Whamcloud - gitweb
LU-12137 llite: use ->iterate_shared() for readdir
[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 = osd_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                         /* below message is checked in sanity.sh test_129 */
499                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s maximum entry limit\n",
500                               osd_name(osd), parent->i_ino, PFID(fid),
501                               rc == -ENOSPC ? "has reached" : "is approaching");
502                 else
503                         /* below message is checked in sanity.sh test_129 */
504                         CWARN("%s: directory (inode: %lu, FID: unknown) %s 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 = osd_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_18_23, 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 *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(struct filter_fid_18_23)) {
722                 struct filter_fid_18_23 *ff_old = (void *)ff;
723
724                 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
725                 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
726                 /*
727                  * XXX: use 0 as the index for compatibility, the caller will
728                  * handle index related issues when necessary.
729                  */
730                 if (!rc)
731                         ostid_to_fid(fid, ostid, 0);
732         } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
733                 rc = 1;
734         } else if (rc >= 0) {
735                 rc = -EINVAL;
736         }
737
738         return rc;
739 }
740
741 static int osd_lma_self_repair(struct osd_thread_info *info,
742                                struct osd_device *osd, struct inode *inode,
743                                const struct lu_fid *fid, __u32 compat)
744 {
745         handle_t *jh;
746         int rc;
747
748         LASSERT(current->journal_info == NULL);
749
750         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
751                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
752         if (IS_ERR(jh)) {
753                 rc = PTR_ERR(jh);
754                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
755                       osd_name(osd), rc);
756                 return rc;
757         }
758
759         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
760         if (rc != 0)
761                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
762                       osd_name(osd), rc);
763         ldiskfs_journal_stop(jh);
764         return rc;
765 }
766
767 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
768 {
769         struct osd_thread_info *info = osd_oti_get(env);
770         struct osd_device *osd = osd_obj2dev(obj);
771         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
772         struct lustre_mdt_attrs *lma = &loa->loa_lma;
773         struct inode *inode = obj->oo_inode;
774         struct dentry *dentry = &info->oti_obj_dentry;
775         struct lu_fid *fid = NULL;
776         const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
777         int rc;
778
779         ENTRY;
780
781         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
782                              (void *)loa, sizeof(*loa));
783         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
784                 fid = &lma->lma_self_fid;
785                 rc = osd_get_idif(info, inode, dentry, fid);
786                 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
787                         /*
788                          * For the given OST-object, if it has neither LMA nor
789                          * FID in XATTR_NAME_FID, then the given FID (which is
790                          * contained in the @obj, from client RPC for locating
791                          * the OST-object) is trusted. We use it to generate
792                          * the LMA.
793                          */
794                         osd_lma_self_repair(info, osd, inode, rfid,
795                                             LMAC_FID_ON_OST);
796                         RETURN(0);
797                 }
798         }
799
800         if (rc < 0)
801                 RETURN(rc);
802
803         if (rc > 0) {
804                 rc = 0;
805                 lustre_lma_swab(lma);
806                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
807                              (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
808                               S_ISREG(inode->i_mode)))) {
809                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
810                               "fid = "DFID", ino = %lu\n", osd_name(osd),
811                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
812                               PFID(rfid), inode->i_ino);
813                         rc = -EOPNOTSUPP;
814                 } else {
815                         fid = &lma->lma_self_fid;
816                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
817                             osd->od_is_ost)
818                                 obj->oo_pfid_in_lma = 1;
819                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
820                             !osd->od_is_ost)
821                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
822                 }
823         }
824
825         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
826                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
827                         struct ost_id   *oi   = &info->oti_ostid;
828                         struct lu_fid   *fid1 = &info->oti_fid3;
829                         __u32            idx  = fid_idif_ost_idx(rfid);
830
831                         /*
832                          * For old IDIF, the OST index is not part of the IDIF,
833                          * Means that different OSTs may have the same IDIFs.
834                          * Under such case, we need to make some compatible
835                          * check to make sure to trigger OI scrub properly.
836                          */
837                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
838                                 /* Given @rfid is new, LMA is old. */
839                                 fid_to_ostid(fid, oi);
840                                 ostid_to_fid(fid1, oi, idx);
841                                 if (lu_fid_eq(fid1, rfid)) {
842                                         if (osd->od_index_in_idif)
843                                                 osd_lma_self_repair(info, osd,
844                                                         inode, rfid,
845                                                         LMAC_FID_ON_OST);
846                                         RETURN(0);
847                                 }
848                         }
849                 }
850
851                 rc = -EREMCHG;
852         }
853
854         RETURN(rc);
855 }
856
857 struct osd_check_lmv_buf {
858 #ifdef HAVE_DIR_CONTEXT
859         /* please keep it as first member */
860         struct dir_context ctx;
861 #endif
862         struct osd_thread_info *oclb_info;
863         struct osd_device *oclb_dev;
864         struct osd_idmap_cache *oclb_oic;
865         int oclb_items;
866         bool oclb_found;
867 };
868
869 /**
870  * It is called internally by ->readdir() to filter out the
871  * local slave object's FID of the striped directory.
872  *
873  * \retval      1 found the local slave's FID
874  * \retval      0 continue to check next item
875  * \retval      -ve for failure
876  */
877 #ifdef HAVE_FILLDIR_USE_CTX
878 static int osd_stripe_dir_filldir(struct dir_context *buf,
879 #else
880 static int osd_stripe_dir_filldir(void *buf,
881 #endif
882                                   const char *name, int namelen,
883                                   loff_t offset, __u64 ino, unsigned int d_type)
884 {
885         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
886         struct osd_thread_info *oti = oclb->oclb_info;
887         struct lu_fid *fid = &oti->oti_fid3;
888         struct osd_inode_id *id = &oti->oti_id3;
889         struct osd_device *dev = oclb->oclb_dev;
890         struct osd_idmap_cache *oic = oclb->oclb_oic;
891         struct inode *inode;
892
893         oclb->oclb_items++;
894
895         if (name[0] == '.')
896                 return 0;
897
898         fid_zero(fid);
899         sscanf(name + 1, SFID, RFID(fid));
900         if (!fid_is_sane(fid))
901                 return 0;
902
903         if (osd_remote_fid(oti->oti_env, dev, fid))
904                 return 0;
905
906         osd_id_gen(id, ino, OSD_OII_NOGEN);
907         inode = osd_iget(oti, dev, id);
908         if (IS_ERR(inode))
909                 return PTR_ERR(inode);
910
911         iput(inode);
912         osd_add_oi_cache(oti, dev, id, fid);
913         oic->oic_fid = *fid;
914         oic->oic_lid = *id;
915         oic->oic_dev = dev;
916         osd_oii_insert(dev, oic, true);
917         oclb->oclb_found = true;
918
919         return 1;
920 }
921
922 /*
923  * When lookup item under striped directory, we need to locate the master
924  * MDT-object of the striped directory firstly, then the client will send
925  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
926  * and the item's name. If the system is restored from MDT file level backup,
927  * then before the OI scrub completely built the OI files, the OI mappings of
928  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
929  * not a problem for the master MDT-object. Because when locate the master
930  * MDT-object, we will do name based lookup (for the striped directory itself)
931  * firstly, during such process we can setup the correct OI mapping for the
932  * master MDT-object. But it will be trouble for the slave MDT-object. Because
933  * the client will not trigger name based lookup on the MDT to locate the slave
934  * MDT-object before locating item under the striped directory, then when
935  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
936  * is invalid and does not know what the right OI mapping is, then the MDT has
937  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
938  * the OI file, related OI mapping is unknown yet, please try again later. And
939  * then client will re-try the RPC again and again until related OI mapping has
940  * been updated. That is quite inefficient.
941  *
942  * To resolve above trouble, we will handle it as the following two cases:
943  *
944  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
945  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
946  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
947  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
948  *    directly. Please check osd_fid_lookup().
949  *
950  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
951  *    Under such case, during lookup the master MDT-object, we will lookup the
952  *    slave MDT-object via readdir against the master MDT-object, because the
953  *    slave MDT-objects information are stored as sub-directories with the name
954  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
955  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
956  *    the correct OI mapping for the slave MDT-object.
957  */
958 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
959                          struct inode *inode, struct osd_idmap_cache *oic)
960 {
961         struct lu_buf *buf = &oti->oti_big_buf;
962         struct dentry *dentry = &oti->oti_obj_dentry;
963         struct file *filp = &oti->oti_file;
964         const struct file_operations *fops;
965         struct lmv_mds_md_v1 *lmv1;
966         struct osd_check_lmv_buf oclb = {
967 #ifdef HAVE_DIR_CONTEXT
968                 .ctx.actor = osd_stripe_dir_filldir,
969 #endif
970                 .oclb_info = oti,
971                 .oclb_dev = dev,
972                 .oclb_oic = oic,
973                 .oclb_found = false,
974         };
975         int rc = 0;
976
977         ENTRY;
978
979 again:
980         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
981                              buf->lb_len);
982         if (rc == -ERANGE) {
983                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
984                 if (rc > 0) {
985                         lu_buf_realloc(buf, rc);
986                         if (buf->lb_buf == NULL)
987                                 GOTO(out, rc = -ENOMEM);
988
989                         goto again;
990                 }
991         }
992
993         if (unlikely(rc == 0 || rc == -ENODATA))
994                 GOTO(out, rc = 0);
995
996         if (rc < 0)
997                 GOTO(out, rc);
998
999         if (unlikely(buf->lb_buf == NULL)) {
1000                 lu_buf_realloc(buf, rc);
1001                 if (buf->lb_buf == NULL)
1002                         GOTO(out, rc = -ENOMEM);
1003
1004                 goto again;
1005         }
1006
1007         lmv1 = buf->lb_buf;
1008         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1009                 GOTO(out, rc = 0);
1010
1011         fops = inode->i_fop;
1012         dentry->d_inode = inode;
1013         dentry->d_sb = inode->i_sb;
1014         filp->f_pos = 0;
1015         filp->f_path.dentry = dentry;
1016         filp->f_mode = FMODE_64BITHASH;
1017         filp->f_mapping = inode->i_mapping;
1018         filp->f_op = fops;
1019         filp->private_data = NULL;
1020         set_file_inode(filp, inode);
1021
1022         do {
1023                 oclb.oclb_items = 0;
1024 #ifdef HAVE_DIR_CONTEXT
1025                 oclb.ctx.pos = filp->f_pos;
1026                 rc = fops->iterate_shared(filp, &oclb.ctx);
1027                 filp->f_pos = oclb.ctx.pos;
1028 #else
1029                 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
1030 #endif
1031         } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1032                  filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1033         fops->release(inode, filp);
1034
1035 out:
1036         if (rc < 0)
1037                 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
1038                        DFID": rc = %d\n", osd_ino2name(inode),
1039                        inode->i_ino, inode->i_generation,
1040                        PFID(&oic->oic_fid), rc);
1041         else
1042                 rc = 0;
1043
1044         RETURN(rc);
1045 }
1046
1047 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1048                           const struct lu_fid *fid,
1049                           const struct lu_object_conf *conf)
1050 {
1051         struct osd_thread_info *info;
1052         struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1053         struct osd_device *dev;
1054         struct osd_idmap_cache *oic;
1055         struct osd_inode_id *id;
1056         struct inode *inode = NULL;
1057         struct lustre_scrub *scrub;
1058         struct scrub_file *sf;
1059         __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1060         __u32 saved_ino;
1061         __u32 saved_gen;
1062         int result = 0;
1063         int rc1 = 0;
1064         bool remote = false;
1065         bool trusted = true;
1066         bool updated = false;
1067         bool checked = false;
1068
1069         ENTRY;
1070
1071         LINVRNT(osd_invariant(obj));
1072         LASSERT(obj->oo_inode == NULL);
1073         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1074
1075         dev = osd_dev(ldev);
1076         scrub = &dev->od_scrub.os_scrub;
1077         sf = &scrub->os_file;
1078         info = osd_oti_get(env);
1079         LASSERT(info);
1080         oic = &info->oti_cache;
1081
1082         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1083                 RETURN(-ENOENT);
1084
1085         /*
1086          * For the object is created as locking anchor, or for the object to
1087          * be created on disk. No need to osd_oi_lookup() at here because FID
1088          * shouldn't never be re-used, if it's really a duplicate FID from
1089          * unexpected reason, we should be able to detect it later by calling
1090          * do_create->osd_oi_insert().
1091          */
1092         if (conf && conf->loc_flags & LOC_F_NEW)
1093                 GOTO(out, result = 0);
1094
1095         /* Search order: 1. per-thread cache. */
1096         if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1097                 id = &oic->oic_lid;
1098                 goto iget;
1099         }
1100
1101         id = &info->oti_id;
1102         if (!list_empty(&scrub->os_inconsistent_items)) {
1103                 /* Search order: 2. OI scrub pending list. */
1104                 result = osd_oii_lookup(dev, fid, id);
1105                 if (!result)
1106                         goto iget;
1107         }
1108
1109         /*
1110          * The OI mapping in the OI file can be updated by the OI scrub
1111          * when we locate the inode via FID. So it may be not trustable.
1112          */
1113         trusted = false;
1114
1115         /* Search order: 3. OI files. */
1116         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1117         if (result == -ENOENT) {
1118                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1119                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1120                     !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1121                                       sf->sf_oi_bitmap))
1122                         GOTO(out, result = 0);
1123
1124                 goto trigger;
1125         }
1126
1127         /* -ESTALE is returned if inode of OST object doesn't exist */
1128         if (result == -ESTALE &&
1129             fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1130                 GOTO(out, result = 0);
1131         }
1132
1133         if (result)
1134                 GOTO(out, result);
1135
1136 iget:
1137         obj->oo_inode = NULL;
1138         /* for later passes through checks, not true on first pass */
1139         if (!IS_ERR_OR_NULL(inode))
1140                 iput(inode);
1141
1142         inode = osd_iget_check(info, dev, fid, id, trusted);
1143         if (!IS_ERR(inode)) {
1144                 obj->oo_inode = inode;
1145                 result = 0;
1146                 if (remote)
1147                         goto trigger;
1148
1149                 goto check_lma;
1150         }
1151
1152         result = PTR_ERR(inode);
1153         if (result == -ENOENT || result == -ESTALE)
1154                 GOTO(out, result = 0);
1155
1156         if (result != -EREMCHG)
1157                 GOTO(out, result);
1158
1159 trigger:
1160         /*
1161          * We still have chance to get the valid inode: for the
1162          * object which is referenced by remote name entry, the
1163          * object on the local MDT will be linked under the dir
1164          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1165          *
1166          * We do not know whether the object for the given FID
1167          * is referenced by some remote name entry or not, and
1168          * especially for DNE II, a multiple-linked object may
1169          * have many name entries reside on many MDTs.
1170          *
1171          * To simplify the operation, OSD will not distinguish
1172          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1173          * only happened for the RPC from other MDT during the
1174          * OI scrub, or for the client side RPC with FID only,
1175          * such as FID to path, or from old connected client.
1176          */
1177         if (!remote) {
1178                 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1179                 if (!rc1) {
1180                         remote = true;
1181                         trusted = true;
1182                         flags |= SS_AUTO_PARTIAL;
1183                         flags &= ~SS_AUTO_FULL;
1184                         goto iget;
1185                 }
1186         }
1187
1188         if (thread_is_running(&scrub->os_thread)) {
1189                 if (scrub->os_partial_scan && !scrub->os_in_join)
1190                         goto join;
1191
1192                 if (IS_ERR_OR_NULL(inode) || result)
1193                         GOTO(out, result = -EINPROGRESS);
1194
1195                 LASSERT(remote);
1196                 LASSERT(obj->oo_inode == inode);
1197
1198                 osd_add_oi_cache(info, dev, id, fid);
1199                 osd_oii_insert(dev, oic, true);
1200                 goto found;
1201         }
1202
1203         if (dev->od_auto_scrub_interval == AS_NEVER) {
1204                 if (!remote)
1205                         GOTO(out, result = -EREMCHG);
1206
1207                 LASSERT(!result);
1208                 LASSERT(obj->oo_inode == inode);
1209
1210                 osd_add_oi_cache(info, dev, id, fid);
1211                 goto found;
1212         }
1213
1214 join:
1215         rc1 = osd_scrub_start(env, dev, flags);
1216         LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1217                       "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1218         if (rc1 && rc1 != -EALREADY)
1219                 GOTO(out, result = -EREMCHG);
1220
1221         if (IS_ERR_OR_NULL(inode) || result)
1222                 GOTO(out, result = -EINPROGRESS);
1223
1224         LASSERT(remote);
1225         LASSERT(obj->oo_inode == inode);
1226
1227         osd_add_oi_cache(info, dev, id, fid);
1228         osd_oii_insert(dev, oic, true);
1229         goto found;
1230
1231 check_lma:
1232         checked = true;
1233         if (unlikely(obj->oo_header))
1234                 goto found;
1235
1236         result = osd_check_lma(env, obj);
1237         if (!result)
1238                 goto found;
1239
1240         LASSERTF(id->oii_ino == inode->i_ino &&
1241                  id->oii_gen == inode->i_generation,
1242                  "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1243                  PFID(fid), id->oii_ino, id->oii_gen,
1244                  inode->i_ino, inode->i_generation);
1245
1246         saved_ino = inode->i_ino;
1247         saved_gen = inode->i_generation;
1248
1249         if (unlikely(result == -ENODATA)) {
1250                 /*
1251                  * If the OI scrub updated the OI mapping by race, it
1252                  * must be valid. Trust the inode that has no LMA EA.
1253                  */
1254                 if (updated)
1255                         goto found;
1256
1257                 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1258                 if (!result) {
1259                         /*
1260                          * The OI mapping is still there, the inode is still
1261                          * valid. It is just becaues the inode has no LMA EA.
1262                          */
1263                         if (saved_ino == id->oii_ino &&
1264                             saved_gen == id->oii_gen)
1265                                 goto found;
1266
1267                         /*
1268                          * It is the OI scrub updated the OI mapping by race.
1269                          * The new OI mapping must be valid.
1270                          */
1271                         trusted = true;
1272                         updated = true;
1273                         goto iget;
1274                 }
1275
1276                 /*
1277                  * "result == -ENOENT" means that the OI mappinghas been
1278                  * removed by race, so the inode belongs to other object.
1279                  *
1280                  * Others error can be returned  directly.
1281                  */
1282                 if (result == -ENOENT) {
1283                         LASSERT(trusted);
1284
1285                         obj->oo_inode = NULL;
1286                         result = 0;
1287                 }
1288         }
1289
1290         if (result != -EREMCHG)
1291                 GOTO(out, result);
1292
1293         LASSERT(!updated);
1294
1295         /*
1296          * if two OST objects map to the same inode, and inode mode is
1297          * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1298          * reserved by precreate, and not written yet, in this case, don't
1299          * set inode for the object whose FID mismatch, so that it can create
1300          * inode and not block precreate.
1301          */
1302         if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1303             inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1304                 obj->oo_inode = NULL;
1305                 GOTO(out, result = 0);
1306         }
1307
1308         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1309         /*
1310          * "result == -ENOENT" means the cached OI mapping has been removed
1311          * from the OI file by race, above inode belongs to other object.
1312          */
1313         if (result == -ENOENT) {
1314                 LASSERT(trusted);
1315
1316                 obj->oo_inode = NULL;
1317                 GOTO(out, result = 0);
1318         }
1319
1320         if (result)
1321                 GOTO(out, result);
1322
1323         if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1324                 result = -EREMCHG;
1325                 goto trigger;
1326         }
1327
1328         /*
1329          * It is the OI scrub updated the OI mapping by race.
1330          * The new OI mapping must be valid.
1331          */
1332         trusted = true;
1333         updated = true;
1334         goto iget;
1335
1336 found:
1337         if (!checked) {
1338                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1339                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1340
1341                 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1342                 if (!result) {
1343                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
1344                             dev->od_is_ost)
1345                                 obj->oo_pfid_in_lma = 1;
1346                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1347                             !dev->od_is_ost)
1348                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1349                 } else if (result != -ENODATA) {
1350                         GOTO(out, result);
1351                 }
1352         }
1353
1354         obj->oo_compat_dot_created = 1;
1355         obj->oo_compat_dotdot_created = 1;
1356
1357         if (S_ISDIR(inode->i_mode) &&
1358             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1359                 osd_check_lmv(info, dev, inode, oic);
1360
1361         result = osd_attach_jinode(inode);
1362         if (result)
1363                 GOTO(out, result);
1364
1365         if (!ldiskfs_pdo)
1366                 GOTO(out, result = 0);
1367
1368         LASSERT(!obj->oo_hl_head);
1369         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1370
1371         GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1372
1373 out:
1374         if (result || !obj->oo_inode) {
1375                 if (!IS_ERR_OR_NULL(inode))
1376                         iput(inode);
1377
1378                 obj->oo_inode = NULL;
1379                 if (trusted)
1380                         fid_zero(&oic->oic_fid);
1381         }
1382
1383         LINVRNT(osd_invariant(obj));
1384         return result;
1385 }
1386
1387 /*
1388  * Concurrency: shouldn't matter.
1389  */
1390 static void osd_object_init0(struct osd_object *obj)
1391 {
1392         LASSERT(obj->oo_inode != NULL);
1393         obj->oo_dt.do_body_ops = &osd_body_ops;
1394         obj->oo_dt.do_lu.lo_header->loh_attr |=
1395                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1396 }
1397
1398 /*
1399  * Concurrency: no concurrent access is possible that early in object
1400  * life-cycle.
1401  */
1402 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1403                            const struct lu_object_conf *conf)
1404 {
1405         struct osd_object *obj = osd_obj(l);
1406         int result;
1407
1408         LINVRNT(osd_invariant(obj));
1409
1410         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1411                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1412                 l->lo_header->loh_attr |= LOHA_EXISTS;
1413                 return 0;
1414         }
1415
1416         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1417         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1418         if (result == 0 && obj->oo_inode != NULL) {
1419                 struct osd_thread_info *oti = osd_oti_get(env);
1420                 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1421
1422                 osd_object_init0(obj);
1423                 if (unlikely(obj->oo_header))
1424                         return 0;
1425
1426                 result = osd_get_lma(oti, obj->oo_inode,
1427                                      &oti->oti_obj_dentry, loa);
1428                 if (!result) {
1429                         /*
1430                          * Convert LMAI flags to lustre LMA flags
1431                          * and cache it to oo_lma_flags
1432                          */
1433                         obj->oo_lma_flags =
1434                                 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1435                 } else if (result == -ENODATA) {
1436                         result = 0;
1437                 }
1438         }
1439
1440         LINVRNT(osd_invariant(obj));
1441         return result;
1442 }
1443
1444 /*
1445  * The first part of oxe_buf is xattr name, and is '\0' terminated.
1446  * The left part is for value, binary mode.
1447  */
1448 struct osd_xattr_entry {
1449         struct list_head        oxe_list;
1450         size_t                  oxe_len;
1451         size_t                  oxe_namelen;
1452         bool                    oxe_exist;
1453         struct rcu_head         oxe_rcu;
1454         char                    oxe_buf[0];
1455 };
1456
1457 static int osd_oxc_get(struct osd_object *obj, const char *name,
1458                        struct lu_buf *buf)
1459 {
1460         struct osd_xattr_entry *tmp;
1461         struct osd_xattr_entry *oxe = NULL;
1462         size_t namelen = strlen(name);
1463         int rc;
1464
1465         ENTRY;
1466
1467         rcu_read_lock();
1468         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1469                 if (namelen == tmp->oxe_namelen &&
1470                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1471                         oxe = tmp;
1472                         break;
1473                 }
1474         }
1475
1476         if (oxe == NULL)
1477                 GOTO(out, rc = -ENOENT);
1478
1479         if (!oxe->oxe_exist)
1480                 GOTO(out, rc = -ENODATA);
1481
1482         /* vallen */
1483         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1484         LASSERT(rc > 0);
1485
1486         if (buf->lb_buf == NULL)
1487                 GOTO(out, rc);
1488
1489         if (buf->lb_len < rc)
1490                 GOTO(out, rc = -ERANGE);
1491
1492         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1493         EXIT;
1494 out:
1495         rcu_read_unlock();
1496
1497         return rc;
1498 }
1499
1500 static void osd_oxc_free(struct rcu_head *head)
1501 {
1502         struct osd_xattr_entry *oxe;
1503
1504         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1505         OBD_FREE(oxe, oxe->oxe_len);
1506 }
1507
1508 static void osd_oxc_add(struct osd_object *obj, const char *name,
1509                         const char *buf, int buflen)
1510 {
1511         struct osd_xattr_entry *oxe;
1512         struct osd_xattr_entry *old = NULL;
1513         struct osd_xattr_entry *tmp;
1514         size_t namelen = strlen(name);
1515         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1516
1517         OBD_ALLOC(oxe, len);
1518         if (oxe == NULL)
1519                 return;
1520
1521         INIT_LIST_HEAD(&oxe->oxe_list);
1522         oxe->oxe_len = len;
1523         oxe->oxe_namelen = namelen;
1524         memcpy(oxe->oxe_buf, name, namelen);
1525         if (buflen > 0) {
1526                 LASSERT(buf != NULL);
1527                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1528                 oxe->oxe_exist = true;
1529         } else {
1530                 oxe->oxe_exist = false;
1531         }
1532
1533         /* this should be rarely called, just remove old and add new */
1534         spin_lock(&obj->oo_guard);
1535         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1536                 if (namelen == tmp->oxe_namelen &&
1537                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1538                         old = tmp;
1539                         break;
1540                 }
1541         }
1542         if (old != NULL) {
1543                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1544                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1545         } else {
1546                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1547         }
1548         spin_unlock(&obj->oo_guard);
1549 }
1550
1551 static void osd_oxc_del(struct osd_object *obj, const char *name)
1552 {
1553         struct osd_xattr_entry *oxe;
1554         size_t namelen = strlen(name);
1555
1556         spin_lock(&obj->oo_guard);
1557         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1558                 if (namelen == oxe->oxe_namelen &&
1559                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1560                         list_del_rcu(&oxe->oxe_list);
1561                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1562                         break;
1563                 }
1564         }
1565         spin_unlock(&obj->oo_guard);
1566 }
1567
1568 static void osd_oxc_fini(struct osd_object *obj)
1569 {
1570         struct osd_xattr_entry *oxe, *next;
1571
1572         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1573                 list_del(&oxe->oxe_list);
1574                 OBD_FREE(oxe, oxe->oxe_len);
1575         }
1576 }
1577
1578 /*
1579  * Concurrency: no concurrent access is possible that late in object
1580  * life-cycle.
1581  */
1582 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1583 {
1584         struct osd_object *obj = osd_obj(l);
1585         struct lu_object_header *h = obj->oo_header;
1586
1587         LINVRNT(osd_invariant(obj));
1588
1589         osd_oxc_fini(obj);
1590         dt_object_fini(&obj->oo_dt);
1591         if (obj->oo_hl_head != NULL)
1592                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1593         OBD_FREE_PTR(obj);
1594         if (unlikely(h)) {
1595                 lu_object_header_fini(h);
1596                 OBD_FREE_PTR(h);
1597         }
1598 }
1599
1600 /*
1601  * Concurrency: no concurrent access is possible that late in object
1602  * life-cycle.
1603  */
1604 static void osd_index_fini(struct osd_object *o)
1605 {
1606         struct iam_container *bag;
1607
1608         if (o->oo_dir != NULL) {
1609                 bag = &o->oo_dir->od_container;
1610                 if (o->oo_inode != NULL) {
1611                         if (bag->ic_object == o->oo_inode)
1612                                 iam_container_fini(bag);
1613                 }
1614                 OBD_FREE_PTR(o->oo_dir);
1615                 o->oo_dir = NULL;
1616         }
1617 }
1618
1619 /*
1620  * Concurrency: no concurrent access is possible that late in object
1621  * life-cycle (for all existing callers, that is. New callers have to provide
1622  * their own locking.)
1623  */
1624 static int osd_inode_unlinked(const struct inode *inode)
1625 {
1626         return inode->i_nlink == 0;
1627 }
1628
1629 enum {
1630         OSD_TXN_OI_DELETE_CREDITS    = 20,
1631         OSD_TXN_INODE_DELETE_CREDITS = 20
1632 };
1633
1634 /*
1635  * Journal
1636  */
1637
1638 #if OSD_THANDLE_STATS
1639 /**
1640  * Set time when the handle is allocated
1641  */
1642 static void osd_th_alloced(struct osd_thandle *oth)
1643 {
1644         oth->oth_alloced = ktime_get();
1645 }
1646
1647 /**
1648  * Set time when the handle started
1649  */
1650 static void osd_th_started(struct osd_thandle *oth)
1651 {
1652         oth->oth_started = ktime_get();
1653 }
1654
1655 /**
1656  * Check whether the we deal with this handle for too long.
1657  */
1658 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1659                                 ktime_t alloced, ktime_t started,
1660                                 ktime_t closed)
1661 {
1662         ktime_t now = ktime_get();
1663
1664         LASSERT(dev != NULL);
1665
1666         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1667                             ktime_us_delta(started, alloced));
1668         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1669                             ktime_us_delta(closed, started));
1670         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1671                             ktime_us_delta(now, closed));
1672
1673         if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1674                 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1675                                 oth, now, alloced, started, closed);
1676                 libcfs_debug_dumpstack(NULL);
1677         }
1678 }
1679
1680 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1681 {                                                                       \
1682         ktime_t __closed = ktime_get();                                 \
1683         ktime_t __alloced = oth->oth_alloced;                           \
1684         ktime_t __started = oth->oth_started;                           \
1685                                                                         \
1686         expr;                                                           \
1687         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1688 }
1689
1690 #else /* OSD_THANDLE_STATS */
1691
1692 #define osd_th_alloced(h)                  do {} while(0)
1693 #define osd_th_started(h)                  do {} while(0)
1694 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1695
1696 #endif /* OSD_THANDLE_STATS */
1697
1698 /*
1699  * Concurrency: doesn't access mutable data.
1700  */
1701 static int osd_param_is_not_sane(const struct osd_device *dev,
1702                                  const struct thandle *th)
1703 {
1704         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1705
1706         return oh->ot_credits > osd_transaction_size(dev);
1707 }
1708
1709 /*
1710  * Concurrency: shouldn't matter.
1711  */
1712 static void osd_trans_commit_cb(struct super_block *sb,
1713                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1714 {
1715         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1716         struct thandle *th = &oh->ot_super;
1717         struct lu_device *lud = &th->th_dev->dd_lu_dev;
1718         struct dt_txn_commit_cb *dcb, *tmp;
1719
1720         LASSERT(oh->ot_handle == NULL);
1721
1722         if (error)
1723                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1724
1725         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1726         /* call per-transaction callbacks if any */
1727         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1728                                  dcb_linkage) {
1729                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1730                          "commit callback entry: magic=%x name='%s'\n",
1731                          dcb->dcb_magic, dcb->dcb_name);
1732                 list_del_init(&dcb->dcb_linkage);
1733                 dcb->dcb_func(NULL, th, dcb, error);
1734         }
1735
1736         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1737         lu_device_put(lud);
1738         th->th_dev = NULL;
1739
1740         OBD_FREE_PTR(oh);
1741 }
1742
1743 #ifndef HAVE_SB_START_WRITE
1744 # define sb_start_write(sb) do {} while (0)
1745 # define sb_end_write(sb) do {} while (0)
1746 #endif
1747
1748 static struct thandle *osd_trans_create(const struct lu_env *env,
1749                                         struct dt_device *d)
1750 {
1751         struct osd_thread_info *oti = osd_oti_get(env);
1752         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1753         struct osd_thandle *oh;
1754         struct thandle *th;
1755
1756         ENTRY;
1757
1758         if (d->dd_rdonly) {
1759                 CERROR("%s: someone try to start transaction under "
1760                        "readonly mode, should be disabled.\n",
1761                        osd_name(osd_dt_dev(d)));
1762                 dump_stack();
1763                 RETURN(ERR_PTR(-EROFS));
1764         }
1765
1766         /* on pending IO in this thread should left from prev. request */
1767         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1768
1769         sb_start_write(osd_sb(osd_dt_dev(d)));
1770
1771         OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1772         if (!oh) {
1773                 sb_end_write(osd_sb(osd_dt_dev(d)));
1774                 RETURN(ERR_PTR(-ENOMEM));
1775         }
1776
1777         oh->ot_quota_trans = &oti->oti_quota_trans;
1778         memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1779         th = &oh->ot_super;
1780         th->th_dev = d;
1781         th->th_result = 0;
1782         oh->ot_credits = 0;
1783         INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1784         INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1785         INIT_LIST_HEAD(&oh->ot_trunc_locks);
1786         osd_th_alloced(oh);
1787
1788         memset(oti->oti_declare_ops, 0,
1789                sizeof(oti->oti_declare_ops));
1790         memset(oti->oti_declare_ops_cred, 0,
1791                sizeof(oti->oti_declare_ops_cred));
1792         memset(oti->oti_declare_ops_used, 0,
1793                sizeof(oti->oti_declare_ops_used));
1794
1795         oti->oti_ins_cache_depth++;
1796
1797         RETURN(th);
1798 }
1799
1800 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1801 {
1802         struct osd_thread_info *oti = osd_oti_get(env);
1803         struct osd_thandle *oh;
1804
1805         oh = container_of0(th, struct osd_thandle, ot_super);
1806         LASSERT(oh != NULL);
1807
1808         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1809               oti->oti_declare_ops[OSD_OT_CREATE],
1810               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1811               oti->oti_declare_ops_used[OSD_OT_CREATE],
1812               oti->oti_declare_ops[OSD_OT_DESTROY],
1813               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1814               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1815         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1816               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1817               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1818               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1819               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1820               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1821               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1822         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1823               oti->oti_declare_ops[OSD_OT_WRITE],
1824               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1825               oti->oti_declare_ops_used[OSD_OT_WRITE],
1826               oti->oti_declare_ops[OSD_OT_PUNCH],
1827               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1828               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1829               oti->oti_declare_ops[OSD_OT_QUOTA],
1830               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1831               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1832         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1833               oti->oti_declare_ops[OSD_OT_INSERT],
1834               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1835               oti->oti_declare_ops_used[OSD_OT_INSERT],
1836               oti->oti_declare_ops[OSD_OT_DELETE],
1837               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1838               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1839         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1840               oti->oti_declare_ops[OSD_OT_REF_ADD],
1841               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1842               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1843               oti->oti_declare_ops[OSD_OT_REF_DEL],
1844               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1845               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1846 }
1847
1848 /*
1849  * Concurrency: shouldn't matter.
1850  */
1851 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1852                            struct thandle *th)
1853 {
1854         struct osd_thread_info *oti = osd_oti_get(env);
1855         struct osd_device *dev = osd_dt_dev(d);
1856         handle_t *jh;
1857         struct osd_thandle *oh;
1858         int rc;
1859
1860         ENTRY;
1861
1862         LASSERT(current->journal_info == NULL);
1863
1864         oh = container_of0(th, struct osd_thandle, ot_super);
1865         LASSERT(oh != NULL);
1866         LASSERT(oh->ot_handle == NULL);
1867
1868         rc = dt_txn_hook_start(env, d, th);
1869         if (rc != 0)
1870                 GOTO(out, rc);
1871
1872         if (unlikely(osd_param_is_not_sane(dev, th))) {
1873                 static unsigned long last_printed;
1874                 static int last_credits;
1875
1876                 /*
1877                  * don't make noise on a tiny testing systems
1878                  * actual credits misuse will be caught anyway
1879                  */
1880                 if (last_credits != oh->ot_credits &&
1881                     time_after(jiffies, last_printed +
1882                                cfs_time_seconds(60)) &&
1883                     osd_transaction_size(dev) > 512) {
1884                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1885                               oh->ot_credits, osd_transaction_size(dev));
1886                         osd_trans_dump_creds(env, th);
1887                         libcfs_debug_dumpstack(NULL);
1888                         last_credits = oh->ot_credits;
1889                         last_printed = jiffies;
1890                 }
1891                 /*
1892                  * XXX Limit the credits to 'max_transaction_buffers', and
1893                  *     let the underlying filesystem to catch the error if
1894                  *     we really need so many credits.
1895                  *
1896                  *     This should be removed when we can calculate the
1897                  *     credits precisely.
1898                  */
1899                 oh->ot_credits = osd_transaction_size(dev);
1900         } else if (ldiskfs_track_declares_assert != 0) {
1901                 /*
1902                  * reserve few credits to prevent an assertion in JBD
1903                  * our debugging mechanism will be able to detected
1904                  * overuse. this can help to debug single-update
1905                  * transactions
1906                  */
1907                 oh->ot_credits += 10;
1908                 if (unlikely(osd_param_is_not_sane(dev, th)))
1909                         oh->ot_credits = osd_transaction_size(dev);
1910         }
1911
1912         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1913                 GOTO(out, rc = -EIO);
1914
1915         /*
1916          * XXX temporary stuff. Some abstraction layer should
1917          * be used.
1918          */
1919         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1920         osd_th_started(oh);
1921         if (!IS_ERR(jh)) {
1922                 oh->ot_handle = jh;
1923                 LASSERT(oti->oti_txns == 0);
1924
1925                 lu_device_get(&d->dd_lu_dev);
1926                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1927                               "osd-tx", th);
1928                 oti->oti_txns++;
1929                 rc = 0;
1930         } else {
1931                 rc = PTR_ERR(jh);
1932         }
1933 out:
1934         RETURN(rc);
1935 }
1936
1937 static int osd_seq_exists(const struct lu_env *env,
1938                           struct osd_device *osd, u64 seq)
1939 {
1940         struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1941         struct seq_server_site *ss = osd_seq_site(osd);
1942         int rc;
1943
1944         ENTRY;
1945
1946         LASSERT(ss != NULL);
1947         LASSERT(ss->ss_server_fld != NULL);
1948
1949         rc = osd_fld_lookup(env, osd, seq, range);
1950         if (rc != 0) {
1951                 if (rc != -ENOENT)
1952                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1953                                osd_name(osd), seq, rc);
1954                 RETURN(0);
1955         }
1956
1957         RETURN(ss->ss_node_id == range->lsr_index);
1958 }
1959
1960 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1961 {
1962         struct dt_txn_commit_cb *dcb;
1963         struct dt_txn_commit_cb *tmp;
1964
1965         /* call per-transaction stop callbacks if any */
1966         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1967                                  dcb_linkage) {
1968                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1969                          "commit callback entry: magic=%x name='%s'\n",
1970                          dcb->dcb_magic, dcb->dcb_name);
1971                 list_del_init(&dcb->dcb_linkage);
1972                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1973         }
1974 }
1975
1976 /*
1977  * Concurrency: shouldn't matter.
1978  */
1979 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1980                           struct thandle *th)
1981 {
1982         struct osd_thread_info *oti = osd_oti_get(env);
1983         struct osd_thandle *oh;
1984         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1985         struct osd_device *osd = osd_dt_dev(th->th_dev);
1986         struct qsd_instance *qsd = osd_def_qsd(osd);
1987         struct lquota_trans *qtrans;
1988         struct list_head truncates = LIST_HEAD_INIT(truncates);
1989         int rc = 0, remove_agents = 0;
1990
1991         ENTRY;
1992
1993         oh = container_of0(th, struct osd_thandle, ot_super);
1994
1995         remove_agents = oh->ot_remove_agents;
1996
1997         qtrans = oh->ot_quota_trans;
1998         oh->ot_quota_trans = NULL;
1999
2000         /* move locks to local list, stop tx, execute truncates */
2001         list_splice(&oh->ot_trunc_locks, &truncates);
2002
2003         if (oh->ot_handle != NULL) {
2004                 int rc2;
2005
2006                 handle_t *hdl = oh->ot_handle;
2007
2008                 /*
2009                  * add commit callback
2010                  * notice we don't do this in osd_trans_start()
2011                  * as underlying transaction can change during truncate
2012                  */
2013                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2014                                              &oh->ot_jcb);
2015
2016                 LASSERT(oti->oti_txns == 1);
2017                 oti->oti_txns--;
2018
2019                 rc = dt_txn_hook_stop(env, th);
2020                 if (rc != 0)
2021                         CERROR("%s: failed in transaction hook: rc = %d\n",
2022                                osd_name(osd), rc);
2023
2024                 osd_trans_stop_cb(oh, rc);
2025                 /* hook functions might modify th_sync */
2026                 hdl->h_sync = th->th_sync;
2027
2028                 oh->ot_handle = NULL;
2029                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2030                 if (rc2 != 0)
2031                         CERROR("%s: failed to stop transaction: rc = %d\n",
2032                                osd_name(osd), rc2);
2033                 if (!rc)
2034                         rc = rc2;
2035
2036                 osd_process_truncates(&truncates);
2037         } else {
2038                 osd_trans_stop_cb(oh, th->th_result);
2039                 OBD_FREE_PTR(oh);
2040         }
2041
2042         osd_trunc_unlock_all(&truncates);
2043
2044         /* inform the quota slave device that the transaction is stopping */
2045         qsd_op_end(env, qsd, qtrans);
2046
2047         /*
2048          * as we want IO to journal and data IO be concurrent, we don't block
2049          * awaiting data IO completion in osd_do_bio(), instead we wait here
2050          * once transaction is submitted to the journal. all reqular requests
2051          * don't do direct IO (except read/write), thus this wait_event becomes
2052          * no-op for them.
2053          *
2054          * IMPORTANT: we have to wait till any IO submited by the thread is
2055          * completed otherwise iobuf may be corrupted by different request
2056          */
2057         wait_event(iobuf->dr_wait,
2058                        atomic_read(&iobuf->dr_numreqs) == 0);
2059         osd_fini_iobuf(osd, iobuf);
2060         if (!rc)
2061                 rc = iobuf->dr_error;
2062
2063         if (unlikely(remove_agents != 0))
2064                 osd_process_scheduled_agent_removals(env, osd);
2065
2066         oti->oti_ins_cache_depth--;
2067         /* reset OI cache for safety */
2068         if (oti->oti_ins_cache_depth == 0)
2069                 oti->oti_ins_cache_used = 0;
2070
2071         sb_end_write(osd_sb(osd));
2072
2073         RETURN(rc);
2074 }
2075
2076 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2077 {
2078         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
2079                                                ot_super);
2080
2081         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2082         LASSERT(&dcb->dcb_func != NULL);
2083         if (dcb->dcb_flags & DCB_TRANS_STOP)
2084                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2085         else
2086                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2087
2088         return 0;
2089 }
2090
2091 /*
2092  * Called just before object is freed. Releases all resources except for
2093  * object itself (that is released by osd_object_free()).
2094  *
2095  * Concurrency: no concurrent access is possible that late in object
2096  * life-cycle.
2097  */
2098 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2099 {
2100         struct osd_object *obj = osd_obj(l);
2101         struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2102         struct inode *inode = obj->oo_inode;
2103         __u64 projid;
2104         qid_t uid;
2105         qid_t gid;
2106
2107         LINVRNT(osd_invariant(obj));
2108
2109         /*
2110          * If object is unlinked remove fid->ino mapping from object index.
2111          */
2112
2113         osd_index_fini(obj);
2114
2115         if (!inode)
2116                 return;
2117
2118         uid = i_uid_read(inode);
2119         gid = i_gid_read(inode);
2120         projid = i_projid_read(inode);
2121
2122         obj->oo_inode = NULL;
2123         iput(inode);
2124
2125         /* do not rebalance quota if the caller needs to release memory
2126          * otherwise qsd_refresh_usage() may went into a new ldiskfs
2127          * transaction and risk to deadlock - LU-12178 */
2128         if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2129                 return;
2130
2131         if (!obj->oo_header && qsd) {
2132                 struct osd_thread_info *info = osd_oti_get(env);
2133                 struct lquota_id_info *qi = &info->oti_qi;
2134
2135                 /* Release granted quota to master if necessary */
2136                 qi->lqi_id.qid_uid = uid;
2137                 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2138
2139                 qi->lqi_id.qid_uid = gid;
2140                 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2141
2142                 qi->lqi_id.qid_uid = projid;
2143                 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2144         }
2145 }
2146
2147 /*
2148  * Concurrency: ->loo_object_release() is called under site spin-lock.
2149  */
2150 static void osd_object_release(const struct lu_env *env,
2151                                struct lu_object *l)
2152 {
2153         struct osd_object *o = osd_obj(l);
2154
2155         /*
2156          * nobody should be releasing a non-destroyed object with nlink=0
2157          * the API allows this, but ldiskfs doesn't like and then report
2158          * this inode as deleted
2159          */
2160         LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2161                   o->oo_inode->i_nlink == 0));
2162 }
2163
2164 /*
2165  * Concurrency: shouldn't matter.
2166  */
2167 static int osd_object_print(const struct lu_env *env, void *cookie,
2168                             lu_printer_t p, const struct lu_object *l)
2169 {
2170         struct osd_object *o = osd_obj(l);
2171         struct iam_descr *d;
2172
2173         if (o->oo_dir != NULL)
2174                 d = o->oo_dir->od_container.ic_descr;
2175         else
2176                 d = NULL;
2177         return (*p)(env, cookie,
2178                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2179                     o, o->oo_inode,
2180                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
2181                     o->oo_inode ? o->oo_inode->i_generation : 0,
2182                     d ? d->id_ops->id_name : "plain");
2183 }
2184
2185 /*
2186  * Concurrency: shouldn't matter.
2187  */
2188 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2189                 struct obd_statfs *sfs, struct obd_statfs_info *info)
2190 {
2191         struct osd_device *osd = osd_dt_dev(d);
2192         struct super_block *sb = osd_sb(osd);
2193         struct kstatfs *ksfs;
2194         __u64 reserved;
2195         int result = 0;
2196
2197         if (unlikely(osd->od_mnt == NULL))
2198                 return -EINPROGRESS;
2199
2200         /* osd_lproc.c call this without env, allocate ksfs for that case */
2201         if (unlikely(env == NULL)) {
2202                 OBD_ALLOC_PTR(ksfs);
2203                 if (ksfs == NULL)
2204                         return -ENOMEM;
2205         } else {
2206                 ksfs = &osd_oti_get(env)->oti_ksfs;
2207         }
2208
2209         result = sb->s_op->statfs(sb->s_root, ksfs);
2210         if (result)
2211                 goto out;
2212
2213         statfs_pack(sfs, ksfs);
2214         if (unlikely(sb->s_flags & SB_RDONLY))
2215                 sfs->os_state |= OS_STATE_READONLY;
2216
2217         sfs->os_state |= osd->od_nonrotational ? OS_STATE_NONROT : 0;
2218
2219         if (ldiskfs_has_feature_extents(sb))
2220                 sfs->os_maxbytes = sb->s_maxbytes;
2221         else
2222                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2223
2224         /*
2225          * Reserve some space so to avoid fragmenting the filesystem too much.
2226          * Fragmentation not only impacts performance, but can also increase
2227          * metadata overhead significantly, causing grant calculation to be
2228          * wrong.
2229          *
2230          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2231          */
2232         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2233         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2234         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2235                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2236
2237         sfs->os_blocks -= reserved;
2238         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2239         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2240
2241 out:
2242         if (unlikely(env == NULL))
2243                 OBD_FREE_PTR(ksfs);
2244         return result;
2245 }
2246
2247 /**
2248  * Estimate space needed for file creations. We assume the largest filename
2249  * which is 2^64 - 1, hence a filename of 20 chars.
2250  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2251  */
2252 #ifdef __LDISKFS_DIR_REC_LEN
2253 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2254 #else
2255 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2256 #endif
2257
2258 /*
2259  * Concurrency: doesn't access mutable data.
2260  */
2261 static void osd_conf_get(const struct lu_env *env,
2262                          const struct dt_device *dev,
2263                          struct dt_device_param *param)
2264 {
2265         struct osd_device *d = osd_dt_dev(dev);
2266         struct super_block *sb = osd_sb(d);
2267         struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2268         const char *name;
2269         int ea_overhead;
2270
2271         /*
2272          * XXX should be taken from not-yet-existing fs abstraction layer.
2273          */
2274         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2275         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2276         param->ddp_symlink_max  = sb->s_blocksize;
2277         param->ddp_mount_type   = LDD_MT_LDISKFS;
2278         if (ldiskfs_has_feature_extents(sb))
2279                 param->ddp_maxbytes = sb->s_maxbytes;
2280         else
2281                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2282         /*
2283          * inode are statically allocated, so per-inode space consumption
2284          * is the space consumed by the directory entry
2285          */
2286         param->ddp_inodespace     = PER_OBJ_USAGE;
2287         /*
2288          * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2289          * is 128MB) which is unlikely to be hit in real life. Report a smaller
2290          * maximum length to not under-count the actual number of extents
2291          * needed for writing a file if there are sub-optimal block allocations.
2292          */
2293         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2294         /* worst-case extent insertion metadata overhead */
2295         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2296         param->ddp_mntopts = 0;
2297         if (test_opt(sb, XATTR_USER))
2298                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2299         if (test_opt(sb, POSIX_ACL))
2300                 param->ddp_mntopts |= MNTOPT_ACL;
2301
2302         /*
2303          * LOD might calculate the max stripe count based on max_ea_size,
2304          * so we need take account in the overhead as well,
2305          * xattr_header + magic + xattr_entry_head
2306          */
2307         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2308                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2309
2310 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2311         if (ldiskfs_has_feature_ea_inode(sb))
2312                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2313                                                                 ea_overhead;
2314         else
2315 #endif
2316                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2317
2318         if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2319                 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2320
2321         /*
2322          * Preferred RPC size for efficient disk IO.  4MB shows good
2323          * all-around performance for ldiskfs, but use bigalloc chunk size
2324          * by default if larger.
2325          */
2326 #if defined(LDISKFS_CLUSTER_SIZE)
2327         if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2328                 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2329         else
2330 #endif
2331                 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2332
2333         param->ddp_t10_cksum_type = 0;
2334         if (bi) {
2335                 unsigned short interval = blk_integrity_interval(bi);
2336                 name = blk_integrity_name(bi);
2337                 /*
2338                  * Expected values:
2339                  * T10-DIF-TYPE1-CRC
2340                  * T10-DIF-TYPE3-CRC
2341                  * T10-DIF-TYPE1-IP
2342                  * T10-DIF-TYPE3-IP
2343                  */
2344                 if (strncmp(name, "T10-DIF-TYPE",
2345                             sizeof("T10-DIF-TYPE") - 1) == 0) {
2346                         /* also skip "1/3-" at end */
2347                         const int type_off = sizeof("T10-DIF-TYPE.");
2348                         char type_number = name[type_off - 2];
2349
2350                         if (interval != 512 && interval != 4096) {
2351                                 CERROR("%s: unsupported T10PI sector size %u\n",
2352                                        d->od_svname, interval);
2353                         } else if (type_number != '1' && type_number != '3') {
2354                                 CERROR("%s: unsupported T10PI type %s\n",
2355                                        d->od_svname, name);
2356                         } else if (strcmp(name + type_off, "CRC") == 0) {
2357                                 d->od_t10_type = type_number == '1' ?
2358                                         OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2359                                 param->ddp_t10_cksum_type = interval == 512 ?
2360                                         OBD_CKSUM_T10CRC512 :
2361                                         OBD_CKSUM_T10CRC4K;
2362                         } else if (strcmp(name + type_off, "IP") == 0) {
2363                                 d->od_t10_type = type_number == '1' ?
2364                                         OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2365                                 param->ddp_t10_cksum_type = interval == 512 ?
2366                                         OBD_CKSUM_T10IP512 :
2367                                         OBD_CKSUM_T10IP4K;
2368                         } else {
2369                                 CERROR("%s: unsupported checksum type of "
2370                                        "T10PI type '%s'",
2371                                        d->od_svname, name);
2372                         }
2373
2374                 } else {
2375                         CERROR("%s: unsupported T10PI type '%s'",
2376                                d->od_svname, name);
2377                 }
2378         }
2379 }
2380
2381 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2382 {
2383         return osd_sb(osd_dt_dev(d));
2384 }
2385
2386 /*
2387  * Concurrency: shouldn't matter.
2388  */
2389 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2390 {
2391         int rc;
2392         struct super_block *s = osd_sb(osd_dt_dev(d));
2393         ENTRY;
2394
2395         down_read(&s->s_umount);
2396         rc = s->s_op->sync_fs(s, 1);
2397         up_read(&s->s_umount);
2398
2399         CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2400                rc);
2401
2402         return rc;
2403 }
2404
2405 /**
2406  * Start commit for OSD device.
2407  *
2408  * An implementation of dt_commit_async method for OSD device.
2409  * Asychronously starts underlayng fs sync and thereby a transaction
2410  * commit.
2411  *
2412  * \param env environment
2413  * \param d dt device
2414  *
2415  * \see dt_device_operations
2416  */
2417 static int osd_commit_async(const struct lu_env *env,
2418                             struct dt_device *d)
2419 {
2420         struct super_block *s = osd_sb(osd_dt_dev(d));
2421         int rc;
2422
2423         ENTRY;
2424
2425         CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2426         down_read(&s->s_umount);
2427         rc = s->s_op->sync_fs(s, 0);
2428         up_read(&s->s_umount);
2429
2430         RETURN(rc);
2431 }
2432
2433 /* Our own copy of the set readonly functions if present, or NU if not. */
2434 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2435 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2436 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2437
2438 /*
2439  * Concurrency: shouldn't matter.
2440  */
2441 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2442 {
2443         struct super_block *sb = osd_sb(osd_dt_dev(d));
2444         struct block_device *dev = sb->s_bdev;
2445         int rc = -EOPNOTSUPP;
2446
2447         ENTRY;
2448
2449         if (priv_dev_set_rdonly) {
2450                 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2451
2452                 rc = 0;
2453                 CERROR("*** setting %s read-only ***\n",
2454                        osd_dt_dev(d)->od_svname);
2455
2456                 if (sb->s_op->freeze_fs) {
2457                         rc = sb->s_op->freeze_fs(sb);
2458                         if (rc)
2459                                 goto out;
2460                 }
2461
2462                 if (jdev && (jdev != dev)) {
2463                         CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2464                                (long)jdev);
2465                         priv_dev_set_rdonly(jdev);
2466                 }
2467                 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2468                 priv_dev_set_rdonly(dev);
2469
2470                 if (sb->s_op->unfreeze_fs)
2471                         sb->s_op->unfreeze_fs(sb);
2472         }
2473
2474 out:
2475         if (rc)
2476                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2477                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2478
2479         RETURN(rc);
2480 }
2481
2482 /**
2483  * Note: we do not count into QUOTA here.
2484  * If we mount with --data_journal we may need more.
2485  */
2486 const int osd_dto_credits_noquota[DTO_NR] = {
2487         /**
2488          * Insert.
2489          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2490          * SINGLEDATA_TRANS_BLOCKS(8)
2491          * XXX Note: maybe iam need more, since iam have more level than
2492          *           EXT3 htree.
2493          */
2494         [DTO_INDEX_INSERT]  = 16,
2495         /**
2496          * Delete
2497          * just modify a single entry, probably merge few within a block
2498          */
2499         [DTO_INDEX_DELETE]  = 1,
2500         /**
2501          * Used for OI scrub
2502          */
2503         [DTO_INDEX_UPDATE]  = 16,
2504         /**
2505          * 4(inode, inode bits, groups, GDT)
2506          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2507          */
2508         [DTO_OBJECT_CREATE] = 4,
2509         /**
2510          * 4(inode, inode bits, groups, GDT)
2511          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2512          */
2513         [DTO_OBJECT_DELETE] = 4,
2514         /**
2515          * Attr set credits (inode)
2516          */
2517         [DTO_ATTR_SET_BASE] = 1,
2518         /**
2519          * Xattr set. The same as xattr of EXT3.
2520          * DATA_TRANS_BLOCKS(14)
2521          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2522          * are also counted in. Do not know why?
2523          */
2524         [DTO_XATTR_SET]     = 14,
2525         /**
2526          * credits for inode change during write.
2527          */
2528         [DTO_WRITE_BASE]    = 3,
2529         /**
2530          * credits for single block write.
2531          */
2532         [DTO_WRITE_BLOCK]   = 14,
2533         /**
2534          * Attr set credits for chown.
2535          * This is extra credits for setattr, and it is null without quota
2536          */
2537         [DTO_ATTR_SET_CHOWN] = 0
2538 };
2539
2540 static const struct dt_device_operations osd_dt_ops = {
2541         .dt_root_get       = osd_root_get,
2542         .dt_statfs         = osd_statfs,
2543         .dt_trans_create   = osd_trans_create,
2544         .dt_trans_start    = osd_trans_start,
2545         .dt_trans_stop     = osd_trans_stop,
2546         .dt_trans_cb_add   = osd_trans_cb_add,
2547         .dt_conf_get       = osd_conf_get,
2548         .dt_mnt_sb_get     = osd_mnt_sb_get,
2549         .dt_sync           = osd_sync,
2550         .dt_ro             = osd_ro,
2551         .dt_commit_async   = osd_commit_async,
2552 };
2553
2554 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2555                           unsigned int role)
2556 {
2557         struct osd_object *obj = osd_dt_obj(dt);
2558         struct osd_thread_info *oti = osd_oti_get(env);
2559
2560         LINVRNT(osd_invariant(obj));
2561
2562         LASSERT(obj->oo_owner != env);
2563         down_read_nested(&obj->oo_sem, role);
2564
2565         LASSERT(obj->oo_owner == NULL);
2566         oti->oti_r_locks++;
2567 }
2568
2569 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2570                            unsigned int role)
2571 {
2572         struct osd_object *obj = osd_dt_obj(dt);
2573         struct osd_thread_info *oti = osd_oti_get(env);
2574
2575         LINVRNT(osd_invariant(obj));
2576
2577         LASSERT(obj->oo_owner != env);
2578         down_write_nested(&obj->oo_sem, role);
2579
2580         LASSERT(obj->oo_owner == NULL);
2581         obj->oo_owner = env;
2582         oti->oti_w_locks++;
2583 }
2584
2585 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2586 {
2587         struct osd_object *obj = osd_dt_obj(dt);
2588         struct osd_thread_info *oti = osd_oti_get(env);
2589
2590         LINVRNT(osd_invariant(obj));
2591
2592         LASSERT(oti->oti_r_locks > 0);
2593         oti->oti_r_locks--;
2594         up_read(&obj->oo_sem);
2595 }
2596
2597 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2598 {
2599         struct osd_object *obj = osd_dt_obj(dt);
2600         struct osd_thread_info *oti = osd_oti_get(env);
2601
2602         LINVRNT(osd_invariant(obj));
2603
2604         LASSERT(obj->oo_owner == env);
2605         LASSERT(oti->oti_w_locks > 0);
2606         oti->oti_w_locks--;
2607         obj->oo_owner = NULL;
2608         up_write(&obj->oo_sem);
2609 }
2610
2611 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2612 {
2613         struct osd_object *obj = osd_dt_obj(dt);
2614
2615         LINVRNT(osd_invariant(obj));
2616
2617         return obj->oo_owner == env;
2618 }
2619
2620 static void osd_inode_getattr(const struct lu_env *env,
2621                               struct inode *inode, struct lu_attr *attr)
2622 {
2623         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2624                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2625                            LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2626                            LA_BLKSIZE | LA_TYPE;
2627
2628         attr->la_atime = inode->i_atime.tv_sec;
2629         attr->la_mtime = inode->i_mtime.tv_sec;
2630         attr->la_ctime = inode->i_ctime.tv_sec;
2631         attr->la_mode    = inode->i_mode;
2632         attr->la_size    = i_size_read(inode);
2633         attr->la_blocks  = inode->i_blocks;
2634         attr->la_uid     = i_uid_read(inode);
2635         attr->la_gid     = i_gid_read(inode);
2636         attr->la_projid  = i_projid_read(inode);
2637         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2638         attr->la_nlink   = inode->i_nlink;
2639         attr->la_rdev    = inode->i_rdev;
2640         attr->la_blksize = 1 << inode->i_blkbits;
2641         attr->la_blkbits = inode->i_blkbits;
2642         /*
2643          * Ext4 did not transfer inherit flags from raw inode
2644          * to inode flags, and ext4 internally test raw inode
2645          * @i_flags directly. Instead of patching ext4, we do it here.
2646          */
2647         if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2648                 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2649 }
2650
2651 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2652                         struct lu_attr *attr)
2653 {
2654         struct osd_object *obj = osd_dt_obj(dt);
2655
2656         if (unlikely(!dt_object_exists(dt)))
2657                 return -ENOENT;
2658         if (unlikely(obj->oo_destroyed))
2659                 return -ENOENT;
2660
2661         LASSERT(!dt_object_remote(dt));
2662         LINVRNT(osd_invariant(obj));
2663
2664         spin_lock(&obj->oo_guard);
2665         osd_inode_getattr(env, obj->oo_inode, attr);
2666         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2667                 attr->la_valid |= LA_FLAGS;
2668                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2669         }
2670         spin_unlock(&obj->oo_guard);
2671
2672         return 0;
2673 }
2674
2675 static int osd_declare_attr_qid(const struct lu_env *env,
2676                                 struct osd_object *obj,
2677                                 struct osd_thandle *oh, long long bspace,
2678                                 qid_t old_id, qid_t new_id, bool enforce,
2679                                 unsigned int type, bool ignore_edquot)
2680 {
2681         int rc;
2682         struct osd_thread_info *info = osd_oti_get(env);
2683         struct lquota_id_info  *qi = &info->oti_qi;
2684
2685         qi->lqi_type = type;
2686         /* inode accounting */
2687         qi->lqi_is_blk = false;
2688
2689         /* one more inode for the new id ... */
2690         qi->lqi_id.qid_uid = new_id;
2691         qi->lqi_space      = 1;
2692         /* Reserve credits for the new id */
2693         rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2694         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2695                 rc = 0;
2696         if (rc)
2697                 RETURN(rc);
2698
2699         /* and one less inode for the current id */
2700         qi->lqi_id.qid_uid = old_id;
2701         qi->lqi_space = -1;
2702         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2703         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2704                 rc = 0;
2705         if (rc)
2706                 RETURN(rc);
2707
2708         /* block accounting */
2709         qi->lqi_is_blk = true;
2710
2711         /* more blocks for the new id ... */
2712         qi->lqi_id.qid_uid = new_id;
2713         qi->lqi_space      = bspace;
2714         /*
2715          * Credits for the new uid has been reserved, re-use "obj"
2716          * to save credit reservation.
2717          */
2718         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2719         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2720                 rc = 0;
2721         if (rc)
2722                 RETURN(rc);
2723
2724         /* and finally less blocks for the current uid */
2725         qi->lqi_id.qid_uid = old_id;
2726         qi->lqi_space      = -bspace;
2727         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2728         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2729                 rc = 0;
2730
2731         RETURN(rc);
2732 }
2733
2734 static int osd_declare_attr_set(const struct lu_env *env,
2735                                 struct dt_object *dt,
2736                                 const struct lu_attr *attr,
2737                                 struct thandle *handle)
2738 {
2739         struct osd_thandle *oh;
2740         struct osd_object *obj;
2741         qid_t uid;
2742         qid_t gid;
2743         long long bspace;
2744         int rc = 0;
2745         bool enforce;
2746
2747         ENTRY;
2748
2749         LASSERT(dt != NULL);
2750         LASSERT(handle != NULL);
2751
2752         obj = osd_dt_obj(dt);
2753         LASSERT(osd_invariant(obj));
2754
2755         oh = container_of0(handle, struct osd_thandle, ot_super);
2756         LASSERT(oh->ot_handle == NULL);
2757
2758         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2759                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2760
2761         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2762                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2763
2764         if (attr == NULL || obj->oo_inode == NULL)
2765                 RETURN(rc);
2766
2767         bspace   = obj->oo_inode->i_blocks << 9;
2768         bspace   = toqb(bspace);
2769
2770         /*
2771          * Changing ownership is always preformed by super user, it should not
2772          * fail with EDQUOT unless required explicitly.
2773          *
2774          * We still need to call the osd_declare_qid() to calculate the journal
2775          * credits for updating quota accounting files and to trigger quota
2776          * space adjustment once the operation is completed.
2777          */
2778         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2779                 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2780
2781                 if (!ignore_edquot)
2782                         CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2783                                "(the quota space is %lld)\n",
2784                                obj->oo_inode->i_sb->s_id, attr->la_uid,
2785                                attr->la_gid, bspace);
2786
2787                 /* USERQUOTA */
2788                 uid = i_uid_read(obj->oo_inode);
2789                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2790                 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2791                                           attr->la_uid, enforce, USRQUOTA,
2792                                           true);
2793                 if (rc)
2794                         RETURN(rc);
2795
2796                 gid = i_gid_read(obj->oo_inode);
2797                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2798                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2799                                           i_gid_read(obj->oo_inode),
2800                                           attr->la_gid, enforce, GRPQUOTA,
2801                                           ignore_edquot);
2802                 if (rc)
2803                         RETURN(rc);
2804
2805         }
2806 #ifdef HAVE_PROJECT_QUOTA
2807         if (attr->la_valid & LA_PROJID) {
2808                 __u32 projid = i_projid_read(obj->oo_inode);
2809
2810                 enforce = (attr->la_valid & LA_PROJID) &&
2811                                         (attr->la_projid != projid);
2812                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2813                                           (qid_t)projid, (qid_t)attr->la_projid,
2814                                           enforce, PRJQUOTA, true);
2815                 if (rc)
2816                         RETURN(rc);
2817         }
2818 #endif
2819         RETURN(rc);
2820 }
2821
2822 static int osd_inode_setattr(const struct lu_env *env,
2823                              struct inode *inode, const struct lu_attr *attr)
2824 {
2825         __u64 bits = attr->la_valid;
2826
2827         /* Only allow set size for regular file */
2828         if (!S_ISREG(inode->i_mode))
2829                 bits &= ~(LA_SIZE | LA_BLOCKS);
2830
2831         if (bits == 0)
2832                 return 0;
2833
2834         if (bits & LA_ATIME)
2835                 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2836         if (bits & LA_CTIME)
2837                 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2838         if (bits & LA_MTIME)
2839                 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2840         if (bits & LA_SIZE) {
2841                 spin_lock(&inode->i_lock);
2842                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2843                 i_size_write(inode, attr->la_size);
2844                 spin_unlock(&inode->i_lock);
2845         }
2846
2847         /*
2848          * OSD should not change "i_blocks" which is used by quota.
2849          * "i_blocks" should be changed by ldiskfs only.
2850          */
2851         if (bits & LA_MODE)
2852                 inode->i_mode = (inode->i_mode & S_IFMT) |
2853                                 (attr->la_mode & ~S_IFMT);
2854         if (bits & LA_UID)
2855                 i_uid_write(inode, attr->la_uid);
2856         if (bits & LA_GID)
2857                 i_gid_write(inode, attr->la_gid);
2858         if (bits & LA_PROJID)
2859                 i_projid_write(inode, attr->la_projid);
2860         if (bits & LA_NLINK)
2861                 set_nlink(inode, attr->la_nlink);
2862         if (bits & LA_RDEV)
2863                 inode->i_rdev = attr->la_rdev;
2864
2865         if (bits & LA_FLAGS) {
2866                 /* always keep S_NOCMTIME */
2867                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2868                                  S_NOCMTIME;
2869                 /*
2870                  * Ext4 did not transfer inherit flags from
2871                  * @inode->i_flags to raw inode i_flags when writing
2872                  * flags, we do it explictly here.
2873                  */
2874                 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2875                         LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2876                 else
2877                         LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2878         }
2879         return 0;
2880 }
2881
2882 #ifdef HAVE_PROJECT_QUOTA
2883 static int osd_transfer_project(struct inode *inode, __u32 projid)
2884 {
2885         struct super_block *sb = inode->i_sb;
2886         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2887         int err;
2888         kprojid_t kprojid;
2889         struct ldiskfs_iloc iloc;
2890         struct ldiskfs_inode *raw_inode;
2891         struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2892
2893         if (!ldiskfs_has_feature_project(sb)) {
2894                 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2895                         == LDISKFS_DEF_PROJID);
2896                 if (projid != LDISKFS_DEF_PROJID)
2897                         return -EOPNOTSUPP;
2898                 else
2899                         return 0;
2900         }
2901
2902         if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2903                 return -EOPNOTSUPP;
2904
2905         kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2906         if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2907                 return 0;
2908
2909         err = ldiskfs_get_inode_loc(inode, &iloc);
2910         if (err)
2911                 return err;
2912
2913         raw_inode = ldiskfs_raw_inode(&iloc);
2914         if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2915                 err = -EOVERFLOW;
2916                 brelse(iloc.bh);
2917                 return err;
2918         }
2919         brelse(iloc.bh);
2920
2921         dquot_initialize(inode);
2922         transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2923         if (transfer_to[PRJQUOTA]) {
2924                 err = __dquot_transfer(inode, transfer_to);
2925                 dqput(transfer_to[PRJQUOTA]);
2926                 if (err)
2927                         return err;
2928         }
2929
2930         return err;
2931 }
2932 #endif
2933
2934 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2935 {
2936         int rc;
2937
2938         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2939             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2940                 struct iattr iattr;
2941
2942                 dquot_initialize(inode);
2943                 iattr.ia_valid = 0;
2944                 if (attr->la_valid & LA_UID)
2945                         iattr.ia_valid |= ATTR_UID;
2946                 if (attr->la_valid & LA_GID)
2947                         iattr.ia_valid |= ATTR_GID;
2948                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2949                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2950
2951                 rc = dquot_transfer(inode, &iattr);
2952                 if (rc) {
2953                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2954                                "enforcement enabled on the ldiskfs "
2955                                "filesystem?\n", inode->i_sb->s_id, rc);
2956                         return rc;
2957                 }
2958         }
2959
2960         /* Handle project id transfer here properly */
2961         if (attr->la_valid & LA_PROJID &&
2962             attr->la_projid != i_projid_read(inode)) {
2963 #ifdef HAVE_PROJECT_QUOTA
2964                 rc = osd_transfer_project(inode, attr->la_projid);
2965 #else
2966                 rc = -ENOTSUPP;
2967 #endif
2968                 if (rc) {
2969                         CERROR("%s: quota transfer failed: rc = %d. Is project "
2970                                "enforcement enabled on the ldiskfs "
2971                                "filesystem?\n", inode->i_sb->s_id, rc);
2972                         return rc;
2973                 }
2974         }
2975         return 0;
2976 }
2977
2978 static int osd_attr_set(const struct lu_env *env,
2979                         struct dt_object *dt,
2980                         const struct lu_attr *attr,
2981                         struct thandle *handle)
2982 {
2983         struct osd_object *obj = osd_dt_obj(dt);
2984         struct inode *inode;
2985         int rc;
2986
2987         if (!dt_object_exists(dt))
2988                 return -ENOENT;
2989
2990         LASSERT(handle != NULL);
2991         LASSERT(!dt_object_remote(dt));
2992         LASSERT(osd_invariant(obj));
2993
2994         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2995
2996         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
2997             !osd_obj2dev(obj)->od_is_ost) {
2998                 struct osd_thread_info *oti = osd_oti_get(env);
2999                 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3000                 struct lu_fid *fid1 = &oti->oti_fid;
3001                 struct osd_inode_id *id = &oti->oti_id;
3002                 struct iam_path_descr *ipd;
3003                 struct iam_container *bag;
3004                 struct osd_thandle *oh;
3005                 int rc;
3006
3007                 fid_cpu_to_be(fid1, fid0);
3008                 memset(id, 1, sizeof(*id));
3009                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3010                                   fid0)->oi_dir.od_container;
3011                 ipd = osd_idx_ipd_get(env, bag);
3012                 if (unlikely(ipd == NULL))
3013                         RETURN(-ENOMEM);
3014
3015                 oh = container_of0(handle, struct osd_thandle, ot_super);
3016                 rc = iam_update(oh->ot_handle, bag,
3017                                 (const struct iam_key *)fid1,
3018                                 (const struct iam_rec *)id, ipd);
3019                 osd_ipd_put(env, bag, ipd);
3020                 return(rc > 0 ? 0 : rc);
3021         }
3022
3023         inode = obj->oo_inode;
3024
3025         rc = osd_quota_transfer(inode, attr);
3026         if (rc)
3027                 return rc;
3028
3029         spin_lock(&obj->oo_guard);
3030         rc = osd_inode_setattr(env, inode, attr);
3031         spin_unlock(&obj->oo_guard);
3032         if (rc != 0)
3033                 GOTO(out, rc);
3034
3035         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3036
3037         if (!(attr->la_valid & LA_FLAGS))
3038                 GOTO(out, rc);
3039
3040         /* Let's check if there are extra flags need to be set into LMA */
3041         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3042                 struct osd_thread_info *info = osd_oti_get(env);
3043                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3044
3045                 LASSERT(!obj->oo_pfid_in_lma);
3046
3047                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3048                                  &info->oti_ost_attrs);
3049                 if (rc)
3050                         GOTO(out, rc);
3051
3052                 lma->lma_incompat |=
3053                         lustre_to_lma_flags(attr->la_flags);
3054                 lustre_lma_swab(lma);
3055                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3056                                      lma, sizeof(*lma), XATTR_REPLACE);
3057                 if (rc != 0) {
3058                         struct osd_device *osd = osd_obj2dev(obj);
3059
3060                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3061                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3062                               lma->lma_incompat, rc);
3063                 } else {
3064                         obj->oo_lma_flags =
3065                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3066                 }
3067                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3068         }
3069 out:
3070         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3071
3072         return rc;
3073 }
3074
3075 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3076                                            struct osd_object *obj,
3077                                            const char *name, const int namelen)
3078 {
3079         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3080 }
3081
3082 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3083                       umode_t mode, struct dt_allocation_hint *hint,
3084                       struct thandle *th, struct lu_attr *attr)
3085 {
3086         int result;
3087         struct osd_device *osd = osd_obj2dev(obj);
3088         struct osd_thandle *oth;
3089         struct dt_object *parent = NULL;
3090         struct inode *inode;
3091         uid_t owner[2] = {0, 0};
3092
3093         if (attr->la_valid & LA_UID)
3094                 owner[0] = attr->la_uid;
3095         if (attr->la_valid & LA_GID)
3096                 owner[1] = attr->la_gid;
3097
3098         LINVRNT(osd_invariant(obj));
3099         LASSERT(obj->oo_inode == NULL);
3100         LASSERT(obj->oo_hl_head == NULL);
3101
3102         if (S_ISDIR(mode) && ldiskfs_pdo) {
3103                 obj->oo_hl_head =
3104                         ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3105                 if (obj->oo_hl_head == NULL)
3106                         return -ENOMEM;
3107         }
3108
3109         oth = container_of(th, struct osd_thandle, ot_super);
3110         LASSERT(oth->ot_handle->h_transaction != NULL);
3111
3112         if (hint != NULL && hint->dah_parent != NULL &&
3113             !dt_object_remote(hint->dah_parent))
3114                 parent = hint->dah_parent;
3115
3116         inode = ldiskfs_create_inode(oth->ot_handle,
3117                                      parent ? osd_dt_obj(parent)->oo_inode :
3118                                               osd_sb(osd)->s_root->d_inode,
3119                                      mode, owner);
3120         if (!IS_ERR(inode)) {
3121                 /* Do not update file c/mtime in ldiskfs. */
3122                 inode->i_flags |= S_NOCMTIME;
3123
3124                 /*
3125                  * For new created object, it must be consistent,
3126                  * and it is unnecessary to scrub against it.
3127                  */
3128                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3129
3130                 obj->oo_inode = inode;
3131                 result = 0;
3132         } else {
3133                 if (obj->oo_hl_head != NULL) {
3134                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3135                         obj->oo_hl_head = NULL;
3136                 }
3137                 result = PTR_ERR(inode);
3138         }
3139         LINVRNT(osd_invariant(obj));
3140         return result;
3141 }
3142
3143 enum {
3144         OSD_NAME_LEN = 255
3145 };
3146
3147 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3148                      struct lu_attr *attr,
3149                      struct dt_allocation_hint *hint,
3150                      struct dt_object_format *dof,
3151                      struct thandle *th)
3152 {
3153         int result;
3154         struct osd_thandle *oth;
3155         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3156
3157         LASSERT(S_ISDIR(attr->la_mode));
3158
3159         oth = container_of(th, struct osd_thandle, ot_super);
3160         LASSERT(oth->ot_handle->h_transaction != NULL);
3161         result = osd_mkfile(info, obj, mode, hint, th, attr);
3162
3163         return result;
3164 }
3165
3166 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3167                         struct lu_attr *attr,
3168                         struct dt_allocation_hint *hint,
3169                         struct dt_object_format *dof,
3170                         struct thandle *th)
3171 {
3172         int result;
3173         struct osd_thandle *oth;
3174         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3175
3176         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3177
3178         LASSERT(S_ISREG(attr->la_mode));
3179
3180         oth = container_of(th, struct osd_thandle, ot_super);
3181         LASSERT(oth->ot_handle->h_transaction != NULL);
3182
3183         result = osd_mkfile(info, obj, mode, hint, th, attr);
3184         if (result == 0) {
3185                 LASSERT(obj->oo_inode != NULL);
3186                 if (feat->dif_flags & DT_IND_VARKEY)
3187                         result = iam_lvar_create(obj->oo_inode,