Whamcloud - gitweb
LU-11838 lustre: discard LTIME_S macro
[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, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/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/kallsyms.h>
43 #include <linux/module.h>
44 #include <linux/user_namespace.h>
45 #ifdef HAVE_UIDGID_HEADER
46 # include <linux/uidgid.h>
47 #endif
48
49 /* prerequisite for linux/xattr.h */
50 #include <linux/types.h>
51 /* prerequisite for linux/xattr.h */
52 #include <linux/fs.h>
53 /* XATTR_{REPLACE,CREATE} */
54 #include <linux/xattr.h>
55
56 #include <ldiskfs/ldiskfs.h>
57 #include <ldiskfs/xattr.h>
58 #include <ldiskfs/ldiskfs_extents.h>
59 #undef ENTRY
60 /*
61  * struct OBD_{ALLOC,FREE}*()
62  * OBD_FAIL_CHECK
63  */
64 #include <obd_support.h>
65 /* struct ptlrpc_thread */
66 #include <lustre_net.h>
67 #include <lustre_fid.h>
68 /* process_config */
69 #include <uapi/linux/lustre/lustre_param.h>
70
71 #include "osd_internal.h"
72 #include "osd_dynlocks.h"
73
74 /* llo_* api support */
75 #include <md_object.h>
76 #include <lustre_quota.h>
77
78 #include <lustre_linkea.h>
79
80 /* Maximum EA size is limited by LNET_MTU for remote objects */
81 #define OSD_MAX_EA_SIZE 1048364
82
83 int ldiskfs_pdo = 1;
84 module_param(ldiskfs_pdo, int, 0644);
85 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
86
87 int ldiskfs_track_declares_assert;
88 module_param(ldiskfs_track_declares_assert, int, 0644);
89 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
90
91 /* Slab to allocate dynlocks */
92 struct kmem_cache *dynlock_cachep;
93
94 /* Slab to allocate osd_it_ea */
95 struct kmem_cache *osd_itea_cachep;
96
97 static struct lu_kmem_descr ldiskfs_caches[] = {
98         {
99                 .ckd_cache = &dynlock_cachep,
100                 .ckd_name  = "dynlock_cache",
101                 .ckd_size  = sizeof(struct dynlock_handle)
102         },
103         {
104                 .ckd_cache = &osd_itea_cachep,
105                 .ckd_name  = "osd_itea_cache",
106                 .ckd_size  = sizeof(struct osd_it_ea)
107         },
108         {
109                 .ckd_cache = NULL
110         }
111 };
112
113 static const char dot[] = ".";
114 static const char dotdot[] = "..";
115
116 static const struct lu_object_operations      osd_lu_obj_ops;
117 static const struct dt_object_operations      osd_obj_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 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
155 {
156         struct osd_thread_info *oti = osd_oti_get(env);
157
158         return oti->oti_w_locks > 0 && o->oo_owner == env;
159 }
160
161 /*
162  * Concurrency: doesn't access mutable data
163  */
164 static int osd_root_get(const struct lu_env *env,
165                         struct dt_device *dev, struct lu_fid *f)
166 {
167         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
168         return 0;
169 }
170
171 /*
172  * the following set of functions are used to maintain per-thread
173  * cache of FID->ino mapping. this mechanism is needed to resolve
174  * FID to inode at dt_insert() which in turn stores ino in the
175  * directory entries to keep ldiskfs compatible with ext[34].
176  * due to locking-originated restrictions we can't lookup ino
177  * using LU cache (deadlock is possible). lookup using OI is quite
178  * expensive. so instead we maintain this cache and methods like
179  * dt_create() fill it. so in the majority of cases dt_insert() is
180  * able to find needed mapping in lockless manner.
181  */
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184              const struct lu_fid *fid)
185 {
186         struct osd_thread_info *oti = osd_oti_get(env);
187         struct osd_idmap_cache *idc = oti->oti_ins_cache;
188         int i;
189
190         for (i = 0; i < oti->oti_ins_cache_used; i++) {
191                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
192                         continue;
193                 if (idc[i].oic_dev != osd)
194                         continue;
195
196                 return idc + i;
197         }
198
199         return NULL;
200 }
201
202 static struct osd_idmap_cache *
203 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
204             const struct lu_fid *fid)
205 {
206         struct osd_thread_info *oti   = osd_oti_get(env);
207         struct osd_idmap_cache *idc;
208         int i;
209
210         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
211                 i = oti->oti_ins_cache_size * 2;
212                 if (i == 0)
213                         i = OSD_INS_CACHE_SIZE;
214                 OBD_ALLOC(idc, sizeof(*idc) * i);
215                 if (idc == NULL)
216                         return ERR_PTR(-ENOMEM);
217                 if (oti->oti_ins_cache != NULL) {
218                         memcpy(idc, oti->oti_ins_cache,
219                                oti->oti_ins_cache_used * sizeof(*idc));
220                         OBD_FREE(oti->oti_ins_cache,
221                                  oti->oti_ins_cache_used * sizeof(*idc));
222                 }
223                 oti->oti_ins_cache = idc;
224                 oti->oti_ins_cache_size = i;
225         }
226
227         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
228         idc->oic_fid = *fid;
229         idc->oic_dev = osd;
230         idc->oic_lid.oii_ino = 0;
231         idc->oic_lid.oii_gen = 0;
232         idc->oic_remote = 0;
233
234         return idc;
235 }
236
237 /*
238  * lookup mapping for the given fid in the cache, initialize a
239  * new one if not found. the initialization checks whether the
240  * object is local or remote. for local objects, OI is used to
241  * learn ino/generation. the function is used when the caller
242  * has no information about the object, e.g. at dt_insert().
243  */
244 static struct osd_idmap_cache *
245 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
246                      const struct lu_fid *fid)
247 {
248         struct osd_idmap_cache *idc;
249         int rc;
250
251         idc = osd_idc_find(env, osd, fid);
252         LASSERT(!IS_ERR(idc));
253         if (idc != NULL)
254                 return idc;
255
256         CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
257                osd->od_svname, PFID(fid));
258
259         /* new mapping is needed */
260         idc = osd_idc_add(env, osd, fid);
261         if (IS_ERR(idc)) {
262                 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
263                        osd->od_svname, PFID(fid), PTR_ERR(idc));
264                 return idc;
265         }
266
267         /* initialize it */
268         rc = osd_remote_fid(env, osd, fid);
269         if (unlikely(rc < 0))
270                 return ERR_PTR(rc);
271
272         if (rc == 0) {
273                 /* the object is local, lookup in OI */
274                 /* XXX: probably cheaper to lookup in LU first? */
275                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
276                                    &idc->oic_lid, 0);
277                 if (unlikely(rc < 0)) {
278                         CERROR("can't lookup: rc = %d\n", rc);
279                         return ERR_PTR(rc);
280                 }
281         } else {
282                 /* the object is remote */
283                 idc->oic_remote = 1;
284         }
285
286         return idc;
287 }
288
289 /*
290  * lookup mapping for given FID and fill it from the given object.
291  * the object is lolcal by definition.
292  */
293 static int osd_idc_find_and_init(const struct lu_env *env,
294                                  struct osd_device *osd,
295                                  struct osd_object *obj)
296 {
297         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
298         struct osd_idmap_cache *idc;
299
300         idc = osd_idc_find(env, osd, fid);
301         LASSERT(!IS_ERR(idc));
302         if (idc != NULL) {
303                 if (obj->oo_inode == NULL)
304                         return 0;
305                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
306                         LASSERT(idc->oic_lid.oii_ino == 0);
307                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
308                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
309                 }
310                 return 0;
311         }
312
313         CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
314                osd->od_svname, PFID(fid));
315
316         /* new mapping is needed */
317         idc = osd_idc_add(env, osd, fid);
318         if (IS_ERR(idc)) {
319                 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
320                        osd->od_svname, PFID(fid), PTR_ERR(idc));
321                 return PTR_ERR(idc);
322         }
323
324         if (obj->oo_inode != NULL) {
325                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
326                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
327         }
328         return 0;
329 }
330
331 /*
332  * OSD object methods.
333  */
334
335 /*
336  * Concurrency: no concurrent access is possible that early in object
337  * life-cycle.
338  */
339 static struct lu_object *osd_object_alloc(const struct lu_env *env,
340                                           const struct lu_object_header *hdr,
341                                           struct lu_device *d)
342 {
343         struct osd_object *mo;
344
345         OBD_ALLOC_PTR(mo);
346         if (mo != NULL) {
347                 struct lu_object *l;
348                 struct lu_object_header *h;
349                 struct osd_device *o = osd_dev(d);
350
351                 l = &mo->oo_dt.do_lu;
352                 if (unlikely(o->od_in_init)) {
353                         OBD_ALLOC_PTR(h);
354                         if (!h) {
355                                 OBD_FREE_PTR(mo);
356                                 return NULL;
357                         }
358
359                         lu_object_header_init(h);
360                         lu_object_init(l, h, d);
361                         lu_object_add_top(h, l);
362                         mo->oo_header = h;
363                 } else {
364                         dt_object_init(&mo->oo_dt, NULL, d);
365                         mo->oo_header = NULL;
366                 }
367
368                 mo->oo_dt.do_ops = &osd_obj_ops;
369                 l->lo_ops = &osd_lu_obj_ops;
370                 init_rwsem(&mo->oo_sem);
371                 init_rwsem(&mo->oo_ext_idx_sem);
372                 spin_lock_init(&mo->oo_guard);
373                 INIT_LIST_HEAD(&mo->oo_xattr_list);
374                 return l;
375         }
376         return NULL;
377 }
378
379 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
380                 struct dentry *dentry, struct lustre_ost_attrs *loa)
381 {
382         int rc;
383
384         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
385                              (void *)loa, sizeof(*loa));
386         if (rc > 0) {
387                 struct lustre_mdt_attrs *lma = &loa->loa_lma;
388
389                 if (rc < sizeof(*lma))
390                         return -EINVAL;
391
392                 rc = 0;
393                 lustre_loa_swab(loa, true);
394                 /* Check LMA compatibility */
395                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
396                         CWARN("%s: unsupported incompat LMA feature(s) %#x "
397                               "for fid = "DFID", ino = %lu\n",
398                               osd_ino2name(inode),
399                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
400                               PFID(&lma->lma_self_fid), inode->i_ino);
401                         rc = -EOPNOTSUPP;
402                 }
403         } else if (rc == 0) {
404                 rc = -ENODATA;
405         }
406
407         return rc;
408 }
409
410 /*
411  * retrieve object from backend ext fs.
412  **/
413 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
414                        struct osd_inode_id *id)
415 {
416         int rc;
417         struct inode *inode = NULL;
418
419         /*
420          * if we look for an inode withing a running
421          * transaction, then we risk to deadlock
422          * osd_dirent_check_repair() breaks this
423          */
424          /* LASSERT(current->journal_info == NULL); */
425
426         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
427         if (IS_ERR(inode)) {
428                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
429                        id->oii_ino, PTR_ERR(inode));
430         } else if (id->oii_gen != OSD_OII_NOGEN &&
431                    inode->i_generation != id->oii_gen) {
432                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
433                        "i_generation = %u\n",
434                        id->oii_ino, id->oii_gen, inode->i_generation);
435                 iput(inode);
436                 inode = ERR_PTR(-ESTALE);
437         } else if (inode->i_nlink == 0) {
438                 /*
439                  * due to parallel readdir and unlink,
440                  * we can have dead inode here.
441                  */
442                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
443                 iput(inode);
444                 inode = ERR_PTR(-ESTALE);
445         } else if (is_bad_inode(inode)) {
446                 CWARN("%s: bad inode: ino = %u\n",
447                       osd_dev2name(dev), id->oii_ino);
448                 iput(inode);
449                 inode = ERR_PTR(-ENOENT);
450         } else if ((rc = osd_attach_jinode(inode))) {
451                 iput(inode);
452                 inode = ERR_PTR(rc);
453         } else {
454                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
455                 if (id->oii_gen == OSD_OII_NOGEN)
456                         osd_id_gen(id, inode->i_ino, inode->i_generation);
457
458                 /*
459                  * Do not update file c/mtime in ldiskfs.
460                  * NB: we don't have any lock to protect this because we don't
461                  * have reference on osd_object now, but contention with
462                  * another lookup + attr_set can't happen in the tiny window
463                  * between if (...) and set S_NOCMTIME.
464                  */
465                 if (!(inode->i_flags & S_NOCMTIME))
466                         inode->i_flags |= S_NOCMTIME;
467         }
468         return inode;
469 }
470
471 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
472                           handle_t *handle, struct dentry *child,
473                           struct inode *inode, struct htree_lock *hlock)
474 {
475         int rc, rc2;
476
477         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
478         if (rc == -ENOBUFS || rc == -ENOSPC) {
479                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
480                 struct inode *parent = child->d_parent->d_inode;
481                 struct lu_fid *fid = NULL;
482
483                 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
484                 if (!rc2) {
485                         fid = &loa->loa_lma.lma_self_fid;
486                 } else if (rc2 == -ENODATA) {
487                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
488                                 fid = &info->oti_fid3;
489                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
490                         } else if (!osd->od_is_ost && osd->od_index == 0) {
491                                 fid = &info->oti_fid3;
492                                 lu_igif_build(fid, parent->i_ino,
493                                               parent->i_generation);
494                         }
495                 }
496
497                 if (fid != NULL)
498                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
499                               "maximum entry limit\n",
500                               osd_name(osd), parent->i_ino, PFID(fid),
501                               rc == -ENOSPC ? "has reached" : "is approaching");
502                 else
503                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
504                               "maximum entry limit\n",
505                               osd_name(osd), parent->i_ino,
506                               rc == -ENOSPC ? "has reached" : "is approaching");
507
508                 /* ignore such error now */
509                 if (rc == -ENOBUFS)
510                         rc = 0;
511         }
512
513         return rc;
514 }
515
516
517 struct inode *
518 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
519              struct osd_inode_id *id, struct lu_fid *fid)
520 {
521         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
522         struct inode *inode;
523         int rc;
524
525         inode = osd_iget(info, dev, id);
526         if (IS_ERR(inode))
527                 return inode;
528
529         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
530         if (!rc) {
531                 *fid = loa->loa_lma.lma_self_fid;
532         } else if (rc == -ENODATA) {
533                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
534                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
535                 else
536                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
537         } else {
538                 iput(inode);
539                 inode = ERR_PTR(rc);
540         }
541         return inode;
542 }
543
544 static struct inode *osd_iget_check(struct osd_thread_info *info,
545                                     struct osd_device *dev,
546                                     const struct lu_fid *fid,
547                                     struct osd_inode_id *id,
548                                     bool trusted)
549 {
550         struct inode *inode;
551         int rc = 0;
552
553         ENTRY;
554
555         /*
556          * The cached OI mapping is trustable. If we cannot locate the inode
557          * via the cached OI mapping, then return the failure to the caller
558          * directly without further OI checking.
559          */
560
561 again:
562         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
563         if (IS_ERR(inode)) {
564                 rc = PTR_ERR(inode);
565                 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
566                         goto check_oi;
567
568                 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
569                        PFID(fid), id->oii_ino, rc);
570                 GOTO(put, rc);
571         }
572
573         if (is_bad_inode(inode)) {
574                 rc = -ENOENT;
575                 if (!trusted)
576                         goto check_oi;
577
578                 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
579                        PFID(fid), id->oii_ino);
580                 GOTO(put, rc);
581         }
582
583         if (id->oii_gen != OSD_OII_NOGEN &&
584             inode->i_generation != id->oii_gen) {
585                 rc = -ESTALE;
586                 if (!trusted)
587                         goto check_oi;
588
589                 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
590                        "oii_gen = %u, i_generation = %u\n", PFID(fid),
591                        id->oii_ino, id->oii_gen, inode->i_generation);
592                 GOTO(put, rc);
593         }
594
595         if (inode->i_nlink == 0) {
596                 rc = -ENOENT;
597                 if (!trusted)
598                         goto check_oi;
599
600                 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
601                        PFID(fid), id->oii_ino);
602                 GOTO(put, rc);
603         }
604
605         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
606
607 check_oi:
608         if (rc != 0) {
609                 __u32 saved_ino = id->oii_ino;
610                 __u32 saved_gen = id->oii_gen;
611
612                 LASSERT(!trusted);
613                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
614
615                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
616                 /*
617                  * XXX: There are four possible cases:
618                  *      1. rc = 0.
619                  *         Backup/restore caused the OI invalid.
620                  *      2. rc = 0.
621                  *         Someone unlinked the object but NOT removed
622                  *         the OI mapping, such as mount target device
623                  *         as ldiskfs, and modify something directly.
624                  *      3. rc = -ENOENT.
625                  *         Someone just removed the object between the
626                  *         former oi_lookup and the iget. It is normal.
627                  *      4. Other failure cases.
628                  *
629                  *      Generally, when the device is mounted, it will
630                  *      auto check whether the system is restored from
631                  *      file-level backup or not. We trust such detect
632                  *      to distinguish the 1st case from the 2nd case:
633                  *      if the OI files are consistent but may contain
634                  *      stale OI mappings because of case 2, if iget()
635                  *      returns -ENOENT or -ESTALE, then it should be
636                  *      the case 2.
637                  */
638                 if (rc != 0)
639                         /*
640                          * If the OI mapping was in OI file before the
641                          * osd_iget_check(), but now, it is disappear,
642                          * then it must be removed by race. That is a
643                          * normal race case.
644                          */
645                         GOTO(put, rc);
646
647                 /*
648                  * It is the OI scrub updated the OI mapping by race.
649                  * The new OI mapping must be valid.
650                  */
651                 if (saved_ino != id->oii_ino ||
652                     (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
653                         if (!IS_ERR(inode))
654                                 iput(inode);
655
656                         trusted = true;
657                         goto again;
658                 }
659
660                 if (IS_ERR(inode)) {
661                         if (dev->od_scrub.os_scrub.os_file.sf_flags &
662                             SF_INCONSISTENT)
663                                 /*
664                                  * It still can be the case 2, but we cannot
665                                  * distinguish it from the case 1. So return
666                                  * -EREMCHG to block current operation until
667                                  *  OI scrub rebuilt the OI mappings.
668                                  */
669                                 rc = -EREMCHG;
670                         else
671                                 rc = -ENOENT;
672
673                         GOTO(put, rc);
674                 }
675
676                 if (inode->i_generation == id->oii_gen)
677                         rc = -ENOENT;
678                 else
679                         rc = -EREMCHG;
680         } else {
681                 if (id->oii_gen == OSD_OII_NOGEN)
682                         osd_id_gen(id, inode->i_ino, inode->i_generation);
683
684                 /*
685                  * Do not update file c/mtime in ldiskfs.
686                  * NB: we don't have any lock to protect this because we don't
687                  * have reference on osd_object now, but contention with
688                  * another lookup + attr_set can't happen in the tiny window
689                  * between if (...) and set S_NOCMTIME.
690                  */
691                 if (!(inode->i_flags & S_NOCMTIME))
692                         inode->i_flags |= S_NOCMTIME;
693         }
694
695         GOTO(put, rc);
696
697 put:
698         if (rc != 0) {
699                 if (!IS_ERR(inode))
700                         iput(inode);
701
702                 inode = ERR_PTR(rc);
703         }
704
705         return inode;
706 }
707
708 /**
709  * \retval +v: new filter_fid, does not contain self-fid
710  * \retval 0:  filter_fid_old, contains self-fid
711  * \retval -v: other failure cases
712  */
713 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
714                  struct dentry *dentry, struct lu_fid *fid)
715 {
716         struct filter_fid_old *ff = &info->oti_ff;
717         struct ost_id *ostid = &info->oti_ostid;
718         int rc;
719
720         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
721         if (rc == sizeof(*ff)) {
722                 rc = 0;
723                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
724                 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
725                 /*
726                  * XXX: use 0 as the index for compatibility, the caller will
727                  * handle index related issues when necessary.
728                  */
729                 if (!rc)
730                         ostid_to_fid(fid, ostid, 0);
731         } else if (rc == sizeof(struct filter_fid)) {
732                 rc = 1;
733         } else if (rc >= 0) {
734                 rc = -EINVAL;
735         }
736
737         return rc;
738 }
739
740 static int osd_lma_self_repair(struct osd_thread_info *info,
741                                struct osd_device *osd, struct inode *inode,
742                                const struct lu_fid *fid, __u32 compat)
743 {
744         handle_t *jh;
745         int rc;
746
747         LASSERT(current->journal_info == NULL);
748
749         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
750                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
751         if (IS_ERR(jh)) {
752                 rc = PTR_ERR(jh);
753                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
754                       osd_name(osd), rc);
755                 return rc;
756         }
757
758         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
759         if (rc != 0)
760                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
761                       osd_name(osd), rc);
762         ldiskfs_journal_stop(jh);
763         return rc;
764 }
765
766 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
767 {
768         struct osd_thread_info *info = osd_oti_get(env);
769         struct osd_device *osd = osd_obj2dev(obj);
770         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
771         struct lustre_mdt_attrs *lma = &loa->loa_lma;
772         struct inode *inode = obj->oo_inode;
773         struct dentry *dentry = &info->oti_obj_dentry;
774         struct lu_fid *fid = NULL;
775         const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
776         int rc;
777
778         ENTRY;
779
780         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
781                              (void *)loa, sizeof(*loa));
782         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
783                 fid = &lma->lma_self_fid;
784                 rc = osd_get_idif(info, inode, dentry, fid);
785                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
786                         /*
787                          * For the given OST-object, if it has neither LMA nor
788                          * FID in XATTR_NAME_FID, then the given FID (which is
789                          * contained in the @obj, from client RPC for locating
790                          * the OST-object) is trusted. We use it to generate
791                          * the LMA.
792                          */
793                         osd_lma_self_repair(info, osd, inode, rfid,
794                                             LMAC_FID_ON_OST);
795                         RETURN(0);
796                 }
797         }
798
799         if (rc < 0)
800                 RETURN(rc);
801
802         if (rc > 0) {
803                 rc = 0;
804                 lustre_lma_swab(lma);
805                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
806                              (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
807                               S_ISREG(inode->i_mode)))) {
808                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
809                               "fid = "DFID", ino = %lu\n", osd_name(osd),
810                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
811                               PFID(rfid), inode->i_ino);
812                         rc = -EOPNOTSUPP;
813                 } else {
814                         fid = &lma->lma_self_fid;
815                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
816                             osd->od_is_ost)
817                                 obj->oo_pfid_in_lma = 1;
818                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
819                             !osd->od_is_ost)
820                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
821                 }
822         }
823
824         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
825                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
826                         struct ost_id   *oi   = &info->oti_ostid;
827                         struct lu_fid   *fid1 = &info->oti_fid3;
828                         __u32            idx  = fid_idif_ost_idx(rfid);
829
830                         /*
831                          * For old IDIF, the OST index is not part of the IDIF,
832                          * Means that different OSTs may have the same IDIFs.
833                          * Under such case, we need to make some compatible
834                          * check to make sure to trigger OI scrub properly.
835                          */
836                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
837                                 /* Given @rfid is new, LMA is old. */
838                                 fid_to_ostid(fid, oi);
839                                 ostid_to_fid(fid1, oi, idx);
840                                 if (lu_fid_eq(fid1, rfid)) {
841                                         if (osd->od_index_in_idif)
842                                                 osd_lma_self_repair(info, osd,
843                                                         inode, rfid,
844                                                         LMAC_FID_ON_OST);
845                                         RETURN(0);
846                                 }
847                         }
848                 }
849
850                 rc = -EREMCHG;
851         }
852
853         RETURN(rc);
854 }
855
856 struct osd_check_lmv_buf {
857 #ifdef HAVE_DIR_CONTEXT
858         /* please keep it as first member */
859         struct dir_context ctx;
860 #endif
861         struct osd_thread_info *oclb_info;
862         struct osd_device *oclb_dev;
863         struct osd_idmap_cache *oclb_oic;
864         int oclb_items;
865         bool oclb_found;
866 };
867
868 /**
869  * It is called internally by ->readdir() to filter out the
870  * local slave object's FID of the striped directory.
871  *
872  * \retval      1 found the local slave's FID
873  * \retval      0 continue to check next item
874  * \retval      -ve for failure
875  */
876 #ifdef HAVE_FILLDIR_USE_CTX
877 static int osd_stripe_dir_filldir(struct dir_context *buf,
878 #else
879 static int osd_stripe_dir_filldir(void *buf,
880 #endif
881                                   const char *name, int namelen,
882                                   loff_t offset, __u64 ino, unsigned int d_type)
883 {
884         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
885         struct osd_thread_info *oti = oclb->oclb_info;
886         struct lu_fid *fid = &oti->oti_fid3;
887         struct osd_inode_id *id = &oti->oti_id3;
888         struct osd_device *dev = oclb->oclb_dev;
889         struct osd_idmap_cache *oic = oclb->oclb_oic;
890         struct inode *inode;
891
892         oclb->oclb_items++;
893
894         if (name[0] == '.')
895                 return 0;
896
897         fid_zero(fid);
898         sscanf(name + 1, SFID, RFID(fid));
899         if (!fid_is_sane(fid))
900                 return 0;
901
902         if (osd_remote_fid(oti->oti_env, dev, fid))
903                 return 0;
904
905         osd_id_gen(id, ino, OSD_OII_NOGEN);
906         inode = osd_iget(oti, dev, id);
907         if (IS_ERR(inode))
908                 return PTR_ERR(inode);
909
910         iput(inode);
911         osd_add_oi_cache(oti, dev, id, fid);
912         oic->oic_fid = *fid;
913         oic->oic_lid = *id;
914         oic->oic_dev = dev;
915         osd_oii_insert(dev, oic, true);
916         oclb->oclb_found = true;
917
918         return 1;
919 }
920
921 /*
922  * When lookup item under striped directory, we need to locate the master
923  * MDT-object of the striped directory firstly, then the client will send
924  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
925  * and the item's name. If the system is restored from MDT file level backup,
926  * then before the OI scrub completely built the OI files, the OI mappings of
927  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
928  * not a problem for the master MDT-object. Because when locate the master
929  * MDT-object, we will do name based lookup (for the striped directory itself)
930  * firstly, during such process we can setup the correct OI mapping for the
931  * master MDT-object. But it will be trouble for the slave MDT-object. Because
932  * the client will not trigger name based lookup on the MDT to locate the slave
933  * MDT-object before locating item under the striped directory, then when
934  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
935  * is invalid and does not know what the right OI mapping is, then the MDT has
936  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
937  * the OI file, related OI mapping is unknown yet, please try again later. And
938  * then client will re-try the RPC again and again until related OI mapping has
939  * been updated. That is quite inefficient.
940  *
941  * To resolve above trouble, we will handle it as the following two cases:
942  *
943  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
944  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
945  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
946  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
947  *    directly. Please check osd_fid_lookup().
948  *
949  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
950  *    Under such case, during lookup the master MDT-object, we will lookup the
951  *    slave MDT-object via readdir against the master MDT-object, because the
952  *    slave MDT-objects information are stored as sub-directories with the name
953  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
954  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
955  *    the correct OI mapping for the slave MDT-object.
956  */
957 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
958                          struct inode *inode, struct osd_idmap_cache *oic)
959 {
960         struct lu_buf *buf = &oti->oti_big_buf;
961         struct dentry *dentry = &oti->oti_obj_dentry;
962         struct file *filp = &oti->oti_file;
963         const struct file_operations *fops;
964         struct lmv_mds_md_v1 *lmv1;
965         struct osd_check_lmv_buf oclb = {
966 #ifdef HAVE_DIR_CONTEXT
967                 .ctx.actor = osd_stripe_dir_filldir,
968 #endif
969                 .oclb_info = oti,
970                 .oclb_dev = dev,
971                 .oclb_oic = oic,
972                 .oclb_found = false,
973         };
974         int rc = 0;
975
976         ENTRY;
977
978 again:
979         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
980                              buf->lb_len);
981         if (rc == -ERANGE) {
982                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
983                 if (rc > 0) {
984                         lu_buf_realloc(buf, rc);
985                         if (buf->lb_buf == NULL)
986                                 GOTO(out, rc = -ENOMEM);
987
988                         goto again;
989                 }
990         }
991
992         if (unlikely(rc == 0 || rc == -ENODATA))
993                 GOTO(out, rc = 0);
994
995         if (rc < 0)
996                 GOTO(out, rc);
997
998         if (unlikely(buf->lb_buf == NULL)) {
999                 lu_buf_realloc(buf, rc);
1000                 if (buf->lb_buf == NULL)
1001                         GOTO(out, rc = -ENOMEM);
1002
1003                 goto again;
1004         }
1005
1006         lmv1 = buf->lb_buf;
1007         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1008                 GOTO(out, rc = 0);
1009
1010         fops = inode->i_fop;
1011         dentry->d_inode = inode;
1012         dentry->d_sb = inode->i_sb;
1013         filp->f_pos = 0;
1014         filp->f_path.dentry = dentry;
1015         filp->f_mode = FMODE_64BITHASH;
1016         filp->f_mapping = inode->i_mapping;
1017         filp->f_op = fops;
1018         filp->private_data = NULL;
1019         set_file_inode(filp, inode);
1020
1021         do {
1022                 oclb.oclb_items = 0;
1023 #ifdef HAVE_DIR_CONTEXT
1024                 oclb.ctx.pos = filp->f_pos;
1025 #ifdef HAVE_ITERATE_SHARED
1026                 rc = fops->iterate_shared(filp, &oclb.ctx);
1027 #else
1028                 rc = fops->iterate(filp, &oclb.ctx);
1029 #endif
1030                 filp->f_pos = oclb.ctx.pos;
1031 #else
1032                 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
1033 #endif
1034         } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1035                  filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1036         fops->release(inode, filp);
1037
1038 out:
1039         if (rc < 0)
1040                 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
1041                        DFID": rc = %d\n", osd_ino2name(inode),
1042                        inode->i_ino, inode->i_generation,
1043                        PFID(&oic->oic_fid), rc);
1044         else
1045                 rc = 0;
1046
1047         RETURN(rc);
1048 }
1049
1050 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1051                           const struct lu_fid *fid,
1052                           const struct lu_object_conf *conf)
1053 {
1054         struct osd_thread_info *info;
1055         struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1056         struct osd_device *dev;
1057         struct osd_idmap_cache *oic;
1058         struct osd_inode_id *id;
1059         struct inode *inode = NULL;
1060         struct lustre_scrub *scrub;
1061         struct scrub_file *sf;
1062         __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1063         __u32 saved_ino;
1064         __u32 saved_gen;
1065         int result = 0;
1066         int rc1 = 0;
1067         bool remote = false;
1068         bool trusted = true;
1069         bool updated = false;
1070         bool checked = false;
1071
1072         ENTRY;
1073
1074         LINVRNT(osd_invariant(obj));
1075         LASSERT(obj->oo_inode == NULL);
1076         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1077
1078         dev = osd_dev(ldev);
1079         scrub = &dev->od_scrub.os_scrub;
1080         sf = &scrub->os_file;
1081         info = osd_oti_get(env);
1082         LASSERT(info);
1083         oic = &info->oti_cache;
1084
1085         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1086                 RETURN(-ENOENT);
1087
1088         /*
1089          * For the object is created as locking anchor, or for the object to
1090          * be created on disk. No need to osd_oi_lookup() at here because FID
1091          * shouldn't never be re-used, if it's really a duplicate FID from
1092          * unexpected reason, we should be able to detect it later by calling
1093          * do_create->osd_oi_insert().
1094          */
1095         if (conf && conf->loc_flags & LOC_F_NEW)
1096                 GOTO(out, result = 0);
1097
1098         /* Search order: 1. per-thread cache. */
1099         if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1100                 id = &oic->oic_lid;
1101                 goto iget;
1102         }
1103
1104         id = &info->oti_id;
1105         if (!list_empty(&scrub->os_inconsistent_items)) {
1106                 /* Search order: 2. OI scrub pending list. */
1107                 result = osd_oii_lookup(dev, fid, id);
1108                 if (!result)
1109                         goto iget;
1110         }
1111
1112         /*
1113          * The OI mapping in the OI file can be updated by the OI scrub
1114          * when we locate the inode via FID. So it may be not trustable.
1115          */
1116         trusted = false;
1117
1118         /* Search order: 3. OI files. */
1119         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1120         if (result == -ENOENT) {
1121                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1122                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1123                     !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1124                                       sf->sf_oi_bitmap))
1125                         GOTO(out, result = 0);
1126
1127                 goto trigger;
1128         }
1129
1130         if (result)
1131                 GOTO(out, result);
1132
1133 iget:
1134         obj->oo_inode = NULL;
1135         /* for later passes through checks, not true on first pass */
1136         if (!IS_ERR_OR_NULL(inode))
1137                 iput(inode);
1138
1139         inode = osd_iget_check(info, dev, fid, id, trusted);
1140         if (!IS_ERR(inode)) {
1141                 obj->oo_inode = inode;
1142                 result = 0;
1143                 if (remote)
1144                         goto trigger;
1145
1146                 goto check_lma;
1147         }
1148
1149         result = PTR_ERR(inode);
1150         if (result == -ENOENT || result == -ESTALE)
1151                 GOTO(out, result = 0);
1152
1153         if (result != -EREMCHG)
1154                 GOTO(out, result);
1155
1156 trigger:
1157         /*
1158          * We still have chance to get the valid inode: for the
1159          * object which is referenced by remote name entry, the
1160          * object on the local MDT will be linked under the dir
1161          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1162          *
1163          * We do not know whether the object for the given FID
1164          * is referenced by some remote name entry or not, and
1165          * especially for DNE II, a multiple-linked object may
1166          * have many name entries reside on many MDTs.
1167          *
1168          * To simplify the operation, OSD will not distinguish
1169          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1170          * only happened for the RPC from other MDT during the
1171          * OI scrub, or for the client side RPC with FID only,
1172          * such as FID to path, or from old connected client.
1173          */
1174         if (!remote) {
1175                 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1176                 if (!rc1) {
1177                         remote = true;
1178                         trusted = true;
1179                         flags |= SS_AUTO_PARTIAL;
1180                         flags &= ~SS_AUTO_FULL;
1181                         goto iget;
1182                 }
1183         }
1184
1185         if (thread_is_running(&scrub->os_thread)) {
1186                 if (scrub->os_partial_scan && !scrub->os_in_join)
1187                         goto join;
1188
1189                 if (IS_ERR_OR_NULL(inode) || result)
1190                         GOTO(out, result = -EINPROGRESS);
1191
1192                 LASSERT(remote);
1193                 LASSERT(obj->oo_inode == inode);
1194
1195                 osd_add_oi_cache(info, dev, id, fid);
1196                 osd_oii_insert(dev, oic, true);
1197                 goto found;
1198         }
1199
1200         if (dev->od_auto_scrub_interval == AS_NEVER) {
1201                 if (!remote)
1202                         GOTO(out, result = -EREMCHG);
1203
1204                 LASSERT(!result);
1205                 LASSERT(obj->oo_inode == inode);
1206
1207                 osd_add_oi_cache(info, dev, id, fid);
1208                 goto found;
1209         }
1210
1211 join:
1212         rc1 = osd_scrub_start(env, dev, flags);
1213         LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1214                       "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1215         if (rc1 && rc1 != -EALREADY)
1216                 GOTO(out, result = -EREMCHG);
1217
1218         if (IS_ERR_OR_NULL(inode) || result)
1219                 GOTO(out, result = -EINPROGRESS);
1220
1221         LASSERT(remote);
1222         LASSERT(obj->oo_inode == inode);
1223
1224         osd_add_oi_cache(info, dev, id, fid);
1225         osd_oii_insert(dev, oic, true);
1226         goto found;
1227
1228 check_lma:
1229         checked = true;
1230         if (unlikely(obj->oo_header))
1231                 goto found;
1232
1233         result = osd_check_lma(env, obj);
1234         if (!result)
1235                 goto found;
1236
1237         LASSERTF(id->oii_ino == inode->i_ino &&
1238                  id->oii_gen == inode->i_generation,
1239                  "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1240                  PFID(fid), id->oii_ino, id->oii_gen,
1241                  inode->i_ino, inode->i_generation);
1242
1243         saved_ino = inode->i_ino;
1244         saved_gen = inode->i_generation;
1245
1246         if (unlikely(result == -ENODATA)) {
1247                 /*
1248                  * If the OI scrub updated the OI mapping by race, it
1249                  * must be valid. Trust the inode that has no LMA EA.
1250                  */
1251                 if (updated)
1252                         goto found;
1253
1254                 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1255                 if (!result) {
1256                         /*
1257                          * The OI mapping is still there, the inode is still
1258                          * valid. It is just becaues the inode has no LMA EA.
1259                          */
1260                         if (saved_ino == id->oii_ino &&
1261                             saved_gen == id->oii_gen)
1262                                 goto found;
1263
1264                         /*
1265                          * It is the OI scrub updated the OI mapping by race.
1266                          * The new OI mapping must be valid.
1267                          */
1268                         trusted = true;
1269                         updated = true;
1270                         goto iget;
1271                 }
1272
1273                 /*
1274                  * "result == -ENOENT" means that the OI mappinghas been
1275                  * removed by race, so the inode belongs to other object.
1276                  *
1277                  * Others error can be returned  directly.
1278                  */
1279                 if (result == -ENOENT) {
1280                         LASSERT(trusted);
1281
1282                         obj->oo_inode = NULL;
1283                         result = 0;
1284                 }
1285         }
1286
1287         if (result != -EREMCHG)
1288                 GOTO(out, result);
1289
1290         LASSERT(!updated);
1291
1292         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1293         /*
1294          * "result == -ENOENT" means the cached OI mapping has been removed
1295          * from the OI file by race, above inode belongs to other object.
1296          */
1297         if (result == -ENOENT) {
1298                 LASSERT(trusted);
1299
1300                 obj->oo_inode = NULL;
1301                 GOTO(out, result = 0);
1302         }
1303
1304         if (result)
1305                 GOTO(out, result);
1306
1307         if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1308                 result = -EREMCHG;
1309                 goto trigger;
1310         }
1311
1312         /*
1313          * It is the OI scrub updated the OI mapping by race.
1314          * The new OI mapping must be valid.
1315          */
1316         trusted = true;
1317         updated = true;
1318         goto iget;
1319
1320 found:
1321         if (!checked) {
1322                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1323                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1324
1325                 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1326                 if (!result) {
1327                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
1328                             dev->od_is_ost)
1329                                 obj->oo_pfid_in_lma = 1;
1330                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1331                             !dev->od_is_ost)
1332                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1333                 } else if (result != -ENODATA) {
1334                         GOTO(out, result);
1335                 }
1336         }
1337
1338         obj->oo_compat_dot_created = 1;
1339         obj->oo_compat_dotdot_created = 1;
1340
1341         if (S_ISDIR(inode->i_mode) &&
1342             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1343                 osd_check_lmv(info, dev, inode, oic);
1344
1345         result = osd_attach_jinode(inode);
1346         if (result)
1347                 GOTO(out, result);
1348
1349         if (!ldiskfs_pdo)
1350                 GOTO(out, result = 0);
1351
1352         LASSERT(!obj->oo_hl_head);
1353         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1354
1355         GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1356
1357 out:
1358         if (result || !obj->oo_inode) {
1359                 if (!IS_ERR_OR_NULL(inode))
1360                         iput(inode);
1361
1362                 obj->oo_inode = NULL;
1363                 if (trusted)
1364                         fid_zero(&oic->oic_fid);
1365         }
1366
1367         LINVRNT(osd_invariant(obj));
1368         return result;
1369 }
1370
1371 /*
1372  * Concurrency: shouldn't matter.
1373  */
1374 static void osd_object_init0(struct osd_object *obj)
1375 {
1376         LASSERT(obj->oo_inode != NULL);
1377         obj->oo_dt.do_body_ops = &osd_body_ops;
1378         obj->oo_dt.do_lu.lo_header->loh_attr |=
1379                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1380 }
1381
1382 /*
1383  * Concurrency: no concurrent access is possible that early in object
1384  * life-cycle.
1385  */
1386 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1387                            const struct lu_object_conf *conf)
1388 {
1389         struct osd_object *obj = osd_obj(l);
1390         int result;
1391
1392         LINVRNT(osd_invariant(obj));
1393
1394         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1395                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1396                 l->lo_header->loh_attr |= LOHA_EXISTS;
1397                 return 0;
1398         }
1399
1400         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1401         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1402         if (result == 0 && obj->oo_inode != NULL) {
1403                 struct osd_thread_info *oti = osd_oti_get(env);
1404                 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1405
1406                 osd_object_init0(obj);
1407                 if (unlikely(obj->oo_header))
1408                         return 0;
1409
1410                 result = osd_get_lma(oti, obj->oo_inode,
1411                                      &oti->oti_obj_dentry, loa);
1412                 if (!result) {
1413                         /*
1414                          * Convert LMAI flags to lustre LMA flags
1415                          * and cache it to oo_lma_flags
1416                          */
1417                         obj->oo_lma_flags =
1418                                 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1419                 } else if (result == -ENODATA) {
1420                         result = 0;
1421                 }
1422         }
1423
1424         LINVRNT(osd_invariant(obj));
1425         return result;
1426 }
1427
1428 /*
1429  * The first part of oxe_buf is xattr name, and is '\0' terminated.
1430  * The left part is for value, binary mode.
1431  */
1432 struct osd_xattr_entry {
1433         struct list_head        oxe_list;
1434         size_t                  oxe_len;
1435         size_t                  oxe_namelen;
1436         bool                    oxe_exist;
1437         struct rcu_head         oxe_rcu;
1438         char                    oxe_buf[0];
1439 };
1440
1441 static int osd_oxc_get(struct osd_object *obj, const char *name,
1442                        struct lu_buf *buf)
1443 {
1444         struct osd_xattr_entry *tmp;
1445         struct osd_xattr_entry *oxe = NULL;
1446         size_t namelen = strlen(name);
1447         int rc;
1448
1449         ENTRY;
1450
1451         rcu_read_lock();
1452         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1453                 if (namelen == tmp->oxe_namelen &&
1454                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1455                         oxe = tmp;
1456                         break;
1457                 }
1458         }
1459
1460         if (oxe == NULL)
1461                 GOTO(out, rc = -ENOENT);
1462
1463         if (!oxe->oxe_exist)
1464                 GOTO(out, rc = -ENODATA);
1465
1466         /* vallen */
1467         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1468         LASSERT(rc > 0);
1469
1470         if (buf->lb_buf == NULL)
1471                 GOTO(out, rc);
1472
1473         if (buf->lb_len < rc)
1474                 GOTO(out, rc = -ERANGE);
1475
1476         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1477         EXIT;
1478 out:
1479         rcu_read_unlock();
1480
1481         return rc;
1482 }
1483
1484 static void osd_oxc_free(struct rcu_head *head)
1485 {
1486         struct osd_xattr_entry *oxe;
1487
1488         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1489         OBD_FREE(oxe, oxe->oxe_len);
1490 }
1491
1492 static void osd_oxc_add(struct osd_object *obj, const char *name,
1493                         const char *buf, int buflen)
1494 {
1495         struct osd_xattr_entry *oxe;
1496         struct osd_xattr_entry *old = NULL;
1497         struct osd_xattr_entry *tmp;
1498         size_t namelen = strlen(name);
1499         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1500
1501         OBD_ALLOC(oxe, len);
1502         if (oxe == NULL)
1503                 return;
1504
1505         INIT_LIST_HEAD(&oxe->oxe_list);
1506         oxe->oxe_len = len;
1507         oxe->oxe_namelen = namelen;
1508         memcpy(oxe->oxe_buf, name, namelen);
1509         if (buflen > 0) {
1510                 LASSERT(buf != NULL);
1511                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1512                 oxe->oxe_exist = true;
1513         } else {
1514                 oxe->oxe_exist = false;
1515         }
1516
1517         /* this should be rarely called, just remove old and add new */
1518         spin_lock(&obj->oo_guard);
1519         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1520                 if (namelen == tmp->oxe_namelen &&
1521                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1522                         old = tmp;
1523                         break;
1524                 }
1525         }
1526         if (old != NULL) {
1527                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1528                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1529         } else {
1530                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1531         }
1532         spin_unlock(&obj->oo_guard);
1533 }
1534
1535 static void osd_oxc_del(struct osd_object *obj, const char *name)
1536 {
1537         struct osd_xattr_entry *oxe;
1538         size_t namelen = strlen(name);
1539
1540         spin_lock(&obj->oo_guard);
1541         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1542                 if (namelen == oxe->oxe_namelen &&
1543                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1544                         list_del_rcu(&oxe->oxe_list);
1545                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1546                         break;
1547                 }
1548         }
1549         spin_unlock(&obj->oo_guard);
1550 }
1551
1552 static void osd_oxc_fini(struct osd_object *obj)
1553 {
1554         struct osd_xattr_entry *oxe, *next;
1555
1556         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1557                 list_del(&oxe->oxe_list);
1558                 OBD_FREE(oxe, oxe->oxe_len);
1559         }
1560 }
1561
1562 /*
1563  * Concurrency: no concurrent access is possible that late in object
1564  * life-cycle.
1565  */
1566 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1567 {
1568         struct osd_object *obj = osd_obj(l);
1569         struct lu_object_header *h = obj->oo_header;
1570
1571         LINVRNT(osd_invariant(obj));
1572
1573         osd_oxc_fini(obj);
1574         dt_object_fini(&obj->oo_dt);
1575         if (obj->oo_hl_head != NULL)
1576                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1577         OBD_FREE_PTR(obj);
1578         if (unlikely(h)) {
1579                 lu_object_header_fini(h);
1580                 OBD_FREE_PTR(h);
1581         }
1582 }
1583
1584 /*
1585  * Concurrency: no concurrent access is possible that late in object
1586  * life-cycle.
1587  */
1588 static void osd_index_fini(struct osd_object *o)
1589 {
1590         struct iam_container *bag;
1591
1592         if (o->oo_dir != NULL) {
1593                 bag = &o->oo_dir->od_container;
1594                 if (o->oo_inode != NULL) {
1595                         if (bag->ic_object == o->oo_inode)
1596                                 iam_container_fini(bag);
1597                 }
1598                 OBD_FREE_PTR(o->oo_dir);
1599                 o->oo_dir = NULL;
1600         }
1601 }
1602
1603 /*
1604  * Concurrency: no concurrent access is possible that late in object
1605  * life-cycle (for all existing callers, that is. New callers have to provide
1606  * their own locking.)
1607  */
1608 static int osd_inode_unlinked(const struct inode *inode)
1609 {
1610         return inode->i_nlink == 0;
1611 }
1612
1613 enum {
1614         OSD_TXN_OI_DELETE_CREDITS    = 20,
1615         OSD_TXN_INODE_DELETE_CREDITS = 20
1616 };
1617
1618 /*
1619  * Journal
1620  */
1621
1622 #if OSD_THANDLE_STATS
1623 /**
1624  * Set time when the handle is allocated
1625  */
1626 static void osd_th_alloced(struct osd_thandle *oth)
1627 {
1628         oth->oth_alloced = ktime_get();
1629 }
1630
1631 /**
1632  * Set time when the handle started
1633  */
1634 static void osd_th_started(struct osd_thandle *oth)
1635 {
1636         oth->oth_started = ktime_get();
1637 }
1638
1639 /**
1640  * Check whether the we deal with this handle for too long.
1641  */
1642 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1643                                 ktime_t alloced, ktime_t started,
1644                                 ktime_t closed)
1645 {
1646         ktime_t now = ktime_get();
1647
1648         LASSERT(dev != NULL);
1649
1650         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1651                             ktime_us_delta(started, alloced));
1652         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1653                             ktime_us_delta(closed, started));
1654         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1655                             ktime_us_delta(now, closed));
1656
1657         if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1658                 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1659                                 oth, now, alloced, started, closed);
1660                 libcfs_debug_dumpstack(NULL);
1661         }
1662 }
1663
1664 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1665 {                                                                       \
1666         ktime_t __closed = ktime_get();                                 \
1667         ktime_t __alloced = oth->oth_alloced;                           \
1668         ktime_t __started = oth->oth_started;                           \
1669                                                                         \
1670         expr;                                                           \
1671         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1672 }
1673
1674 #else /* OSD_THANDLE_STATS */
1675
1676 #define osd_th_alloced(h)                  do {} while(0)
1677 #define osd_th_started(h)                  do {} while(0)
1678 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1679
1680 #endif /* OSD_THANDLE_STATS */
1681
1682 /*
1683  * Concurrency: doesn't access mutable data.
1684  */
1685 static int osd_param_is_not_sane(const struct osd_device *dev,
1686                                  const struct thandle *th)
1687 {
1688         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1689
1690         return oh->ot_credits > osd_transaction_size(dev);
1691 }
1692
1693 /*
1694  * Concurrency: shouldn't matter.
1695  */
1696 static void osd_trans_commit_cb(struct super_block *sb,
1697                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1698 {
1699         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1700         struct thandle *th = &oh->ot_super;
1701         struct lu_device *lud = &th->th_dev->dd_lu_dev;
1702         struct dt_txn_commit_cb *dcb, *tmp;
1703
1704         LASSERT(oh->ot_handle == NULL);
1705
1706         if (error)
1707                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1708
1709         dt_txn_hook_commit(th);
1710
1711         /* call per-transaction callbacks if any */
1712         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1713                                  dcb_linkage) {
1714                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1715                          "commit callback entry: magic=%x name='%s'\n",
1716                          dcb->dcb_magic, dcb->dcb_name);
1717                 list_del_init(&dcb->dcb_linkage);
1718                 dcb->dcb_func(NULL, th, dcb, error);
1719         }
1720
1721         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1722         lu_device_put(lud);
1723         th->th_dev = NULL;
1724
1725         OBD_FREE_PTR(oh);
1726 }
1727
1728 #ifndef HAVE_SB_START_WRITE
1729 # define sb_start_write(sb) do {} while (0)
1730 # define sb_end_write(sb) do {} while (0)
1731 #endif
1732
1733 static struct thandle *osd_trans_create(const struct lu_env *env,
1734                                         struct dt_device *d)
1735 {
1736         struct osd_thread_info *oti = osd_oti_get(env);
1737         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1738         struct osd_thandle *oh;
1739         struct thandle *th;
1740
1741         ENTRY;
1742
1743         if (d->dd_rdonly) {
1744                 CERROR("%s: someone try to start transaction under "
1745                        "readonly mode, should be disabled.\n",
1746                        osd_name(osd_dt_dev(d)));
1747                 dump_stack();
1748                 RETURN(ERR_PTR(-EROFS));
1749         }
1750
1751         /* on pending IO in this thread should left from prev. request */
1752         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1753
1754         sb_start_write(osd_sb(osd_dt_dev(d)));
1755
1756         OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1757         if (!oh) {
1758                 sb_end_write(osd_sb(osd_dt_dev(d)));
1759                 RETURN(ERR_PTR(-ENOMEM));
1760         }
1761
1762         oh->ot_quota_trans = &oti->oti_quota_trans;
1763         memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1764         th = &oh->ot_super;
1765         th->th_dev = d;
1766         th->th_result = 0;
1767         oh->ot_credits = 0;
1768         INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1769         INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1770         INIT_LIST_HEAD(&oh->ot_trunc_locks);
1771         osd_th_alloced(oh);
1772
1773         memset(oti->oti_declare_ops, 0,
1774                sizeof(oti->oti_declare_ops));
1775         memset(oti->oti_declare_ops_cred, 0,
1776                sizeof(oti->oti_declare_ops_cred));
1777         memset(oti->oti_declare_ops_used, 0,
1778                sizeof(oti->oti_declare_ops_used));
1779
1780         oti->oti_ins_cache_depth++;
1781
1782         RETURN(th);
1783 }
1784
1785 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1786 {
1787         struct osd_thread_info *oti = osd_oti_get(env);
1788         struct osd_thandle *oh;
1789
1790         oh = container_of0(th, struct osd_thandle, ot_super);
1791         LASSERT(oh != NULL);
1792
1793         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1794               oti->oti_declare_ops[OSD_OT_CREATE],
1795               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1796               oti->oti_declare_ops_used[OSD_OT_CREATE],
1797               oti->oti_declare_ops[OSD_OT_DESTROY],
1798               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1799               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1800         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1801               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1802               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1803               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1804               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1805               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1806               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1807         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1808               oti->oti_declare_ops[OSD_OT_WRITE],
1809               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1810               oti->oti_declare_ops_used[OSD_OT_WRITE],
1811               oti->oti_declare_ops[OSD_OT_PUNCH],
1812               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1813               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1814               oti->oti_declare_ops[OSD_OT_QUOTA],
1815               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1816               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1817         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1818               oti->oti_declare_ops[OSD_OT_INSERT],
1819               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1820               oti->oti_declare_ops_used[OSD_OT_INSERT],
1821               oti->oti_declare_ops[OSD_OT_DELETE],
1822               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1823               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1824         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1825               oti->oti_declare_ops[OSD_OT_REF_ADD],
1826               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1827               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1828               oti->oti_declare_ops[OSD_OT_REF_DEL],
1829               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1830               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1831 }
1832
1833 /*
1834  * Concurrency: shouldn't matter.
1835  */
1836 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1837                            struct thandle *th)
1838 {
1839         struct osd_thread_info *oti = osd_oti_get(env);
1840         struct osd_device *dev = osd_dt_dev(d);
1841         handle_t *jh;
1842         struct osd_thandle *oh;
1843         int rc;
1844
1845         ENTRY;
1846
1847         LASSERT(current->journal_info == NULL);
1848
1849         oh = container_of0(th, struct osd_thandle, ot_super);
1850         LASSERT(oh != NULL);
1851         LASSERT(oh->ot_handle == NULL);
1852
1853         rc = dt_txn_hook_start(env, d, th);
1854         if (rc != 0)
1855                 GOTO(out, rc);
1856
1857         if (unlikely(osd_param_is_not_sane(dev, th))) {
1858                 static unsigned long last_printed;
1859                 static int last_credits;
1860
1861                 /*
1862                  * don't make noise on a tiny testing systems
1863                  * actual credits misuse will be caught anyway
1864                  */
1865                 if (last_credits != oh->ot_credits &&
1866                     time_after(jiffies, last_printed +
1867                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1868                     osd_transaction_size(dev) > 512) {
1869                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1870                               oh->ot_credits, osd_transaction_size(dev));
1871                         osd_trans_dump_creds(env, th);
1872                         libcfs_debug_dumpstack(NULL);
1873                         last_credits = oh->ot_credits;
1874                         last_printed = jiffies;
1875                 }
1876                 /*
1877                  * XXX Limit the credits to 'max_transaction_buffers', and
1878                  *     let the underlying filesystem to catch the error if
1879                  *     we really need so many credits.
1880                  *
1881                  *     This should be removed when we can calculate the
1882                  *     credits precisely.
1883                  */
1884                 oh->ot_credits = osd_transaction_size(dev);
1885         } else if (ldiskfs_track_declares_assert != 0) {
1886                 /*
1887                  * reserve few credits to prevent an assertion in JBD
1888                  * our debugging mechanism will be able to detected
1889                  * overuse. this can help to debug single-update
1890                  * transactions
1891                  */
1892                 oh->ot_credits += 10;
1893                 if (unlikely(osd_param_is_not_sane(dev, th)))
1894                         oh->ot_credits = osd_transaction_size(dev);
1895         }
1896
1897         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1898                 GOTO(out, rc = -EIO);
1899
1900         /*
1901          * XXX temporary stuff. Some abstraction layer should
1902          * be used.
1903          */
1904         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1905         osd_th_started(oh);
1906         if (!IS_ERR(jh)) {
1907                 oh->ot_handle = jh;
1908                 LASSERT(oti->oti_txns == 0);
1909
1910                 lu_device_get(&d->dd_lu_dev);
1911                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1912                               "osd-tx", th);
1913                 oti->oti_txns++;
1914                 rc = 0;
1915         } else {
1916                 rc = PTR_ERR(jh);
1917         }
1918 out:
1919         RETURN(rc);
1920 }
1921
1922 static int osd_seq_exists(const struct lu_env *env,
1923                           struct osd_device *osd, u64 seq)
1924 {
1925         struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1926         struct seq_server_site *ss = osd_seq_site(osd);
1927         int rc;
1928
1929         ENTRY;
1930
1931         LASSERT(ss != NULL);
1932         LASSERT(ss->ss_server_fld != NULL);
1933
1934         rc = osd_fld_lookup(env, osd, seq, range);
1935         if (rc != 0) {
1936                 if (rc != -ENOENT)
1937                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1938                                osd_name(osd), seq, rc);
1939                 RETURN(0);
1940         }
1941
1942         RETURN(ss->ss_node_id == range->lsr_index);
1943 }
1944
1945 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1946 {
1947         struct dt_txn_commit_cb *dcb;
1948         struct dt_txn_commit_cb *tmp;
1949
1950         /* call per-transaction stop callbacks if any */
1951         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1952                                  dcb_linkage) {
1953                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1954                          "commit callback entry: magic=%x name='%s'\n",
1955                          dcb->dcb_magic, dcb->dcb_name);
1956                 list_del_init(&dcb->dcb_linkage);
1957                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1958         }
1959 }
1960
1961 /*
1962  * Concurrency: shouldn't matter.
1963  */
1964 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1965                           struct thandle *th)
1966 {
1967         struct osd_thread_info *oti = osd_oti_get(env);
1968         struct osd_thandle *oh;
1969         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1970         struct osd_device *osd = osd_dt_dev(th->th_dev);
1971         struct qsd_instance *qsd = osd_def_qsd(osd);
1972         struct lquota_trans *qtrans;
1973         struct list_head truncates = LIST_HEAD_INIT(truncates);
1974         int rc = 0, remove_agents = 0;
1975
1976         ENTRY;
1977
1978         oh = container_of0(th, struct osd_thandle, ot_super);
1979
1980         remove_agents = oh->ot_remove_agents;
1981
1982         qtrans = oh->ot_quota_trans;
1983         oh->ot_quota_trans = NULL;
1984
1985         /* move locks to local list, stop tx, execute truncates */
1986         list_splice(&oh->ot_trunc_locks, &truncates);
1987
1988         if (oh->ot_handle != NULL) {
1989                 int rc2;
1990
1991                 handle_t *hdl = oh->ot_handle;
1992
1993                 /*
1994                  * add commit callback
1995                  * notice we don't do this in osd_trans_start()
1996                  * as underlying transaction can change during truncate
1997                  */
1998                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1999                                              &oh->ot_jcb);
2000
2001                 LASSERT(oti->oti_txns == 1);
2002                 oti->oti_txns--;
2003
2004                 rc = dt_txn_hook_stop(env, th);
2005                 if (rc != 0)
2006                         CERROR("%s: failed in transaction hook: rc = %d\n",
2007                                osd_name(osd), rc);
2008
2009                 osd_trans_stop_cb(oh, rc);
2010                 /* hook functions might modify th_sync */
2011                 hdl->h_sync = th->th_sync;
2012
2013                 oh->ot_handle = NULL;
2014                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2015                 if (rc2 != 0)
2016                         CERROR("%s: failed to stop transaction: rc = %d\n",
2017                                osd_name(osd), rc2);
2018                 if (!rc)
2019                         rc = rc2;
2020
2021                 osd_process_truncates(&truncates);
2022         } else {
2023                 osd_trans_stop_cb(oh, th->th_result);
2024                 OBD_FREE_PTR(oh);
2025         }
2026
2027         osd_trunc_unlock_all(&truncates);
2028
2029         /* inform the quota slave device that the transaction is stopping */
2030         qsd_op_end(env, qsd, qtrans);
2031
2032         /*
2033          * as we want IO to journal and data IO be concurrent, we don't block
2034          * awaiting data IO completion in osd_do_bio(), instead we wait here
2035          * once transaction is submitted to the journal. all reqular requests
2036          * don't do direct IO (except read/write), thus this wait_event becomes
2037          * no-op for them.
2038          *
2039          * IMPORTANT: we have to wait till any IO submited by the thread is
2040          * completed otherwise iobuf may be corrupted by different request
2041          */
2042         wait_event(iobuf->dr_wait,
2043                        atomic_read(&iobuf->dr_numreqs) == 0);
2044         osd_fini_iobuf(osd, iobuf);
2045         if (!rc)
2046                 rc = iobuf->dr_error;
2047
2048         if (unlikely(remove_agents != 0))
2049                 osd_process_scheduled_agent_removals(env, osd);
2050
2051         oti->oti_ins_cache_depth--;
2052         /* reset OI cache for safety */
2053         if (oti->oti_ins_cache_depth == 0)
2054                 oti->oti_ins_cache_used = 0;
2055
2056         sb_end_write(osd_sb(osd));
2057
2058         RETURN(rc);
2059 }
2060
2061 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2062 {
2063         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
2064                                                ot_super);
2065
2066         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2067         LASSERT(&dcb->dcb_func != NULL);
2068         if (dcb->dcb_flags & DCB_TRANS_STOP)
2069                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2070         else
2071                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2072
2073         return 0;
2074 }
2075
2076 /*
2077  * Called just before object is freed. Releases all resources except for
2078  * object itself (that is released by osd_object_free()).
2079  *
2080  * Concurrency: no concurrent access is possible that late in object
2081  * life-cycle.
2082  */
2083 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2084 {
2085         struct osd_object *obj = osd_obj(l);
2086         struct inode *inode = obj->oo_inode;
2087
2088         LINVRNT(osd_invariant(obj));
2089
2090         /*
2091          * If object is unlinked remove fid->ino mapping from object index.
2092          */
2093
2094         osd_index_fini(obj);
2095         if (inode != NULL) {
2096                 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2097                 qid_t                uid = i_uid_read(inode);
2098                 qid_t                gid = i_gid_read(inode);
2099
2100                 obj->oo_inode = NULL;
2101                 iput(inode);
2102                 if (!obj->oo_header && qsd) {
2103                         struct osd_thread_info *info = osd_oti_get(env);
2104                         struct lquota_id_info *qi = &info->oti_qi;
2105
2106                         /* Release granted quota to master if necessary */
2107                         qi->lqi_id.qid_uid = uid;
2108                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2109
2110                         qi->lqi_id.qid_uid = gid;
2111                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2112
2113                         qi->lqi_id.qid_uid = i_projid_read(inode);
2114                         qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2115                 }
2116         }
2117 }
2118
2119 /*
2120  * Concurrency: ->loo_object_release() is called under site spin-lock.
2121  */
2122 static void osd_object_release(const struct lu_env *env,
2123                                struct lu_object *l)
2124 {
2125         struct osd_object *o = osd_obj(l);
2126
2127         /*
2128          * nobody should be releasing a non-destroyed object with nlink=0
2129          * the API allows this, but ldiskfs doesn't like and then report
2130          * this inode as deleted
2131          */
2132         LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2133                   o->oo_inode->i_nlink == 0));
2134 }
2135
2136 /*
2137  * Concurrency: shouldn't matter.
2138  */
2139 static int osd_object_print(const struct lu_env *env, void *cookie,
2140                             lu_printer_t p, const struct lu_object *l)
2141 {
2142         struct osd_object *o = osd_obj(l);
2143         struct iam_descr *d;
2144
2145         if (o->oo_dir != NULL)
2146                 d = o->oo_dir->od_container.ic_descr;
2147         else
2148                 d = NULL;
2149         return (*p)(env, cookie,
2150                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2151                     o, o->oo_inode,
2152                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
2153                     o->oo_inode ? o->oo_inode->i_generation : 0,
2154                     d ? d->id_ops->id_name : "plain");
2155 }
2156
2157 /*
2158  * Concurrency: shouldn't matter.
2159  */
2160 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2161                 struct obd_statfs *sfs)
2162 {
2163         struct osd_device *osd = osd_dt_dev(d);
2164         struct super_block *sb = osd_sb(osd);
2165         struct kstatfs *ksfs;
2166         __u64 reserved;
2167         int result = 0;
2168
2169         if (unlikely(osd->od_mnt == NULL))
2170                 return -EINPROGRESS;
2171
2172         /* osd_lproc.c call this without env, allocate ksfs for that case */
2173         if (unlikely(env == NULL)) {
2174                 OBD_ALLOC_PTR(ksfs);
2175                 if (ksfs == NULL)
2176                         return -ENOMEM;
2177         } else {
2178                 ksfs = &osd_oti_get(env)->oti_ksfs;
2179         }
2180
2181         result = sb->s_op->statfs(sb->s_root, ksfs);
2182         if (result)
2183                 goto out;
2184
2185         statfs_pack(sfs, ksfs);
2186         if (unlikely(sb->s_flags & MS_RDONLY))
2187                 sfs->os_state |= OS_STATE_READONLY;
2188         if (ldiskfs_has_feature_extents(sb))
2189                 sfs->os_maxbytes = sb->s_maxbytes;
2190         else
2191                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2192
2193         /*
2194          * Reserve some space so to avoid fragmenting the filesystem too much.
2195          * Fragmentation not only impacts performance, but can also increase
2196          * metadata overhead significantly, causing grant calculation to be
2197          * wrong.
2198          *
2199          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2200          */
2201         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2202         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2203         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2204                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2205
2206         sfs->os_blocks -= reserved;
2207         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2208         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2209
2210 out:
2211         if (unlikely(env == NULL))
2212                 OBD_FREE_PTR(ksfs);
2213         return result;
2214 }
2215
2216 /**
2217  * Estimate space needed for file creations. We assume the largest filename
2218  * which is 2^64 - 1, hence a filename of 20 chars.
2219  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2220  */
2221 #ifdef __LDISKFS_DIR_REC_LEN
2222 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2223 #else
2224 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2225 #endif
2226
2227 /*
2228  * Concurrency: doesn't access mutable data.
2229  */
2230 static void osd_conf_get(const struct lu_env *env,
2231                          const struct dt_device *dev,
2232                          struct dt_device_param *param)
2233 {
2234         struct osd_device *d = osd_dt_dev(dev);
2235         struct super_block *sb = osd_sb(d);
2236         struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2237         const char *name;
2238         int ea_overhead;
2239
2240         /*
2241          * XXX should be taken from not-yet-existing fs abstraction layer.
2242          */
2243         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2244         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2245         param->ddp_symlink_max  = sb->s_blocksize;
2246         param->ddp_mount_type   = LDD_MT_LDISKFS;
2247         if (ldiskfs_has_feature_extents(sb))
2248                 param->ddp_maxbytes = sb->s_maxbytes;
2249         else
2250                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2251         /*
2252          * inode are statically allocated, so per-inode space consumption
2253          * is the space consumed by the directory entry
2254          */
2255         param->ddp_inodespace     = PER_OBJ_USAGE;
2256         /*
2257          * EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
2258          * = 128MB) which is unlikely to be hit in real life. Report a smaller
2259          * maximum length to not under count the actual number of extents
2260          * needed for writing a file.
2261          */
2262         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2263         /* worst-case extent insertion metadata overhead */
2264         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2265         param->ddp_mntopts = 0;
2266         if (test_opt(sb, XATTR_USER))
2267                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2268         if (test_opt(sb, POSIX_ACL))
2269                 param->ddp_mntopts |= MNTOPT_ACL;
2270
2271         /*
2272          * LOD might calculate the max stripe count based on max_ea_size,
2273          * so we need take account in the overhead as well,
2274          * xattr_header + magic + xattr_entry_head
2275          */
2276         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2277                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2278
2279 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2280         if (ldiskfs_has_feature_ea_inode(sb))
2281                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2282                                                                 ea_overhead;
2283         else
2284 #endif
2285                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2286
2287         if (param->ddp_max_ea_size > OSD_MAX_EA_SIZE)
2288                 param->ddp_max_ea_size = OSD_MAX_EA_SIZE;
2289
2290         /*
2291          * Preferred RPC size for efficient disk IO.  4MB shows good
2292          * all-around performance for ldiskfs, but use bigalloc chunk size
2293          * by default if larger.
2294          */
2295 #if defined(LDISKFS_CLUSTER_SIZE)
2296         if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2297                 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2298         else
2299 #endif
2300                 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2301
2302         param->ddp_t10_cksum_type = 0;
2303         if (bi) {
2304                 unsigned short interval = blk_integrity_interval(bi);
2305                 name = blk_integrity_name(bi);
2306                 /*
2307                  * Expected values:
2308                  * T10-DIF-TYPE1-CRC
2309                  * T10-DIF-TYPE3-CRC
2310                  * T10-DIF-TYPE1-IP
2311                  * T10-DIF-TYPE3-IP
2312                  */
2313                 if (strncmp(name, "T10-DIF-TYPE",
2314                             sizeof("T10-DIF-TYPE") - 1) == 0) {
2315                         /* also skip "1/3-" at end */
2316                         const int type_off = sizeof("T10-DIF-TYPE.");
2317                         char type_number = name[type_off - 2];
2318
2319                         if (interval != 512 && interval != 4096) {
2320                                 CERROR("%s: unsupported T10PI sector size %u\n",
2321                                        d->od_svname, interval);
2322                         } else if (type_number != '1' && type_number != '3') {
2323                                 CERROR("%s: unsupported T10PI type %s\n",
2324                                        d->od_svname, name);
2325                         } else if (strcmp(name + type_off, "CRC") == 0) {
2326                                 d->od_t10_type = type_number == '1' ?
2327                                         OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2328                                 param->ddp_t10_cksum_type = interval == 512 ?
2329                                         OBD_CKSUM_T10CRC512 :
2330                                         OBD_CKSUM_T10CRC4K;
2331                         } else if (strcmp(name + type_off, "IP") == 0) {
2332                                 d->od_t10_type = type_number == '1' ?
2333                                         OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2334                                 param->ddp_t10_cksum_type = interval == 512 ?
2335                                         OBD_CKSUM_T10IP512 :
2336                                         OBD_CKSUM_T10IP4K;
2337                         } else {
2338                                 CERROR("%s: unsupported checksum type of "
2339                                        "T10PI type '%s'",
2340                                        d->od_svname, name);
2341                         }
2342
2343                 } else {
2344                         CERROR("%s: unsupported T10PI type '%s'",
2345                                d->od_svname, name);
2346                 }
2347         }
2348 }
2349
2350 /*
2351  * Concurrency: shouldn't matter.
2352  */
2353 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2354 {
2355         int rc;
2356         struct super_block *s = osd_sb(osd_dt_dev(d));
2357         ENTRY;
2358
2359         down_read(&s->s_umount);
2360         rc = s->s_op->sync_fs(s, 1);
2361         up_read(&s->s_umount);
2362
2363         CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2364                rc);
2365
2366         return rc;
2367 }
2368
2369 /**
2370  * Start commit for OSD device.
2371  *
2372  * An implementation of dt_commit_async method for OSD device.
2373  * Asychronously starts underlayng fs sync and thereby a transaction
2374  * commit.
2375  *
2376  * \param env environment
2377  * \param d dt device
2378  *
2379  * \see dt_device_operations
2380  */
2381 static int osd_commit_async(const struct lu_env *env,
2382                             struct dt_device *d)
2383 {
2384         struct super_block *s = osd_sb(osd_dt_dev(d));
2385         int rc;
2386
2387         ENTRY;
2388
2389         CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2390         down_read(&s->s_umount);
2391         rc = s->s_op->sync_fs(s, 0);
2392         up_read(&s->s_umount);
2393
2394         RETURN(rc);
2395 }
2396
2397 /* Our own copy of the set readonly functions if present, or NU if not. */
2398 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2399 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2400 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2401
2402 /*
2403  * Concurrency: shouldn't matter.
2404  */
2405 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2406 {
2407         struct super_block *sb = osd_sb(osd_dt_dev(d));
2408         struct block_device *dev = sb->s_bdev;
2409         int rc = -EOPNOTSUPP;
2410
2411         ENTRY;
2412
2413         if (priv_dev_set_rdonly) {
2414                 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2415
2416                 rc = 0;
2417                 CERROR("*** setting %s read-only ***\n",
2418                        osd_dt_dev(d)->od_svname);
2419
2420                 if (sb->s_op->freeze_fs) {
2421                         rc = sb->s_op->freeze_fs(sb);
2422                         if (rc)
2423                                 goto out;
2424                 }
2425
2426                 if (jdev && (jdev != dev)) {
2427                         CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2428                                (long)jdev);
2429                         priv_dev_set_rdonly(jdev);
2430                 }
2431                 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2432                 priv_dev_set_rdonly(dev);
2433
2434                 if (sb->s_op->unfreeze_fs)
2435                         sb->s_op->unfreeze_fs(sb);
2436         }
2437
2438 out:
2439         if (rc)
2440                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2441                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2442
2443         RETURN(rc);
2444 }
2445
2446 /**
2447  * Note: we do not count into QUOTA here.
2448  * If we mount with --data_journal we may need more.
2449  */
2450 const int osd_dto_credits_noquota[DTO_NR] = {
2451         /**
2452          * Insert.
2453          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2454          * SINGLEDATA_TRANS_BLOCKS(8)
2455          * XXX Note: maybe iam need more, since iam have more level than
2456          *           EXT3 htree.
2457          */
2458         [DTO_INDEX_INSERT]  = 16,
2459         /**
2460          * Delete
2461          * just modify a single entry, probably merge few within a block
2462          */
2463         [DTO_INDEX_DELETE]  = 1,
2464         /**
2465          * Used for OI scrub
2466          */
2467         [DTO_INDEX_UPDATE]  = 16,
2468         /**
2469          * 4(inode, inode bits, groups, GDT)
2470          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2471          */
2472         [DTO_OBJECT_CREATE] = 4,
2473         /**
2474          * 4(inode, inode bits, groups, GDT)
2475          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2476          */
2477         [DTO_OBJECT_DELETE] = 4,
2478         /**
2479          * Attr set credits (inode)
2480          */
2481         [DTO_ATTR_SET_BASE] = 1,
2482         /**
2483          * Xattr set. The same as xattr of EXT3.
2484          * DATA_TRANS_BLOCKS(14)
2485          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2486          * are also counted in. Do not know why?
2487          */
2488         [DTO_XATTR_SET]     = 14,
2489         /**
2490          * credits for inode change during write.
2491          */
2492         [DTO_WRITE_BASE]    = 3,
2493         /**
2494          * credits for single block write.
2495          */
2496         [DTO_WRITE_BLOCK]   = 14,
2497         /**
2498          * Attr set credits for chown.
2499          * This is extra credits for setattr, and it is null without quota
2500          */
2501         [DTO_ATTR_SET_CHOWN] = 0
2502 };
2503
2504 static const struct dt_device_operations osd_dt_ops = {
2505         .dt_root_get       = osd_root_get,
2506         .dt_statfs         = osd_statfs,
2507         .dt_trans_create   = osd_trans_create,
2508         .dt_trans_start    = osd_trans_start,
2509         .dt_trans_stop     = osd_trans_stop,
2510         .dt_trans_cb_add   = osd_trans_cb_add,
2511         .dt_conf_get       = osd_conf_get,
2512         .dt_sync           = osd_sync,
2513         .dt_ro             = osd_ro,
2514         .dt_commit_async   = osd_commit_async,
2515 };
2516
2517 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2518                           unsigned int role)
2519 {
2520         struct osd_object *obj = osd_dt_obj(dt);
2521         struct osd_thread_info *oti = osd_oti_get(env);
2522
2523         LINVRNT(osd_invariant(obj));
2524
2525         LASSERT(obj->oo_owner != env);
2526         down_read_nested(&obj->oo_sem, role);
2527
2528         LASSERT(obj->oo_owner == NULL);
2529         oti->oti_r_locks++;
2530 }
2531
2532 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2533                            unsigned int role)
2534 {
2535         struct osd_object *obj = osd_dt_obj(dt);
2536         struct osd_thread_info *oti = osd_oti_get(env);
2537
2538         LINVRNT(osd_invariant(obj));
2539
2540         LASSERT(obj->oo_owner != env);
2541         down_write_nested(&obj->oo_sem, role);
2542
2543         LASSERT(obj->oo_owner == NULL);
2544         obj->oo_owner = env;
2545         oti->oti_w_locks++;
2546 }
2547
2548 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2549 {
2550         struct osd_object *obj = osd_dt_obj(dt);
2551         struct osd_thread_info *oti = osd_oti_get(env);
2552
2553         LINVRNT(osd_invariant(obj));
2554
2555         LASSERT(oti->oti_r_locks > 0);
2556         oti->oti_r_locks--;
2557         up_read(&obj->oo_sem);
2558 }
2559
2560 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2561 {
2562         struct osd_object *obj = osd_dt_obj(dt);
2563         struct osd_thread_info *oti = osd_oti_get(env);
2564
2565         LINVRNT(osd_invariant(obj));
2566
2567         LASSERT(obj->oo_owner == env);
2568         LASSERT(oti->oti_w_locks > 0);
2569         oti->oti_w_locks--;
2570         obj->oo_owner = NULL;
2571         up_write(&obj->oo_sem);
2572 }
2573
2574 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2575 {
2576         struct osd_object *obj = osd_dt_obj(dt);
2577
2578         LINVRNT(osd_invariant(obj));
2579
2580         return obj->oo_owner == env;
2581 }
2582
2583 static struct timespec *osd_inode_time(const struct lu_env *env,
2584                                        struct inode *inode, __u64 seconds)
2585 {
2586         struct osd_thread_info *oti = osd_oti_get(env);
2587         struct timespec *t = &oti->oti_time;
2588
2589         t->tv_sec = seconds;
2590         t->tv_nsec = 0;
2591         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2592         return t;
2593 }
2594
2595 static void osd_inode_getattr(const struct lu_env *env,
2596                               struct inode *inode, struct lu_attr *attr)
2597 {
2598         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2599                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2600                            LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2601                            LA_BLKSIZE | LA_TYPE;
2602
2603         attr->la_atime = inode->i_atime.tv_sec;
2604         attr->la_mtime = inode->i_mtime.tv_sec;
2605         attr->la_ctime = inode->i_ctime.tv_sec;
2606         attr->la_mode    = inode->i_mode;
2607         attr->la_size    = i_size_read(inode);
2608         attr->la_blocks  = inode->i_blocks;
2609         attr->la_uid     = i_uid_read(inode);
2610         attr->la_gid     = i_gid_read(inode);
2611         attr->la_projid  = i_projid_read(inode);
2612         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2613         attr->la_nlink   = inode->i_nlink;
2614         attr->la_rdev    = inode->i_rdev;
2615         attr->la_blksize = 1 << inode->i_blkbits;
2616         attr->la_blkbits = inode->i_blkbits;
2617         /*
2618          * Ext4 did not transfer inherit flags from raw inode
2619          * to inode flags, and ext4 internally test raw inode
2620          * @i_flags directly. Instead of patching ext4, we do it here.
2621          */
2622         if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2623                 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2624 }
2625
2626 static int osd_attr_get(const struct lu_env *env,
2627                         struct dt_object *dt,
2628                         struct lu_attr *attr)
2629 {
2630         struct osd_object *obj = osd_dt_obj(dt);
2631
2632         if (unlikely(!dt_object_exists(dt)))
2633                 return -ENOENT;
2634         if (unlikely(obj->oo_destroyed))
2635                 return -ENOENT;
2636
2637         LASSERT(!dt_object_remote(dt));
2638         LINVRNT(osd_invariant(obj));
2639
2640         spin_lock(&obj->oo_guard);
2641         osd_inode_getattr(env, obj->oo_inode, attr);
2642         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2643                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2644         spin_unlock(&obj->oo_guard);
2645
2646         return 0;
2647 }
2648
2649 static int osd_declare_attr_qid(const struct lu_env *env,
2650                                 struct osd_object *obj,
2651                                 struct osd_thandle *oh, long long bspace,
2652                                 qid_t old_id, qid_t new_id, bool enforce,
2653                                 unsigned int type, bool ignore_edquot)
2654 {
2655         int rc;
2656         struct osd_thread_info *info = osd_oti_get(env);
2657         struct lquota_id_info  *qi = &info->oti_qi;
2658
2659         qi->lqi_type = type;
2660         /* inode accounting */
2661         qi->lqi_is_blk = false;
2662
2663         /* one more inode for the new id ... */
2664         qi->lqi_id.qid_uid = new_id;
2665         qi->lqi_space      = 1;
2666         /* Reserve credits for the new id */
2667         rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2668         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2669                 rc = 0;
2670         if (rc)
2671                 RETURN(rc);
2672
2673         /* and one less inode for the current id */
2674         qi->lqi_id.qid_uid = old_id;
2675         qi->lqi_space = -1;
2676         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2677         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2678                 rc = 0;
2679         if (rc)
2680                 RETURN(rc);
2681
2682         /* block accounting */
2683         qi->lqi_is_blk = true;
2684
2685         /* more blocks for the new id ... */
2686         qi->lqi_id.qid_uid = new_id;
2687         qi->lqi_space      = bspace;
2688         /*
2689          * Credits for the new uid has been reserved, re-use "obj"
2690          * to save credit reservation.
2691          */
2692         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2693         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2694                 rc = 0;
2695         if (rc)
2696                 RETURN(rc);
2697
2698         /* and finally less blocks for the current uid */
2699         qi->lqi_id.qid_uid = old_id;
2700         qi->lqi_space      = -bspace;
2701         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2702         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2703                 rc = 0;
2704
2705         RETURN(rc);
2706 }
2707
2708 static int osd_declare_attr_set(const struct lu_env *env,
2709                                 struct dt_object *dt,
2710                                 const struct lu_attr *attr,
2711                                 struct thandle *handle)
2712 {
2713         struct osd_thandle *oh;
2714         struct osd_object *obj;
2715         qid_t uid;
2716         qid_t gid;
2717         long long bspace;
2718         int rc = 0;
2719         bool enforce;
2720
2721         ENTRY;
2722
2723         LASSERT(dt != NULL);
2724         LASSERT(handle != NULL);
2725
2726         obj = osd_dt_obj(dt);
2727         LASSERT(osd_invariant(obj));
2728
2729         oh = container_of0(handle, struct osd_thandle, ot_super);
2730         LASSERT(oh->ot_handle == NULL);
2731
2732         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2733                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2734
2735         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2736                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2737
2738         if (attr == NULL || obj->oo_inode == NULL)
2739                 RETURN(rc);
2740
2741         bspace   = obj->oo_inode->i_blocks << 9;
2742         bspace   = toqb(bspace);
2743
2744         /*
2745          * Changing ownership is always preformed by super user, it should not
2746          * fail with EDQUOT unless required explicitly.
2747          *
2748          * We still need to call the osd_declare_qid() to calculate the journal
2749          * credits for updating quota accounting files and to trigger quota
2750          * space adjustment once the operation is completed.
2751          */
2752         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2753                 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2754
2755                 if (!ignore_edquot)
2756                         CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2757                                "(the quota space is %lld)\n",
2758                                obj->oo_inode->i_sb->s_id, attr->la_uid,
2759                                attr->la_gid, bspace);
2760
2761                 /* USERQUOTA */
2762                 uid = i_uid_read(obj->oo_inode);
2763                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2764                 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2765                                           attr->la_uid, enforce, USRQUOTA,
2766                                           true);
2767                 if (rc)
2768                         RETURN(rc);
2769
2770                 gid = i_gid_read(obj->oo_inode);
2771                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2772                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2773                                           i_gid_read(obj->oo_inode),
2774                                           attr->la_gid, enforce, GRPQUOTA,
2775                                           ignore_edquot);
2776                 if (rc)
2777                         RETURN(rc);
2778
2779         }
2780 #ifdef HAVE_PROJECT_QUOTA
2781         if (attr->la_valid & LA_PROJID) {
2782                 __u32 projid = i_projid_read(obj->oo_inode);
2783
2784                 enforce = (attr->la_valid & LA_PROJID) &&
2785                                         (attr->la_projid != projid);
2786                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2787                                           (qid_t)projid, (qid_t)attr->la_projid,
2788                                           enforce, PRJQUOTA, true);
2789                 if (rc)
2790                         RETURN(rc);
2791         }
2792 #endif
2793         RETURN(rc);
2794 }
2795
2796 static int osd_inode_setattr(const struct lu_env *env,
2797                              struct inode *inode, const struct lu_attr *attr)
2798 {
2799         __u64 bits = attr->la_valid;
2800
2801         /* Only allow set size for regular file */
2802         if (!S_ISREG(inode->i_mode))
2803                 bits &= ~(LA_SIZE | LA_BLOCKS);
2804
2805         if (bits == 0)
2806                 return 0;
2807
2808         if (bits & LA_ATIME)
2809                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2810         if (bits & LA_CTIME)
2811                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2812         if (bits & LA_MTIME)
2813                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2814         if (bits & LA_SIZE) {
2815                 spin_lock(&inode->i_lock);
2816                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2817                 i_size_write(inode, attr->la_size);
2818                 spin_unlock(&inode->i_lock);
2819         }
2820
2821         /*
2822          * OSD should not change "i_blocks" which is used by quota.
2823          * "i_blocks" should be changed by ldiskfs only.
2824          */
2825         if (bits & LA_MODE)
2826                 inode->i_mode = (inode->i_mode & S_IFMT) |
2827                                 (attr->la_mode & ~S_IFMT);
2828         if (bits & LA_UID)
2829                 i_uid_write(inode, attr->la_uid);
2830         if (bits & LA_GID)
2831                 i_gid_write(inode, attr->la_gid);
2832         if (bits & LA_PROJID)
2833                 i_projid_write(inode, attr->la_projid);
2834         if (bits & LA_NLINK)
2835                 set_nlink(inode, attr->la_nlink);
2836         if (bits & LA_RDEV)
2837                 inode->i_rdev = attr->la_rdev;
2838
2839         if (bits & LA_FLAGS) {
2840                 /* always keep S_NOCMTIME */
2841                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2842                                  S_NOCMTIME;
2843                 /*
2844                  * Ext4 did not transfer inherit flags from
2845                  * @inode->i_flags to raw inode i_flags when writing
2846                  * flags, we do it explictly here.
2847                  */
2848                 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2849                         LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2850                 else
2851                         LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2852         }
2853         return 0;
2854 }
2855
2856 #ifdef HAVE_PROJECT_QUOTA
2857 static int osd_transfer_project(struct inode *inode, __u32 projid)
2858 {
2859         struct super_block *sb = inode->i_sb;
2860         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2861         int err;
2862         kprojid_t kprojid;
2863         struct ldiskfs_iloc iloc;
2864         struct ldiskfs_inode *raw_inode;
2865         struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2866
2867         if (!ldiskfs_has_feature_project(sb)) {
2868                 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2869                         == LDISKFS_DEF_PROJID);
2870                 if (projid != LDISKFS_DEF_PROJID)
2871                         return -EOPNOTSUPP;
2872                 else
2873                         return 0;
2874         }
2875
2876         if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2877                 return -EOPNOTSUPP;
2878
2879         kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2880         if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2881                 return 0;
2882
2883         err = ldiskfs_get_inode_loc(inode, &iloc);
2884         if (err)
2885                 return err;
2886
2887         raw_inode = ldiskfs_raw_inode(&iloc);
2888         if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2889                 err = -EOVERFLOW;
2890                 brelse(iloc.bh);
2891                 return err;
2892         }
2893         brelse(iloc.bh);
2894
2895         dquot_initialize(inode);
2896         transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2897         if (transfer_to[PRJQUOTA]) {
2898                 err = __dquot_transfer(inode, transfer_to);
2899                 dqput(transfer_to[PRJQUOTA]);
2900                 if (err)
2901                         return err;
2902         }
2903
2904         return err;
2905 }
2906 #endif
2907
2908 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2909 {
2910         int rc;
2911
2912         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2913             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2914                 struct iattr iattr;
2915
2916                 ll_vfs_dq_init(inode);
2917                 iattr.ia_valid = 0;
2918                 if (attr->la_valid & LA_UID)
2919                         iattr.ia_valid |= ATTR_UID;
2920                 if (attr->la_valid & LA_GID)
2921                         iattr.ia_valid |= ATTR_GID;
2922                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2923                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2924
2925                 rc = ll_vfs_dq_transfer(inode, &iattr);
2926                 if (rc) {
2927                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2928                                "enforcement enabled on the ldiskfs "
2929                                "filesystem?\n", inode->i_sb->s_id, rc);
2930                         return rc;
2931                 }
2932         }
2933
2934         /* Handle project id transfer here properly */
2935         if (attr->la_valid & LA_PROJID &&
2936             attr->la_projid != i_projid_read(inode)) {
2937 #ifdef HAVE_PROJECT_QUOTA
2938                 rc = osd_transfer_project(inode, attr->la_projid);
2939 #else
2940                 rc = -ENOTSUPP;
2941 #endif
2942                 if (rc) {
2943                         CERROR("%s: quota transfer failed: rc = %d. Is project "
2944                                "enforcement enabled on the ldiskfs "
2945                                "filesystem?\n", inode->i_sb->s_id, rc);
2946                         return rc;
2947                 }
2948         }
2949         return 0;
2950 }
2951
2952 static int osd_attr_set(const struct lu_env *env,
2953                         struct dt_object *dt,
2954                         const struct lu_attr *attr,
2955                         struct thandle *handle)
2956 {
2957         struct osd_object *obj = osd_dt_obj(dt);
2958         struct inode *inode;
2959         int rc;
2960
2961         if (!dt_object_exists(dt))
2962                 return -ENOENT;
2963
2964         LASSERT(handle != NULL);
2965         LASSERT(!dt_object_remote(dt));
2966         LASSERT(osd_invariant(obj));
2967
2968         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2969
2970         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
2971             !osd_obj2dev(obj)->od_is_ost) {
2972                 struct osd_thread_info *oti = osd_oti_get(env);
2973                 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2974                 struct lu_fid *fid1 = &oti->oti_fid;
2975                 struct osd_inode_id *id = &oti->oti_id;
2976                 struct iam_path_descr *ipd;
2977                 struct iam_container *bag;
2978                 struct osd_thandle *oh;
2979                 int rc;
2980
2981                 fid_cpu_to_be(fid1, fid0);
2982                 memset(id, 1, sizeof(*id));
2983                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2984                                   fid0)->oi_dir.od_container;
2985                 ipd = osd_idx_ipd_get(env, bag);
2986                 if (unlikely(ipd == NULL))
2987                         RETURN(-ENOMEM);
2988
2989                 oh = container_of0(handle, struct osd_thandle, ot_super);
2990                 rc = iam_update(oh->ot_handle, bag,
2991                                 (const struct iam_key *)fid1,
2992                                 (const struct iam_rec *)id, ipd);
2993                 osd_ipd_put(env, bag, ipd);
2994                 return(rc > 0 ? 0 : rc);
2995         }
2996
2997         inode = obj->oo_inode;
2998
2999         rc = osd_quota_transfer(inode, attr);
3000         if (rc)
3001                 return rc;
3002
3003         spin_lock(&obj->oo_guard);
3004         rc = osd_inode_setattr(env, inode, attr);
3005         spin_unlock(&obj->oo_guard);
3006         if (rc != 0)
3007                 GOTO(out, rc);
3008
3009         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3010
3011         if (!(attr->la_valid & LA_FLAGS))
3012                 GOTO(out, rc);
3013
3014         /* Let's check if there are extra flags need to be set into LMA */
3015         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3016                 struct osd_thread_info *info = osd_oti_get(env);
3017                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3018
3019                 LASSERT(!obj->oo_pfid_in_lma);
3020
3021                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3022                                  &info->oti_ost_attrs);
3023                 if (rc)
3024                         GOTO(out, rc);
3025
3026                 lma->lma_incompat |=
3027                         lustre_to_lma_flags(attr->la_flags);
3028                 lustre_lma_swab(lma);
3029                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3030                                      lma, sizeof(*lma), XATTR_REPLACE);
3031                 if (rc != 0) {
3032                         struct osd_device *osd = osd_obj2dev(obj);
3033
3034                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3035                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3036                               lma->lma_incompat, rc);
3037                 } else {
3038                         obj->oo_lma_flags =
3039                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3040                 }
3041                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3042         }
3043 out:
3044         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3045
3046         return rc;
3047 }
3048
3049 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3050                                            struct osd_object *obj,
3051                                            const char *name, const int namelen)
3052 {
3053         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3054 }
3055
3056 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3057                       umode_t mode, struct dt_allocation_hint *hint,
3058                       struct thandle *th)
3059 {
3060         int result;
3061         struct osd_device *osd = osd_obj2dev(obj);
3062         struct osd_thandle *oth;
3063         struct dt_object *parent = NULL;
3064         struct inode *inode;
3065
3066         LINVRNT(osd_invariant(obj));
3067         LASSERT(obj->oo_inode == NULL);
3068         LASSERT(obj->oo_hl_head == NULL);
3069
3070         if (S_ISDIR(mode) && ldiskfs_pdo) {
3071                 obj->oo_hl_head =
3072                         ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3073                 if (obj->oo_hl_head == NULL)
3074                         return -ENOMEM;
3075         }
3076
3077         oth = container_of(th, struct osd_thandle, ot_super);
3078         LASSERT(oth->ot_handle->h_transaction != NULL);
3079
3080         if (hint != NULL && hint->dah_parent != NULL &&
3081             !dt_object_remote(hint->dah_parent))
3082                 parent = hint->dah_parent;
3083
3084         inode = ldiskfs_create_inode(oth->ot_handle,
3085                                      parent ? osd_dt_obj(parent)->oo_inode :
3086                                               osd_sb(osd)->s_root->d_inode,
3087                                      mode);
3088         if (!IS_ERR(inode)) {
3089                 /* Do not update file c/mtime in ldiskfs. */
3090                 inode->i_flags |= S_NOCMTIME;
3091
3092                 /*
3093                  * For new created object, it must be consistent,
3094                  * and it is unnecessary to scrub against it.
3095                  */
3096                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3097
3098                 obj->oo_inode = inode;
3099                 result = 0;
3100         } else {
3101                 if (obj->oo_hl_head != NULL) {
3102                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3103                         obj->oo_hl_head = NULL;
3104                 }
3105                 result = PTR_ERR(inode);
3106         }
3107         LINVRNT(osd_invariant(obj));
3108         return result;
3109 }
3110
3111 enum {
3112         OSD_NAME_LEN = 255
3113 };
3114
3115 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3116                      struct lu_attr *attr,
3117                      struct dt_allocation_hint *hint,
3118                      struct dt_object_format *dof,
3119                      struct thandle *th)
3120 {
3121         int result;
3122         struct osd_thandle *oth;
3123         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3124
3125         LASSERT(S_ISDIR(attr->la_mode));
3126
3127         oth = container_of(th, struct osd_thandle, ot_super);
3128         LASSERT(oth->ot_handle->h_transaction != NULL);
3129         result = osd_mkfile(info, obj, mode, hint, th);
3130
3131         return result;
3132 }
3133
3134 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3135                         struct lu_attr *attr,
3136                         struct dt_allocation_hint *hint,
3137                         struct dt_object_format *dof,
3138                         struct thandle *th)
3139 {
3140         int result;
3141         struct osd_thandle *oth;
3142         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3143
3144         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3145
3146         LASSERT(S_ISREG(attr->la_mode));
3147
3148         oth = container_of(th, struct osd_thandle, ot_super);
3149         LASSERT(oth->ot_handle->h_transaction != NULL);
3150
3151         result = osd_mkfile(info, obj, mode, hint, th);
3152         if (result == 0) {
3153                 LASSERT(obj->oo_inode != NULL);
3154                 if (feat->dif_flags & DT_IND_VARKEY)
3155                         result = iam_lvar_create(obj->oo_inode,
3156                                                  feat->dif_keysize_max,
3157                                                  feat->dif_ptrsize,
3158                                                  feat->dif_recsize_max,
3159                                                  oth->ot_handle);
3160                 else
3161                         result = iam_lfix_create(obj->oo_inode,
3162                                                  feat->dif_keysize_max,
3163                                                  feat->dif_ptrsize,
3164                                                  feat->dif_recsize_max,
3165                                                  oth->ot_handle);
3166         }
3167         return result;
3168 }
3169
3170 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3171                      struct lu_attr *attr,
3172                      struct dt_allocation_hint *hint,
3173                      struct dt_object_format *dof,
3174                      struct thandle *th)
3175 {
3176         LASSERT(S_ISREG(attr->la_mode));
3177         return osd_mkfile(info, obj, (attr->la_mode &
3178                          (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
3179 }
3180