Whamcloud - gitweb
LU-8681 osd: ingore ENODATA during unlink agent parent
[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, 2015, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/osd/osd_handler.c
33  *
34  * Top-level entry points into osd module
35  *
36  * Author: Nikita Danilov <nikita@clusterfs.com>
37  *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
38  */
39
40 #define DEBUG_SUBSYSTEM S_OSD
41
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #ifdef HAVE_UIDGID_HEADER
45 # include <linux/uidgid.h>
46 #endif
47
48 /* LUSTRE_VERSION_CODE */
49 #include <lustre_ver.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/types.h>
52 /* prerequisite for linux/xattr.h */
53 #include <linux/fs.h>
54 /* XATTR_{REPLACE,CREATE} */
55 #include <linux/xattr.h>
56
57 #include <ldiskfs/ldiskfs.h>
58 #include <ldiskfs/xattr.h>
59 #include <ldiskfs/ldiskfs_extents.h>
60 #undef ENTRY
61 /*
62  * struct OBD_{ALLOC,FREE}*()
63  * OBD_FAIL_CHECK
64  */
65 #include <obd_support.h>
66 /* struct ptlrpc_thread */
67 #include <lustre_net.h>
68 #include <lustre_fid.h>
69 /* process_config */
70 #include <lustre_param.h>
71
72 #include "osd_internal.h"
73 #include "osd_dynlocks.h"
74
75 /* llo_* api support */
76 #include <md_object.h>
77 #include <lustre_quota.h>
78
79 #include <lustre_linkea.h>
80
81 int ldiskfs_pdo = 1;
82 module_param(ldiskfs_pdo, int, 0644);
83 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
84
85 int ldiskfs_track_declares_assert;
86 module_param(ldiskfs_track_declares_assert, int, 0644);
87 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
88
89 /* Slab to allocate dynlocks */
90 struct kmem_cache *dynlock_cachep;
91
92 /* Slab to allocate osd_it_ea */
93 struct kmem_cache *osd_itea_cachep;
94
95 static struct lu_kmem_descr ldiskfs_caches[] = {
96         {
97                 .ckd_cache = &dynlock_cachep,
98                 .ckd_name  = "dynlock_cache",
99                 .ckd_size  = sizeof(struct dynlock_handle)
100         },
101         {
102                 .ckd_cache = &osd_itea_cachep,
103                 .ckd_name  = "osd_itea_cache",
104                 .ckd_size  = sizeof(struct osd_it_ea)
105         },
106         {
107                 .ckd_cache = NULL
108         }
109 };
110
111 static const char dot[] = ".";
112 static const char dotdot[] = "..";
113 static const char remote_obj_dir[] = "REM_OBJ_DIR";
114
115 static const struct lu_object_operations      osd_lu_obj_ops;
116 static const struct dt_object_operations      osd_obj_ops;
117 static const struct dt_object_operations      osd_obj_ea_ops;
118 static const struct dt_object_operations      osd_obj_otable_it_ops;
119 static const struct dt_index_operations       osd_index_iam_ops;
120 static const struct dt_index_operations       osd_index_ea_ops;
121
122 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
123                           const struct lu_fid *fid);
124 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
125                                                 struct osd_device *osd);
126
127 int osd_trans_declare_op2rb[] = {
128         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
129         [OSD_OT_PUNCH]          = OSD_OT_MAX,
130         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
131         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
132         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
133         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
134         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
135         [OSD_OT_WRITE]          = OSD_OT_WRITE,
136         [OSD_OT_INSERT]         = OSD_OT_DELETE,
137         [OSD_OT_DELETE]         = OSD_OT_INSERT,
138         [OSD_OT_QUOTA]          = OSD_OT_MAX,
139 };
140
141 static int osd_has_index(const struct osd_object *obj)
142 {
143         return obj->oo_dt.do_index_ops != NULL;
144 }
145
146 static int osd_object_invariant(const struct lu_object *l)
147 {
148         return osd_invariant(osd_obj(l));
149 }
150
151 /*
152  * Concurrency: doesn't matter
153  */
154
155 /*
156  * Concurrency: doesn't matter
157  */
158 static int osd_write_locked(const struct lu_env *env, struct osd_object *o)
159 {
160         struct osd_thread_info *oti = osd_oti_get(env);
161         return oti->oti_w_locks > 0 && o->oo_owner == env;
162 }
163
164 /*
165  * Concurrency: doesn't access mutable data
166  */
167 static int osd_root_get(const struct lu_env *env,
168                         struct dt_device *dev, struct lu_fid *f)
169 {
170         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
171         return 0;
172 }
173
174 /*
175  * the following set of functions are used to maintain per-thread
176  * cache of FID->ino mapping. this mechanism is needed to resolve
177  * FID to inode at dt_insert() which in turn stores ino in the
178  * directory entries to keep ldiskfs compatible with ext[34].
179  * due to locking-originated restrictions we can't lookup ino
180  * using LU cache (deadlock is possible). lookup using OI is quite
181  * expensive. so instead we maintain this cache and methods like
182  * dt_create() fill it. so in the majority of cases dt_insert() is
183  * able to find needed mapping in lockless manner.
184  */
185 static struct osd_idmap_cache *
186 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
187              const struct lu_fid *fid)
188 {
189         struct osd_thread_info  *oti   = osd_oti_get(env);
190         struct osd_idmap_cache  *idc    = oti->oti_ins_cache;
191         int i;
192         for (i = 0; i < oti->oti_ins_cache_used; i++) {
193                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
194                         continue;
195                 if (idc[i].oic_dev != osd)
196                         continue;
197
198                 return idc + i;
199         }
200
201         return NULL;
202 }
203
204 static struct osd_idmap_cache *
205 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
206             const struct lu_fid *fid)
207 {
208         struct osd_thread_info  *oti   = osd_oti_get(env);
209         struct osd_idmap_cache  *idc;
210         int i;
211
212         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
213                 i = oti->oti_ins_cache_size * 2;
214                 if (i == 0)
215                         i = OSD_INS_CACHE_SIZE;
216                 OBD_ALLOC(idc, sizeof(*idc) * i);
217                 if (idc == NULL)
218                         return ERR_PTR(-ENOMEM);
219                 if (oti->oti_ins_cache != NULL) {
220                         memcpy(idc, oti->oti_ins_cache,
221                                oti->oti_ins_cache_used * sizeof(*idc));
222                         OBD_FREE(oti->oti_ins_cache,
223                                  oti->oti_ins_cache_used * sizeof(*idc));
224                 }
225                 oti->oti_ins_cache = idc;
226                 oti->oti_ins_cache_size = i;
227         }
228
229         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
230         idc->oic_fid = *fid;
231         idc->oic_dev = osd;
232         idc->oic_lid.oii_ino = 0;
233         idc->oic_lid.oii_gen = 0;
234         idc->oic_remote = 0;
235
236         return idc;
237 }
238
239 /*
240  * lookup mapping for the given fid in the cache, initialize a
241  * new one if not found. the initialization checks whether the
242  * object is local or remote. for local objects, OI is used to
243  * learn ino/generation. the function is used when the caller
244  * has no information about the object, e.g. at dt_insert().
245  */
246 static struct osd_idmap_cache *
247 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
248                      const struct lu_fid *fid)
249 {
250         struct osd_idmap_cache *idc;
251         int rc;
252
253         idc = osd_idc_find(env, osd, fid);
254         LASSERT(!IS_ERR(idc));
255         if (idc != NULL)
256                 return idc;
257
258         /* new mapping is needed */
259         idc = osd_idc_add(env, osd, fid);
260         if (IS_ERR(idc))
261                 return idc;
262
263         /* initialize it */
264         rc = osd_remote_fid(env, osd, fid);
265         if (unlikely(rc < 0))
266                 return ERR_PTR(rc);
267
268         if (rc == 0) {
269                 /* the object is local, lookup in OI */
270                 /* XXX: probably cheaper to lookup in LU first? */
271                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
272                                    &idc->oic_lid, 0);
273                 if (unlikely(rc < 0)) {
274                         CERROR("can't lookup: rc = %d\n", rc);
275                         return ERR_PTR(rc);
276                 }
277         } else {
278                 /* the object is remote */
279                 idc->oic_remote = 1;
280         }
281
282         return idc;
283 }
284
285 /*
286  * lookup mapping for given FID and fill it from the given object.
287  * the object is lolcal by definition.
288  */
289 static int osd_idc_find_and_init(const struct lu_env *env,
290                                  struct osd_device *osd,
291                                  struct osd_object *obj)
292 {
293         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
294         struct osd_idmap_cache  *idc;
295
296         idc = osd_idc_find(env, osd, fid);
297         LASSERT(!IS_ERR(idc));
298         if (idc != NULL) {
299                 if (obj->oo_inode == NULL)
300                         return 0;
301                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
302                         LASSERT(idc->oic_lid.oii_ino == 0);
303                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
304                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
305                 }
306                 return 0;
307         }
308
309         /* new mapping is needed */
310         idc = osd_idc_add(env, osd, fid);
311         if (IS_ERR(idc))
312                 return PTR_ERR(idc);
313
314         if (obj->oo_inode != NULL) {
315                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
316                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
317         }
318         return 0;
319 }
320
321 /*
322  * OSD object methods.
323  */
324
325 /*
326  * Concurrency: no concurrent access is possible that early in object
327  * life-cycle.
328  */
329 static struct lu_object *osd_object_alloc(const struct lu_env *env,
330                                           const struct lu_object_header *hdr,
331                                           struct lu_device *d)
332 {
333         struct osd_object *mo;
334
335         OBD_ALLOC_PTR(mo);
336         if (mo != NULL) {
337                 struct lu_object *l;
338
339                 l = &mo->oo_dt.do_lu;
340                 dt_object_init(&mo->oo_dt, NULL, d);
341                 mo->oo_dt.do_ops = &osd_obj_ea_ops;
342                 l->lo_ops = &osd_lu_obj_ops;
343                 init_rwsem(&mo->oo_sem);
344                 init_rwsem(&mo->oo_ext_idx_sem);
345                 spin_lock_init(&mo->oo_guard);
346                 INIT_LIST_HEAD(&mo->oo_xattr_list);
347                 return l;
348         } else {
349                 return NULL;
350         }
351 }
352
353 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
354                 struct dentry *dentry, struct lustre_mdt_attrs *lma)
355 {
356         int rc;
357
358         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
359         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
360                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
361         if (rc > 0) {
362                 if ((void *)lma != (void *)info->oti_mdt_attrs_old)
363                         memcpy(lma, info->oti_mdt_attrs_old, sizeof(*lma));
364                 rc = 0;
365                 lustre_lma_swab(lma);
366                 /* Check LMA compatibility */
367                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
368                         CWARN("%.16s: unsupported incompat LMA feature(s) %#x "
369                               "for fid = "DFID", ino = %lu\n",
370                               LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
371                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
372                               PFID(&lma->lma_self_fid), inode->i_ino);
373                         rc = -EOPNOTSUPP;
374                 }
375         } else if (rc == 0) {
376                 rc = -ENODATA;
377         }
378
379         return rc;
380 }
381
382 /*
383  * retrieve object from backend ext fs.
384  **/
385 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
386                        struct osd_inode_id *id)
387 {
388         struct inode *inode = NULL;
389
390         /* if we look for an inode withing a running
391          * transaction, then we risk to deadlock */
392         /* osd_dirent_check_repair() breaks this */
393         /*LASSERT(current->journal_info == NULL);*/
394
395         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
396         if (IS_ERR(inode)) {
397                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
398                        id->oii_ino, PTR_ERR(inode));
399         } else if (id->oii_gen != OSD_OII_NOGEN &&
400                    inode->i_generation != id->oii_gen) {
401                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
402                        "i_generation = %u\n",
403                        id->oii_ino, id->oii_gen, inode->i_generation);
404                 iput(inode);
405                 inode = ERR_PTR(-ESTALE);
406         } else if (inode->i_nlink == 0) {
407                 /* due to parallel readdir and unlink,
408                 * we can have dead inode here. */
409                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
410                 iput(inode);
411                 inode = ERR_PTR(-ESTALE);
412         } else if (is_bad_inode(inode)) {
413                 CWARN("%.16s: bad inode: ino = %u\n",
414                 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, id->oii_ino);
415                 iput(inode);
416                 inode = ERR_PTR(-ENOENT);
417         } else {
418                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
419                 if (id->oii_gen == OSD_OII_NOGEN)
420                         osd_id_gen(id, inode->i_ino, inode->i_generation);
421
422                 /* Do not update file c/mtime in ldiskfs.
423                  * NB: we don't have any lock to protect this because we don't
424                  * have reference on osd_object now, but contention with
425                  * another lookup + attr_set can't happen in the tiny window
426                  * between if (...) and set S_NOCMTIME. */
427                 if (!(inode->i_flags & S_NOCMTIME))
428                         inode->i_flags |= S_NOCMTIME;
429         }
430         return inode;
431 }
432
433 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
434                           handle_t *handle, struct dentry *child,
435                           struct inode *inode, struct htree_lock *hlock)
436 {
437         int rc, rc2;
438
439         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
440         if (rc == -ENOBUFS || rc == -ENOSPC) {
441                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
442                 struct inode *parent = child->d_parent->d_inode;
443                 struct lu_fid *fid = NULL;
444
445                 rc2 = osd_get_lma(info, parent, child->d_parent, lma);
446                 if (rc2 == 0) {
447                         fid = &lma->lma_self_fid;
448                 } else if (rc2 == -ENODATA) {
449                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
450                                 fid = &info->oti_fid3;
451                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
452                         } else if (!osd->od_is_ost && osd->od_index == 0) {
453                                 fid = &info->oti_fid3;
454                                 lu_igif_build(fid, parent->i_ino,
455                                               parent->i_generation);
456                         }
457                 }
458
459                 if (fid != NULL)
460                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
461                               "maximum entry limit\n",
462                               osd_name(osd), parent->i_ino, PFID(fid),
463                               rc == -ENOSPC ? "has reached" : "is approaching");
464                 else
465                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
466                               "maximum entry limit\n",
467                               osd_name(osd), parent->i_ino,
468                               rc == -ENOSPC ? "has reached" : "is approaching");
469
470                 /* ignore such error now */
471                 if (rc == -ENOBUFS)
472                         rc = 0;
473         }
474
475         return rc;
476 }
477
478
479 static struct inode *
480 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
481              struct osd_inode_id *id, struct lu_fid *fid)
482 {
483         struct lustre_mdt_attrs *lma   = &info->oti_mdt_attrs;
484         struct inode            *inode;
485         int                      rc;
486
487         inode = osd_iget(info, dev, id);
488         if (IS_ERR(inode))
489                 return inode;
490
491         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
492         if (rc == 0) {
493                 *fid = lma->lma_self_fid;
494         } else if (rc == -ENODATA) {
495                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
496                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
497                 else
498                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
499         } else {
500                 iput(inode);
501                 inode = ERR_PTR(rc);
502         }
503         return inode;
504 }
505
506 static struct inode *osd_iget_check(struct osd_thread_info *info,
507                                     struct osd_device *dev,
508                                     const struct lu_fid *fid,
509                                     struct osd_inode_id *id,
510                                     bool cached)
511 {
512         struct inode    *inode;
513         int              rc     = 0;
514         ENTRY;
515
516         /* The cached OI mapping is trustable. If we cannot locate the inode
517          * via the cached OI mapping, then return the failure to the caller
518          * directly without further OI checking. */
519
520         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
521         if (IS_ERR(inode)) {
522                 rc = PTR_ERR(inode);
523                 if (cached || (rc != -ENOENT && rc != -ESTALE)) {
524                         CDEBUG(D_INODE, "no inode: ino = %u, rc = %d\n",
525                                id->oii_ino, rc);
526
527                         GOTO(put, rc);
528                 }
529
530                 goto check_oi;
531         }
532
533         if (is_bad_inode(inode)) {
534                 rc = -ENOENT;
535                 if (cached) {
536                         CDEBUG(D_INODE, "bad inode: ino = %u\n", id->oii_ino);
537
538                         GOTO(put, rc);
539                 }
540
541                 goto check_oi;
542         }
543
544         if (id->oii_gen != OSD_OII_NOGEN &&
545             inode->i_generation != id->oii_gen) {
546                 rc = -ESTALE;
547                 if (cached) {
548                         CDEBUG(D_INODE, "unmatched inode: ino = %u, "
549                                "oii_gen = %u, i_generation = %u\n",
550                                id->oii_ino, id->oii_gen, inode->i_generation);
551
552                         GOTO(put, rc);
553                 }
554
555                 goto check_oi;
556         }
557
558         if (inode->i_nlink == 0) {
559                 rc = -ENOENT;
560                 if (cached) {
561                         CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
562
563                         GOTO(put, rc);
564                 }
565
566                 goto check_oi;
567         }
568
569         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
570
571 check_oi:
572         if (rc != 0) {
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                 if ((!IS_ERR(inode) && inode->i_generation != 0 &&
604                      inode->i_generation == id->oii_gen) ||
605                     (IS_ERR(inode) && !(dev->od_scrub.os_file.sf_flags &
606                                         SF_INCONSISTENT)))
607                         rc = -ENOENT;
608                 else
609                         rc = -EREMCHG;
610         } else {
611                 if (id->oii_gen == OSD_OII_NOGEN)
612                         osd_id_gen(id, inode->i_ino, inode->i_generation);
613
614                 /* Do not update file c/mtime in ldiskfs.
615                  * NB: we don't have any lock to protect this because we don't
616                  * have reference on osd_object now, but contention with
617                  * another lookup + attr_set can't happen in the tiny window
618                  * between if (...) and set S_NOCMTIME. */
619                 if (!(inode->i_flags & S_NOCMTIME))
620                         inode->i_flags |= S_NOCMTIME;
621         }
622
623         GOTO(put, rc);
624
625 put:
626         if (rc != 0) {
627                 if (!IS_ERR(inode))
628                         iput(inode);
629
630                 inode = ERR_PTR(rc);
631         }
632
633         return inode;
634 }
635
636 /**
637  * \retval +v: new filter_fid, does not contain self-fid
638  * \retval 0:  filter_fid_old, contains self-fid
639  * \retval -v: other failure cases
640  */
641 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
642                  struct dentry *dentry, struct lu_fid *fid)
643 {
644         struct filter_fid_old   *ff     = &info->oti_ff;
645         struct ost_id           *ostid  = &info->oti_ostid;
646         int                      rc;
647
648         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
649         if (rc == sizeof(*ff)) {
650                 rc = 0;
651                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
652                 ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
653                 /* XXX: use 0 as the index for compatibility, the caller will
654                  *      handle index related issues when necessarry. */
655                 ostid_to_fid(fid, ostid, 0);
656         } else if (rc == sizeof(struct filter_fid)) {
657                 rc = 1;
658         } else if (rc >= 0) {
659                 rc = -EINVAL;
660         }
661
662         return rc;
663 }
664
665 static int osd_lma_self_repair(struct osd_thread_info *info,
666                                struct osd_device *osd, struct inode *inode,
667                                const struct lu_fid *fid, __u32 compat)
668 {
669         handle_t *jh;
670         int       rc;
671
672         LASSERT(current->journal_info == NULL);
673
674         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
675                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
676         if (IS_ERR(jh)) {
677                 rc = PTR_ERR(jh);
678                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
679                       osd_name(osd), rc);
680                 return rc;
681         }
682
683         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
684         if (rc != 0)
685                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
686                       osd_name(osd), rc);
687         ldiskfs_journal_stop(jh);
688         return rc;
689 }
690
691 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
692 {
693         struct osd_thread_info  *info   = osd_oti_get(env);
694         struct osd_device       *osd    = osd_obj2dev(obj);
695         struct lustre_mdt_attrs *lma    = &info->oti_mdt_attrs;
696         struct inode            *inode  = obj->oo_inode;
697         struct dentry           *dentry = &info->oti_obj_dentry;
698         struct lu_fid           *fid    = NULL;
699         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
700         int                      rc;
701         ENTRY;
702
703         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
704         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
705                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
706         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
707                 fid = &lma->lma_self_fid;
708                 rc = osd_get_idif(info, inode, dentry, fid);
709                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
710                         /* For the given OST-object, if it has neither LMA nor
711                          * FID in XATTR_NAME_FID, then the given FID (which is
712                          * contained in the @obj, from client RPC for locating
713                          * the OST-object) is trusted. We use it to generate
714                          * the LMA. */
715                         osd_lma_self_repair(info, osd, inode, rfid,
716                                             LMAC_FID_ON_OST);
717                         RETURN(0);
718                 }
719         }
720
721         if (rc < 0)
722                 RETURN(rc);
723
724         if (rc > 0) {
725                 rc = 0;
726                 lustre_lma_swab(lma);
727                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
728                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
729                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
730                               "fid = "DFID", ino = %lu\n", osd_name(osd),
731                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
732                               PFID(rfid), inode->i_ino);
733                         rc = -EOPNOTSUPP;
734                 } else {
735                         fid = &lma->lma_self_fid;
736                 }
737         }
738
739         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
740                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
741                         struct ost_id   *oi   = &info->oti_ostid;
742                         struct lu_fid   *fid1 = &info->oti_fid3;
743                         __u32            idx  = fid_idif_ost_idx(rfid);
744
745                         /* For old IDIF, the OST index is not part of the IDIF,
746                          * Means that different OSTs may have the same IDIFs.
747                          * Under such case, we need to make some compatible
748                          * check to make sure to trigger OI scrub properly. */
749                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
750                                 /* Given @rfid is new, LMA is old. */
751                                 fid_to_ostid(fid, oi);
752                                 ostid_to_fid(fid1, oi, idx);
753                                 if (lu_fid_eq(fid1, rfid)) {
754                                         if (osd->od_index_in_idif)
755                                                 osd_lma_self_repair(info, osd,
756                                                         inode, rfid,
757                                                         LMAC_FID_ON_OST);
758                                         RETURN(0);
759                                 }
760                         }
761                 }
762
763                 rc = -EREMCHG;
764         }
765
766         RETURN(rc);
767 }
768
769 struct osd_check_lmv_buf {
770 #ifdef HAVE_DIR_CONTEXT
771         /* please keep it as first member */
772         struct dir_context       ctx;
773 #endif
774         struct osd_thread_info  *oclb_info;
775         struct osd_device       *oclb_dev;
776         struct osd_idmap_cache  *oclb_oic;
777 };
778
779 /**
780  * It is called internally by ->readdir() to filter out the
781  * local slave object's FID of the striped directory.
782  *
783  * \retval      1 found the local slave's FID
784  * \retval      0 continue to check next item
785  * \retval      -ve for failure
786  */
787 #ifdef HAVE_FILLDIR_USE_CTX
788 static int osd_stripe_dir_filldir(struct dir_context *buf,
789 #else
790 static int osd_stripe_dir_filldir(void *buf,
791 #endif
792                                   const char *name, int namelen,
793                                   loff_t offset, __u64 ino, unsigned d_type)
794 {
795         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
796         struct osd_thread_info *oti = oclb->oclb_info;
797         struct lu_fid *fid = &oti->oti_fid3;
798         struct osd_inode_id *id = &oti->oti_id3;
799         struct osd_device *dev = oclb->oclb_dev;
800         struct osd_idmap_cache *oic = oclb->oclb_oic;
801         struct inode *inode;
802         int rc;
803
804         if (name[0] == '.')
805                 return 0;
806
807         fid_zero(fid);
808         sscanf(name + 1, SFID, RFID(fid));
809         if (!fid_is_sane(fid))
810                 return 0;
811
812         if (osd_remote_fid(oti->oti_env, dev, fid))
813                 return 0;
814
815         osd_id_gen(id, ino, OSD_OII_NOGEN);
816         inode = osd_iget(oti, dev, id);
817         if (IS_ERR(inode))
818                 return PTR_ERR(inode);
819
820         iput(inode);
821         osd_add_oi_cache(oti, dev, id, fid);
822         oic->oic_fid = *fid;
823         oic->oic_lid = *id;
824         oic->oic_dev = dev;
825         rc = osd_oii_insert(dev, oic, true);
826
827         return rc == 0 ? 1 : rc;
828 }
829
830 /* When lookup item under striped directory, we need to locate the master
831  * MDT-object of the striped directory firstly, then the client will send
832  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
833  * and the item's name. If the system is restored from MDT file level backup,
834  * then before the OI scrub completely built the OI files, the OI mappings of
835  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
836  * not a problem for the master MDT-object. Because when locate the master
837  * MDT-object, we will do name based lookup (for the striped directory itself)
838  * firstly, during such process we can setup the correct OI mapping for the
839  * master MDT-object. But it will be trouble for the slave MDT-object. Because
840  * the client will not trigger name based lookup on the MDT to locate the slave
841  * MDT-object before locating item under the striped directory, then when
842  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
843  * is invalid and does not know what the right OI mapping is, then the MDT has
844  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
845  * the OI file, related OI mapping is unknown yet, please try again later. And
846  * then client will re-try the RPC again and again until related OI mapping has
847  * been updated. That is quite inefficient.
848  *
849  * To resolve above trouble, we will handle it as the following two cases:
850  *
851  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
852  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
853  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
854  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
855  *    directly. Please check osd_fid_lookup().
856  *
857  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
858  *    Under such case, during lookup the master MDT-object, we will lookup the
859  *    slave MDT-object via readdir against the master MDT-object, because the
860  *    slave MDT-objects information are stored as sub-directories with the name
861  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
862  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
863  *    the correct OI mapping for the slave MDT-object. */
864 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
865                          struct inode *inode, struct osd_idmap_cache *oic)
866 {
867         struct lu_buf *buf = &oti->oti_big_buf;
868         struct dentry *dentry = &oti->oti_obj_dentry;
869         struct file *filp = &oti->oti_file;
870         const struct file_operations *fops;
871         struct lmv_mds_md_v1 *lmv1;
872         struct osd_check_lmv_buf oclb = {
873 #ifdef HAVE_DIR_CONTEXT
874                 .ctx.actor = osd_stripe_dir_filldir,
875 #endif
876                 .oclb_info = oti,
877                 .oclb_dev = dev,
878                 .oclb_oic = oic
879         };
880         int rc = 0;
881         ENTRY;
882
883 again:
884         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
885                              buf->lb_len);
886         if (rc == -ERANGE) {
887                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
888                 if (rc > 0) {
889                         lu_buf_realloc(buf, rc);
890                         if (buf->lb_buf == NULL)
891                                 GOTO(out, rc = -ENOMEM);
892
893                         goto again;
894                 }
895         }
896
897         if (unlikely(rc == 0 || rc == -ENODATA))
898                 GOTO(out, rc = 0);
899
900         if (rc < 0)
901                 GOTO(out, rc);
902
903         if (unlikely(buf->lb_buf == NULL)) {
904                 lu_buf_realloc(buf, rc);
905                 if (buf->lb_buf == NULL)
906                         GOTO(out, rc = -ENOMEM);
907
908                 goto again;
909         }
910
911         lmv1 = buf->lb_buf;
912         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
913                 GOTO(out, rc = 0);
914
915         fops = inode->i_fop;
916         dentry->d_inode = inode;
917         dentry->d_sb = inode->i_sb;
918         filp->f_pos = 0;
919         filp->f_path.dentry = dentry;
920         filp->f_mode = FMODE_64BITHASH;
921         filp->f_mapping = inode->i_mapping;
922         filp->f_op = fops;
923         filp->private_data = NULL;
924         set_file_inode(filp, inode);
925
926 #ifdef HAVE_DIR_CONTEXT
927         oclb.ctx.pos = filp->f_pos;
928         rc = fops->iterate(filp, &oclb.ctx);
929         filp->f_pos = oclb.ctx.pos;
930 #else
931         rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
932 #endif
933         fops->release(inode, filp);
934
935 out:
936         if (rc < 0)
937                 CDEBUG(D_LFSCK, "%.16s: fail to check LMV EA, inode = %lu/%u,"
938                        DFID": rc = %d\n",
939                        LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
940                        inode->i_ino, inode->i_generation,
941                        PFID(&oic->oic_fid), rc);
942         else
943                 rc = 0;
944
945         RETURN(rc);
946 }
947
948 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
949                           const struct lu_fid *fid,
950                           const struct lu_object_conf *conf)
951 {
952         struct osd_thread_info *info;
953         struct lu_device       *ldev   = obj->oo_dt.do_lu.lo_dev;
954         struct osd_device      *dev;
955         struct osd_idmap_cache *oic;
956         struct osd_inode_id    *id;
957         struct osd_inode_id    *tid;
958         struct inode           *inode = NULL;
959         struct osd_scrub       *scrub;
960         struct scrub_file      *sf;
961         __u32                   flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT |
962                                         SS_AUTO_FULL;
963         __u32                   saved_ino;
964         __u32                   saved_gen;
965         int                     result  = 0;
966         int                     rc1     = 0;
967         bool                    cached  = true;
968         bool                    remote  = false;
969         ENTRY;
970
971         LINVRNT(osd_invariant(obj));
972         LASSERT(obj->oo_inode == NULL);
973         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
974
975         dev = osd_dev(ldev);
976         scrub = &dev->od_scrub;
977         sf = &scrub->os_file;
978         info = osd_oti_get(env);
979         LASSERT(info);
980         oic = &info->oti_cache;
981
982         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
983                 RETURN(-ENOENT);
984
985         /* For the object is created as locking anchor, or for the object to
986          * be created on disk. No need to osd_oi_lookup() at here because FID
987          * shouldn't never be re-used, if it's really a duplicate FID from
988          * unexpected reason, we should be able to detect it later by calling
989          * do_create->osd_oi_insert(). */
990         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
991                 GOTO(out, result = 0);
992
993         /* Search order: 1. per-thread cache. */
994         if (lu_fid_eq(fid, &oic->oic_fid) &&
995             likely(oic->oic_dev == dev)) {
996                 id = &oic->oic_lid;
997                 goto iget;
998         }
999
1000         id = &info->oti_id;
1001         if (!list_empty(&scrub->os_inconsistent_items)) {
1002                 /* Search order: 2. OI scrub pending list. */
1003                 result = osd_oii_lookup(dev, fid, id);
1004                 if (result == 0)
1005                         goto iget;
1006         }
1007
1008         cached = false;
1009         /* Search order: 3. OI files. */
1010         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1011         if (result == -ENOENT) {
1012                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1013                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1014                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1015                                       sf->sf_oi_bitmap))
1016                         GOTO(out, result = 0);
1017
1018                 goto trigger;
1019         }
1020
1021         if (result != 0)
1022                 GOTO(out, result);
1023
1024 iget:
1025         inode = osd_iget_check(info, dev, fid, id, cached);
1026         if (IS_ERR(inode)) {
1027                 result = PTR_ERR(inode);
1028                 if (result == -ENOENT || result == -ESTALE)
1029                         GOTO(out, result = 0);
1030
1031                 if (result == -EREMCHG) {
1032
1033 trigger:
1034                         /* We still have chance to get the valid inode: for the
1035                          * object which is referenced by remote name entry, the
1036                          * object on the local MDT will be linked under the dir
1037                          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1038                          *
1039                          * We do not know whether the object for the given FID
1040                          * is referenced by some remote name entry or not, and
1041                          * especially for DNE II, a multiple-linked object may
1042                          * have many name entries reside on many MDTs.
1043                          *
1044                          * To simplify the operation, OSD will not distinguish
1045                          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1046                          * only happened for the RPC from other MDT during the
1047                          * OI scrub, or for the client side RPC with FID only,
1048                          * such as FID to path, or from old connected client. */
1049                         if (!remote &&
1050                             !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1051                                 rc1 = osd_lookup_in_remote_parent(info, dev,
1052                                                                   fid, id);
1053                                 if (rc1 == 0) {
1054                                         remote = true;
1055                                         cached = true;
1056                                         flags |= SS_AUTO_PARTIAL;
1057                                         flags &= ~SS_AUTO_FULL;
1058                                         goto iget;
1059                                 }
1060                         }
1061
1062                         if (thread_is_running(&scrub->os_thread)) {
1063                                 if (scrub->os_partial_scan &&
1064                                     !scrub->os_in_join) {
1065                                         goto join;
1066                                 } else {
1067                                         if (inode != NULL && !IS_ERR(inode)) {
1068                                                 LASSERT(remote);
1069
1070                                                 osd_add_oi_cache(info, dev, id,
1071                                                                  fid);
1072                                                 osd_oii_insert(dev, oic, true);
1073                                         } else {
1074                                                 result = -EINPROGRESS;
1075                                         }
1076                                 }
1077                         } else if (!dev->od_noscrub) {
1078
1079 join:
1080                                 rc1 = osd_scrub_start(dev, flags);
1081                                 LCONSOLE_WARN("%.16s: trigger OI scrub by RPC "
1082                                               "for the "DFID" with flags 0x%x,"
1083                                               " rc = %d\n", osd_name(dev),
1084                                               PFID(fid), flags, rc1);
1085                                 if (rc1 == 0 || rc1 == -EALREADY) {
1086                                         if (inode != NULL && !IS_ERR(inode)) {
1087                                                 LASSERT(remote);
1088
1089                                                 osd_add_oi_cache(info, dev, id,
1090                                                                  fid);
1091                                                 osd_oii_insert(dev, oic, true);
1092                                         } else {
1093                                                 result = -EINPROGRESS;
1094                                         }
1095                                 } else {
1096                                         result = -EREMCHG;
1097                                 }
1098                         } else {
1099                                 result = -EREMCHG;
1100                         }
1101                 }
1102
1103                 if (inode == NULL || IS_ERR(inode))
1104                         GOTO(out, result);
1105         } else if (remote) {
1106                 goto trigger;
1107         }
1108
1109         obj->oo_inode = inode;
1110         LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
1111
1112         result = osd_check_lma(env, obj);
1113         if (result == 0)
1114                 goto found;
1115
1116         tid = &info->oti_id3;
1117         LASSERT(tid != id);
1118
1119         if (result == -ENODATA) {
1120                 if (!cached)
1121                         /* The current OI mapping is from the OI file,
1122                          * since the inode has been found via
1123                          * osd_iget_check(), no need recheck OI. */
1124                         goto found;
1125
1126                 result = osd_oi_lookup(info, dev, fid, tid, OI_CHECK_FLD);
1127                 if (result == 0) {
1128                         LASSERTF(tid->oii_ino == id->oii_ino &&
1129                                  tid->oii_gen == id->oii_gen,
1130                                  "OI mapping changed(1): %u/%u => %u/%u",
1131                                  tid->oii_ino, tid->oii_gen,
1132                                  id->oii_ino, id->oii_gen);
1133
1134                         LASSERTF(tid->oii_ino == inode->i_ino &&
1135                                  tid->oii_gen == inode->i_generation,
1136                                  "locate wrong inode(1): %u/%u => %ld/%u",
1137                                  tid->oii_ino, tid->oii_gen,
1138                                  inode->i_ino, inode->i_generation);
1139
1140                         /* "result == 0" means the cached OI mapping is still in
1141                          * the OI file, so the target the inode is valid. */
1142                         goto found;
1143                 }
1144
1145                 /* "result == -ENOENT" means that the OI mappinghas been removed
1146                  * by race, the target inode belongs to other object.
1147                  *
1148                  * Others error can be returned  directly. */
1149                 if (result == -ENOENT)
1150                         result = 0;
1151         }
1152
1153         saved_ino = inode->i_ino;
1154         saved_gen = inode->i_generation;
1155         iput(inode);
1156         inode = NULL;
1157         obj->oo_inode = NULL;
1158
1159         if (result != -EREMCHG)
1160                 GOTO(out, result);
1161
1162         if (!cached)
1163                 /* The current OI mapping is from the OI file,
1164                  * since the inode has been found via
1165                  * osd_iget_check(), no need recheck OI. */
1166                 goto trigger;
1167
1168         result = osd_oi_lookup(info, dev, fid, tid, OI_CHECK_FLD);
1169         /* "result == -ENOENT" means the cached OI mapping has been removed from
1170          * the OI file by race, above target inode belongs to other object.
1171          *
1172          * Others error can be returned directly. */
1173         if (result != 0)
1174                 GOTO(out, result = (result == -ENOENT ? 0 : result));
1175
1176         LASSERTF(tid->oii_ino == id->oii_ino && tid->oii_gen == id->oii_gen,
1177                  "OI mapping changed(2): %u/%u => %u/%u",
1178                  tid->oii_ino, tid->oii_gen, id->oii_ino, id->oii_gen);
1179
1180         LASSERTF(tid->oii_ino == saved_ino && tid->oii_gen == saved_gen,
1181                  "locate wrong inode(2): %u/%u => %u/%u",
1182                  tid->oii_ino, tid->oii_gen, saved_ino, saved_gen);
1183
1184         goto trigger;
1185
1186 found:
1187         obj->oo_compat_dot_created = 1;
1188         obj->oo_compat_dotdot_created = 1;
1189
1190         if (S_ISDIR(inode->i_mode) &&
1191             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1192                 osd_check_lmv(info, dev, inode, oic);
1193
1194         if (!ldiskfs_pdo)
1195                 GOTO(out, result = 0);
1196
1197         LASSERT(obj->oo_hl_head == NULL);
1198         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1199         if (obj->oo_hl_head == NULL) {
1200                 obj->oo_inode = NULL;
1201                 iput(inode);
1202                 GOTO(out, result = -ENOMEM);
1203         }
1204         GOTO(out, result = 0);
1205
1206 out:
1207         if (result != 0 && cached)
1208                 fid_zero(&oic->oic_fid);
1209
1210         LINVRNT(osd_invariant(obj));
1211         return result;
1212 }
1213
1214 /*
1215  * Concurrency: shouldn't matter.
1216  */
1217 static void osd_object_init0(struct osd_object *obj)
1218 {
1219         LASSERT(obj->oo_inode != NULL);
1220         obj->oo_dt.do_body_ops = &osd_body_ops;
1221         obj->oo_dt.do_lu.lo_header->loh_attr |=
1222                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1223 }
1224
1225 /*
1226  * Concurrency: no concurrent access is possible that early in object
1227  * life-cycle.
1228  */
1229 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1230                            const struct lu_object_conf *conf)
1231 {
1232         struct osd_object *obj = osd_obj(l);
1233         int result;
1234
1235         LINVRNT(osd_invariant(obj));
1236
1237         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1238                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1239                 l->lo_header->loh_attr |= LOHA_EXISTS;
1240                 return 0;
1241         }
1242
1243         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1244         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1245         if (result == 0 && obj->oo_inode != NULL) {
1246                 struct osd_thread_info *oti = osd_oti_get(env);
1247                 struct lustre_mdt_attrs *lma = &oti->oti_mdt_attrs;
1248
1249                 osd_object_init0(obj);
1250                 result = osd_get_lma(oti, obj->oo_inode,
1251                                      &oti->oti_obj_dentry, lma);
1252                 if (result == 0) {
1253                         /* Convert LMAI flags to lustre LMA flags
1254                          * and cache it to oo_lma_flags */
1255                         obj->oo_lma_flags =
1256                                 lma_to_lustre_flags(lma->lma_incompat);
1257                 } else if (result == -ENODATA) {
1258                         result = 0;
1259                 }
1260         }
1261
1262         LINVRNT(osd_invariant(obj));
1263         return result;
1264 }
1265
1266 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1267  * The left part is for value, binary mode. */
1268 struct osd_xattr_entry {
1269         struct list_head        oxe_list;
1270         size_t                  oxe_len;
1271         size_t                  oxe_namelen;
1272         bool                    oxe_exist;
1273         struct rcu_head         oxe_rcu;
1274         char                    oxe_buf[0];
1275 };
1276
1277 static int osd_oxc_get(struct osd_object *obj, const char *name,
1278                        struct lu_buf *buf)
1279 {
1280         struct osd_xattr_entry *tmp;
1281         struct osd_xattr_entry *oxe = NULL;
1282         size_t namelen = strlen(name);
1283         int rc;
1284         ENTRY;
1285
1286         rcu_read_lock();
1287         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1288                 if (namelen == tmp->oxe_namelen &&
1289                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1290                         oxe = tmp;
1291                         break;
1292                 }
1293         }
1294
1295         if (oxe == NULL)
1296                 GOTO(out, rc = -ENOENT);
1297
1298         if (!oxe->oxe_exist)
1299                 GOTO(out, rc = -ENODATA);
1300
1301         /* vallen */
1302         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1303         LASSERT(rc > 0);
1304
1305         if (buf->lb_buf == NULL)
1306                 GOTO(out, rc);
1307
1308         if (buf->lb_len < rc)
1309                 GOTO(out, rc = -ERANGE);
1310
1311         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1312         EXIT;
1313 out:
1314         rcu_read_unlock();
1315
1316         return rc;
1317 }
1318
1319 static void osd_oxc_free(struct rcu_head *head)
1320 {
1321         struct osd_xattr_entry *oxe;
1322
1323         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1324         OBD_FREE(oxe, oxe->oxe_len);
1325 }
1326
1327 static void osd_oxc_add(struct osd_object *obj, const char *name,
1328                         const char *buf, int buflen)
1329 {
1330         struct osd_xattr_entry *oxe;
1331         struct osd_xattr_entry *old = NULL;
1332         struct osd_xattr_entry *tmp;
1333         size_t namelen = strlen(name);
1334         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1335
1336         OBD_ALLOC(oxe, len);
1337         if (oxe == NULL)
1338                 return;
1339
1340         INIT_LIST_HEAD(&oxe->oxe_list);
1341         oxe->oxe_len = len;
1342         oxe->oxe_namelen = namelen;
1343         memcpy(oxe->oxe_buf, name, namelen);
1344         if (buflen > 0) {
1345                 LASSERT(buf != NULL);
1346                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1347                 oxe->oxe_exist = true;
1348         } else {
1349                 oxe->oxe_exist = false;
1350         }
1351
1352         /* this should be rarely called, just remove old and add new */
1353         spin_lock(&obj->oo_guard);
1354         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1355                 if (namelen == tmp->oxe_namelen &&
1356                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1357                         old = tmp;
1358                         break;
1359                 }
1360         }
1361         if (old != NULL) {
1362                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1363                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1364         } else {
1365                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1366         }
1367         spin_unlock(&obj->oo_guard);
1368 }
1369
1370 static void osd_oxc_del(struct osd_object *obj, const char *name)
1371 {
1372         struct osd_xattr_entry *oxe;
1373         size_t namelen = strlen(name);
1374
1375         spin_lock(&obj->oo_guard);
1376         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1377                 if (namelen == oxe->oxe_namelen &&
1378                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1379                         list_del_rcu(&oxe->oxe_list);
1380                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1381                         break;
1382                 }
1383         }
1384         spin_unlock(&obj->oo_guard);
1385 }
1386
1387 static void osd_oxc_fini(struct osd_object *obj)
1388 {
1389         struct osd_xattr_entry *oxe, *next;
1390
1391         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1392                 list_del(&oxe->oxe_list);
1393                 OBD_FREE(oxe, oxe->oxe_len);
1394         }
1395 }
1396
1397 /*
1398  * Concurrency: no concurrent access is possible that late in object
1399  * life-cycle.
1400  */
1401 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1402 {
1403         struct osd_object *obj = osd_obj(l);
1404
1405         LINVRNT(osd_invariant(obj));
1406
1407         osd_oxc_fini(obj);
1408         dt_object_fini(&obj->oo_dt);
1409         if (obj->oo_hl_head != NULL)
1410                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1411         OBD_FREE_PTR(obj);
1412 }
1413
1414 /*
1415  * Concurrency: no concurrent access is possible that late in object
1416  * life-cycle.
1417  */
1418 static void osd_index_fini(struct osd_object *o)
1419 {
1420         struct iam_container *bag;
1421
1422         if (o->oo_dir != NULL) {
1423                 bag = &o->oo_dir->od_container;
1424                 if (o->oo_inode != NULL) {
1425                         if (bag->ic_object == o->oo_inode)
1426                                 iam_container_fini(bag);
1427                 }
1428                 OBD_FREE_PTR(o->oo_dir);
1429                 o->oo_dir = NULL;
1430         }
1431 }
1432
1433 /*
1434  * Concurrency: no concurrent access is possible that late in object
1435  * life-cycle (for all existing callers, that is. New callers have to provide
1436  * their own locking.)
1437  */
1438 static int osd_inode_unlinked(const struct inode *inode)
1439 {
1440         return inode->i_nlink == 0;
1441 }
1442
1443 enum {
1444         OSD_TXN_OI_DELETE_CREDITS    = 20,
1445         OSD_TXN_INODE_DELETE_CREDITS = 20
1446 };
1447
1448 /*
1449  * Journal
1450  */
1451
1452 #if OSD_THANDLE_STATS
1453 /**
1454  * Set time when the handle is allocated
1455  */
1456 static void osd_th_alloced(struct osd_thandle *oth)
1457 {
1458         oth->oth_alloced = cfs_time_current();
1459 }
1460
1461 /**
1462  * Set time when the handle started
1463  */
1464 static void osd_th_started(struct osd_thandle *oth)
1465 {
1466         oth->oth_started = cfs_time_current();
1467 }
1468
1469 /**
1470  * Helper function to convert time interval to microseconds packed in
1471  * long int.
1472  */
1473 static long interval_to_usec(cfs_time_t start, cfs_time_t end)
1474 {
1475         struct timeval val;
1476
1477         cfs_duration_usec(cfs_time_sub(end, start), &val);
1478         return val.tv_sec * 1000000 + val.tv_usec;
1479 }
1480
1481 /**
1482  * Check whether the we deal with this handle for too long.
1483  */
1484 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1485                                 cfs_time_t alloced, cfs_time_t started,
1486                                 cfs_time_t closed)
1487 {
1488         cfs_time_t now = cfs_time_current();
1489
1490         LASSERT(dev != NULL);
1491
1492         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1493                             interval_to_usec(alloced, started));
1494         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1495                             interval_to_usec(started, closed));
1496         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1497                             interval_to_usec(closed, now));
1498
1499         if (cfs_time_before(cfs_time_add(alloced, cfs_time_seconds(30)), now)) {
1500                 CWARN("transaction handle %p was open for too long: "
1501                       "now "CFS_TIME_T" ,"
1502                       "alloced "CFS_TIME_T" ,"
1503                       "started "CFS_TIME_T" ,"
1504                       "closed "CFS_TIME_T"\n",
1505                       oth, now, alloced, started, closed);
1506                 libcfs_debug_dumpstack(NULL);
1507         }
1508 }
1509
1510 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1511 {                                                                       \
1512         cfs_time_t __closed = cfs_time_current();                       \
1513         cfs_time_t __alloced = oth->oth_alloced;                        \
1514         cfs_time_t __started = oth->oth_started;                        \
1515                                                                         \
1516         expr;                                                           \
1517         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1518 }
1519
1520 #else /* OSD_THANDLE_STATS */
1521
1522 #define osd_th_alloced(h)                  do {} while(0)
1523 #define osd_th_started(h)                  do {} while(0)
1524 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1525
1526 #endif /* OSD_THANDLE_STATS */
1527
1528 /*
1529  * Concurrency: doesn't access mutable data.
1530  */
1531 static int osd_param_is_not_sane(const struct osd_device *dev,
1532                                  const struct thandle *th)
1533 {
1534         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1535
1536         return oh->ot_credits > osd_transaction_size(dev);
1537 }
1538
1539 /*
1540  * Concurrency: shouldn't matter.
1541  */
1542 static void osd_trans_commit_cb(struct super_block *sb,
1543                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1544 {
1545         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1546         struct thandle     *th  = &oh->ot_super;
1547         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
1548         struct dt_txn_commit_cb *dcb, *tmp;
1549
1550         LASSERT(oh->ot_handle == NULL);
1551
1552         if (error)
1553                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1554
1555         dt_txn_hook_commit(th);
1556
1557         /* call per-transaction callbacks if any */
1558         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1559                                  dcb_linkage) {
1560                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1561                          "commit callback entry: magic=%x name='%s'\n",
1562                          dcb->dcb_magic, dcb->dcb_name);
1563                 list_del_init(&dcb->dcb_linkage);
1564                 dcb->dcb_func(NULL, th, dcb, error);
1565         }
1566
1567         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1568         lu_device_put(lud);
1569         th->th_dev = NULL;
1570
1571         lu_context_exit(&th->th_ctx);
1572         lu_context_fini(&th->th_ctx);
1573         OBD_FREE_PTR(oh);
1574 }
1575
1576 #ifndef HAVE_SB_START_WRITE
1577 # define sb_start_write(sb) do {} while (0)
1578 # define sb_end_write(sb) do {} while (0)
1579 #endif
1580
1581 static struct thandle *osd_trans_create(const struct lu_env *env,
1582                                         struct dt_device *d)
1583 {
1584         struct osd_thread_info  *oti = osd_oti_get(env);
1585         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1586         struct osd_thandle      *oh;
1587         struct thandle          *th;
1588         ENTRY;
1589
1590         /* on pending IO in this thread should left from prev. request */
1591         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1592
1593         sb_start_write(osd_sb(osd_dt_dev(d)));
1594
1595         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1596         if (oh != NULL) {
1597                 oh->ot_quota_trans = &oti->oti_quota_trans;
1598                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1599                 th = &oh->ot_super;
1600                 th->th_dev = d;
1601                 th->th_result = 0;
1602                 th->th_tags = LCT_TX_HANDLE;
1603                 oh->ot_credits = 0;
1604                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1605                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1606                 osd_th_alloced(oh);
1607
1608                 memset(oti->oti_declare_ops, 0,
1609                        sizeof(oti->oti_declare_ops));
1610                 memset(oti->oti_declare_ops_cred, 0,
1611                        sizeof(oti->oti_declare_ops_cred));
1612                 memset(oti->oti_declare_ops_used, 0,
1613                        sizeof(oti->oti_declare_ops_used));
1614         } else {
1615                 sb_end_write(osd_sb(osd_dt_dev(d)));
1616                 th = ERR_PTR(-ENOMEM);
1617         }
1618         RETURN(th);
1619 }
1620
1621 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1622 {
1623         struct osd_thread_info  *oti = osd_oti_get(env);
1624         struct osd_thandle      *oh;
1625
1626         oh = container_of0(th, struct osd_thandle, ot_super);
1627         LASSERT(oh != NULL);
1628
1629         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1630               oti->oti_declare_ops[OSD_OT_CREATE],
1631               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1632               oti->oti_declare_ops_used[OSD_OT_CREATE],
1633               oti->oti_declare_ops[OSD_OT_DESTROY],
1634               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1635               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1636         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1637               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1638               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1639               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1640               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1641               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1642               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1643         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1644               oti->oti_declare_ops[OSD_OT_WRITE],
1645               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1646               oti->oti_declare_ops_used[OSD_OT_WRITE],
1647               oti->oti_declare_ops[OSD_OT_PUNCH],
1648               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1649               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1650               oti->oti_declare_ops[OSD_OT_QUOTA],
1651               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1652               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1653         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1654               oti->oti_declare_ops[OSD_OT_INSERT],
1655               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1656               oti->oti_declare_ops_used[OSD_OT_INSERT],
1657               oti->oti_declare_ops[OSD_OT_DELETE],
1658               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1659               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1660         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1661               oti->oti_declare_ops[OSD_OT_REF_ADD],
1662               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1663               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1664               oti->oti_declare_ops[OSD_OT_REF_DEL],
1665               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1666               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1667 }
1668
1669 /*
1670  * Concurrency: shouldn't matter.
1671  */
1672 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1673                            struct thandle *th)
1674 {
1675         struct osd_thread_info *oti = osd_oti_get(env);
1676         struct osd_device  *dev = osd_dt_dev(d);
1677         handle_t           *jh;
1678         struct osd_thandle *oh;
1679         int rc;
1680
1681         ENTRY;
1682
1683         LASSERT(current->journal_info == NULL);
1684
1685         oh = container_of0(th, struct osd_thandle, ot_super);
1686         LASSERT(oh != NULL);
1687         LASSERT(oh->ot_handle == NULL);
1688
1689         rc = dt_txn_hook_start(env, d, th);
1690         if (rc != 0)
1691                 GOTO(out, rc);
1692
1693         if (unlikely(osd_param_is_not_sane(dev, th))) {
1694                 static unsigned long last_printed;
1695                 static int last_credits;
1696
1697                 /* don't make noise on a tiny testing systems
1698                  * actual credits misuse will be caught anyway */
1699                 if (last_credits != oh->ot_credits &&
1700                     time_after(jiffies, last_printed +
1701                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1702                     osd_transaction_size(dev) > 512) {
1703                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1704                               oh->ot_credits, osd_transaction_size(dev));
1705                         osd_trans_dump_creds(env, th);
1706                         libcfs_debug_dumpstack(NULL);
1707                         last_credits = oh->ot_credits;
1708                         last_printed = jiffies;
1709                 }
1710                 /* XXX Limit the credits to 'max_transaction_buffers', and
1711                  *     let the underlying filesystem to catch the error if
1712                  *     we really need so many credits.
1713                  *
1714                  *     This should be removed when we can calculate the
1715                  *     credits precisely. */
1716                 oh->ot_credits = osd_transaction_size(dev);
1717         }
1718
1719         /*
1720          * XXX temporary stuff. Some abstraction layer should
1721          * be used.
1722          */
1723         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1724         osd_th_started(oh);
1725         if (!IS_ERR(jh)) {
1726                 oh->ot_handle = jh;
1727                 LASSERT(oti->oti_txns == 0);
1728                 lu_context_init(&th->th_ctx, th->th_tags);
1729                 lu_context_enter(&th->th_ctx);
1730
1731                 lu_device_get(&d->dd_lu_dev);
1732                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1733                               "osd-tx", th);
1734                 oti->oti_txns++;
1735                 rc = 0;
1736         } else {
1737                 rc = PTR_ERR(jh);
1738         }
1739 out:
1740         RETURN(rc);
1741 }
1742
1743 static int osd_seq_exists(const struct lu_env *env,
1744                           struct osd_device *osd, u64 seq)
1745 {
1746         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1747         struct seq_server_site  *ss = osd_seq_site(osd);
1748         int                     rc;
1749         ENTRY;
1750
1751         LASSERT(ss != NULL);
1752         LASSERT(ss->ss_server_fld != NULL);
1753
1754         rc = osd_fld_lookup(env, osd, seq, range);
1755         if (rc != 0) {
1756                 if (rc != -ENOENT)
1757                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1758                                osd_name(osd), seq, rc);
1759                 RETURN(0);
1760         }
1761
1762         RETURN(ss->ss_node_id == range->lsr_index);
1763 }
1764
1765 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1766 {
1767         struct dt_txn_commit_cb *dcb;
1768         struct dt_txn_commit_cb *tmp;
1769
1770         /* call per-transaction stop callbacks if any */
1771         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1772                                  dcb_linkage) {
1773                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1774                          "commit callback entry: magic=%x name='%s'\n",
1775                          dcb->dcb_magic, dcb->dcb_name);
1776                 list_del_init(&dcb->dcb_linkage);
1777                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1778         }
1779 }
1780
1781 /*
1782  * Concurrency: shouldn't matter.
1783  */
1784 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1785                           struct thandle *th)
1786 {
1787         int                     rc = 0, remove_agents = 0;
1788         struct osd_thandle     *oh;
1789         struct osd_thread_info *oti = osd_oti_get(env);
1790         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1791         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1792         struct qsd_instance    *qsd = osd->od_quota_slave;
1793         struct lquota_trans    *qtrans;
1794         ENTRY;
1795
1796         oh = container_of0(th, struct osd_thandle, ot_super);
1797
1798         /* reset OI cache for safety */
1799         oti->oti_ins_cache_used = 0;
1800
1801         remove_agents = oh->ot_remove_agents;
1802
1803         qtrans = oh->ot_quota_trans;
1804         oh->ot_quota_trans = NULL;
1805
1806         if (oh->ot_handle != NULL) {
1807                 handle_t *hdl = oh->ot_handle;
1808
1809                 /*
1810                  * add commit callback
1811                  * notice we don't do this in osd_trans_start()
1812                  * as underlying transaction can change during truncate
1813                  */
1814                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1815                                          &oh->ot_jcb);
1816
1817                 LASSERT(oti->oti_txns == 1);
1818                 oti->oti_txns--;
1819
1820                 rc = dt_txn_hook_stop(env, th);
1821                 if (rc != 0)
1822                         CERROR("%s: failed in transaction hook: rc = %d\n",
1823                                osd_name(osd), rc);
1824
1825                 osd_trans_stop_cb(oh, rc);
1826                 /* hook functions might modify th_sync */
1827                 hdl->h_sync = th->th_sync;
1828
1829                 oh->ot_handle = NULL;
1830                 OSD_CHECK_SLOW_TH(oh, osd, rc = ldiskfs_journal_stop(hdl));
1831                 if (rc != 0)
1832                         CERROR("%s: failed to stop transaction: rc = %d\n",
1833                                osd_name(osd), rc);
1834         } else {
1835                 osd_trans_stop_cb(oh, th->th_result);
1836                 OBD_FREE_PTR(oh);
1837         }
1838
1839         /* inform the quota slave device that the transaction is stopping */
1840         qsd_op_end(env, qsd, qtrans);
1841
1842         /* as we want IO to journal and data IO be concurrent, we don't block
1843          * awaiting data IO completion in osd_do_bio(), instead we wait here
1844          * once transaction is submitted to the journal. all reqular requests
1845          * don't do direct IO (except read/write), thus this wait_event becomes
1846          * no-op for them.
1847          *
1848          * IMPORTANT: we have to wait till any IO submited by the thread is
1849          * completed otherwise iobuf may be corrupted by different request
1850          */
1851         wait_event(iobuf->dr_wait,
1852                        atomic_read(&iobuf->dr_numreqs) == 0);
1853         osd_fini_iobuf(osd, iobuf);
1854         if (!rc)
1855                 rc = iobuf->dr_error;
1856
1857         if (unlikely(remove_agents != 0))
1858                 osd_process_scheduled_agent_removals(env, osd);
1859
1860         sb_end_write(osd_sb(osd));
1861
1862         RETURN(rc);
1863 }
1864
1865 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1866 {
1867         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1868                                                ot_super);
1869
1870         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1871         LASSERT(&dcb->dcb_func != NULL);
1872         if (dcb->dcb_flags & DCB_TRANS_STOP)
1873                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1874         else
1875                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1876
1877         return 0;
1878 }
1879
1880 /*
1881  * Called just before object is freed. Releases all resources except for
1882  * object itself (that is released by osd_object_free()).
1883  *
1884  * Concurrency: no concurrent access is possible that late in object
1885  * life-cycle.
1886  */
1887 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1888 {
1889         struct osd_object *obj   = osd_obj(l);
1890         struct inode      *inode = obj->oo_inode;
1891
1892         LINVRNT(osd_invariant(obj));
1893
1894         /*
1895          * If object is unlinked remove fid->ino mapping from object index.
1896          */
1897
1898         osd_index_fini(obj);
1899         if (inode != NULL) {
1900                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1901                 qid_t                    uid = i_uid_read(inode);
1902                 qid_t                    gid = i_gid_read(inode);
1903
1904                 iput(inode);
1905                 obj->oo_inode = NULL;
1906
1907                 if (qsd != NULL) {
1908                         struct osd_thread_info  *info = osd_oti_get(env);
1909                         struct lquota_id_info   *qi = &info->oti_qi;
1910
1911                         /* Release granted quota to master if necessary */
1912                         qi->lqi_id.qid_uid = uid;
1913                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1914
1915                         qi->lqi_id.qid_uid = gid;
1916                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1917                 }
1918         }
1919 }
1920
1921 /*
1922  * Concurrency: ->loo_object_release() is called under site spin-lock.
1923  */
1924 static void osd_object_release(const struct lu_env *env,
1925                                struct lu_object *l)
1926 {
1927         struct osd_object *o = osd_obj(l);
1928         /* nobody should be releasing a non-destroyed object with nlink=0
1929          * the API allows this, but ldiskfs doesn't like and then report
1930          * this inode as deleted */
1931         if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1932                 LBUG();
1933 }
1934
1935 /*
1936  * Concurrency: shouldn't matter.
1937  */
1938 static int osd_object_print(const struct lu_env *env, void *cookie,
1939                             lu_printer_t p, const struct lu_object *l)
1940 {
1941         struct osd_object *o = osd_obj(l);
1942         struct iam_descr  *d;
1943
1944         if (o->oo_dir != NULL)
1945                 d = o->oo_dir->od_container.ic_descr;
1946         else
1947                 d = NULL;
1948         return (*p)(env, cookie,
1949                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1950                     o, o->oo_inode,
1951                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1952                     o->oo_inode ? o->oo_inode->i_generation : 0,
1953                     d ? d->id_ops->id_name : "plain");
1954 }
1955
1956 /*
1957  * Concurrency: shouldn't matter.
1958  */
1959 int osd_statfs(const struct lu_env *env, struct dt_device *d,
1960                struct obd_statfs *sfs)
1961 {
1962         struct osd_device       *osd = osd_dt_dev(d);
1963         struct super_block      *sb = osd_sb(osd);
1964         struct kstatfs          *ksfs;
1965         __u64                    reserved;
1966         int                      result = 0;
1967
1968         if (unlikely(osd->od_mnt == NULL))
1969                 return -EINPROGRESS;
1970
1971         /* osd_lproc.c call this without env, allocate ksfs for that case */
1972         if (unlikely(env == NULL)) {
1973                 OBD_ALLOC_PTR(ksfs);
1974                 if (ksfs == NULL)
1975                         return -ENOMEM;
1976         } else {
1977                 ksfs = &osd_oti_get(env)->oti_ksfs;
1978         }
1979
1980         result = sb->s_op->statfs(sb->s_root, ksfs);
1981         if (result)
1982                 goto out;
1983
1984         statfs_pack(sfs, ksfs);
1985         if (unlikely(sb->s_flags & MS_RDONLY))
1986                 sfs->os_state |= OS_STATE_READONLY;
1987         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1988                                          LDISKFS_FEATURE_INCOMPAT_EXTENTS))
1989                 sfs->os_maxbytes = sb->s_maxbytes;
1990         else
1991                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
1992
1993         /*
1994          * Reserve some space so to avoid fragmenting the filesystem too much.
1995          * Fragmentation not only impacts performance, but can also increase
1996          * metadata overhead significantly, causing grant calculation to be
1997          * wrong.
1998          *
1999          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2000          */
2001         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2002         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2003         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2004                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2005
2006         sfs->os_blocks -= reserved;
2007         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2008         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2009
2010 out:
2011         if (unlikely(env == NULL))
2012                 OBD_FREE_PTR(ksfs);
2013         return result;
2014 }
2015
2016 /**
2017  * Estimate space needed for file creations. We assume the largest filename
2018  * which is 2^64 - 1, hence a filename of 20 chars.
2019  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2020  */
2021 #ifdef __LDISKFS_DIR_REC_LEN
2022 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2023 #else
2024 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2025 #endif
2026
2027 /*
2028  * Concurrency: doesn't access mutable data.
2029  */
2030 static void osd_conf_get(const struct lu_env *env,
2031                          const struct dt_device *dev,
2032                          struct dt_device_param *param)
2033 {
2034         struct super_block *sb = osd_sb(osd_dt_dev(dev));
2035         int                ea_overhead;
2036
2037         /*
2038          * XXX should be taken from not-yet-existing fs abstraction layer.
2039          */
2040         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2041         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2042         param->ddp_symlink_max  = sb->s_blocksize;
2043         param->ddp_mount_type     = LDD_MT_LDISKFS;
2044         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2045                 param->ddp_maxbytes = sb->s_maxbytes;
2046         else
2047                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2048         /* inode are statically allocated, so per-inode space consumption
2049          * is the space consumed by the directory entry */
2050         param->ddp_inodespace     = PER_OBJ_USAGE;
2051         /* EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
2052          * = 128MB) which is unlikely to be hit in real life. Report a smaller
2053          * maximum length to not under count the actual number of extents
2054          * needed for writing a file. */
2055         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2056         /* worst-case extent insertion metadata overhead */
2057         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2058         param->ddp_mntopts      = 0;
2059         if (test_opt(sb, XATTR_USER))
2060                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2061         if (test_opt(sb, POSIX_ACL))
2062                 param->ddp_mntopts |= MNTOPT_ACL;
2063
2064         /* LOD might calculate the max stripe count based on max_ea_size,
2065          * so we need take account in the overhead as well,
2066          * xattr_header + magic + xattr_entry_head */
2067         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2068                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2069
2070 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2071         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2072                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2073                                                                 ea_overhead;
2074         else
2075 #endif
2076                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2077 }
2078
2079 /*
2080  * Concurrency: shouldn't matter.
2081  */
2082 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2083 {
2084         int rc;
2085
2086         CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2087
2088         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2089
2090         CDEBUG(D_CACHE, "synced OSD %s: rc = %d\n",
2091                LUSTRE_OSD_LDISKFS_NAME, rc);
2092
2093         return rc;
2094 }
2095
2096 /**
2097  * Start commit for OSD device.
2098  *
2099  * An implementation of dt_commit_async method for OSD device.
2100  * Asychronously starts underlayng fs sync and thereby a transaction
2101  * commit.
2102  *
2103  * \param env environment
2104  * \param d dt device
2105  *
2106  * \see dt_device_operations
2107  */
2108 static int osd_commit_async(const struct lu_env *env,
2109                             struct dt_device *d)
2110 {
2111         struct super_block *s = osd_sb(osd_dt_dev(d));
2112         ENTRY;
2113
2114         CDEBUG(D_HA, "async commit OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2115         RETURN(s->s_op->sync_fs(s, 0));
2116 }
2117
2118 /*
2119  * Concurrency: shouldn't matter.
2120  */
2121
2122 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2123 {
2124         struct super_block *sb = osd_sb(osd_dt_dev(d));
2125         struct block_device *dev = sb->s_bdev;
2126 #ifdef HAVE_DEV_SET_RDONLY
2127         struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2128         int rc = 0;
2129 #else
2130         int rc = -EOPNOTSUPP;
2131 #endif
2132         ENTRY;
2133
2134 #ifdef HAVE_DEV_SET_RDONLY
2135         CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
2136
2137         if (sb->s_op->freeze_fs) {
2138                 rc = sb->s_op->freeze_fs(sb);
2139                 if (rc)
2140                         goto out;
2141         }
2142
2143         if (jdev && (jdev != dev)) {
2144                 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2145                        (long)jdev);
2146                 dev_set_rdonly(jdev);
2147         }
2148         CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2149         dev_set_rdonly(dev);
2150
2151         if (sb->s_op->unfreeze_fs)
2152                 sb->s_op->unfreeze_fs(sb);
2153
2154 out:
2155 #endif
2156         if (rc)
2157                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2158                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2159
2160         RETURN(rc);
2161 }
2162
2163 /**
2164  * Note: we do not count into QUOTA here.
2165  * If we mount with --data_journal we may need more.
2166  */
2167 const int osd_dto_credits_noquota[DTO_NR] = {
2168         /**
2169          * Insert.
2170          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2171          * SINGLEDATA_TRANS_BLOCKS(8)
2172          * XXX Note: maybe iam need more, since iam have more level than
2173          *           EXT3 htree.
2174          */
2175         [DTO_INDEX_INSERT]  = 16,
2176         /**
2177          * Delete
2178          * just modify a single entry, probably merge few within a block
2179          */
2180         [DTO_INDEX_DELETE]  = 1,
2181         /**
2182          * Used for OI scrub
2183          */
2184         [DTO_INDEX_UPDATE]  = 16,
2185         /**
2186          * 4(inode, inode bits, groups, GDT)
2187          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2188          */
2189         [DTO_OBJECT_CREATE] = 4,
2190         /**
2191          * 4(inode, inode bits, groups, GDT)
2192          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2193          */
2194         [DTO_OBJECT_DELETE] = 4,
2195         /**
2196          * Attr set credits (inode)
2197          */
2198         [DTO_ATTR_SET_BASE] = 1,
2199         /**
2200          * Xattr set. The same as xattr of EXT3.
2201          * DATA_TRANS_BLOCKS(14)
2202          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2203          * are also counted in. Do not know why?
2204          */
2205         [DTO_XATTR_SET]     = 14,
2206         /**
2207          * credits for inode change during write.
2208          */
2209         [DTO_WRITE_BASE]    = 3,
2210         /**
2211          * credits for single block write.
2212          */
2213         [DTO_WRITE_BLOCK]   = 14,
2214         /**
2215          * Attr set credits for chown.
2216          * This is extra credits for setattr, and it is null without quota
2217          */
2218         [DTO_ATTR_SET_CHOWN] = 0
2219 };
2220
2221 static const struct dt_device_operations osd_dt_ops = {
2222         .dt_root_get       = osd_root_get,
2223         .dt_statfs         = osd_statfs,
2224         .dt_trans_create   = osd_trans_create,
2225         .dt_trans_start    = osd_trans_start,
2226         .dt_trans_stop     = osd_trans_stop,
2227         .dt_trans_cb_add   = osd_trans_cb_add,
2228         .dt_conf_get       = osd_conf_get,
2229         .dt_sync           = osd_sync,
2230         .dt_ro             = osd_ro,
2231         .dt_commit_async   = osd_commit_async,
2232 };
2233
2234 static void osd_object_read_lock(const struct lu_env *env,
2235                                  struct dt_object *dt, unsigned role)
2236 {
2237         struct osd_object *obj = osd_dt_obj(dt);
2238         struct osd_thread_info *oti = osd_oti_get(env);
2239
2240         LINVRNT(osd_invariant(obj));
2241
2242         LASSERT(obj->oo_owner != env);
2243         down_read_nested(&obj->oo_sem, role);
2244
2245         LASSERT(obj->oo_owner == NULL);
2246         oti->oti_r_locks++;
2247 }
2248
2249 static void osd_object_write_lock(const struct lu_env *env,
2250                                   struct dt_object *dt, unsigned role)
2251 {
2252         struct osd_object *obj = osd_dt_obj(dt);
2253         struct osd_thread_info *oti = osd_oti_get(env);
2254
2255         LINVRNT(osd_invariant(obj));
2256
2257         LASSERT(obj->oo_owner != env);
2258         down_write_nested(&obj->oo_sem, role);
2259
2260         LASSERT(obj->oo_owner == NULL);
2261         obj->oo_owner = env;
2262         oti->oti_w_locks++;
2263 }
2264
2265 static void osd_object_read_unlock(const struct lu_env *env,
2266                                    struct dt_object *dt)
2267 {
2268         struct osd_object *obj = osd_dt_obj(dt);
2269         struct osd_thread_info *oti = osd_oti_get(env);
2270
2271         LINVRNT(osd_invariant(obj));
2272
2273         LASSERT(oti->oti_r_locks > 0);
2274         oti->oti_r_locks--;
2275         up_read(&obj->oo_sem);
2276 }
2277
2278 static void osd_object_write_unlock(const struct lu_env *env,
2279                                     struct dt_object *dt)
2280 {
2281         struct osd_object *obj = osd_dt_obj(dt);
2282         struct osd_thread_info *oti = osd_oti_get(env);
2283
2284         LINVRNT(osd_invariant(obj));
2285
2286         LASSERT(obj->oo_owner == env);
2287         LASSERT(oti->oti_w_locks > 0);
2288         oti->oti_w_locks--;
2289         obj->oo_owner = NULL;
2290         up_write(&obj->oo_sem);
2291 }
2292
2293 static int osd_object_write_locked(const struct lu_env *env,
2294                                    struct dt_object *dt)
2295 {
2296         struct osd_object *obj = osd_dt_obj(dt);
2297
2298         LINVRNT(osd_invariant(obj));
2299
2300         return obj->oo_owner == env;
2301 }
2302
2303 static struct timespec *osd_inode_time(const struct lu_env *env,
2304                                        struct inode *inode, __u64 seconds)
2305 {
2306         struct osd_thread_info  *oti = osd_oti_get(env);
2307         struct timespec         *t   = &oti->oti_time;
2308
2309         t->tv_sec = seconds;
2310         t->tv_nsec = 0;
2311         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2312         return t;
2313 }
2314
2315 static void osd_inode_getattr(const struct lu_env *env,
2316                               struct inode *inode, struct lu_attr *attr)
2317 {
2318         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2319                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2320                            LA_FLAGS | LA_NLINK | LA_RDEV | LA_BLKSIZE |
2321                            LA_TYPE;
2322
2323         attr->la_atime   = LTIME_S(inode->i_atime);
2324         attr->la_mtime   = LTIME_S(inode->i_mtime);
2325         attr->la_ctime   = LTIME_S(inode->i_ctime);
2326         attr->la_mode    = inode->i_mode;
2327         attr->la_size    = i_size_read(inode);
2328         attr->la_blocks  = inode->i_blocks;
2329         attr->la_uid     = i_uid_read(inode);
2330         attr->la_gid     = i_gid_read(inode);
2331         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2332         attr->la_nlink   = inode->i_nlink;
2333         attr->la_rdev    = inode->i_rdev;
2334         attr->la_blksize = 1 << inode->i_blkbits;
2335         attr->la_blkbits = inode->i_blkbits;
2336 }
2337
2338 static int osd_attr_get(const struct lu_env *env,
2339                         struct dt_object *dt,
2340                         struct lu_attr *attr)
2341 {
2342         struct osd_object *obj = osd_dt_obj(dt);
2343
2344         if (unlikely(!dt_object_exists(dt)))
2345                 return -ENOENT;
2346         if (unlikely(obj->oo_destroyed))
2347                 return -ENOENT;
2348
2349         LASSERT(!dt_object_remote(dt));
2350         LINVRNT(osd_invariant(obj));
2351
2352         spin_lock(&obj->oo_guard);
2353         osd_inode_getattr(env, obj->oo_inode, attr);
2354         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2355                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2356         spin_unlock(&obj->oo_guard);
2357
2358         return 0;
2359 }
2360
2361 static int osd_declare_attr_set(const struct lu_env *env,
2362                                 struct dt_object *dt,
2363                                 const struct lu_attr *attr,
2364                                 struct thandle *handle)
2365 {
2366         struct osd_thandle     *oh;
2367         struct osd_object      *obj;
2368         struct osd_thread_info *info = osd_oti_get(env);
2369         struct lquota_id_info  *qi = &info->oti_qi;
2370         qid_t                   uid;
2371         qid_t                   gid;
2372         long long               bspace;
2373         int                     rc = 0;
2374         bool                    enforce;
2375         ENTRY;
2376
2377         LASSERT(dt != NULL);
2378         LASSERT(handle != NULL);
2379
2380         obj = osd_dt_obj(dt);
2381         LASSERT(osd_invariant(obj));
2382
2383         oh = container_of0(handle, struct osd_thandle, ot_super);
2384         LASSERT(oh->ot_handle == NULL);
2385
2386         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2387                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2388
2389         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2390                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2391
2392         if (attr == NULL || obj->oo_inode == NULL)
2393                 RETURN(rc);
2394
2395         bspace   = obj->oo_inode->i_blocks;
2396         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2397         bspace   = toqb(bspace);
2398
2399         /* Changing ownership is always preformed by super user, it should not
2400          * fail with EDQUOT.
2401          *
2402          * We still need to call the osd_declare_qid() to calculate the journal
2403          * credits for updating quota accounting files and to trigger quota
2404          * space adjustment once the operation is completed.*/
2405         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2406                 /* USERQUOTA */
2407                 uid = i_uid_read(obj->oo_inode);
2408                 qi->lqi_type = USRQUOTA;
2409                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2410                 /* inode accounting */
2411                 qi->lqi_is_blk = false;
2412
2413                 /* one more inode for the new uid ... */
2414                 qi->lqi_id.qid_uid = attr->la_uid;
2415                 qi->lqi_space      = 1;
2416                 /* Reserve credits for the new uid */
2417                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2418                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2419                         rc = 0;
2420                 if (rc)
2421                         RETURN(rc);
2422
2423                 /* and one less inode for the current uid */
2424                 qi->lqi_id.qid_uid = uid;
2425                 qi->lqi_space      = -1;
2426                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2427                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2428                         rc = 0;
2429                 if (rc)
2430                         RETURN(rc);
2431
2432                 /* block accounting */
2433                 qi->lqi_is_blk = true;
2434
2435                 /* more blocks for the new uid ... */
2436                 qi->lqi_id.qid_uid = attr->la_uid;
2437                 qi->lqi_space      = bspace;
2438                 /*
2439                  * Credits for the new uid has been reserved, re-use "obj"
2440                  * to save credit reservation.
2441                  */
2442                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2443                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2444                         rc = 0;
2445                 if (rc)
2446                         RETURN(rc);
2447
2448                 /* and finally less blocks for the current uid */
2449                 qi->lqi_id.qid_uid = uid;
2450                 qi->lqi_space      = -bspace;
2451                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2452                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2453                         rc = 0;
2454                 if (rc)
2455                         RETURN(rc);
2456
2457                 /* GROUP QUOTA */
2458                 gid = i_gid_read(obj->oo_inode);
2459                 qi->lqi_type = GRPQUOTA;
2460                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2461
2462                 /* inode accounting */
2463                 qi->lqi_is_blk = false;
2464
2465                 /* one more inode for the new gid ... */
2466                 qi->lqi_id.qid_gid = attr->la_gid;
2467                 qi->lqi_space      = 1;
2468                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2469                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2470                         rc = 0;
2471                 if (rc)
2472                         RETURN(rc);
2473
2474                 /* and one less inode for the current gid */
2475                 qi->lqi_id.qid_gid = gid;
2476                 qi->lqi_space      = -1;
2477                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2478                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2479                         rc = 0;
2480                 if (rc)
2481                         RETURN(rc);
2482
2483                 /* block accounting */
2484                 qi->lqi_is_blk = true;
2485
2486                 /* more blocks for the new gid ... */
2487                 qi->lqi_id.qid_gid = attr->la_gid;
2488                 qi->lqi_space      = bspace;
2489                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2490                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2491                         rc = 0;
2492                 if (rc)
2493                         RETURN(rc);
2494
2495                 /* and finally less blocks for the current gid */
2496                 qi->lqi_id.qid_gid = gid;
2497                 qi->lqi_space      = -bspace;
2498                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2499                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2500                         rc = 0;
2501                 if (rc)
2502                         RETURN(rc);
2503         }
2504
2505         RETURN(rc);
2506 }
2507
2508 static int osd_inode_setattr(const struct lu_env *env,
2509                              struct inode *inode, const struct lu_attr *attr)
2510 {
2511         __u64 bits = attr->la_valid;
2512
2513         /* Only allow set size for regular file */
2514         if (!S_ISREG(inode->i_mode))
2515                 bits &= ~(LA_SIZE | LA_BLOCKS);
2516
2517         if (bits == 0)
2518                 return 0;
2519
2520         if (bits & LA_ATIME)
2521                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2522         if (bits & LA_CTIME)
2523                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2524         if (bits & LA_MTIME)
2525                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2526         if (bits & LA_SIZE) {
2527                 spin_lock(&inode->i_lock);
2528                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2529                 i_size_write(inode, attr->la_size);
2530                 spin_unlock(&inode->i_lock);
2531         }
2532
2533         /* OSD should not change "i_blocks" which is used by quota.
2534          * "i_blocks" should be changed by ldiskfs only. */
2535         if (bits & LA_MODE)
2536                 inode->i_mode = (inode->i_mode & S_IFMT) |
2537                                 (attr->la_mode & ~S_IFMT);
2538         if (bits & LA_UID)
2539                 i_uid_write(inode, attr->la_uid);
2540         if (bits & LA_GID)
2541                 i_gid_write(inode, attr->la_gid);
2542         if (bits & LA_NLINK)
2543                 set_nlink(inode, attr->la_nlink);
2544         if (bits & LA_RDEV)
2545                 inode->i_rdev = attr->la_rdev;
2546
2547         if (bits & LA_FLAGS) {
2548                 /* always keep S_NOCMTIME */
2549                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2550                                  S_NOCMTIME;
2551         }
2552         return 0;
2553 }
2554
2555 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2556 {
2557         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2558             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2559                 struct iattr    iattr;
2560                 int             rc;
2561
2562                 ll_vfs_dq_init(inode);
2563                 iattr.ia_valid = 0;
2564                 if (attr->la_valid & LA_UID)
2565                         iattr.ia_valid |= ATTR_UID;
2566                 if (attr->la_valid & LA_GID)
2567                         iattr.ia_valid |= ATTR_GID;
2568                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2569                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2570
2571                 rc = ll_vfs_dq_transfer(inode, &iattr);
2572                 if (rc) {
2573                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2574                                "enforcement enabled on the ldiskfs "
2575                                "filesystem?\n", inode->i_sb->s_id, rc);
2576                         return rc;
2577                 }
2578         }
2579         return 0;
2580 }
2581
2582 static int osd_attr_set(const struct lu_env *env,
2583                         struct dt_object *dt,
2584                         const struct lu_attr *attr,
2585                         struct thandle *handle)
2586 {
2587         struct osd_object *obj = osd_dt_obj(dt);
2588         struct inode      *inode;
2589         int rc;
2590
2591         if (!dt_object_exists(dt))
2592                 return -ENOENT;
2593
2594         LASSERT(handle != NULL);
2595         LASSERT(!dt_object_remote(dt));
2596         LASSERT(osd_invariant(obj));
2597
2598         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2599
2600         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2601                 struct osd_thread_info  *oti  = osd_oti_get(env);
2602                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
2603                 struct lu_fid           *fid1 = &oti->oti_fid;
2604                 struct osd_inode_id     *id   = &oti->oti_id;
2605                 struct iam_path_descr   *ipd;
2606                 struct iam_container    *bag;
2607                 struct osd_thandle      *oh;
2608                 int                      rc;
2609
2610                 fid_cpu_to_be(fid1, fid0);
2611                 memset(id, 1, sizeof(*id));
2612                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2613                                   fid0)->oi_dir.od_container;
2614                 ipd = osd_idx_ipd_get(env, bag);
2615                 if (unlikely(ipd == NULL))
2616                         RETURN(-ENOMEM);
2617
2618                 oh = container_of0(handle, struct osd_thandle, ot_super);
2619                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2620                                 (const struct iam_rec *)id, ipd);
2621                 osd_ipd_put(env, bag, ipd);
2622                 return(rc > 0 ? 0 : rc);
2623         }
2624
2625         inode = obj->oo_inode;
2626
2627         rc = osd_quota_transfer(inode, attr);
2628         if (rc)
2629                 return rc;
2630
2631         spin_lock(&obj->oo_guard);
2632         rc = osd_inode_setattr(env, inode, attr);
2633         spin_unlock(&obj->oo_guard);
2634         if (rc != 0)
2635                 GOTO(out, rc);
2636
2637         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2638
2639         if (!(attr->la_valid & LA_FLAGS))
2640                 GOTO(out, rc);
2641
2642         /* Let's check if there are extra flags need to be set into LMA */
2643         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2644                 struct osd_thread_info *info = osd_oti_get(env);
2645                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
2646
2647                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
2648                 if (rc != 0)
2649                         GOTO(out, rc);
2650
2651                 lma->lma_incompat |=
2652                         lustre_to_lma_flags(attr->la_flags);
2653                 lustre_lma_swab(lma);
2654                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2655                                      lma, sizeof(*lma), XATTR_REPLACE);
2656                 if (rc != 0) {
2657                         struct osd_device *osd = osd_obj2dev(obj);
2658
2659                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2660                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2661                               lma->lma_incompat, rc);
2662                 } else {
2663                         obj->oo_lma_flags =
2664                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2665                 }
2666                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2667         }
2668 out:
2669         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2670
2671         return rc;
2672 }
2673
2674 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2675                                            struct osd_object *obj,
2676                                            const char *name, const int namelen)
2677 {
2678         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2679 }
2680
2681 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2682                       umode_t mode, struct dt_allocation_hint *hint,
2683                       struct thandle *th)
2684 {
2685         int result;
2686         struct osd_device  *osd = osd_obj2dev(obj);
2687         struct osd_thandle *oth;
2688         struct dt_object   *parent = NULL;
2689         struct inode       *inode;
2690
2691         LINVRNT(osd_invariant(obj));
2692         LASSERT(obj->oo_inode == NULL);
2693         LASSERT(obj->oo_hl_head == NULL);
2694
2695         if (S_ISDIR(mode) && ldiskfs_pdo) {
2696                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2697                 if (obj->oo_hl_head == NULL)
2698                         return -ENOMEM;
2699         }
2700
2701         oth = container_of(th, struct osd_thandle, ot_super);
2702         LASSERT(oth->ot_handle->h_transaction != NULL);
2703
2704         if (hint != NULL && hint->dah_parent != NULL &&
2705             !dt_object_remote(hint->dah_parent))
2706                 parent = hint->dah_parent;
2707
2708         inode = ldiskfs_create_inode(oth->ot_handle,
2709                                      parent ? osd_dt_obj(parent)->oo_inode :
2710                                               osd_sb(osd)->s_root->d_inode,
2711                                      mode);
2712         if (!IS_ERR(inode)) {
2713                 /* Do not update file c/mtime in ldiskfs. */
2714                 inode->i_flags |= S_NOCMTIME;
2715
2716                 /* For new created object, it must be consistent,
2717                  * and it is unnecessary to scrub against it. */
2718                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2719
2720                 obj->oo_inode = inode;
2721                 result = 0;
2722         } else {
2723                 if (obj->oo_hl_head != NULL) {
2724                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2725                         obj->oo_hl_head = NULL;
2726                 }
2727                 result = PTR_ERR(inode);
2728         }
2729         LINVRNT(osd_invariant(obj));
2730         return result;
2731 }
2732
2733 enum {
2734         OSD_NAME_LEN = 255
2735 };
2736
2737 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2738                      struct lu_attr *attr,
2739                      struct dt_allocation_hint *hint,
2740                      struct dt_object_format *dof,
2741                      struct thandle *th)
2742 {
2743         int result;
2744         struct osd_thandle *oth;
2745         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX));
2746
2747         LASSERT(S_ISDIR(attr->la_mode));
2748
2749         oth = container_of(th, struct osd_thandle, ot_super);
2750         LASSERT(oth->ot_handle->h_transaction != NULL);
2751         result = osd_mkfile(info, obj, mode, hint, th);
2752
2753         return result;
2754 }
2755
2756 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2757                         struct lu_attr *attr,
2758                         struct dt_allocation_hint *hint,
2759                         struct dt_object_format *dof,
2760                         struct thandle *th)
2761 {
2762         int result;
2763         struct osd_thandle *oth;
2764         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2765
2766         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2767
2768         LASSERT(S_ISREG(attr->la_mode));
2769
2770         oth = container_of(th, struct osd_thandle, ot_super);
2771         LASSERT(oth->ot_handle->h_transaction != NULL);
2772
2773         result = osd_mkfile(info, obj, mode, hint, th);
2774         if (result == 0) {
2775                 LASSERT(obj->oo_inode != NULL);
2776                 if (feat->dif_flags & DT_IND_VARKEY)
2777                         result = iam_lvar_create(obj->oo_inode,
2778                                                  feat->dif_keysize_max,
2779                                                  feat->dif_ptrsize,
2780                                                  feat->dif_recsize_max,
2781                                                  oth->ot_handle);
2782                 else
2783                         result = iam_lfix_create(obj->oo_inode,
2784                                                  feat->dif_keysize_max,
2785                                                  feat->dif_ptrsize,
2786                                                  feat->dif_recsize_max,
2787                                                  oth->ot_handle);
2788
2789         }
2790         return result;
2791 }
2792
2793 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2794                      struct lu_attr *attr,
2795                      struct dt_allocation_hint *hint,
2796                      struct dt_object_format *dof,
2797                      struct thandle *th)
2798 {
2799         LASSERT(S_ISREG(attr->la_mode));
2800         return osd_mkfile(info, obj, (attr->la_mode &
2801                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2802 }
2803
2804 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2805                      struct lu_attr *attr,
2806                      struct dt_allocation_hint *hint,
2807                      struct dt_object_format *dof,
2808                      struct thandle *th)
2809 {
2810         LASSERT(S_ISLNK(attr->la_mode));
2811         return osd_mkfile(info, obj, (attr->la_mode &
2812                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2813 }
2814
2815 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2816                      struct lu_attr *attr,
2817                      struct dt_allocation_hint *hint,
2818                      struct dt_object_format *dof,
2819                      struct thandle *th)
2820 {
2821         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2822         int result;
2823
2824         LINVRNT(osd_invariant(obj));
2825         LASSERT(obj->oo_inode == NULL);
2826         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2827                 S_ISFIFO(mode) || S_ISSOCK(mode));
2828
2829         result = osd_mkfile(info, obj, mode, hint, th);
2830         if (result == 0) {
2831                 LASSERT(obj->oo_inode != NULL);
2832                 /*
2833                  * This inode should be marked dirty for i_rdev.  Currently
2834                  * that is done in the osd_attr_init().
2835                  */
2836                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2837                                    attr->la_rdev);
2838         }
2839         LINVRNT(osd_invariant(obj));
2840         return result;
2841 }
2842
2843 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2844                               struct lu_attr *,
2845                               struct dt_allocation_hint *hint,
2846                               struct dt_object_format *dof,
2847                               struct thandle *);
2848
2849 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2850 {
2851         osd_obj_type_f result;
2852
2853         switch (type) {
2854         case DFT_DIR:
2855                 result = osd_mkdir;
2856                 break;
2857         case DFT_REGULAR:
2858                 result = osd_mkreg;
2859                 break;
2860         case DFT_SYM:
2861                 result = osd_mksym;
2862                 break;
2863         case DFT_NODE:
2864                 result = osd_mknod;
2865                 break;
2866         case DFT_INDEX:
2867                 result = osd_mk_index;
2868                 break;
2869
2870         default:
2871                 LBUG();
2872                 break;
2873         }
2874         return result;
2875 }
2876
2877
2878 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2879                         struct dt_object *parent, struct dt_object *child,
2880                         umode_t child_mode)
2881 {
2882         LASSERT(ah);
2883
2884         ah->dah_parent = parent;
2885         ah->dah_mode = child_mode;
2886
2887         if (parent != NULL && !dt_object_remote(parent)) {
2888                 /* will help to find FID->ino at dt_insert("..") */
2889                 struct osd_object *pobj = osd_dt_obj(parent);
2890                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2891         }
2892 }
2893
2894 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2895                           struct lu_attr *attr, struct dt_object_format *dof)
2896 {
2897         struct inode   *inode = obj->oo_inode;
2898         __u64           valid = attr->la_valid;
2899         int             result;
2900
2901         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2902
2903         if (dof->dof_type != DFT_NODE)
2904                 attr->la_valid &= ~LA_RDEV;
2905         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2906                 attr->la_valid &= ~LA_ATIME;
2907         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2908                 attr->la_valid &= ~LA_CTIME;
2909         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2910                 attr->la_valid &= ~LA_MTIME;
2911
2912         result = osd_quota_transfer(inode, attr);
2913         if (result)
2914                 return;
2915
2916         if (attr->la_valid != 0) {
2917                 result = osd_inode_setattr(info->oti_env, inode, attr);
2918                 /*
2919                  * The osd_inode_setattr() should always succeed here.  The
2920                  * only error that could be returned is EDQUOT when we are
2921                  * trying to change the UID or GID of the inode. However, this
2922                  * should not happen since quota enforcement is no longer
2923                  * enabled on ldiskfs (lquota takes care of it).
2924                  */
2925                 LASSERTF(result == 0, "%d\n", result);
2926                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2927         }
2928
2929         attr->la_valid = valid;
2930 }
2931
2932 /**
2933  * Helper function for osd_object_create()
2934  *
2935  * \retval 0, on success
2936  */
2937 static int __osd_object_create(struct osd_thread_info *info,
2938                                struct osd_object *obj, struct lu_attr *attr,
2939                                struct dt_allocation_hint *hint,
2940                                struct dt_object_format *dof,
2941                                struct thandle *th)
2942 {
2943         int     result;
2944         __u32   umask;
2945
2946         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
2947
2948         /* we drop umask so that permissions we pass are not affected */
2949         umask = current->fs->umask;
2950         current->fs->umask = 0;
2951
2952         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
2953                                                   th);
2954         if (likely(obj->oo_inode != NULL)) {
2955                 LASSERT(obj->oo_inode->i_state & I_NEW);
2956
2957                 /* Unlock the inode before attr initialization to avoid
2958                  * unnecessary dqget operations. LU-6378 */
2959                 unlock_new_inode(obj->oo_inode);
2960         }
2961
2962         if (likely(result == 0)) {
2963                 osd_attr_init(info, obj, attr, dof);
2964                 osd_object_init0(obj);
2965         }
2966
2967         /* restore previous umask value */
2968         current->fs->umask = umask;
2969
2970         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
2971
2972         return result;
2973 }
2974
2975 /**
2976  * Helper function for osd_object_create()
2977  *
2978  * \retval 0, on success
2979  */
2980 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
2981                            const struct lu_fid *fid, struct thandle *th)
2982 {
2983         struct osd_thread_info *info = osd_oti_get(env);
2984         struct osd_inode_id    *id   = &info->oti_id;
2985         struct osd_device      *osd  = osd_obj2dev(obj);
2986         struct osd_thandle     *oh;
2987         int                     rc;
2988
2989         LASSERT(obj->oo_inode != NULL);
2990
2991         oh = container_of0(th, struct osd_thandle, ot_super);
2992         LASSERT(oh->ot_handle);
2993         osd_trans_exec_op(env, th, OSD_OT_INSERT);
2994
2995         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
2996         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
2997                            OI_CHECK_FLD, NULL);
2998         osd_trans_exec_check(env, th, OSD_OT_INSERT);
2999
3000         return rc;
3001 }
3002
3003 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3004                    u64 seq, struct lu_seq_range *range)
3005 {
3006         struct seq_server_site  *ss = osd_seq_site(osd);
3007
3008         if (fid_seq_is_idif(seq)) {
3009                 fld_range_set_ost(range);
3010                 range->lsr_index = idif_ost_idx(seq);
3011                 return 0;
3012         }
3013
3014         if (!fid_seq_in_fldb(seq)) {
3015                 fld_range_set_mdt(range);
3016                 if (ss != NULL)
3017                         /* FIXME: If ss is NULL, it suppose not get lsr_index
3018                          * at all */
3019                         range->lsr_index = ss->ss_node_id;
3020                 return 0;
3021         }
3022
3023         LASSERT(ss != NULL);
3024         fld_range_set_any(range);
3025         /* OSD will only do local fld lookup */
3026         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3027 }
3028
3029 /*
3030  * Concurrency: no external locking is necessary.
3031  */
3032 static int osd_declare_object_create(const struct lu_env *env,
3033                                      struct dt_object *dt,
3034                                      struct lu_attr *attr,
3035                                      struct dt_allocation_hint *hint,
3036                                      struct dt_object_format *dof,
3037                                      struct thandle *handle)
3038 {
3039         struct osd_thandle      *oh;
3040         int                      rc;
3041         ENTRY;
3042
3043         LASSERT(handle != NULL);
3044
3045         oh = container_of0(handle, struct osd_thandle, ot_super);
3046         LASSERT(oh->ot_handle == NULL);
3047
3048         /* EA object consumes more credits than regular object: osd_mk_index
3049          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3050          * leaf_node, could involves the block, block bitmap, groups, GDT
3051          * change for each block, so add 4 * 2 credits in that case. */
3052         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3053                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3054                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3055         /* Reuse idle OI block may cause additional one OI block
3056          * to be changed. */
3057         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3058                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3059
3060         if (!attr)
3061                 RETURN(0);
3062
3063         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid, 1, oh,
3064                                    osd_dt_obj(dt), false, NULL, false);
3065         if (rc != 0)
3066                 RETURN(rc);
3067
3068         /* will help to find FID->ino mapping at dt_insert() */
3069         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3070                                    osd_dt_obj(dt));
3071
3072         RETURN(rc);
3073 }
3074
3075 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
3076                              struct lu_attr *attr,
3077                              struct dt_allocation_hint *hint,
3078                              struct dt_object_format *dof, struct thandle *th)
3079 {
3080         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3081         struct osd_object       *obj    = osd_dt_obj(dt);
3082         struct osd_thread_info  *info   = osd_oti_get(env);
3083         int result;
3084         ENTRY;
3085
3086         if (dt_object_exists(dt))
3087                 return -EEXIST;
3088
3089         LINVRNT(osd_invariant(obj));
3090         LASSERT(!dt_object_remote(dt));
3091         LASSERT(osd_write_locked(env, obj));
3092         LASSERT(th != NULL);
3093
3094         if (unlikely(fid_is_acct(fid)))
3095                 /* Quota files can't be created from the kernel any more,
3096                  * 'tune2fs -O quota' will take care of creating them */
3097                 RETURN(-EPERM);
3098
3099         result = __osd_object_create(info, obj, attr, hint, dof, th);
3100         if (result == 0) {
3101                 result = __osd_oi_insert(env, obj, fid, th);
3102                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3103                         obj->oo_dt.do_body_ops = &osd_body_ops;
3104         }
3105         LASSERT(ergo(result == 0,
3106                 dt_object_exists(dt) && !dt_object_remote(dt)));
3107
3108         LASSERT(osd_invariant(obj));
3109         RETURN(result);
3110 }
3111
3112 /**
3113  * Called to destroy on-disk representation of the object
3114  *
3115  * Concurrency: must be locked
3116  */
3117 static int osd_declare_object_destroy(const struct lu_env *env,
3118                                       struct dt_object *dt,
3119                                       struct thandle *th)
3120 {
3121         struct osd_object  *obj = osd_dt_obj(dt);
3122         struct inode       *inode = obj->oo_inode;
3123         struct osd_thandle *oh;
3124         int                 rc;
3125         ENTRY;
3126
3127         if (inode == NULL)
3128                 RETURN(-ENOENT);
3129
3130         oh = container_of0(th, struct osd_thandle, ot_super);
3131         LASSERT(oh->ot_handle == NULL);
3132
3133         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3134                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3135         /* Recycle idle OI leaf may cause additional three OI blocks
3136          * to be changed. */
3137         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3138                 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3139                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3140         /* one less inode */
3141         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3142                                    -1, oh, obj, false, NULL, false);
3143         if (rc)
3144                 RETURN(rc);
3145         /* data to be truncated */
3146         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3147                                    0, oh, obj, true, NULL, false);
3148         if (rc)
3149                 RETURN(rc);