->f_path.dentry = dentry;
945         filp->f_mode = FMODE_64BITHASH;
946         filp->f_mapping = inode->i_mapping;
947         filp->f_op = fops;
948         filp->private_data = NULL;
949         set_file_inode(filp, inode);
950
951 #ifdef HAVE_DIR_CONTEXT
952         oclb.ctx.pos = filp->f_pos;
953         rc = fops->iterate(filp, &oclb.ctx);
954         filp->f_pos = oclb.ctx.pos;
955 #else
956         rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
957 #endif
958         fops->release(inode, filp);
959
960 out:
961         if (rc < 0)
962                 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
963                        DFID": rc = %d\n", osd_ino2name(inode),
964                        inode->i_ino, inode->i_generation,
965                        PFID(&oic->oic_fid), rc);
966         else
967                 rc = 0;
968
969         RETURN(rc);
970 }
971
972 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
973                           const struct lu_fid *fid,
974                           const struct lu_object_conf *conf)
975 {
976         struct osd_thread_info *info;
977         struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
978         struct osd_device *dev;
979         struct osd_idmap_cache *oic;
980         struct osd_inode_id *id;
981         struct inode *inode = NULL;
982         struct osd_scrub *scrub;
983         struct scrub_file *sf;
984         __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
985         __u32 saved_ino;
986         __u32 saved_gen;
987         int result = 0;
988         int rc1 = 0;
989         bool remote = false;
990         bool trusted = true;
991         bool updated = false;
992         ENTRY;
993
994         LINVRNT(osd_invariant(obj));
995         LASSERT(obj->oo_inode == NULL);
996         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
997
998         dev = osd_dev(ldev);
999         scrub = &dev->od_scrub;
1000         sf = &scrub->os_file;
1001         info = osd_oti_get(env);
1002         LASSERT(info);
1003         oic = &info->oti_cache;
1004
1005         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1006                 RETURN(-ENOENT);
1007
1008         /* For the object is created as locking anchor, or for the object to
1009          * be created on disk. No need to osd_oi_lookup() at here because FID
1010          * shouldn't never be re-used, if it's really a duplicate FID from
1011          * unexpected reason, we should be able to detect it later by calling
1012          * do_create->osd_oi_insert(). */
1013         if (conf && conf->loc_flags & LOC_F_NEW)
1014                 GOTO(out, result = 0);
1015
1016         /* Search order: 1. per-thread cache. */
1017         if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1018                 id = &oic->oic_lid;
1019                 goto iget;
1020         }
1021
1022         id = &info->oti_id;
1023         if (!list_empty(&scrub->os_inconsistent_items)) {
1024                 /* Search order: 2. OI scrub pending list. */
1025                 result = osd_oii_lookup(dev, fid, id);
1026                 if (!result)
1027                         goto iget;
1028         }
1029
1030         /* The OI mapping in the OI file can be updated by the OI scrub
1031          * when we locate the inode via FID. So it may be not trustable. */
1032         trusted = false;
1033
1034         /* Search order: 3. OI files. */
1035         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1036         if (result == -ENOENT) {
1037                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1038                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1039                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1040                                       sf->sf_oi_bitmap))
1041                         GOTO(out, result = 0);
1042
1043                 goto trigger;
1044         }
1045
1046         if (result)
1047                 GOTO(out, result);
1048
1049 iget:
1050         obj->oo_inode = NULL;
1051         /* for later passes through checks, not true on first pass */
1052         if (!IS_ERR_OR_NULL(inode))
1053                 iput(inode);
1054
1055         inode = osd_iget_check(info, dev, fid, id, trusted);
1056         if (!IS_ERR(inode)) {
1057                 obj->oo_inode = inode;
1058                 result = 0;
1059                 if (remote)
1060                         goto trigger;
1061
1062                 goto check_lma;
1063         }
1064
1065         result = PTR_ERR(inode);
1066         if (result == -ENOENT || result == -ESTALE)
1067                 GOTO(out, result = 0);
1068
1069         if (result != -EREMCHG)
1070                 GOTO(out, result);
1071
1072 trigger:
1073         /* We still have chance to get the valid inode: for the
1074          * object which is referenced by remote name entry, the
1075          * object on the local MDT will be linked under the dir
1076          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1077          *
1078          * We do not know whether the object for the given FID
1079          * is referenced by some remote name entry or not, and
1080          * especially for DNE II, a multiple-linked object may
1081          * have many name entries reside on many MDTs.
1082          *
1083          * To simplify the operation, OSD will not distinguish
1084          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1085          * only happened for the RPC from other MDT during the
1086          * OI scrub, or for the client side RPC with FID only,
1087          * such as FID to path, or from old connected client. */
1088         if (!remote && !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1089                 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1090                 if (!rc1) {
1091                         remote = true;
1092                         trusted = true;
1093                         flags |= SS_AUTO_PARTIAL;
1094                         flags &= ~SS_AUTO_FULL;
1095                         goto iget;
1096                 }
1097         }
1098
1099         if (thread_is_running(&scrub->os_thread)) {
1100                 if (scrub->os_partial_scan && !scrub->os_in_join)
1101                         goto join;
1102
1103                 if (IS_ERR_OR_NULL(inode) || result)
1104                         GOTO(out, result = -EINPROGRESS);
1105
1106                 LASSERT(remote);
1107                 LASSERT(obj->oo_inode == inode);
1108
1109                 osd_add_oi_cache(info, dev, id, fid);
1110                 osd_oii_insert(dev, oic, true);
1111                 goto found;
1112         }
1113
1114         if (dev->od_noscrub) {
1115                 if (!remote)
1116                         GOTO(out, result = -EREMCHG);
1117
1118                 LASSERT(!result);
1119                 LASSERT(obj->oo_inode == inode);
1120
1121                 osd_add_oi_cache(info, dev, id, fid);
1122                 goto found;
1123         }
1124
1125 join:
1126         rc1 = osd_scrub_start(dev, flags);
1127         LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1128                       "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1129         if (rc1 && rc1 != -EALREADY)
1130                 GOTO(out, result = -EREMCHG);
1131
1132         if (IS_ERR_OR_NULL(inode) || result)
1133                 GOTO(out, result = -EINPROGRESS);
1134
1135         LASSERT(remote);
1136         LASSERT(obj->oo_inode == inode);
1137
1138         osd_add_oi_cache(info, dev, id, fid);
1139         osd_oii_insert(dev, oic, true);
1140         goto found;
1141
1142 check_lma:
1143         result = osd_check_lma(env, obj);
1144         if (!result)
1145                 goto found;
1146
1147         LASSERTF(id->oii_ino == inode->i_ino &&
1148                  id->oii_gen == inode->i_generation,
1149                  "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1150                  PFID(fid), id->oii_ino, id->oii_gen,
1151                  inode->i_ino, inode->i_generation);
1152
1153         saved_ino = inode->i_ino;
1154         saved_gen = inode->i_generation;
1155
1156         if (unlikely(result == -ENODATA)) {
1157                 /* If the OI scrub updated the OI mapping by race, it
1158                  * must be valid. Trust the inode that has no LMA EA. */
1159                 if (updated)
1160                         goto found;
1161
1162                 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1163                 if (!result) {
1164                         /* The OI mapping is still there, the inode is still
1165                          * valid. It is just becaues the inode has no LMA EA. */
1166                         if (saved_ino == id->oii_ino &&
1167                             saved_gen == id->oii_gen)
1168                                 goto found;
1169
1170                         /* It is the OI scrub updated the OI mapping by race.
1171                          * The new OI mapping must be valid. */
1172                         trusted = true;
1173                         updated = true;
1174                         goto iget;
1175                 }
1176
1177                 /* "result == -ENOENT" means that the OI mappinghas been
1178                  * removed by race, so the inode belongs to other object.
1179                  *
1180                  * Others error can be returned  directly. */
1181                 if (result == -ENOENT) {
1182                         LASSERT(trusted);
1183
1184                         obj->oo_inode = NULL;
1185                         result = 0;
1186                 }
1187         }
1188
1189         if (result != -EREMCHG)
1190                 GOTO(out, result);
1191
1192         LASSERT(!updated);
1193
1194         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1195         /* "result == -ENOENT" means the cached OI mapping has been removed
1196          * from the OI file by race, above inode belongs to other object. */
1197         if (result == -ENOENT) {
1198                 LASSERT(trusted);
1199
1200                 obj->oo_inode = NULL;
1201                 GOTO(out, result = 0);
1202         }
1203
1204         if (result)
1205                 GOTO(out, result);
1206
1207         if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1208                 result = -EREMCHG;
1209                 goto trigger;
1210         }
1211
1212         /* It is the OI scrub updated the OI mapping by race.
1213          * The new OI mapping must be valid. */
1214         trusted = true;
1215         updated = true;
1216         goto iget;
1217
1218 found:
1219         obj->oo_compat_dot_created = 1;
1220         obj->oo_compat_dotdot_created = 1;
1221
1222         if (S_ISDIR(inode->i_mode) &&
1223             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1224                 osd_check_lmv(info, dev, inode, oic);
1225
1226         result = osd_attach_jinode(inode);
1227         if (result)
1228                 GOTO(out, result);
1229
1230         if (!ldiskfs_pdo)
1231                 GOTO(out, result = 0);
1232
1233         LASSERT(!obj->oo_hl_head);
1234         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1235
1236         GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1237
1238 out:
1239         if (result || !obj->oo_inode) {
1240                 if (!IS_ERR_OR_NULL(inode))
1241                         iput(inode);
1242
1243                 obj->oo_inode = NULL;
1244                 if (trusted)
1245                         fid_zero(&oic->oic_fid);
1246         }
1247
1248         LINVRNT(osd_invariant(obj));
1249         return result;
1250 }
1251
1252 /*
1253  * Concurrency: shouldn't matter.
1254  */
1255 static void osd_object_init0(struct osd_object *obj)
1256 {
1257         LASSERT(obj->oo_inode != NULL);
1258         obj->oo_dt.do_body_ops = &osd_body_ops;
1259         obj->oo_dt.do_lu.lo_header->loh_attr |=
1260                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1261 }
1262
1263 /*
1264  * Concurrency: no concurrent access is possible that early in object
1265  * life-cycle.
1266  */
1267 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1268                            const struct lu_object_conf *conf)
1269 {
1270         struct osd_object *obj = osd_obj(l);
1271         int result;
1272
1273         LINVRNT(osd_invariant(obj));
1274
1275         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1276                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1277                 l->lo_header->loh_attr |= LOHA_EXISTS;
1278                 return 0;
1279         }
1280
1281         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1282         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1283         if (result == 0 && obj->oo_inode != NULL) {
1284                 struct osd_thread_info *oti = osd_oti_get(env);
1285                 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1286
1287                 osd_object_init0(obj);
1288                 result = osd_get_lma(oti, obj->oo_inode,
1289                                      &oti->oti_obj_dentry, loa);
1290                 if (!result) {
1291                         /* Convert LMAI flags to lustre LMA flags
1292                          * and cache it to oo_lma_flags */
1293                         obj->oo_lma_flags =
1294                                 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1295                 } else if (result == -ENODATA) {
1296                         result = 0;
1297                 }
1298         }
1299
1300         LINVRNT(osd_invariant(obj));
1301         return result;
1302 }
1303
1304 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1305  * The left part is for value, binary mode. */
1306 struct osd_xattr_entry {
1307         struct list_head        oxe_list;
1308         size_t                  oxe_len;
1309         size_t                  oxe_namelen;
1310         bool                    oxe_exist;
1311         struct rcu_head         oxe_rcu;
1312         char                    oxe_buf[0];
1313 };
1314
1315 static int osd_oxc_get(struct osd_object *obj, const char *name,
1316                        struct lu_buf *buf)
1317 {
1318         struct osd_xattr_entry *tmp;
1319         struct osd_xattr_entry *oxe = NULL;
1320         size_t namelen = strlen(name);
1321         int rc;
1322         ENTRY;
1323
1324         rcu_read_lock();
1325         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1326                 if (namelen == tmp->oxe_namelen &&
1327                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1328                         oxe = tmp;
1329                         break;
1330                 }
1331         }
1332
1333         if (oxe == NULL)
1334                 GOTO(out, rc = -ENOENT);
1335
1336         if (!oxe->oxe_exist)
1337                 GOTO(out, rc = -ENODATA);
1338
1339         /* vallen */
1340         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1341         LASSERT(rc > 0);
1342
1343         if (buf->lb_buf == NULL)
1344                 GOTO(out, rc);
1345
1346         if (buf->lb_len < rc)
1347                 GOTO(out, rc = -ERANGE);
1348
1349         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1350         EXIT;
1351 out:
1352         rcu_read_unlock();
1353
1354         return rc;
1355 }
1356
1357 static void osd_oxc_free(struct rcu_head *head)
1358 {
1359         struct osd_xattr_entry *oxe;
1360
1361         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1362         OBD_FREE(oxe, oxe->oxe_len);
1363 }
1364
1365 static void osd_oxc_add(struct osd_object *obj, const char *name,
1366                         const char *buf, int buflen)
1367 {
1368         struct osd_xattr_entry *oxe;
1369         struct osd_xattr_entry *old = NULL;
1370         struct osd_xattr_entry *tmp;
1371         size_t namelen = strlen(name);
1372         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1373
1374         OBD_ALLOC(oxe, len);
1375         if (oxe == NULL)
1376                 return;
1377
1378         INIT_LIST_HEAD(&oxe->oxe_list);
1379         oxe->oxe_len = len;
1380         oxe->oxe_namelen = namelen;
1381         memcpy(oxe->oxe_buf, name, namelen);
1382         if (buflen > 0) {
1383                 LASSERT(buf != NULL);
1384                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1385                 oxe->oxe_exist = true;
1386         } else {
1387                 oxe->oxe_exist = false;
1388         }
1389
1390         /* this should be rarely called, just remove old and add new */
1391         spin_lock(&obj->oo_guard);
1392         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1393                 if (namelen == tmp->oxe_namelen &&
1394                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1395                         old = tmp;
1396                         break;
1397                 }
1398         }
1399         if (old != NULL) {
1400                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1401                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1402         } else {
1403                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1404         }
1405         spin_unlock(&obj->oo_guard);
1406 }
1407
1408 static void osd_oxc_del(struct osd_object *obj, const char *name)
1409 {
1410         struct osd_xattr_entry *oxe;
1411         size_t namelen = strlen(name);
1412
1413         spin_lock(&obj->oo_guard);
1414         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1415                 if (namelen == oxe->oxe_namelen &&
1416                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1417                         list_del_rcu(&oxe->oxe_list);
1418                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1419                         break;
1420                 }
1421         }
1422         spin_unlock(&obj->oo_guard);
1423 }
1424
1425 static void osd_oxc_fini(struct osd_object *obj)
1426 {
1427         struct osd_xattr_entry *oxe, *next;
1428
1429         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1430                 list_del(&oxe->oxe_list);
1431                 OBD_FREE(oxe, oxe->oxe_len);
1432         }
1433 }
1434
1435 /*
1436  * Concurrency: no concurrent access is possible that late in object
1437  * life-cycle.
1438  */
1439 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1440 {
1441         struct osd_object *obj = osd_obj(l);
1442
1443         LINVRNT(osd_invariant(obj));
1444
1445         osd_oxc_fini(obj);
1446         dt_object_fini(&obj->oo_dt);
1447         if (obj->oo_hl_head != NULL)
1448                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1449         OBD_FREE_PTR(obj);
1450 }
1451
1452 /*
1453  * Concurrency: no concurrent access is possible that late in object
1454  * life-cycle.
1455  */
1456 static void osd_index_fini(struct osd_object *o)
1457 {
1458         struct iam_container *bag;
1459
1460         if (o->oo_dir != NULL) {
1461                 bag = &o->oo_dir->od_container;
1462                 if (o->oo_inode != NULL) {
1463                         if (bag->ic_object == o->oo_inode)
1464                                 iam_container_fini(bag);
1465                 }
1466                 OBD_FREE_PTR(o->oo_dir);
1467                 o->oo_dir = NULL;
1468         }
1469 }
1470
1471 /*
1472  * Concurrency: no concurrent access is possible that late in object
1473  * life-cycle (for all existing callers, that is. New callers have to provide
1474  * their own locking.)
1475  */
1476 static int osd_inode_unlinked(const struct inode *inode)
1477 {
1478         return inode->i_nlink == 0;
1479 }
1480
1481 enum {
1482         OSD_TXN_OI_DELETE_CREDITS    = 20,
1483         OSD_TXN_INODE_DELETE_CREDITS = 20
1484 };
1485
1486 /*
1487  * Journal
1488  */
1489
1490 #if OSD_THANDLE_STATS
1491 /**
1492  * Set time when the handle is allocated
1493  */
1494 static void osd_th_alloced(struct osd_thandle *oth)
1495 {
1496         oth->oth_alloced = ktime_get();
1497 }
1498
1499 /**
1500  * Set time when the handle started
1501  */
1502 static void osd_th_started(struct osd_thandle *oth)
1503 {
1504         oth->oth_started = ktime_get();
1505 }
1506
1507 /**
1508  * Check whether the we deal with this handle for too long.
1509  */
1510 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1511                                 ktime_t alloced, ktime_t started,
1512                                 ktime_t closed)
1513 {
1514         ktime_t now = ktime_get();
1515
1516         LASSERT(dev != NULL);
1517
1518         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1519                             ktime_us_delta(started, alloced));
1520         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1521                             ktime_us_delta(closed, started));
1522         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1523                             ktime_us_delta(now, closed));
1524
1525         if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1526                 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1527                       oth, now, alloced, started, closed);
1528                 libcfs_debug_dumpstack(NULL);
1529         }
1530 }
1531
1532 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1533 {                                                                       \
1534         ktime_t __closed = ktime_get();                                 \
1535         ktime_t __alloced = oth->oth_alloced;                           \
1536         ktime_t __started = oth->oth_started;                           \
1537                                                                         \
1538         expr;                                                           \
1539         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1540 }
1541
1542 #else /* OSD_THANDLE_STATS */
1543
1544 #define osd_th_alloced(h)                  do {} while(0)
1545 #define osd_th_started(h)                  do {} while(0)
1546 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1547
1548 #endif /* OSD_THANDLE_STATS */
1549
1550 /*
1551  * Concurrency: doesn't access mutable data.
1552  */
1553 static int osd_param_is_not_sane(const struct osd_device *dev,
1554                                  const struct thandle *th)
1555 {
1556         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1557
1558         return oh->ot_credits > osd_transaction_size(dev);
1559 }
1560
1561 /*
1562  * Concurrency: shouldn't matter.
1563  */
1564 static void osd_trans_commit_cb(struct super_block *sb,
1565                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1566 {
1567         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1568         struct thandle     *th  = &oh->ot_super;
1569         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
1570         struct dt_txn_commit_cb *dcb, *tmp;
1571
1572         LASSERT(oh->ot_handle == NULL);
1573
1574         if (error)
1575                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1576
1577         dt_txn_hook_commit(th);
1578
1579         /* call per-transaction callbacks if any */
1580         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1581                                  dcb_linkage) {
1582                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1583                          "commit callback entry: magic=%x name='%s'\n",
1584                          dcb->dcb_magic, dcb->dcb_name);
1585                 list_del_init(&dcb->dcb_linkage);
1586                 dcb->dcb_func(NULL, th, dcb, error);
1587         }
1588
1589         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1590         lu_device_put(lud);
1591         th->th_dev = NULL;
1592
1593         lu_context_exit(&th->th_ctx);
1594         lu_context_fini(&th->th_ctx);
1595         OBD_FREE_PTR(oh);
1596 }
1597
1598 #ifndef HAVE_SB_START_WRITE
1599 # define sb_start_write(sb) do {} while (0)
1600 # define sb_end_write(sb) do {} while (0)
1601 #endif
1602
1603 static struct thandle *osd_trans_create(const struct lu_env *env,
1604                                         struct dt_device *d)
1605 {
1606         struct osd_thread_info  *oti = osd_oti_get(env);
1607         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1608         struct osd_thandle      *oh;
1609         struct thandle          *th;
1610         ENTRY;
1611
1612         if (d->dd_rdonly) {
1613                 CERROR("%s: someone try to start transaction under "
1614                        "readonly mode, should be disabled.\n",
1615                        osd_name(osd_dt_dev(d)));
1616                 dump_stack();
1617                 RETURN(ERR_PTR(-EROFS));
1618         }
1619
1620         /* on pending IO in this thread should left from prev. request */
1621         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1622
1623         sb_start_write(osd_sb(osd_dt_dev(d)));
1624
1625         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1626         if (oh != NULL) {
1627                 oh->ot_quota_trans = &oti->oti_quota_trans;
1628                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1629                 th = &oh->ot_super;
1630                 th->th_dev = d;
1631                 th->th_result = 0;
1632                 th->th_tags = LCT_TX_HANDLE;
1633                 oh->ot_credits = 0;
1634                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1635                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1636                 osd_th_alloced(oh);
1637
1638                 memset(oti->oti_declare_ops, 0,
1639                        sizeof(oti->oti_declare_ops));
1640                 memset(oti->oti_declare_ops_cred, 0,
1641                        sizeof(oti->oti_declare_ops_cred));
1642                 memset(oti->oti_declare_ops_used, 0,
1643                        sizeof(oti->oti_declare_ops_used));
1644         } else {
1645                 sb_end_write(osd_sb(osd_dt_dev(d)));
1646                 th = ERR_PTR(-ENOMEM);
1647         }
1648         RETURN(th);
1649 }
1650
1651 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1652 {
1653         struct osd_thread_info  *oti = osd_oti_get(env);
1654         struct osd_thandle      *oh;
1655
1656         oh = container_of0(th, struct osd_thandle, ot_super);
1657         LASSERT(oh != NULL);
1658
1659         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1660               oti->oti_declare_ops[OSD_OT_CREATE],
1661               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1662               oti->oti_declare_ops_used[OSD_OT_CREATE],
1663               oti->oti_declare_ops[OSD_OT_DESTROY],
1664               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1665               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1666         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1667               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1668               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1669               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1670               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1671               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1672               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1673         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1674               oti->oti_declare_ops[OSD_OT_WRITE],
1675               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1676               oti->oti_declare_ops_used[OSD_OT_WRITE],
1677               oti->oti_declare_ops[OSD_OT_PUNCH],
1678               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1679               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1680               oti->oti_declare_ops[OSD_OT_QUOTA],
1681               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1682               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1683         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1684               oti->oti_declare_ops[OSD_OT_INSERT],
1685               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1686               oti->oti_declare_ops_used[OSD_OT_INSERT],
1687               oti->oti_declare_ops[OSD_OT_DELETE],
1688               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1689               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1690         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1691               oti->oti_declare_ops[OSD_OT_REF_ADD],
1692               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1693               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1694               oti->oti_declare_ops[OSD_OT_REF_DEL],
1695               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1696               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1697 }
1698
1699 /*
1700  * Concurrency: shouldn't matter.
1701  */
1702 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1703                            struct thandle *th)
1704 {
1705         struct osd_thread_info *oti = osd_oti_get(env);
1706         struct osd_device  *dev = osd_dt_dev(d);
1707         handle_t           *jh;
1708         struct osd_thandle *oh;
1709         int rc;
1710
1711         ENTRY;
1712
1713         LASSERT(current->journal_info == NULL);
1714
1715         oh = container_of0(th, struct osd_thandle, ot_super);
1716         LASSERT(oh != NULL);
1717         LASSERT(oh->ot_handle == NULL);
1718
1719         rc = dt_txn_hook_start(env, d, th);
1720         if (rc != 0)
1721                 GOTO(out, rc);
1722
1723         if (unlikely(osd_param_is_not_sane(dev, th))) {
1724                 static unsigned long last_printed;
1725                 static int last_credits;
1726
1727                 /* don't make noise on a tiny testing systems
1728                  * actual credits misuse will be caught anyway */
1729                 if (last_credits != oh->ot_credits &&
1730                     time_after(jiffies, last_printed +
1731                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1732                     osd_transaction_size(dev) > 512) {
1733                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1734                               oh->ot_credits, osd_transaction_size(dev));
1735                         osd_trans_dump_creds(env, th);
1736                         libcfs_debug_dumpstack(NULL);
1737                         last_credits = oh->ot_credits;
1738                         last_printed = jiffies;
1739                 }
1740                 /* XXX Limit the credits to 'max_transaction_buffers', and
1741                  *     let the underlying filesystem to catch the error if
1742                  *     we really need so many credits.
1743                  *
1744                  *     This should be removed when we can calculate the
1745                  *     credits precisely. */
1746                 oh->ot_credits = osd_transaction_size(dev);
1747         } else if (ldiskfs_track_declares_assert != 0) {
1748                 /* reserve few credits to prevent an assertion in JBD
1749                  * our debugging mechanism will be able to detected
1750                  * overuse. this can help to debug single-update
1751                  * transactions */
1752                 oh->ot_credits += 10;
1753                 if (unlikely(osd_param_is_not_sane(dev, th)))
1754                         oh->ot_credits = osd_transaction_size(dev);
1755         }
1756
1757         /*
1758          * XXX temporary stuff. Some abstraction layer should
1759          * be used.
1760          */
1761         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1762         osd_th_started(oh);
1763         if (!IS_ERR(jh)) {
1764                 oh->ot_handle = jh;
1765                 LASSERT(oti->oti_txns == 0);
1766                 lu_context_init(&th->th_ctx, th->th_tags);
1767                 lu_context_enter(&th->th_ctx);
1768
1769                 lu_device_get(&d->dd_lu_dev);
1770                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1771                               "osd-tx", th);
1772                 oti->oti_txns++;
1773                 rc = 0;
1774         } else {
1775                 rc = PTR_ERR(jh);
1776         }
1777 out:
1778         RETURN(rc);
1779 }
1780
1781 static int osd_seq_exists(const struct lu_env *env,
1782                           struct osd_device *osd, u64 seq)
1783 {
1784         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1785         struct seq_server_site  *ss = osd_seq_site(osd);
1786         int                     rc;
1787         ENTRY;
1788
1789         LASSERT(ss != NULL);
1790         LASSERT(ss->ss_server_fld != NULL);
1791
1792         rc = osd_fld_lookup(env, osd, seq, range);
1793         if (rc != 0) {
1794                 if (rc != -ENOENT)
1795                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1796                                osd_name(osd), seq, rc);
1797                 RETURN(0);
1798         }
1799
1800         RETURN(ss->ss_node_id == range->lsr_index);
1801 }
1802
1803 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1804 {
1805         struct dt_txn_commit_cb *dcb;
1806         struct dt_txn_commit_cb *tmp;
1807
1808         /* call per-transaction stop callbacks if any */
1809         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1810                                  dcb_linkage) {
1811                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1812                          "commit callback entry: magic=%x name='%s'\n",
1813                          dcb->dcb_magic, dcb->dcb_name);
1814                 list_del_init(&dcb->dcb_linkage);
1815                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1816         }
1817 }
1818
1819 /*
1820  * Concurrency: shouldn't matter.
1821  */
1822 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1823                           struct thandle *th)
1824 {
1825         int                     rc = 0, remove_agents = 0;
1826         struct osd_thandle     *oh;
1827         struct osd_thread_info *oti = osd_oti_get(env);
1828         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1829         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1830         struct qsd_instance    *qsd = osd->od_quota_slave;
1831         struct lquota_trans    *qtrans;
1832         ENTRY;
1833
1834         oh = container_of0(th, struct osd_thandle, ot_super);
1835
1836         /* reset OI cache for safety */
1837         oti->oti_ins_cache_used = 0;
1838
1839         remove_agents = oh->ot_remove_agents;
1840
1841         qtrans = oh->ot_quota_trans;
1842         oh->ot_quota_trans = NULL;
1843
1844         if (oh->ot_handle != NULL) {
1845                 int rc2;
1846                 handle_t *hdl = oh->ot_handle;
1847
1848                 /*
1849                  * add commit callback
1850                  * notice we don't do this in osd_trans_start()
1851                  * as underlying transaction can change during truncate
1852                  */
1853                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1854                                          &oh->ot_jcb);
1855
1856                 LASSERT(oti->oti_txns == 1);
1857                 oti->oti_txns--;
1858
1859                 rc = dt_txn_hook_stop(env, th);
1860                 if (rc != 0)
1861                         CERROR("%s: failed in transaction hook: rc = %d\n",
1862                                osd_name(osd), rc);
1863
1864                 osd_trans_stop_cb(oh, rc);
1865                 /* hook functions might modify th_sync */
1866                 hdl->h_sync = th->th_sync;
1867
1868                 oh->ot_handle = NULL;
1869                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1870                 if (rc2 != 0)
1871                         CERROR("%s: failed to stop transaction: rc = %d\n",
1872                                osd_name(osd), rc2);
1873                 if (!rc)
1874                         rc = rc2;
1875         } else {
1876                 osd_trans_stop_cb(oh, th->th_result);
1877                 OBD_FREE_PTR(oh);
1878         }
1879
1880         /* inform the quota slave device that the transaction is stopping */
1881         qsd_op_end(env, qsd, qtrans);
1882
1883         /* as we want IO to journal and data IO be concurrent, we don't block
1884          * awaiting data IO completion in osd_do_bio(), instead we wait here
1885          * once transaction is submitted to the journal. all reqular requests
1886          * don't do direct IO (except read/write), thus this wait_event becomes
1887          * no-op for them.
1888          *
1889          * IMPORTANT: we have to wait till any IO submited by the thread is
1890          * completed otherwise iobuf may be corrupted by different request
1891          */
1892         wait_event(iobuf->dr_wait,
1893                        atomic_read(&iobuf->dr_numreqs) == 0);
1894         osd_fini_iobuf(osd, iobuf);
1895         if (!rc)
1896                 rc = iobuf->dr_error;
1897
1898         if (unlikely(remove_agents != 0))
1899                 osd_process_scheduled_agent_removals(env, osd);
1900
1901         sb_end_write(osd_sb(osd));
1902
1903         RETURN(rc);
1904 }
1905
1906 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1907 {
1908         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1909                                                ot_super);
1910
1911         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1912         LASSERT(&dcb->dcb_func != NULL);
1913         if (dcb->dcb_flags & DCB_TRANS_STOP)
1914                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1915         else
1916                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1917
1918         return 0;
1919 }
1920
1921 /*
1922  * Called just before object is freed. Releases all resources except for
1923  * object itself (that is released by osd_object_free()).
1924  *
1925  * Concurrency: no concurrent access is possible that late in object
1926  * life-cycle.
1927  */
1928 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1929 {
1930         struct osd_object *obj   = osd_obj(l);
1931         struct inode      *inode = obj->oo_inode;
1932
1933         LINVRNT(osd_invariant(obj));
1934
1935         /*
1936          * If object is unlinked remove fid->ino mapping from object index.
1937          */
1938
1939         osd_index_fini(obj);
1940         if (inode != NULL) {
1941                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1942                 qid_t                    uid = i_uid_read(inode);
1943                 qid_t                    gid = i_gid_read(inode);
1944
1945                 iput(inode);
1946                 obj->oo_inode = NULL;
1947
1948                 if (qsd != NULL) {
1949                         struct osd_thread_info  *info = osd_oti_get(env);
1950                         struct lquota_id_info   *qi = &info->oti_qi;
1951
1952                         /* Release granted quota to master if necessary */
1953                         qi->lqi_id.qid_uid = uid;
1954                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1955
1956                         qi->lqi_id.qid_uid = gid;
1957                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1958
1959                         qi->lqi_id.qid_uid = i_projid_read(inode);
1960                         qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1961                 }
1962         }
1963 }
1964
1965 /*
1966  * Concurrency: ->loo_object_release() is called under site spin-lock.
1967  */
1968 static void osd_object_release(const struct lu_env *env,
1969                                struct lu_object *l)
1970 {
1971         struct osd_object *o = osd_obj(l);
1972         /* nobody should be releasing a non-destroyed object with nlink=0
1973          * the API allows this, but ldiskfs doesn't like and then report
1974          * this inode as deleted */
1975         if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1976                 LBUG();
1977 }
1978
1979 /*
1980  * Concurrency: shouldn't matter.
1981  */
1982 static int osd_object_print(const struct lu_env *env, void *cookie,
1983                             lu_printer_t p, const struct lu_object *l)
1984 {
1985         struct osd_object *o = osd_obj(l);
1986         struct iam_descr  *d;
1987
1988         if (o->oo_dir != NULL)
1989                 d = o->oo_dir->od_container.ic_descr;
1990         else
1991                 d = NULL;
1992         return (*p)(env, cookie,
1993                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1994                     o, o->oo_inode,
1995                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1996                     o->oo_inode ? o->oo_inode->i_generation : 0,
1997                     d ? d->id_ops->id_name : "plain");
1998 }
1999
2000 /*
2001  * Concurrency: shouldn't matter.
2002  */
2003 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2004                struct obd_statfs *sfs)
2005 {
2006         struct osd_device       *osd = osd_dt_dev(d);
2007         struct super_block      *sb = osd_sb(osd);
2008         struct kstatfs          *ksfs;
2009         __u64                    reserved;
2010         int                      result = 0;
2011
2012         if (unlikely(osd->od_mnt == NULL))
2013                 return -EINPROGRESS;
2014
2015         /* osd_lproc.c call this without env, allocate ksfs for that case */
2016         if (unlikely(env == NULL)) {
2017                 OBD_ALLOC_PTR(ksfs);
2018                 if (ksfs == NULL)
2019                         return -ENOMEM;
2020         } else {
2021                 ksfs = &osd_oti_get(env)->oti_ksfs;
2022         }
2023
2024         result = sb->s_op->statfs(sb->s_root, ksfs);
2025         if (result)
2026                 goto out;
2027
2028         statfs_pack(sfs, ksfs);
2029         if (unlikely(sb->s_flags & MS_RDONLY))
2030                 sfs->os_state |= OS_STATE_READONLY;
2031         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2032                                          LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2033                 sfs->os_maxbytes = sb->s_maxbytes;
2034         else
2035                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2036
2037         /*
2038          * Reserve some space so to avoid fragmenting the filesystem too much.
2039          * Fragmentation not only impacts performance, but can also increase
2040          * metadata overhead significantly, causing grant calculation to be
2041          * wrong.
2042          *
2043          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2044          */
2045         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2046         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2047         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2048                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2049
2050         sfs->os_blocks -= reserved;
2051         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2052         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2053
2054 out:
2055         if (unlikely(env == NULL))
2056                 OBD_FREE_PTR(ksfs);
2057         return result;
2058 }
2059
2060 /**
2061  * Estimate space needed for file creations. We assume the largest filename
2062  * which is 2^64 - 1, hence a filename of 20 chars.
2063  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2064  */
2065 #ifdef __LDISKFS_DIR_REC_LEN
2066 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2067 #else
2068 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2069 #endif
2070
2071 /*
2072  * Concurrency: doesn't access mutable data.
2073  */
2074 static void osd_conf_get(const struct lu_env *env,
2075                          const struct dt_device *dev,
2076                          struct dt_device_param *param)
2077 {
2078         struct super_block *sb = osd_sb(osd_dt_dev(dev));
2079         int                ea_overhead;
2080
2081         /*
2082          * XXX should be taken from not-yet-existing fs abstraction layer.
2083          */
2084         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2085         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2086         param->ddp_symlink_max  = sb->s_blocksize;
2087         param->ddp_mount_type     = LDD_MT_LDISKFS;
2088         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2089                 param->ddp_maxbytes = sb->s_maxbytes;
2090         else
2091                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2092         /* inode are statically allocated, so per-inode space consumption
2093          * is the space consumed by the directory entry */
2094         param->ddp_inodespace     = PER_OBJ_USAGE;
2095         /* EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
2096          * = 128MB) which is unlikely to be hit in real life. Report a smaller
2097          * maximum length to not under count the actual number of extents
2098          * needed for writing a file. */
2099         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2100         /* worst-case extent insertion metadata overhead */
2101         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2102         param->ddp_mntopts      = 0;
2103         if (test_opt(sb, XATTR_USER))
2104                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2105         if (test_opt(sb, POSIX_ACL))
2106                 param->ddp_mntopts |= MNTOPT_ACL;
2107
2108         /* LOD might calculate the max stripe count based on max_ea_size,
2109          * so we need take account in the overhead as well,
2110          * xattr_header + magic + xattr_entry_head */
2111         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2112                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2113
2114 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2115         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2116                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2117                                                                 ea_overhead;
2118         else
2119 #endif
2120                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2121 }
2122
2123 /*
2124  * Concurrency: shouldn't matter.
2125  */
2126 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2127 {
2128         int rc;
2129
2130         CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2131
2132         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2133
2134         CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2135                rc);
2136
2137         return rc;
2138 }
2139
2140 /**
2141  * Start commit for OSD device.
2142  *
2143  * An implementation of dt_commit_async method for OSD device.
2144  * Asychronously starts underlayng fs sync and thereby a transaction
2145  * commit.
2146  *
2147  * \param env environment
2148  * \param d dt device
2149  *
2150  * \see dt_device_operations
2151  */
2152 static int osd_commit_async(const struct lu_env *env,
2153                             struct dt_device *d)
2154 {
2155         struct super_block *s = osd_sb(osd_dt_dev(d));
2156         ENTRY;
2157
2158         CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2159         RETURN(s->s_op->sync_fs(s, 0));
2160 }
2161
2162 /* Our own copy of the set readonly functions if present, or NU if not. */
2163 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2164 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2165 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2166
2167 /*
2168  * Concurrency: shouldn't matter.
2169  */
2170 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2171 {
2172         struct super_block *sb = osd_sb(osd_dt_dev(d));
2173         struct block_device *dev = sb->s_bdev;
2174         int rc = -EOPNOTSUPP;
2175         ENTRY;
2176
2177         if (priv_dev_set_rdonly) {
2178                 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2179
2180                 rc = 0;
2181                 CERROR("*** setting %s read-only ***\n",
2182                        osd_dt_dev(d)->od_svname);
2183
2184                 if (sb->s_op->freeze_fs) {
2185                         rc = sb->s_op->freeze_fs(sb);
2186                         if (rc)
2187                                 goto out;
2188                 }
2189
2190                 if (jdev && (jdev != dev)) {
2191                         CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2192                                (long)jdev);
2193                         priv_dev_set_rdonly(jdev);
2194                 }
2195                 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2196                 priv_dev_set_rdonly(dev);
2197
2198                 if (sb->s_op->unfreeze_fs)
2199                         sb->s_op->unfreeze_fs(sb);
2200         }
2201
2202 out:
2203         if (rc)
2204                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2205                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2206
2207         RETURN(rc);
2208 }
2209
2210 /**
2211  * Note: we do not count into QUOTA here.
2212  * If we mount with --data_journal we may need more.
2213  */
2214 const int osd_dto_credits_noquota[DTO_NR] = {
2215         /**
2216          * Insert.
2217          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2218          * SINGLEDATA_TRANS_BLOCKS(8)
2219          * XXX Note: maybe iam need more, since iam have more level than
2220          *           EXT3 htree.
2221          */
2222         [DTO_INDEX_INSERT]  = 16,
2223         /**
2224          * Delete
2225          * just modify a single entry, probably merge few within a block
2226          */
2227         [DTO_INDEX_DELETE]  = 1,
2228         /**
2229          * Used for OI scrub
2230          */
2231         [DTO_INDEX_UPDATE]  = 16,
2232         /**
2233          * 4(inode, inode bits, groups, GDT)
2234          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2235          */
2236         [DTO_OBJECT_CREATE] = 4,
2237         /**
2238          * 4(inode, inode bits, groups, GDT)
2239          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2240          */
2241         [DTO_OBJECT_DELETE] = 4,
2242         /**
2243          * Attr set credits (inode)
2244          */
2245         [DTO_ATTR_SET_BASE] = 1,
2246         /**
2247          * Xattr set. The same as xattr of EXT3.
2248          * DATA_TRANS_BLOCKS(14)
2249          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2250          * are also counted in. Do not know why?
2251          */
2252         [DTO_XATTR_SET]     = 14,
2253         /**
2254          * credits for inode change during write.
2255          */
2256         [DTO_WRITE_BASE]    = 3,
2257         /**
2258          * credits for single block write.
2259          */
2260         [DTO_WRITE_BLOCK]   = 14,
2261         /**
2262          * Attr set credits for chown.
2263          * This is extra credits for setattr, and it is null without quota
2264          */
2265         [DTO_ATTR_SET_CHOWN] = 0
2266 };
2267
2268 static const struct dt_device_operations osd_dt_ops = {
2269         .dt_root_get       = osd_root_get,
2270         .dt_statfs         = osd_statfs,
2271         .dt_trans_create   = osd_trans_create,
2272         .dt_trans_start    = osd_trans_start,
2273         .dt_trans_stop     = osd_trans_stop,
2274         .dt_trans_cb_add   = osd_trans_cb_add,
2275         .dt_conf_get       = osd_conf_get,
2276         .dt_sync           = osd_sync,
2277         .dt_ro             = osd_ro,
2278         .dt_commit_async   = osd_commit_async,
2279 };
2280
2281 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2282                           unsigned role)
2283 {
2284         struct osd_object *obj = osd_dt_obj(dt);
2285         struct osd_thread_info *oti = osd_oti_get(env);
2286
2287         LINVRNT(osd_invariant(obj));
2288
2289         LASSERT(obj->oo_owner != env);
2290         down_read_nested(&obj->oo_sem, role);
2291
2292         LASSERT(obj->oo_owner == NULL);
2293         oti->oti_r_locks++;
2294 }
2295
2296 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2297                            unsigned role)
2298 {
2299         struct osd_object *obj = osd_dt_obj(dt);
2300         struct osd_thread_info *oti = osd_oti_get(env);
2301
2302         LINVRNT(osd_invariant(obj));
2303
2304         LASSERT(obj->oo_owner != env);
2305         down_write_nested(&obj->oo_sem, role);
2306
2307         LASSERT(obj->oo_owner == NULL);
2308         obj->oo_owner = env;
2309         oti->oti_w_locks++;
2310 }
2311
2312 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2313 {
2314         struct osd_object *obj = osd_dt_obj(dt);
2315         struct osd_thread_info *oti = osd_oti_get(env);
2316
2317         LINVRNT(osd_invariant(obj));
2318
2319         LASSERT(oti->oti_r_locks > 0);
2320         oti->oti_r_locks--;
2321         up_read(&obj->oo_sem);
2322 }
2323
2324 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2325 {
2326         struct osd_object *obj = osd_dt_obj(dt);
2327         struct osd_thread_info *oti = osd_oti_get(env);
2328
2329         LINVRNT(osd_invariant(obj));
2330
2331         LASSERT(obj->oo_owner == env);
2332         LASSERT(oti->oti_w_locks > 0);
2333         oti->oti_w_locks--;
2334         obj->oo_owner = NULL;
2335         up_write(&obj->oo_sem);
2336 }
2337
2338 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2339 {
2340         struct osd_object *obj = osd_dt_obj(dt);
2341
2342         LINVRNT(osd_invariant(obj));
2343
2344         return obj->oo_owner == env;
2345 }
2346
2347 static struct timespec *osd_inode_time(const struct lu_env *env,
2348                                        struct inode *inode, __u64 seconds)
2349 {
2350         struct osd_thread_info  *oti = osd_oti_get(env);
2351         struct timespec         *t   = &oti->oti_time;
2352
2353         t->tv_sec = seconds;
2354         t->tv_nsec = 0;
2355         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2356         return t;
2357 }
2358
2359 static void osd_inode_getattr(const struct lu_env *env,
2360                               struct inode *inode, struct lu_attr *attr)
2361 {
2362         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2363                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2364                            LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2365                            LA_BLKSIZE | LA_TYPE;
2366
2367         attr->la_atime   = LTIME_S(inode->i_atime);
2368         attr->la_mtime   = LTIME_S(inode->i_mtime);
2369         attr->la_ctime   = LTIME_S(inode->i_ctime);
2370         attr->la_mode    = inode->i_mode;
2371         attr->la_size    = i_size_read(inode);
2372         attr->la_blocks  = inode->i_blocks;
2373         attr->la_uid     = i_uid_read(inode);
2374         attr->la_gid     = i_gid_read(inode);
2375         attr->la_projid  = i_projid_read(inode);
2376         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2377         attr->la_nlink   = inode->i_nlink;
2378         attr->la_rdev    = inode->i_rdev;
2379         attr->la_blksize = 1 << inode->i_blkbits;
2380         attr->la_blkbits = inode->i_blkbits;
2381         /*
2382          * Ext4 did not transfer inherit flags from raw inode
2383          * to inode flags, and ext4 internally test raw inode
2384          * @i_flags directly. Instead of patching ext4, we do it here.
2385          */
2386         if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2387                 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2388 }
2389
2390 static int osd_attr_get(const struct lu_env *env,
2391                         struct dt_object *dt,
2392                         struct lu_attr *attr)
2393 {
2394         struct osd_object *obj = osd_dt_obj(dt);
2395
2396         if (unlikely(!dt_object_exists(dt)))
2397                 return -ENOENT;
2398         if (unlikely(obj->oo_destroyed))
2399                 return -ENOENT;
2400
2401         LASSERT(!dt_object_remote(dt));
2402         LINVRNT(osd_invariant(obj));
2403
2404         spin_lock(&obj->oo_guard);
2405         osd_inode_getattr(env, obj->oo_inode, attr);
2406         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2407                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2408         spin_unlock(&obj->oo_guard);
2409
2410         return 0;
2411 }
2412
2413 static int osd_declare_attr_qid(const struct lu_env *env,
2414                                 struct osd_object *obj,
2415                                 struct osd_thandle *oh, long long bspace,
2416                                 qid_t old_id, qid_t new_id, bool enforce,
2417                                 unsigned type)
2418 {
2419         int rc;
2420         struct osd_thread_info *info = osd_oti_get(env);
2421         struct lquota_id_info  *qi = &info->oti_qi;
2422
2423         qi->lqi_type = type;
2424         /* inode accounting */
2425         qi->lqi_is_blk = false;
2426
2427         /* one more inode for the new id ... */
2428         qi->lqi_id.qid_uid = new_id;
2429         qi->lqi_space      = 1;
2430         /* Reserve credits for the new id */
2431         rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2432         if (rc == -EDQUOT || rc == -EINPROGRESS)
2433                 rc = 0;
2434         if (rc)
2435                 RETURN(rc);
2436
2437         /* and one less inode for the current id */
2438         qi->lqi_id.qid_uid = old_id;
2439         qi->lqi_space      = -1;
2440         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2441         if (rc == -EDQUOT || rc == -EINPROGRESS)
2442                 rc = 0;
2443         if (rc)
2444                 RETURN(rc);
2445
2446         /* block accounting */
2447         qi->lqi_is_blk = true;
2448
2449         /* more blocks for the new id ... */
2450         qi->lqi_id.qid_uid = new_id;
2451         qi->lqi_space      = bspace;
2452         /*
2453          * Credits for the new uid has been reserved, re-use "obj"
2454          * to save credit reservation.
2455          */
2456         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2457         if (rc == -EDQUOT || rc == -EINPROGRESS)
2458                 rc = 0;
2459         if (rc)
2460                 RETURN(rc);
2461
2462         /* and finally less blocks for the current uid */
2463         qi->lqi_id.qid_uid = old_id;
2464         qi->lqi_space      = -bspace;
2465         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2466         if (rc == -EDQUOT || rc == -EINPROGRESS)
2467                 rc = 0;
2468
2469         RETURN(rc);
2470 }
2471
2472 static int osd_declare_attr_set(const struct lu_env *env,
2473                                 struct dt_object *dt,
2474                                 const struct lu_attr *attr,
2475                                 struct thandle *handle)
2476 {
2477         struct osd_thandle     *oh;
2478         struct osd_object      *obj;
2479         qid_t                   uid;
2480         qid_t                   gid;
2481         long long               bspace;
2482         int                     rc = 0;
2483         bool                    enforce;
2484         ENTRY;
2485
2486         LASSERT(dt != NULL);
2487         LASSERT(handle != NULL);
2488
2489         obj = osd_dt_obj(dt);
2490         LASSERT(osd_invariant(obj));
2491
2492         oh = container_of0(handle, struct osd_thandle, ot_super);
2493         LASSERT(oh->ot_handle == NULL);
2494
2495         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2496                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2497
2498         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2499                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2500
2501         if (attr == NULL || obj->oo_inode == NULL)
2502                 RETURN(rc);
2503
2504         bspace   = obj->oo_inode->i_blocks;
2505         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2506         bspace   = toqb(bspace);
2507
2508         /* Changing ownership is always preformed by super user, it should not
2509          * fail with EDQUOT.
2510          *
2511          * We still need to call the osd_declare_qid() to calculate the journal
2512          * credits for updating quota accounting files and to trigger quota
2513          * space adjustment once the operation is completed.*/
2514         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2515                 /* USERQUOTA */
2516                 uid = i_uid_read(obj->oo_inode);
2517                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2518                 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2519                                           attr->la_uid, enforce, USRQUOTA);
2520                 if (rc)
2521                         RETURN(rc);
2522
2523                 gid = i_gid_read(obj->oo_inode);
2524                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2525                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2526                                           i_gid_read(obj->oo_inode),
2527                                           attr->la_gid, enforce, GRPQUOTA);
2528                 if (rc)
2529                         RETURN(rc);
2530
2531         }
2532 #ifdef HAVE_PROJECT_QUOTA
2533         if (attr->la_valid & LA_PROJID) {
2534                 __u32 projid = i_projid_read(obj->oo_inode);
2535                 enforce = (attr->la_valid & LA_PROJID) &&
2536                                         (attr->la_projid != projid);
2537                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2538                                           (qid_t)projid, (qid_t)attr->la_projid,
2539                                           enforce, PRJQUOTA);
2540                 if (rc)
2541                         RETURN(rc);
2542         }
2543 #endif
2544         RETURN(rc);
2545 }
2546
2547 static int osd_inode_setattr(const struct lu_env *env,
2548                              struct inode *inode, const struct lu_attr *attr)
2549 {
2550         __u64 bits = attr->la_valid;
2551
2552         /* Only allow set size for regular file */
2553         if (!S_ISREG(inode->i_mode))
2554                 bits &= ~(LA_SIZE | LA_BLOCKS);
2555
2556         if (bits == 0)
2557                 return 0;
2558
2559         if (bits & LA_ATIME)
2560                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2561         if (bits & LA_CTIME)
2562                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2563         if (bits & LA_MTIME)
2564                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2565         if (bits & LA_SIZE) {
2566                 spin_lock(&inode->i_lock);
2567                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2568                 i_size_write(inode, attr->la_size);
2569                 spin_unlock(&inode->i_lock);
2570         }
2571
2572         /* OSD should not change "i_blocks" which is used by quota.
2573          * "i_blocks" should be changed by ldiskfs only. */
2574         if (bits & LA_MODE)
2575                 inode->i_mode = (inode->i_mode & S_IFMT) |
2576                                 (attr->la_mode & ~S_IFMT);
2577         if (bits & LA_UID)
2578                 i_uid_write(inode, attr->la_uid);
2579         if (bits & LA_GID)
2580                 i_gid_write(inode, attr->la_gid);
2581         if (bits & LA_PROJID)
2582                 i_projid_write(inode, attr->la_projid);
2583         if (bits & LA_NLINK)
2584                 set_nlink(inode, attr->la_nlink);
2585         if (bits & LA_RDEV)
2586                 inode->i_rdev = attr->la_rdev;
2587
2588         if (bits & LA_FLAGS) {
2589                 /* always keep S_NOCMTIME */
2590                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2591                                  S_NOCMTIME;
2592                 /*
2593                  * Ext4 did not transfer inherit flags from
2594                  * @inode->i_flags to raw inode i_flags when writing
2595                  * flags, we do it explictly here.
2596                  */
2597                 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2598                         LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2599         }
2600         return 0;
2601 }
2602
2603 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2604 {
2605         int rc;
2606
2607         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2608             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2609                 struct iattr    iattr;
2610
2611                 ll_vfs_dq_init(inode);
2612                 iattr.ia_valid = 0;
2613                 if (attr->la_valid & LA_UID)
2614                         iattr.ia_valid |= ATTR_UID;
2615                 if (attr->la_valid & LA_GID)
2616                         iattr.ia_valid |= ATTR_GID;
2617                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2618                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2619
2620                 rc = ll_vfs_dq_transfer(inode, &iattr);
2621                 if (rc) {
2622                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2623                                "enforcement enabled on the ldiskfs "
2624                                "filesystem?\n", inode->i_sb->s_id, rc);
2625                         return rc;
2626                 }
2627         }
2628
2629 #ifdef HAVE_PROJECT_QUOTA
2630         /* Handle project id transfer here properly */
2631         if (attr->la_valid & LA_PROJID &&
2632             attr->la_projid != i_projid_read(inode)) {
2633                 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2634                 if (rc) {
2635                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2636                                "enforcement enabled on the ldiskfs "
2637                                "filesystem?\n", inode->i_sb->s_id, rc);
2638                         return rc;
2639                 }
2640         }
2641 #endif
2642         return 0;
2643 }
2644
2645 static int osd_attr_set(const struct lu_env *env,
2646                         struct dt_object *dt,
2647                         const struct lu_attr *attr,
2648                         struct thandle *handle)
2649 {
2650         struct osd_object *obj = osd_dt_obj(dt);
2651         struct inode      *inode;
2652         int rc;
2653
2654         if (!dt_object_exists(dt))
2655                 return -ENOENT;
2656
2657         LASSERT(handle != NULL);
2658         LASSERT(!dt_object_remote(dt));
2659         LASSERT(osd_invariant(obj));
2660
2661         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2662
2663         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2664                 struct osd_thread_info  *oti  = osd_oti_get(env);
2665                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
2666                 struct lu_fid           *fid1 = &oti->oti_fid;
2667                 struct osd_inode_id     *id   = &oti->oti_id;
2668                 struct iam_path_descr   *ipd;
2669                 struct iam_container    *bag;
2670                 struct osd_thandle      *oh;
2671                 int                      rc;
2672
2673                 fid_cpu_to_be(fid1, fid0);
2674                 memset(id, 1, sizeof(*id));
2675                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2676                                   fid0)->oi_dir.od_container;
2677                 ipd = osd_idx_ipd_get(env, bag);
2678                 if (unlikely(ipd == NULL))
2679                         RETURN(-ENOMEM);
2680
2681                 oh = container_of0(handle, struct osd_thandle, ot_super);
2682                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2683                                 (const struct iam_rec *)id, ipd);
2684                 osd_ipd_put(env, bag, ipd);
2685                 return(rc > 0 ? 0 : rc);
2686         }
2687
2688         inode = obj->oo_inode;
2689
2690         rc = osd_quota_transfer(inode, attr);
2691         if (rc)
2692                 return rc;
2693
2694         spin_lock(&obj->oo_guard);
2695         rc = osd_inode_setattr(env, inode, attr);
2696         spin_unlock(&obj->oo_guard);
2697         if (rc != 0)
2698                 GOTO(out, rc);
2699
2700         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2701
2702         if (!(attr->la_valid & LA_FLAGS))
2703                 GOTO(out, rc);
2704
2705         /* Let's check if there are extra flags need to be set into LMA */
2706         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2707                 struct osd_thread_info *info = osd_oti_get(env);
2708                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2709
2710                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2711                                  &info->oti_ost_attrs);
2712                 if (rc)
2713                         GOTO(out, rc);
2714
2715                 lma->lma_incompat |=
2716                         lustre_to_lma_flags(attr->la_flags);
2717                 lustre_lma_swab(lma);
2718                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2719                                      lma, sizeof(*lma), XATTR_REPLACE);
2720                 if (rc != 0) {
2721                         struct osd_device *osd = osd_obj2dev(obj);
2722
2723                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2724                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2725                               lma->lma_incompat, rc);
2726                 } else {
2727                         obj->oo_lma_flags =
2728                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2729                 }
2730                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2731         }
2732 out:
2733         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2734
2735         return rc;
2736 }
2737
2738 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2739                                            struct osd_object *obj,
2740                                            const char *name, const int namelen)
2741 {
2742         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2743 }
2744
2745 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2746                       umode_t mode, struct dt_allocation_hint *hint,
2747                       struct thandle *th)
2748 {
2749         int result;
2750         struct osd_device  *osd = osd_obj2dev(obj);
2751         struct osd_thandle *oth;
2752         struct dt_object   *parent = NULL;
2753         struct inode       *inode;
2754
2755         LINVRNT(osd_invariant(obj));
2756         LASSERT(obj->oo_inode == NULL);
2757         LASSERT(obj->oo_hl_head == NULL);
2758
2759         if (S_ISDIR(mode) && ldiskfs_pdo) {
2760                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2761                 if (obj->oo_hl_head == NULL)
2762                         return -ENOMEM;
2763         }
2764
2765         oth = container_of(th, struct osd_thandle, ot_super);
2766         LASSERT(oth->ot_handle->h_transaction != NULL);
2767
2768         if (hint != NULL && hint->dah_parent != NULL &&
2769             !dt_object_remote(hint->dah_parent))
2770                 parent = hint->dah_parent;
2771
2772         inode = ldiskfs_create_inode(oth->ot_handle,
2773                                      parent ? osd_dt_obj(parent)->oo_inode :
2774                                               osd_sb(osd)->s_root->d_inode,
2775                                      mode);
2776         if (!IS_ERR(inode)) {
2777                 /* Do not update file c/mtime in ldiskfs. */
2778                 inode->i_flags |= S_NOCMTIME;
2779
2780                 /* For new created object, it must be consistent,
2781                  * and it is unnecessary to scrub against it. */
2782                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2783
2784                 obj->oo_inode = inode;
2785                 result = 0;
2786         } else {
2787                 if (obj->oo_hl_head != NULL) {
2788                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2789                         obj->oo_hl_head = NULL;
2790                 }
2791                 result = PTR_ERR(inode);
2792         }
2793         LINVRNT(osd_invariant(obj));
2794         return result;
2795 }
2796
2797 enum {
2798         OSD_NAME_LEN = 255
2799 };
2800
2801 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2802                      struct lu_attr *attr,
2803                      struct dt_allocation_hint *hint,
2804                      struct dt_object_format *dof,
2805                      struct thandle *th)
2806 {
2807         int result;
2808         struct osd_thandle *oth;
2809         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2810
2811         LASSERT(S_ISDIR(attr->la_mode));
2812
2813         oth = container_of(th, struct osd_thandle, ot_super);
2814         LASSERT(oth->ot_handle->h_transaction != NULL);
2815         result = osd_mkfile(info, obj, mode, hint, th);
2816
2817         return result;
2818 }
2819
2820 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2821                         struct lu_attr *attr,
2822                         struct dt_allocation_hint *hint,
2823                         struct dt_object_format *dof,
2824                         struct thandle *th)
2825 {
2826         int result;
2827         struct osd_thandle *oth;
2828         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2829
2830         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2831
2832         LASSERT(S_ISREG(attr->la_mode));
2833
2834         oth = container_of(th, struct osd_thandle, ot_super);
2835         LASSERT(oth->ot_handle->h_transaction != NULL);
2836
2837         result = osd_mkfile(info, obj, mode, hint, th);
2838         if (result == 0) {
2839                 LASSERT(obj->oo_inode != NULL);
2840                 if (feat->dif_flags & DT_IND_VARKEY)
2841                         result = iam_lvar_create(obj->oo_inode,
2842                                                  feat->dif_keysize_max,
2843                                                  feat->dif_ptrsize,
2844                                                  feat->dif_recsize_max,
2845                                                  oth->ot_handle);
2846                 else
2847                         result = iam_lfix_create(obj->oo_inode,
2848                                                  feat->dif_keysize_max,
2849                                                  feat->dif_ptrsize,
2850                                                  feat->dif_recsize_max,
2851                                                  oth->ot_handle);
2852
2853         }
2854         return result;
2855 }
2856
2857 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2858                      struct lu_attr *attr,
2859                      struct dt_allocation_hint *hint,
2860                      struct dt_object_format *dof,
2861                      struct thandle *th)
2862 {
2863         LASSERT(S_ISREG(attr->la_mode));
2864         return osd_mkfile(info, obj, (attr->la_mode &
2865                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2866 }
2867
2868 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2869                      struct lu_attr *attr,
2870                      struct dt_allocation_hint *hint,
2871                      struct dt_object_format *dof,
2872                      struct thandle *th)
2873 {
2874         LASSERT(S_ISLNK(attr->la_mode));
2875         return osd_mkfile(info, obj, (attr->la_mode &
2876                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2877 }
2878
2879 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2880                      struct lu_attr *attr,
2881                      struct dt_allocation_hint *hint,
2882                      struct dt_object_format *dof,
2883                      struct thandle *th)
2884 {
2885         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2886         int result;
2887
2888         LINVRNT(osd_invariant(obj));
2889         LASSERT(obj->oo_inode == NULL);
2890         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2891                 S_ISFIFO(mode) || S_ISSOCK(mode));
2892
2893         result = osd_mkfile(info, obj, mode, hint, th);
2894         if (result == 0) {
2895                 LASSERT(obj->oo_inode != NULL);
2896                 /*
2897                  * This inode should be marked dirty for i_rdev.  Currently
2898                  * that is done in the osd_attr_init().
2899                  */
2900                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2901                                    attr->la_rdev);
2902         }
2903         LINVRNT(osd_invariant(obj));
2904         return result;
2905 }
2906
2907 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2908                               struct lu_attr *,
2909                               struct dt_allocation_hint *hint,
2910                               struct dt_object_format *dof,
2911                               struct thandle *);
2912
2913 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2914 {
2915         osd_obj_type_f result;
2916
2917         switch (type) {
2918         case DFT_DIR:
2919                 result = osd_mkdir;
2920                 break;
2921         case DFT_REGULAR:
2922                 result = osd_mkreg;
2923                 break;
2924         case DFT_SYM:
2925                 result = osd_mksym;
2926                 break;
2927         case DFT_NODE:
2928                 result = osd_mknod;
2929                 break;
2930         case DFT_INDEX:
2931                 result = osd_mk_index;
2932                 break;
2933
2934         default:
2935                 LBUG();
2936                 break;
2937         }
2938         return result;
2939 }
2940
2941
2942 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2943                         struct dt_object *parent, struct dt_object *child,
2944                         umode_t child_mode)
2945 {
2946         LASSERT(ah);
2947
2948         ah->dah_parent = parent;
2949         ah->dah_mode = child_mode;
2950
2951         if (parent != NULL && !dt_object_remote(parent)) {
2952                 /* will help to find FID->ino at dt_insert("..") */
2953                 struct osd_object *pobj = osd_dt_obj(parent);
2954                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2955         }
2956 }
2957
2958 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2959                           struct lu_attr *attr, struct dt_object_format *dof)
2960 {
2961         struct inode   *inode = obj->oo_inode;
2962         __u64           valid = attr->la_valid;
2963         int             result;
2964
2965         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2966
2967         if (dof->dof_type != DFT_NODE)
2968                 attr->la_valid &= ~LA_RDEV;
2969         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2970                 attr->la_valid &= ~LA_ATIME;
2971         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2972                 attr->la_valid &= ~LA_CTIME;
2973         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2974                 attr->la_valid &= ~LA_MTIME;
2975
2976         result = osd_quota_transfer(inode, attr);
2977         if (result)
2978                 return;
2979
2980         if (attr->la_valid != 0) {
2981                 result = osd_inode_setattr(info->oti_env, inode, attr);
2982                 /*
2983                  * The osd_inode_setattr() should always succeed here.  The
2984                  * only error that could be returned is EDQUOT when we are
2985                  * trying to change the UID or GID of the inode. However, this
2986                  * should not happen since quota enforcement is no longer
2987                  * enabled on ldiskfs (lquota takes care of it).
2988                  */
2989                 LASSERTF(result == 0, "%d\n", result);
2990                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2991         }
2992
2993         attr->la_valid = valid;
2994 }
2995
2996 /**
2997  * Helper function for osd_create()
2998  *
2999  * \retval 0, on success
3000  */
3001 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3002                         struct lu_attr *attr, struct dt_allocation_hint *hint,
3003                         struct dt_object_format *dof, struct thandle *th)
3004 {
3005         int     result;
3006         __u32   umask;
3007
3008         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3009
3010         /* we drop umask so that permissions we pass are not affected */
3011         umask = current->fs->umask;
3012         current->fs->umask = 0;
3013
3014         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3015                                                   th);
3016         if (likely(obj->oo_inode != NULL)) {
3017                 LASSERT(obj->oo_inode->i_state & I_NEW);
3018
3019                 /* Unlock the inode before attr initialization to avoid
3020                  * unnecessary dqget operations. LU-6378 */
3021                 unlock_new_inode(obj->oo_inode);
3022         }
3023
3024         if (likely(result == 0)) {
3025                 osd_attr_init(info, obj, attr, dof);
3026                 osd_object_init0(obj);
3027         }
3028
3029         /* restore previous umask value */
3030         current->fs->umask = umask;
3031
3032         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3033
3034         return result;
3035 }
3036
3037 /**
3038  * Helper function for osd_create()
3039  *
3040  * \retval 0, on success
3041  */
3042 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3043                            const struct lu_fid *fid, struct thandle *th)
3044 {
3045         struct osd_thread_info *info = osd_oti_get(env);
3046         struct osd_inode_id    *id   = &info->oti_id;
3047         struct osd_device      *osd  = osd_obj2dev(obj);
3048         struct osd_thandle     *oh;
3049         int                     rc;
3050
3051         LASSERT(obj->oo_inode != NULL);
3052
3053         oh = container_of0(th, struct osd_thandle, ot_super);
3054         LASSERT(oh->ot_handle);
3055         osd_trans_exec_op(env, th, OSD_OT_INSERT);
3056
3057         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3058         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3059                            OI_CHECK_FLD, NULL);
3060         osd_trans_exec_check(env, th, OSD_OT_INSERT);
3061
3062         return rc;
3063 }
3064
3065 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3066                    u64 seq, struct lu_seq_range *range)
3067 {
3068         struct seq_server_site  *ss = osd_seq_site(osd);
3069
3070         if (fid_seq_is_idif(seq)) {
3071                 fld_range_set_ost(range);
3072                 range->lsr_index = idif_ost_idx(seq);
3073                 return 0;
3074         }
3075
3076         if (!fid_seq_in_fldb(seq)) {
3077                 fld_range_set_mdt(range);
3078                 if (ss != NULL)
3079                         /* FIXME: If ss is NULL, it suppose not get lsr_index
3080                          * at all */
3081                         range->lsr_index = ss->ss_node_id;
3082                 return 0;
3083         }
3084
3085         LASSERT(ss != NULL);
3086         fld_range_set_any(range);
3087         /* OSD will only do local fld lookup */
3088         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3089 }
3090
3091 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3092                               struct lu_attr *attr,
3093                               struct dt_allocation_hint *hint,
3094                               struct dt_object_format *dof,
3095                               struct thandle *handle)
3096 {
3097         struct osd_thandle *oh;
3098         int rc;
3099         ENTRY;
3100
3101         LASSERT(handle != NULL);
3102
3103         oh = container_of0(handle, struct osd_thandle, ot_super);
3104         LASSERT(oh->ot_handle == NULL);
3105
3106         /* EA object consumes more credits than regular object: osd_mk_index
3107          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3108          * leaf_node, could involves the block, block bitmap, groups, GDT
3109          * change for each block, so add 4 * 2 credits in that case. */
3110         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3111                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3112                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3113         /* Reuse idle OI block may cause additional one OI block
3114          * to be changed. */
3115         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3116                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3117
3118         if (!attr)
3119                 RETURN(0);
3120
3121         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3122                                    attr->la_projid, 1, oh, osd_dt_obj(dt),
3123                                    NULL, OSD_QID_INODE);
3124         if (rc != 0)
3125                 RETURN(rc);
3126
3127         /* will help to find FID->ino mapping at dt_insert() */
3128         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3129                                    osd_dt_obj(dt));
3130
3131         RETURN(rc);
3132 }
3133
3134 /**
3135  * Called to destroy on-disk representation of the object
3136  *
3137  * Concurrency: must be locked
3138  */
3139 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3140                                struct thandle *th)
3141 {
3142         struct osd_object  *obj = osd_dt_obj(dt);
3143         struct inode       *inode = obj->oo_inode;
3144         struct osd_thandle *oh;
3145         int                 rc;
3146         ENTRY;
3147
3148         if (inode == NULL)
3149                 RETURN(-ENOENT);
3150
3151         oh = container_of0(th, struct osd_thandle, ot_super);
3152         LASSERT(oh->ot_handle == NULL);
3153
3154         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3155                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3156         /* Recycle idle OI leaf may cause additional three OI blocks
3157          * to be changed. */
3158         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3159                 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3160                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3161         /* one less inode */
3162         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3163                                    i_projid_read(inode), -1, oh, obj, NULL,
3164                                    OSD_QID_INODE);
3165         if (rc)
3166                 RETURN(rc);
3167         /* data to be truncated */
3168         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3169                                    i_projid_read(inode), 0, oh, obj, NULL,
3170                                    OSD_QID_BLK);
3171         if (rc)
3172                 RETURN(rc);
3173
3174         /* will help to find FID->ino when this object is being
3175          * added to PENDING/ */
3176         rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3177
3178         RETURN(rc);
3179 }
3180
3181 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3182                        struct thandle *th)
3183 {
3184         const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3185         struct osd_object *obj = osd_dt_obj(dt);
3186         struct inode *inode = obj->oo_inode;
3187         struct osd_device *osd = osd_obj2dev(obj);
3188         struct osd_thandle *oh;
3189         int result;
3190         ENTRY;
3191
3192         oh = container_of0(th, struct osd_thandle, ot_super);
3193         LASSERT(oh->ot_handle);
3194         LASSERT(inode);
3195         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3196
3197         if (unlikely(fid_is_acct(fid)))
3198                 RETURN(-EPERM);
3199
3200         if (S_ISDIR(inode->i_mode)) {
3201                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3202                         inode->i_nlink == 2);
3203                 /* it will check/delete the inode from remote parent,
3204                  * how to optimize it? unlink performance impaction XXX */
3205                 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3206                 if (result != 0)
3207                         CERROR("%s: delete inode "DFID": rc = %d\n",
3208                                osd_name(osd), PFID(fid), result);
3209
3210                 spin_lock(&obj->oo_guard);
3211                 clear_nlink(inode);
3212                 spin_unlock(&obj->oo_guard);
3213                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3214         }
3215
3216         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3217
3218         ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3219
3220         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3221                 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3222                                        oh->ot_handle, OI_CHECK_FLD);
3223
3224         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3225         /* XXX: add to ext3 orphan list */
3226         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3227
3228         /* not needed in the cache anymore */
3229         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3230         obj->oo_destroyed = 1;
3231
3232         RETURN(0);
3233 }
3234
3235 /**
3236  * Put the fid into lustre_mdt_attrs, and then place the structure
3237  * inode's ea. This fid should not be altered during the life time
3238  * of the inode.
3239  *
3240  * \retval +ve, on success
3241  * \retval -ve, on error
3242  *
3243  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3244  */
3245 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3246                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
3247 {
3248         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3249         struct lustre_mdt_attrs *lma = &loa->loa_lma;
3250         int rc;
3251         ENTRY;
3252
3253         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3254                 RETURN(0);
3255
3256         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3257                 rc = -ENOMEM;
3258
3259         lustre_loa_init(loa, fid, compat, incompat);
3260         lustre_loa_swab(loa, false);
3261
3262         /* For the OST device with 256 bytes inode size by default,
3263          * the PFID EA will be stored together with LMA EA to avoid
3264          * performance trouble. Otherwise the PFID EA can be stored
3265          * independently. LU-8998 */
3266         if ((compat & LMAC_FID_ON_OST) &&
3267             LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3268                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3269                                      sizeof(*loa), XATTR_CREATE);
3270         else
3271                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3272                                      sizeof(*lma), XATTR_CREATE);
3273         /* LMA may already exist, but we need to check that all the
3274          * desired compat/incompat flags have been added. */
3275         if (unlikely(rc == -EEXIST)) {
3276                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3277                                      XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3278                 if (rc < 0)
3279                         RETURN(rc);
3280
3281                 if (rc < sizeof(*lma))
3282                         RETURN(-EINVAL);
3283
3284                 lustre_loa_swab(loa, true);
3285                 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3286                     ((compat == 0 && incompat == 0) ||
3287                      (!(~lma->lma_compat & compat) &&
3288                       !(~lma->lma_incompat & incompat))))
3289                         RETURN(0);
3290
3291                 lma->lma_self_fid = *fid;
3292                 lma->lma_compat |= compat;
3293                 lma->lma_incompat |= incompat;
3294                 if (rc == sizeof(*lma)) {
3295                         lustre_lma_swab(lma);
3296                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3297                                              sizeof(*lma), XATTR_REPLACE);
3298                 } else {
3299                         lustre_loa_swab(loa, false);
3300                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3301                                              sizeof(*loa), XATTR_REPLACE);
3302                 }
3303         }
3304
3305         RETURN(rc);
3306 }
3307
3308 /**
3309  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3310  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3311  * To have compatilibility with 1.8 ldiskfs driver we need to have
3312  * magic number at start of fid data.
3313  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3314  * its inmemory API.
3315  */
3316 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3317                                          const struct lu_fid *fid)
3318 {
3319         if (!fid_is_namespace_visible(fid) ||
3320             OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3321                 param->edp_magic = 0;
3322                 return;
3323         }
3324
3325         param->edp_magic = LDISKFS_LUFID_MAGIC;
3326         param->edp_len =  sizeof(struct lu_fid) + 1;
3327         fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3328 }
3329
3330 /**
3331  * Try to read the fid from inode ea into dt_rec.
3332  *
3333  * \param fid object fid.
3334  *
3335  * \retval 0 on success
3336  */
3337 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3338                           __u32 ino, struct lu_fid *fid,
3339                           struct osd_inode_id *id)
3340 {
3341         struct osd_thread_info *info  = osd_oti_get(env);
3342         struct inode           *inode;
3343         ENTRY;
3344
3345         osd_id_gen(id, ino, OSD_OII_NOGEN);
3346         inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3347         if (IS_ERR(inode))
3348                 RETURN(PTR_ERR(inode));
3349
3350         iput(inode);
3351         RETURN(0);
3352 }
3353
3354 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3355                                         struct inode *dir,
3356                                         struct inode *parent_dir,
3357                                         const struct lu_fid *dot_fid,
3358                                         const struct lu_fid *dot_dot_fid,
3359                                         struct osd_thandle *oth)
3360 {
3361         struct ldiskfs_dentry_param *dot_ldp;
3362         struct ldiskfs_dentry_param *dot_dot_ldp;
3363         __u32 saved_nlink = dir->i_nlink;
3364         int rc;
3365
3366         dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3367         osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3368
3369         dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3370         dot_ldp->edp_magic = 0;
3371
3372         rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3373                                     dir, dot_ldp, dot_dot_ldp);
3374         /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3375          * the subseqent ref_add() will increase the dir->i_nlink
3376          * as 3. That is incorrect for new created directory.
3377          *
3378          * It looks like hack, because we want to make the OSD API
3379          * to be order-independent for new created directory object
3380          * between dt_insert(..) and ref_add() operations.
3381          *
3382          * Here, we only restore the in-RAM dir-inode's nlink attr,
3383          * becuase if the nlink attr is not 2, then there will be
3384          * ref_add() called following the dt_insert(..), such call
3385          * will make both the in-RAM and on-disk dir-inode's nlink
3386          * attr to be set as 2. LU-7447 */
3387         set_nlink(dir, saved_nlink);
3388         return rc;
3389 }
3390
3391 /**
3392  * Create an local agent inode for remote entry
3393  */
3394 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3395                                                   struct osd_device *osd,
3396                                                   struct osd_object *pobj,
3397                                                   const struct lu_fid *fid,
3398                                                   __u32 type,
3399                                                   struct thandle *th)
3400 {
3401         struct osd_thread_info  *info = osd_oti_get(env);
3402         struct inode            *local;
3403         struct osd_thandle      *oh;
3404         int                     rc;
3405         ENTRY;
3406
3407         LASSERT(th);
3408         oh = container_of(th, struct osd_thandle, ot_super);
3409         LASSERT(oh->ot_handle->h_transaction != NULL);
3410
3411         local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3412         if (IS_ERR(local)) {
3413                 CERROR("%s: create local error %d\n", osd_name(osd),
3414                        (int)PTR_ERR(local));
3415                 RETURN(local);
3416         }
3417
3418         /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3419          * correct gid on remote file, not agent here */
3420         local->i_gid = current_fsgid();
3421         ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3422         unlock_new_inode(local);
3423
3424         /* Agent inode should not have project ID */
3425 #ifdef  HAVE_PROJECT_QUOTA
3426         if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3427                 rc = __ldiskfs_ioctl_setproject(local, 0);
3428                 if (rc) {
3429                         CERROR("%s: quota transfer failed: rc = %d. Is project "
3430                                "quota enforcement enabled on the ldiskfs "
3431                                "filesystem?\n", local->i_sb->s_id, rc);
3432                         RETURN(ERR_PTR(rc));
3433                 }
3434         }
3435 #endif
3436         /* Set special LMA flag for local agent inode */
3437         rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3438         if (rc != 0) {
3439                 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3440                        osd_name(osd), PFID(fid), rc);
3441                 RETURN(ERR_PTR(rc));
3442         }
3443
3444         if (!S_ISDIR(type))
3445                 RETURN(local);
3446
3447         rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3448                                          lu_object_fid(&pobj->oo_dt.do_lu),
3449                                          fid, oh);
3450         if (rc != 0) {
3451                 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3452                         osd_name(osd), PFID(fid), rc);
3453                 RETURN(ERR_PTR(rc));
3454         }
3455
3456         RETURN(local);
3457 }
3458
3459 /**
3460  * when direntry is deleted, we have to take care of possible agent inode
3461  * referenced by that. unfortunately we can't do this at that point:
3462  * iget() within a running transaction leads to deadlock and we better do
3463  * not call that every delete declaration to save performance. so we put
3464  * a potention agent inode on a list and process that once the transaction
3465  * is over. Notice it's not any worse in terms of real orphans as regular
3466  * object destroy doesn't put inodes on the on-disk orphan list. this should
3467  * be addressed separately
3468  */
3469 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3470                                             struct osd_thandle *oh,
3471                                             __u32 ino)
3472 {
3473         struct osd_device      *osd = osd_dt_dev(oh->ot_super.th_dev);
3474         struct osd_obj_orphan *oor;
3475
3476         OBD_ALLOC_PTR(oor);
3477         if (oor == NULL)
3478                 return -ENOMEM;
3479
3480         oor->oor_ino = ino;
3481         oor->oor_env = (struct lu_env *)env;
3482         spin_lock(&osd->od_osfs_lock);
3483         list_add(&oor->oor_list, &osd->od_orphan_list);
3484         spin_unlock(&osd->od_osfs_lock);
3485
3486         oh->ot_remove_agents = 1;
3487
3488         return 0;
3489
3490 }
3491
3492 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3493                                                 struct osd_device *osd)
3494 {
3495         struct osd_thread_info *info = osd_oti_get(env);
3496         struct osd_obj_orphan *oor, *tmp;
3497         struct osd_inode_id id;
3498         struct list_head list;
3499         struct inode *inode;
3500         struct lu_fid fid;
3501         handle_t *jh;
3502         __u32 ino;
3503
3504         INIT_LIST_HEAD(&list);
3505
3506         spin_lock(&osd->od_osfs_lock);
3507         list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3508                 if (oor->oor_env == env) {
3509                         list_del(&oor->oor_list);
3510                         list_add(&oor->oor_list, &list);
3511                 }
3512         }
3513         spin_unlock(&osd->od_osfs_lock);
3514
3515         list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3516
3517                 ino = oor->oor_ino;
3518
3519                 list_del(&oor->oor_list);
3520                 OBD_FREE_PTR(oor);
3521
3522                 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3523                 inode = osd_iget_fid(info, osd, &id, &fid);
3524                 if (IS_ERR(inode))
3525                         continue;
3526
3527                 if (!osd_remote_fid(env, osd, &fid)) {
3528                         iput(inode);
3529                         continue;
3530                 }
3531
3532                 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3533                 clear_nlink(inode);
3534                 mark_inode_dirty(inode);
3535                 ldiskfs_journal_stop(jh);
3536                 iput(inode);
3537         }
3538
3539         return 0;
3540 }
3541
3542 /**
3543  * OSD layer object create function for OST objects (b=11826).
3544  *
3545  * The FID is inserted into inode xattr here.
3546  *
3547  * \retval   0, on success
3548  * \retval -ve, on error
3549  */
3550 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3551                       struct lu_attr *attr, struct dt_allocation_hint *hint,
3552                       struct dt_object_format *dof, struct thandle *th)
3553 {
3554         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3555         struct osd_object       *obj    = osd_dt_obj(dt);
3556         struct osd_thread_info  *info   = osd_oti_get(env);
3557         int                      result, on_ost = 0;
3558
3559         ENTRY;
3560
3561         if (dt_object_exists(dt))
3562                 RETURN(-EEXIST);
3563
3564         LINVRNT(osd_invariant(obj));
3565         LASSERT(!dt_object_remote(dt));
3566         LASSERT(osd_is_write_locked(env, obj));
3567         LASSERT(th != NULL);
3568
3569         if (unlikely(fid_is_acct(fid)))
3570                 /* Quota files can't be created from the kernel any more,
3571                  * 'tune2fs -O quota' will take care of creating them */
3572                 RETURN(-EPERM);
3573
3574         result = __osd_create(info, obj, attr, hint, dof, th);
3575         if (result == 0) {
3576                 if (fid_is_idif(fid) &&
3577                     !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3578                         struct lu_fid *tfid = &info->oti_fid;
3579                         struct ost_id *oi   = &info->oti_ostid;
3580
3581                         fid_to_ostid(fid, oi);
3582                         ostid_to_fid(tfid, oi, 0);
3583                         on_ost = 1;
3584                         result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3585                                                 LMAC_FID_ON_OST, 0);
3586                 } else {
3587                         on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3588                                                fid, OI_CHECK_FLD);
3589                         result = osd_ea_fid_set(info, obj->oo_inode, fid,
3590                                                 on_ost ? LMAC_FID_ON_OST : 0,
3591                                                 0);
3592                 }
3593                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3594                         obj->oo_dt.do_body_ops = &osd_body_ops;
3595         }
3596
3597         if (result == 0)
3598                 result = __osd_oi_insert(env, obj, fid, th);
3599
3600         /* a small optimization - dt_insert() isn't usually applied
3601          * to OST objects, so we don't need to cache OI mapping for
3602          * OST objects */
3603         if (result == 0 && on_ost == 0) {
3604                 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3605                 result = osd_idc_find_and_init(env, osd, obj);
3606                 LASSERT(result == 0);
3607         }
3608
3609         LASSERT(ergo(result == 0,
3610                      dt_object_exists(dt) && !dt_object_remote(dt)));
3611         LINVRNT(osd_invariant(obj));
3612         RETURN(result);
3613 }
3614
3615 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3616                                struct thandle *handle)
3617 {
3618         struct osd_thandle *oh;
3619
3620         /* it's possible that object doesn't exist yet */
3621         LASSERT(handle != NULL);
3622
3623         oh = container_of0(handle, struct osd_thandle, ot_super);
3624         LASSERT(oh->ot_handle == NULL);
3625
3626         osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3627                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3628
3629         return 0;
3630 }
3631
3632 /*
3633  * Concurrency: @dt is write locked.
3634  */
3635 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3636                        struct thandle *th)
3637 {
3638         struct osd_object  *obj = osd_dt_obj(dt);
3639         struct inode       *inode = obj->oo_inode;
3640         struct osd_thandle *oh;
3641         int                 rc = 0;
3642
3643         if (!dt_object_exists(dt) || obj->oo_destroyed)
3644                 return -ENOENT;
3645
3646         LINVRNT(osd_invariant(obj));
3647         LASSERT(!dt_object_remote(dt));
3648         LASSERT(osd_is_write_locked(env, obj));
3649         LASSERT(th != NULL);
3650
3651         oh = container_of0(th, struct osd_thandle, ot_super);
3652         LASSERT(oh->ot_handle != NULL);
3653
3654         osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3655
3656         CDEBUG(D_INODE, DFID" increase nlink %d\n",
3657                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3658         /*
3659          * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3660          * (65000) subdirectories by storing "1" in i_nlink if the link count
3661          * would otherwise overflow. Directory tranversal tools understand
3662          * that (st_nlink == 1) indicates that the filesystem dose not track
3663          * hard links count on the directory, and will not abort subdirectory
3664          * scanning early once (st_nlink - 2) subdirs have been found.
3665          *
3666          * This also has to properly handle the case of inodes with nlink == 0
3667          * in case they are being linked into the PENDING directory
3668