Whamcloud - gitweb
5db753bfb13875b9dd04f16d19f4b0e596e26565
[fs/lustre-release.git] / lustre / lfsck / lfsck_namespace.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,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License version 2 for more details.  A copy is
14  * included in the COPYING file that accompanied this code.
15
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write to the Free Software
18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2013, 2015, Intel Corporation.
24  */
25 /*
26  * lustre/lfsck/lfsck_namespace.c
27  *
28  * Author: Fan, Yong <fan.yong@intel.com>
29  */
30
31 #define DEBUG_SUBSYSTEM S_LFSCK
32
33 #include <lustre/lustre_idl.h>
34 #include <lu_object.h>
35 #include <dt_object.h>
36 #include <md_object.h>
37 #include <lustre_fid.h>
38 #include <lustre_lib.h>
39 #include <lustre_net.h>
40 #include <lustre/lustre_user.h>
41
42 #include "lfsck_internal.h"
43
44 #define LFSCK_NAMESPACE_MAGIC_V1        0xA0629D03
45 #define LFSCK_NAMESPACE_MAGIC_V2        0xA0621A0B
46
47 /* For Lustre-2.x (x <= 6), the namespace LFSCK used LFSCK_NAMESPACE_MAGIC_V1
48  * as the trace file magic. When downgrade to such old release, the old LFSCK
49  * will not recognize the new LFSCK_NAMESPACE_MAGIC_V2 in the new trace file,
50  * then it will reset the whole LFSCK, and will not cause start failure. The
51  * similar case will happen when upgrade from such old release. */
52 #define LFSCK_NAMESPACE_MAGIC           LFSCK_NAMESPACE_MAGIC_V2
53
54 enum lfsck_nameentry_check {
55         LFSCK_NAMEENTRY_DEAD            = 1, /* The object has been unlinked. */
56         LFSCK_NAMEENTRY_REMOVED         = 2, /* The entry has been removed. */
57         LFSCK_NAMEENTRY_RECREATED       = 3, /* The entry has been recreated. */
58 };
59
60 static struct lfsck_namespace_req *
61 lfsck_namespace_assistant_req_init(struct lfsck_instance *lfsck,
62                                    struct lfsck_assistant_object *lso,
63                                    struct lu_dirent *ent, __u16 type)
64 {
65         struct lfsck_namespace_req *lnr;
66         int                         size;
67
68         size = sizeof(*lnr) + (ent->lde_namelen & ~3) + 4;
69         OBD_ALLOC(lnr, size);
70         if (lnr == NULL)
71                 return ERR_PTR(-ENOMEM);
72
73         INIT_LIST_HEAD(&lnr->lnr_lar.lar_list);
74         lnr->lnr_lar.lar_parent = lfsck_assistant_object_get(lso);
75         lnr->lnr_lmv = lfsck_lmv_get(lfsck->li_lmv);
76         lnr->lnr_fid = ent->lde_fid;
77         lnr->lnr_dir_cookie = ent->lde_hash;
78         lnr->lnr_attr = ent->lde_attrs;
79         lnr->lnr_size = size;
80         lnr->lnr_type = type;
81         lnr->lnr_namelen = ent->lde_namelen;
82         memcpy(lnr->lnr_name, ent->lde_name, ent->lde_namelen);
83
84         return lnr;
85 }
86
87 static void lfsck_namespace_assistant_req_fini(const struct lu_env *env,
88                                                struct lfsck_assistant_req *lar)
89 {
90         struct lfsck_namespace_req *lnr =
91                         container_of0(lar, struct lfsck_namespace_req, lnr_lar);
92
93         if (lnr->lnr_lmv != NULL)
94                 lfsck_lmv_put(env, lnr->lnr_lmv);
95
96         lfsck_assistant_object_put(env, lar->lar_parent);
97         OBD_FREE(lnr, lnr->lnr_size);
98 }
99
100 static void lfsck_namespace_le_to_cpu(struct lfsck_namespace *dst,
101                                       struct lfsck_namespace *src)
102 {
103         dst->ln_magic = le32_to_cpu(src->ln_magic);
104         dst->ln_status = le32_to_cpu(src->ln_status);
105         dst->ln_flags = le32_to_cpu(src->ln_flags);
106         dst->ln_success_count = le32_to_cpu(src->ln_success_count);
107         dst->ln_run_time_phase1 = le32_to_cpu(src->ln_run_time_phase1);
108         dst->ln_run_time_phase2 = le32_to_cpu(src->ln_run_time_phase2);
109         dst->ln_time_last_complete = le64_to_cpu(src->ln_time_last_complete);
110         dst->ln_time_latest_start = le64_to_cpu(src->ln_time_latest_start);
111         dst->ln_time_last_checkpoint =
112                                 le64_to_cpu(src->ln_time_last_checkpoint);
113         lfsck_position_le_to_cpu(&dst->ln_pos_latest_start,
114                                  &src->ln_pos_latest_start);
115         lfsck_position_le_to_cpu(&dst->ln_pos_last_checkpoint,
116                                  &src->ln_pos_last_checkpoint);
117         lfsck_position_le_to_cpu(&dst->ln_pos_first_inconsistent,
118                                  &src->ln_pos_first_inconsistent);
119         dst->ln_items_checked = le64_to_cpu(src->ln_items_checked);
120         dst->ln_items_repaired = le64_to_cpu(src->ln_items_repaired);
121         dst->ln_items_failed = le64_to_cpu(src->ln_items_failed);
122         dst->ln_dirs_checked = le64_to_cpu(src->ln_dirs_checked);
123         dst->ln_objs_checked_phase2 = le64_to_cpu(src->ln_objs_checked_phase2);
124         dst->ln_objs_repaired_phase2 =
125                                 le64_to_cpu(src->ln_objs_repaired_phase2);
126         dst->ln_objs_failed_phase2 = le64_to_cpu(src->ln_objs_failed_phase2);
127         dst->ln_objs_nlink_repaired = le64_to_cpu(src->ln_objs_nlink_repaired);
128         fid_le_to_cpu(&dst->ln_fid_latest_scanned_phase2,
129                       &src->ln_fid_latest_scanned_phase2);
130         dst->ln_dirent_repaired = le64_to_cpu(src->ln_dirent_repaired);
131         dst->ln_linkea_repaired = le64_to_cpu(src->ln_linkea_repaired);
132         dst->ln_mul_linked_checked = le64_to_cpu(src->ln_mul_linked_checked);
133         dst->ln_mul_linked_repaired = le64_to_cpu(src->ln_mul_linked_repaired);
134         dst->ln_unknown_inconsistency =
135                                 le64_to_cpu(src->ln_unknown_inconsistency);
136         dst->ln_unmatched_pairs_repaired =
137                                 le64_to_cpu(src->ln_unmatched_pairs_repaired);
138         dst->ln_dangling_repaired = le64_to_cpu(src->ln_dangling_repaired);
139         dst->ln_mul_ref_repaired = le64_to_cpu(src->ln_mul_ref_repaired);
140         dst->ln_bad_type_repaired = le64_to_cpu(src->ln_bad_type_repaired);
141         dst->ln_lost_dirent_repaired =
142                                 le64_to_cpu(src->ln_lost_dirent_repaired);
143         dst->ln_striped_dirs_scanned =
144                                 le64_to_cpu(src->ln_striped_dirs_scanned);
145         dst->ln_striped_dirs_repaired =
146                                 le64_to_cpu(src->ln_striped_dirs_repaired);
147         dst->ln_striped_dirs_failed =
148                                 le64_to_cpu(src->ln_striped_dirs_failed);
149         dst->ln_striped_dirs_disabled =
150                                 le64_to_cpu(src->ln_striped_dirs_disabled);
151         dst->ln_striped_dirs_skipped =
152                                 le64_to_cpu(src->ln_striped_dirs_skipped);
153         dst->ln_striped_shards_scanned =
154                                 le64_to_cpu(src->ln_striped_shards_scanned);
155         dst->ln_striped_shards_repaired =
156                                 le64_to_cpu(src->ln_striped_shards_repaired);
157         dst->ln_striped_shards_failed =
158                                 le64_to_cpu(src->ln_striped_shards_failed);
159         dst->ln_striped_shards_skipped =
160                                 le64_to_cpu(src->ln_striped_shards_skipped);
161         dst->ln_name_hash_repaired = le64_to_cpu(src->ln_name_hash_repaired);
162         dst->ln_local_lpf_scanned = le64_to_cpu(src->ln_local_lpf_scanned);
163         dst->ln_local_lpf_moved = le64_to_cpu(src->ln_local_lpf_moved);
164         dst->ln_local_lpf_skipped = le64_to_cpu(src->ln_local_lpf_skipped);
165         dst->ln_local_lpf_failed = le64_to_cpu(src->ln_local_lpf_failed);
166         dst->ln_bitmap_size = le32_to_cpu(src->ln_bitmap_size);
167 }
168
169 static void lfsck_namespace_cpu_to_le(struct lfsck_namespace *dst,
170                                       struct lfsck_namespace *src)
171 {
172         dst->ln_magic = cpu_to_le32(src->ln_magic);
173         dst->ln_status = cpu_to_le32(src->ln_status);
174         dst->ln_flags = cpu_to_le32(src->ln_flags);
175         dst->ln_success_count = cpu_to_le32(src->ln_success_count);
176         dst->ln_run_time_phase1 = cpu_to_le32(src->ln_run_time_phase1);
177         dst->ln_run_time_phase2 = cpu_to_le32(src->ln_run_time_phase2);
178         dst->ln_time_last_complete = cpu_to_le64(src->ln_time_last_complete);
179         dst->ln_time_latest_start = cpu_to_le64(src->ln_time_latest_start);
180         dst->ln_time_last_checkpoint =
181                                 cpu_to_le64(src->ln_time_last_checkpoint);
182         lfsck_position_cpu_to_le(&dst->ln_pos_latest_start,
183                                  &src->ln_pos_latest_start);
184         lfsck_position_cpu_to_le(&dst->ln_pos_last_checkpoint,
185                                  &src->ln_pos_last_checkpoint);
186         lfsck_position_cpu_to_le(&dst->ln_pos_first_inconsistent,
187                                  &src->ln_pos_first_inconsistent);
188         dst->ln_items_checked = cpu_to_le64(src->ln_items_checked);
189         dst->ln_items_repaired = cpu_to_le64(src->ln_items_repaired);
190         dst->ln_items_failed = cpu_to_le64(src->ln_items_failed);
191         dst->ln_dirs_checked = cpu_to_le64(src->ln_dirs_checked);
192         dst->ln_objs_checked_phase2 = cpu_to_le64(src->ln_objs_checked_phase2);
193         dst->ln_objs_repaired_phase2 =
194                                 cpu_to_le64(src->ln_objs_repaired_phase2);
195         dst->ln_objs_failed_phase2 = cpu_to_le64(src->ln_objs_failed_phase2);
196         dst->ln_objs_nlink_repaired = cpu_to_le64(src->ln_objs_nlink_repaired);
197         fid_cpu_to_le(&dst->ln_fid_latest_scanned_phase2,
198                       &src->ln_fid_latest_scanned_phase2);
199         dst->ln_dirent_repaired = cpu_to_le64(src->ln_dirent_repaired);
200         dst->ln_linkea_repaired = cpu_to_le64(src->ln_linkea_repaired);
201         dst->ln_mul_linked_checked = cpu_to_le64(src->ln_mul_linked_checked);
202         dst->ln_mul_linked_repaired = cpu_to_le64(src->ln_mul_linked_repaired);
203         dst->ln_unknown_inconsistency =
204                                 cpu_to_le64(src->ln_unknown_inconsistency);
205         dst->ln_unmatched_pairs_repaired =
206                                 cpu_to_le64(src->ln_unmatched_pairs_repaired);
207         dst->ln_dangling_repaired = cpu_to_le64(src->ln_dangling_repaired);
208         dst->ln_mul_ref_repaired = cpu_to_le64(src->ln_mul_ref_repaired);
209         dst->ln_bad_type_repaired = cpu_to_le64(src->ln_bad_type_repaired);
210         dst->ln_lost_dirent_repaired =
211                                 cpu_to_le64(src->ln_lost_dirent_repaired);
212         dst->ln_striped_dirs_scanned =
213                                 cpu_to_le64(src->ln_striped_dirs_scanned);
214         dst->ln_striped_dirs_repaired =
215                                 cpu_to_le64(src->ln_striped_dirs_repaired);
216         dst->ln_striped_dirs_failed =
217                                 cpu_to_le64(src->ln_striped_dirs_failed);
218         dst->ln_striped_dirs_disabled =
219                                 cpu_to_le64(src->ln_striped_dirs_disabled);
220         dst->ln_striped_dirs_skipped =
221                                 cpu_to_le64(src->ln_striped_dirs_skipped);
222         dst->ln_striped_shards_scanned =
223                                 cpu_to_le64(src->ln_striped_shards_scanned);
224         dst->ln_striped_shards_repaired =
225                                 cpu_to_le64(src->ln_striped_shards_repaired);
226         dst->ln_striped_shards_failed =
227                                 cpu_to_le64(src->ln_striped_shards_failed);
228         dst->ln_striped_shards_skipped =
229                                 cpu_to_le64(src->ln_striped_shards_skipped);
230         dst->ln_name_hash_repaired = cpu_to_le64(src->ln_name_hash_repaired);
231         dst->ln_local_lpf_scanned = cpu_to_le64(src->ln_local_lpf_scanned);
232         dst->ln_local_lpf_moved = cpu_to_le64(src->ln_local_lpf_moved);
233         dst->ln_local_lpf_skipped = cpu_to_le64(src->ln_local_lpf_skipped);
234         dst->ln_local_lpf_failed = cpu_to_le64(src->ln_local_lpf_failed);
235         dst->ln_bitmap_size = cpu_to_le32(src->ln_bitmap_size);
236 }
237
238 static void lfsck_namespace_record_failure(const struct lu_env *env,
239                                            struct lfsck_instance *lfsck,
240                                            struct lfsck_namespace *ns)
241 {
242         struct lfsck_position pos;
243
244         ns->ln_items_failed++;
245         lfsck_pos_fill(env, lfsck, &pos, false);
246         if (lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent) ||
247             lfsck_pos_is_eq(&pos, &ns->ln_pos_first_inconsistent) < 0) {
248                 ns->ln_pos_first_inconsistent = pos;
249
250                 CDEBUG(D_LFSCK, "%s: namespace LFSCK hit first non-repaired "
251                        "inconsistency at the pos ["LPU64", "DFID", "LPX64"]\n",
252                        lfsck_lfsck2name(lfsck),
253                        ns->ln_pos_first_inconsistent.lp_oit_cookie,
254                        PFID(&ns->ln_pos_first_inconsistent.lp_dir_parent),
255                        ns->ln_pos_first_inconsistent.lp_dir_cookie);
256         }
257 }
258
259 /**
260  * Load the MDT bitmap from the lfsck_namespace trace file.
261  *
262  * \param[in] env       pointer to the thread context
263  * \param[in] com       pointer to the lfsck component
264  *
265  * \retval              0 for success
266  * \retval              negative error number on failure or data corruption
267  */
268 static int lfsck_namespace_load_bitmap(const struct lu_env *env,
269                                        struct lfsck_component *com)
270 {
271         struct dt_object                *obj    = com->lc_obj;
272         struct lfsck_assistant_data     *lad    = com->lc_data;
273         struct lfsck_namespace          *ns     = com->lc_file_ram;
274         cfs_bitmap_t                    *bitmap = lad->lad_bitmap;
275         ssize_t                          size;
276         __u32                            nbits;
277         int                              rc;
278         ENTRY;
279
280         if (com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size >
281             ns->ln_bitmap_size)
282                 nbits = com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size;
283         else
284                 nbits = ns->ln_bitmap_size;
285
286         if (unlikely(nbits < BITS_PER_LONG))
287                 nbits = BITS_PER_LONG;
288
289         if (nbits > bitmap->size) {
290                 __u32 new_bits = bitmap->size;
291                 cfs_bitmap_t *new_bitmap;
292
293                 while (new_bits < nbits)
294                         new_bits <<= 1;
295
296                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
297                 if (new_bitmap == NULL)
298                         RETURN(-ENOMEM);
299
300                 lad->lad_bitmap = new_bitmap;
301                 CFS_FREE_BITMAP(bitmap);
302                 bitmap = new_bitmap;
303         }
304
305         if (ns->ln_bitmap_size == 0) {
306                 lad->lad_incomplete = 0;
307                 CFS_RESET_BITMAP(bitmap);
308
309                 RETURN(0);
310         }
311
312         size = (ns->ln_bitmap_size + 7) >> 3;
313         rc = dt_xattr_get(env, obj,
314                           lfsck_buf_get(env, bitmap->data, size),
315                           XATTR_NAME_LFSCK_BITMAP);
316         if (rc != size)
317                 RETURN(rc >= 0 ? -EINVAL : rc);
318
319         if (cfs_bitmap_check_empty(bitmap))
320                 lad->lad_incomplete = 0;
321         else
322                 lad->lad_incomplete = 1;
323
324         RETURN(0);
325 }
326
327 /**
328  * Load namespace LFSCK statistics information from the trace file.
329  *
330  * For old release (Lustre-2.6 or older), the statistics information was
331  * stored as XATTR_NAME_LFSCK_NAMESPACE_OLD EA. But in Lustre-2.7, we need
332  * more statistics information. To avoid confusing old MDT when downgrade,
333  * Lustre-2.7 stores the namespace LFSCK statistics information as new
334  * XATTR_NAME_LFSCK_NAMESPACE EA.
335  *
336  * \param[in] env       pointer to the thread context
337  * \param[in] com       pointer to the lfsck component
338  *
339  * \retval              0 for success
340  * \retval              negative error number on failure
341  */
342 static int lfsck_namespace_load(const struct lu_env *env,
343                                 struct lfsck_component *com)
344 {
345         int len = com->lc_file_size;
346         int rc;
347
348         rc = dt_xattr_get(env, com->lc_obj,
349                           lfsck_buf_get(env, com->lc_file_disk, len),
350                           XATTR_NAME_LFSCK_NAMESPACE);
351         if (rc == len) {
352                 struct lfsck_namespace *ns = com->lc_file_ram;
353
354                 lfsck_namespace_le_to_cpu(ns,
355                                 (struct lfsck_namespace *)com->lc_file_disk);
356                 if (ns->ln_magic != LFSCK_NAMESPACE_MAGIC) {
357                         CDEBUG(D_LFSCK, "%s: invalid lfsck_namespace magic "
358                                "%#x != %#x\n", lfsck_lfsck2name(com->lc_lfsck),
359                                ns->ln_magic, LFSCK_NAMESPACE_MAGIC);
360                         rc = -ESTALE;
361                 } else {
362                         rc = 0;
363                 }
364         } else if (rc != -ENODATA) {
365                 CDEBUG(D_LFSCK, "%s: fail to load lfsck_namespace, "
366                        "expected = %d: rc = %d\n",
367                        lfsck_lfsck2name(com->lc_lfsck), len, rc);
368                 if (rc >= 0)
369                         rc = -ESTALE;
370         } else {
371                 /* Check whether it is old trace file or not.
372                  * If yes, it should be reset via returning -ESTALE. */
373                 rc = dt_xattr_get(env, com->lc_obj,
374                                   lfsck_buf_get(env, com->lc_file_disk, len),
375                                   XATTR_NAME_LFSCK_NAMESPACE_OLD);
376                 if (rc >= 0)
377                         rc = -ESTALE;
378         }
379
380         return rc;
381 }
382
383 static int lfsck_namespace_store(const struct lu_env *env,
384                                  struct lfsck_component *com, bool init)
385 {
386         struct dt_object                *obj    = com->lc_obj;
387         struct lfsck_instance           *lfsck  = com->lc_lfsck;
388         struct lfsck_namespace          *ns     = com->lc_file_ram;
389         struct lfsck_assistant_data     *lad    = com->lc_data;
390         struct dt_device                *dev    = lfsck_obj2dev(obj);
391         cfs_bitmap_t                    *bitmap = NULL;
392         struct thandle                  *handle;
393         __u32                            nbits  = 0;
394         int                              len    = com->lc_file_size;
395         int                              rc;
396 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
397         struct lu_buf            tbuf   = { &len, sizeof(len) };
398 #endif
399         ENTRY;
400
401         if (lad != NULL) {
402                 bitmap = lad->lad_bitmap;
403                 nbits = bitmap->size;
404
405                 LASSERT(nbits > 0);
406                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
407         }
408
409         ns->ln_bitmap_size = nbits;
410         lfsck_namespace_cpu_to_le((struct lfsck_namespace *)com->lc_file_disk,
411                                   ns);
412         handle = dt_trans_create(env, dev);
413         if (IS_ERR(handle))
414                 GOTO(log, rc = PTR_ERR(handle));
415
416         rc = dt_declare_xattr_set(env, obj,
417                                   lfsck_buf_get(env, com->lc_file_disk, len),
418                                   XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
419         if (rc != 0)
420                 GOTO(out, rc);
421
422         if (bitmap != NULL) {
423                 rc = dt_declare_xattr_set(env, obj,
424                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
425                                 XATTR_NAME_LFSCK_BITMAP, 0, handle);
426                 if (rc != 0)
427                         GOTO(out, rc);
428         }
429
430 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
431         /* To be compatible with old Lustre-2.x MDT (x <= 6), generate dummy
432          * XATTR_NAME_LFSCK_NAMESPACE_OLD EA, then when downgrade to Lustre-2.x,
433          * the old LFSCK will find "invalid" XATTR_NAME_LFSCK_NAMESPACE_OLD EA,
434          * then reset the namespace LFSCK trace file. */
435         if (init) {
436                 rc = dt_declare_xattr_set(env, obj, &tbuf,
437                                           XATTR_NAME_LFSCK_NAMESPACE_OLD,
438                                           LU_XATTR_CREATE, handle);
439                 if (rc != 0)
440                         GOTO(out, rc);
441         }
442 #endif
443
444         rc = dt_trans_start_local(env, dev, handle);
445         if (rc != 0)
446                 GOTO(out, rc);
447
448         rc = dt_xattr_set(env, obj,
449                           lfsck_buf_get(env, com->lc_file_disk, len),
450                           XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
451         if (rc == 0 && bitmap != NULL)
452                 rc = dt_xattr_set(env, obj,
453                                   lfsck_buf_get(env, bitmap->data, nbits >> 3),
454                                   XATTR_NAME_LFSCK_BITMAP, 0, handle);
455
456 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
457         if (rc == 0 && init)
458                 rc = dt_xattr_set(env, obj, &tbuf,
459                                   XATTR_NAME_LFSCK_NAMESPACE_OLD,
460                                   LU_XATTR_CREATE, handle);
461 #endif
462
463         GOTO(out, rc);
464
465 out:
466         dt_trans_stop(env, dev, handle);
467
468 log:
469         if (rc != 0)
470                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_namespace: rc = %d\n",
471                        lfsck_lfsck2name(lfsck), rc);
472         return rc;
473 }
474
475 static struct dt_object *
476 lfsck_namespace_load_one_trace_file(const struct lu_env *env,
477                                     struct lfsck_component *com,
478                                     struct dt_object *parent,
479                                     const char *name, bool reset)
480 {
481         struct lfsck_instance   *lfsck = com->lc_lfsck;
482         struct dt_object        *obj;
483         int                      rc;
484
485         if (reset) {
486                 rc = local_object_unlink(env, lfsck->li_bottom, parent, name);
487                 if (rc != 0 && rc != -ENOENT)
488                         return ERR_PTR(rc);
489         }
490
491         obj = local_index_find_or_create(env, lfsck->li_los, parent, name,
492                                          S_IFREG | S_IRUGO | S_IWUSR,
493                                          &dt_lfsck_features);
494
495         return obj;
496 }
497
498 static int lfsck_namespace_load_sub_trace_files(const struct lu_env *env,
499                                                 struct lfsck_component *com,
500                                                 bool reset)
501 {
502         char                            *name = lfsck_env_info(env)->lti_key;
503         struct lfsck_sub_trace_obj      *lsto;
504         struct dt_object                *obj;
505         int                              rc;
506         int                              i;
507
508         for (i = 0, lsto = &com->lc_sub_trace_objs[0];
509              i < LFSCK_STF_COUNT; i++, lsto++) {
510                 snprintf(name, NAME_MAX, "%s_%02d", LFSCK_NAMESPACE, i);
511                 if (lsto->lsto_obj != NULL) {
512                         if (!reset)
513                                 continue;
514
515                         lfsck_object_put(env, lsto->lsto_obj);
516                         lsto->lsto_obj = NULL;
517                 }
518
519                 obj = lfsck_namespace_load_one_trace_file(env, com,
520                                 com->lc_lfsck->li_lfsck_dir, name, reset);
521                 if (IS_ERR(obj))
522                         return PTR_ERR(obj);
523
524                 lsto->lsto_obj = obj;
525                 rc = obj->do_ops->do_index_try(env, obj, &dt_lfsck_features);
526                 if (rc != 0)
527                         return rc;
528         }
529
530         return 0;
531 }
532
533 static int lfsck_namespace_init(const struct lu_env *env,
534                                 struct lfsck_component *com)
535 {
536         struct lfsck_namespace *ns = com->lc_file_ram;
537         int rc;
538
539         memset(ns, 0, sizeof(*ns));
540         ns->ln_magic = LFSCK_NAMESPACE_MAGIC;
541         ns->ln_status = LS_INIT;
542         down_write(&com->lc_sem);
543         rc = lfsck_namespace_store(env, com, true);
544         up_write(&com->lc_sem);
545         if (rc == 0)
546                 rc = lfsck_namespace_load_sub_trace_files(env, com, true);
547
548         return rc;
549 }
550
551 /**
552  * Update the namespace LFSCK trace file for the given @fid
553  *
554  * \param[in] env       pointer to the thread context
555  * \param[in] com       pointer to the lfsck component
556  * \param[in] fid       the fid which flags to be updated in the lfsck
557  *                      trace file
558  * \param[in] add       true if add new flags, otherwise remove flags
559  *
560  * \retval              0 for success or nothing to be done
561  * \retval              negative error number on failure
562  */
563 int lfsck_namespace_trace_update(const struct lu_env *env,
564                                  struct lfsck_component *com,
565                                  const struct lu_fid *fid,
566                                  const __u8 flags, bool add)
567 {
568         struct lfsck_instance   *lfsck  = com->lc_lfsck;
569         struct dt_object        *obj;
570         struct lu_fid           *key    = &lfsck_env_info(env)->lti_fid3;
571         struct dt_device        *dev;
572         struct thandle          *th     = NULL;
573         int                      idx;
574         int                      rc     = 0;
575         __u8                     old    = 0;
576         __u8                     new    = 0;
577         ENTRY;
578
579         LASSERT(flags != 0);
580
581         if (unlikely(!fid_is_sane(fid)))
582                 RETURN(0);
583
584         idx = lfsck_sub_trace_file_fid2idx(fid);
585         obj = com->lc_sub_trace_objs[idx].lsto_obj;
586         dev = lfsck_obj2dev(obj);
587         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
588         fid_cpu_to_be(key, fid);
589         rc = dt_lookup(env, obj, (struct dt_rec *)&old,
590                        (const struct dt_key *)key);
591         if (rc == -ENOENT) {
592                 if (!add)
593                         GOTO(unlock, rc = 0);
594
595                 old = 0;
596                 new = flags;
597         } else if (rc == 0) {
598                 if (add) {
599                         if ((old & flags) == flags)
600                                 GOTO(unlock, rc = 0);
601
602                         new = old | flags;
603                 } else {
604                         if ((old & flags) == 0)
605                                 GOTO(unlock, rc = 0);
606
607                         new = old & ~flags;
608                 }
609         } else {
610                 GOTO(log, rc);
611         }
612
613         th = dt_trans_create(env, dev);
614         if (IS_ERR(th))
615                 GOTO(log, rc = PTR_ERR(th));
616
617         if (old != 0) {
618                 rc = dt_declare_delete(env, obj,
619                                        (const struct dt_key *)key, th);
620                 if (rc != 0)
621                         GOTO(log, rc);
622         }
623
624         if (new != 0) {
625                 rc = dt_declare_insert(env, obj,
626                                        (const struct dt_rec *)&new,
627                                        (const struct dt_key *)key, th);
628                 if (rc != 0)
629                         GOTO(log, rc);
630         }
631
632         rc = dt_trans_start_local(env, dev, th);
633         if (rc != 0)
634                 GOTO(log, rc);
635
636         if (old != 0) {
637                 rc = dt_delete(env, obj, (const struct dt_key *)key, th);
638                 if (rc != 0)
639                         GOTO(log, rc);
640         }
641
642         if (new != 0) {
643                 rc = dt_insert(env, obj, (const struct dt_rec *)&new,
644                                (const struct dt_key *)key, th, 1);
645                 if (rc != 0)
646                         GOTO(log, rc);
647         }
648
649         GOTO(log, rc);
650
651 log:
652         if (th != NULL && !IS_ERR(th))
653                 dt_trans_stop(env, dev, th);
654
655         CDEBUG(D_LFSCK, "%s: namespace LFSCK %s flags for "DFID" in the "
656                "trace file, flags %x, old %x, new %x: rc = %d\n",
657                lfsck_lfsck2name(lfsck), add ? "add" : "del", PFID(fid),
658                (__u32)flags, (__u32)old, (__u32)new, rc);
659
660 unlock:
661         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
662
663         return rc;
664 }
665
666 int lfsck_namespace_check_exist(const struct lu_env *env,
667                                 struct dt_object *dir,
668                                 struct dt_object *obj, const char *name)
669 {
670         struct lu_fid    *fid = &lfsck_env_info(env)->lti_fid;
671         int               rc;
672         ENTRY;
673
674         if (unlikely(lfsck_is_dead_obj(obj)))
675                 RETURN(LFSCK_NAMEENTRY_DEAD);
676
677         rc = dt_lookup(env, dir, (struct dt_rec *)fid,
678                        (const struct dt_key *)name);
679         if (rc == -ENOENT)
680                 RETURN(LFSCK_NAMEENTRY_REMOVED);
681
682         if (rc < 0)
683                 RETURN(rc);
684
685         if (!lu_fid_eq(fid, lfsck_dto2fid(obj)))
686                 RETURN(LFSCK_NAMEENTRY_RECREATED);
687
688         RETURN(0);
689 }
690
691 static int lfsck_declare_namespace_exec_dir(const struct lu_env *env,
692                                             struct dt_object *obj,
693                                             struct thandle *handle)
694 {
695         int rc;
696
697         /* For destroying all invalid linkEA entries. */
698         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, handle);
699         if (rc != 0)
700                 return rc;
701
702         /* For insert new linkEA entry. */
703         rc = dt_declare_xattr_set(env, obj,
704                         lfsck_buf_get_const(env, NULL, DEFAULT_LINKEA_SIZE),
705                         XATTR_NAME_LINK, 0, handle);
706         return rc;
707 }
708
709 int __lfsck_links_read(const struct lu_env *env, struct dt_object *obj,
710                        struct linkea_data *ldata)
711 {
712         int rc;
713
714         if (ldata->ld_buf->lb_buf == NULL)
715                 return -ENOMEM;
716
717         if (!dt_object_exists(obj))
718                 return -ENOENT;
719
720         rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
721         if (rc == -ERANGE) {
722                 /* Buf was too small, figure out what we need. */
723                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LINK);
724                 if (rc <= 0)
725                         return rc;
726
727                 lu_buf_realloc(ldata->ld_buf, rc);
728                 if (ldata->ld_buf->lb_buf == NULL)
729                         return -ENOMEM;
730
731                 rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
732         }
733
734         if (rc > 0)
735                 rc = linkea_init(ldata);
736
737         return rc;
738 }
739
740 /**
741  * Remove linkEA for the given object.
742  *
743  * The caller should take the ldlm lock before the calling.
744  *
745  * \param[in] env       pointer to the thread context
746  * \param[in] com       pointer to the lfsck component
747  * \param[in] obj       pointer to the dt_object to be handled
748  *
749  * \retval              0 for repaired cases
750  * \retval              negative error number on failure
751  */
752 static int lfsck_namespace_links_remove(const struct lu_env *env,
753                                         struct lfsck_component *com,
754                                         struct dt_object *obj)
755 {
756         struct lfsck_instance           *lfsck  = com->lc_lfsck;
757         struct dt_device                *dev    = lfsck_obj2dev(obj);
758         struct thandle                  *th     = NULL;
759         int                              rc     = 0;
760         ENTRY;
761
762         LASSERT(dt_object_remote(obj) == 0);
763
764         th = dt_trans_create(env, dev);
765         if (IS_ERR(th))
766                 GOTO(log, rc = PTR_ERR(th));
767
768         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, th);
769         if (rc != 0)
770                 GOTO(stop, rc);
771
772         rc = dt_trans_start_local(env, dev, th);
773         if (rc != 0)
774                 GOTO(stop, rc);
775
776         dt_write_lock(env, obj, 0);
777         if (unlikely(lfsck_is_dead_obj(obj)))
778                 GOTO(unlock, rc = -ENOENT);
779
780         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
781                 GOTO(unlock, rc = 0);
782
783         rc = dt_xattr_del(env, obj, XATTR_NAME_LINK, th);
784
785         GOTO(unlock, rc);
786
787 unlock:
788         dt_write_unlock(env, obj);
789
790 stop:
791         dt_trans_stop(env, dev, th);
792
793 log:
794         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove invalid linkEA "
795                "for the object "DFID": rc = %d\n",
796                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
797
798         if (rc == 0) {
799                 struct lfsck_namespace *ns = com->lc_file_ram;
800
801                 ns->ln_flags |= LF_INCONSISTENT;
802         }
803
804         return rc;
805 }
806
807 static int lfsck_links_write(const struct lu_env *env, struct dt_object *obj,
808                              struct linkea_data *ldata, struct thandle *handle)
809 {
810         const struct lu_buf *buf = lfsck_buf_get_const(env,
811                                                        ldata->ld_buf->lb_buf,
812                                                        ldata->ld_leh->leh_len);
813
814         return dt_xattr_set(env, obj, buf, XATTR_NAME_LINK, 0, handle);
815 }
816
817 static void lfsck_namespace_unpack_linkea_entry(struct linkea_data *ldata,
818                                                 struct lu_name *cname,
819                                                 struct lu_fid *pfid,
820                                                 char *buf)
821 {
822         linkea_entry_unpack(ldata->ld_lee, &ldata->ld_reclen, cname, pfid);
823         /* To guarantee the 'name' is terminated with '0'. */
824         memcpy(buf, cname->ln_name, cname->ln_namelen);
825         buf[cname->ln_namelen] = 0;
826         cname->ln_name = buf;
827 }
828
829 static int lfsck_namespace_filter_linkea_entry(struct linkea_data *ldata,
830                                                struct lu_name *cname,
831                                                struct lu_fid *pfid,
832                                                bool remove)
833 {
834         struct link_ea_entry    *oldlee;
835         int                      oldlen;
836         int                      repeated = 0;
837
838         oldlee = ldata->ld_lee;
839         oldlen = ldata->ld_reclen;
840         linkea_next_entry(ldata);
841         while (ldata->ld_lee != NULL) {
842                 ldata->ld_reclen = (ldata->ld_lee->lee_reclen[0] << 8) |
843                                    ldata->ld_lee->lee_reclen[1];
844                 if (unlikely(ldata->ld_reclen == oldlen &&
845                              memcmp(ldata->ld_lee, oldlee, oldlen) == 0)) {
846                         repeated++;
847                         if (!remove)
848                                 break;
849
850                         linkea_del_buf(ldata, cname);
851                 } else {
852                         linkea_next_entry(ldata);
853                 }
854         }
855         ldata->ld_lee = oldlee;
856         ldata->ld_reclen = oldlen;
857
858         return repeated;
859 }
860
861 /**
862  * Insert orphan into .lustre/lost+found/MDTxxxx/ locally.
863  *
864  * Add the specified orphan MDT-object to the .lustre/lost+found/MDTxxxx/
865  * with the given type to generate the name, the detailed rules for name
866  * have been described as following.
867  *
868  * The function also generates the linkEA corresponding to the name entry
869  * under the .lustre/lost+found/MDTxxxx/ for the orphan MDT-object.
870  *
871  * \param[in] env       pointer to the thread context
872  * \param[in] com       pointer to the lfsck component
873  * \param[in] orphan    pointer to the orphan MDT-object
874  * \param[in] infix     additional information for the orphan name, such as
875  *                      the FID for original
876  * \param[in] type      the type for describing why the orphan MDT-object is
877  *                      created. The rules are as following:
878  *
879  *  type "D":           The MDT-object is a directory, it may knows its parent
880  *                      but because there is no valid linkEA, the LFSCK cannot
881  *                      know where to put it back to the namespace.
882  *  type "O":           The MDT-object has no linkEA, and there is no name
883  *                      entry that references the MDT-object.
884  *
885  *  type "S":           The orphan MDT-object is a shard of a striped directory
886  *
887  * \see lfsck_layout_recreate_parent() for more types.
888  *
889  * The orphan name will be like:
890  * ${FID}-${infix}-${type}-${conflict_version}
891  *
892  * \param[out] count    if some others inserted some linkEA entries by race,
893  *                      then return the linkEA entries count.
894  *
895  * \retval              positive number for repaired cases
896  * \retval              0 if needs to repair nothing
897  * \retval              negative error number on failure
898  */
899 static int lfsck_namespace_insert_orphan(const struct lu_env *env,
900                                          struct lfsck_component *com,
901                                          struct dt_object *orphan,
902                                          const char *infix, const char *type,
903                                          int *count)
904 {
905         struct lfsck_thread_info        *info   = lfsck_env_info(env);
906         struct lu_name                  *cname  = &info->lti_name;
907         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
908         struct lu_attr                  *la     = &info->lti_la2;
909         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
910         const struct lu_fid             *pfid;
911         struct lu_fid                    tfid;
912         struct lfsck_instance           *lfsck  = com->lc_lfsck;
913         struct dt_device                *dev    = lfsck_obj2dev(orphan);
914         struct dt_object                *parent;
915         struct thandle                  *th     = NULL;
916         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
917         struct lustre_handle             clh    = { 0 };
918         struct linkea_data               ldata  = { NULL };
919         struct lu_buf                    linkea_buf;
920         int                              namelen;
921         int                              idx    = 0;
922         int                              rc     = 0;
923         bool                             exist  = false;
924         ENTRY;
925
926         cname->ln_name = NULL;
927         if (unlikely(lfsck->li_lpf_obj == NULL))
928                 GOTO(log, rc = -ENXIO);
929
930         parent = lfsck->li_lpf_obj;
931         pfid = lfsck_dto2fid(parent);
932
933 again:
934         do {
935                 namelen = snprintf(info->lti_key, NAME_MAX, DFID"%s-%s-%d",
936                                    PFID(cfid), infix, type, idx++);
937                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
938                                (const struct dt_key *)info->lti_key);
939                 if (rc != 0 && rc != -ENOENT)
940                         GOTO(log, rc);
941
942                 if (unlikely(rc == 0 && lu_fid_eq(cfid, &tfid)))
943                         exist = true;
944         } while (rc == 0 && !exist);
945
946         rc = lfsck_lock(env, lfsck, parent, info->lti_key, pllh,
947                         MDS_INODELOCK_UPDATE, LCK_PW);
948         if (rc != 0)
949                 GOTO(log, rc);
950
951         /* Re-check whether the name conflict with othrs after taken
952          * the ldlm lock. */
953         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
954                        (const struct dt_key *)info->lti_key);
955         if (rc == 0) {
956                 if (!lu_fid_eq(cfid, &tfid)) {
957                         exist = false;
958                         lfsck_unlock(pllh);
959                         goto again;
960                 }
961
962                 exist = true;
963         } else if (rc != -ENOENT) {
964                 GOTO(log, rc);
965         } else {
966                 exist = false;
967         }
968
969         cname->ln_name = info->lti_key;
970         cname->ln_namelen = namelen;
971         rc = linkea_data_new(&ldata, &info->lti_linkea_buf2);
972         if (rc != 0)
973                 GOTO(log, rc);
974
975         rc = linkea_add_buf(&ldata, cname, pfid);
976         if (rc != 0)
977                 GOTO(log, rc);
978
979         rc = lfsck_ibits_lock(env, lfsck, orphan, &clh,
980                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_LOOKUP |
981                               MDS_INODELOCK_XATTR, LCK_EX);
982         if (rc != 0)
983                 GOTO(log, rc);
984
985         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
986                        ldata.ld_leh->leh_len);
987         th = dt_trans_create(env, dev);
988         if (IS_ERR(th))
989                 GOTO(log, rc = PTR_ERR(th));
990
991         if (S_ISDIR(lfsck_object_type(orphan))) {
992                 rc = dt_declare_delete(env, orphan,
993                                        (const struct dt_key *)dotdot, th);
994                 if (rc != 0)
995                         GOTO(stop, rc);
996
997                 rec->rec_type = S_IFDIR;
998                 rec->rec_fid = pfid;
999                 rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1000                                        (const struct dt_key *)dotdot, th);
1001                 if (rc != 0)
1002                         GOTO(stop, rc);
1003         }
1004
1005         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1006                                   XATTR_NAME_LINK, 0, th);
1007         if (rc != 0)
1008                 GOTO(stop, rc);
1009
1010         if (!exist) {
1011                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1012                 rec->rec_fid = cfid;
1013                 rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1014                                        (const struct dt_key *)cname->ln_name,
1015                                        th);
1016                 if (rc != 0)
1017                         GOTO(stop, rc);
1018
1019                 if (S_ISDIR(rec->rec_type)) {
1020                         rc = dt_declare_ref_add(env, parent, th);
1021                         if (rc != 0)
1022                                 GOTO(stop, rc);
1023                 }
1024         }
1025
1026         memset(la, 0, sizeof(*la));
1027         la->la_ctime = cfs_time_current_sec();
1028         la->la_valid = LA_CTIME;
1029         rc = dt_declare_attr_set(env, orphan, la, th);
1030         if (rc != 0)
1031                 GOTO(stop, rc);
1032
1033         rc = dt_trans_start_local(env, dev, th);
1034         if (rc != 0)
1035                 GOTO(stop, rc);
1036
1037         dt_write_lock(env, orphan, 0);
1038         rc = lfsck_links_read(env, orphan, &ldata);
1039         if (likely((rc == -ENODATA) || (rc == -EINVAL) ||
1040                    (rc == 0 && ldata.ld_leh->leh_reccount == 0))) {
1041                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1042                         GOTO(unlock, rc = 1);
1043
1044                 if (S_ISDIR(lfsck_object_type(orphan))) {
1045                         rc = dt_delete(env, orphan,
1046                                        (const struct dt_key *)dotdot, th);
1047                         if (rc != 0)
1048                                 GOTO(unlock, rc);
1049
1050                         rec->rec_type = S_IFDIR;
1051                         rec->rec_fid = pfid;
1052                         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1053                                        (const struct dt_key *)dotdot, th, 1);
1054                         if (rc != 0)
1055                                 GOTO(unlock, rc);
1056                 }
1057
1058                 rc = dt_xattr_set(env, orphan, &linkea_buf, XATTR_NAME_LINK, 0,
1059                                   th);
1060         } else {
1061                 if (rc == 0 && count != NULL)
1062                         *count = ldata.ld_leh->leh_reccount;
1063
1064                 GOTO(unlock, rc);
1065         }
1066         dt_write_unlock(env, orphan);
1067
1068         if (rc == 0 && !exist) {
1069                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1070                 rec->rec_fid = cfid;
1071                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1072                                (const struct dt_key *)cname->ln_name, th, 1);
1073                 if (rc == 0 && S_ISDIR(rec->rec_type)) {
1074                         dt_write_lock(env, parent, 0);
1075                         rc = dt_ref_add(env, parent, th);
1076                         dt_write_unlock(env, parent);
1077                 }
1078         }
1079
1080         if (rc == 0)
1081                 rc = dt_attr_set(env, orphan, la, th);
1082
1083         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1084
1085 unlock:
1086         dt_write_unlock(env, orphan);
1087
1088 stop:
1089         dt_trans_stop(env, dev, th);
1090
1091 log:
1092         lfsck_ibits_unlock(&clh, LCK_EX);
1093         lfsck_unlock(pllh);
1094         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert orphan for the "
1095                "object "DFID", name = %s: rc = %d\n",
1096                lfsck_lfsck2name(lfsck), PFID(cfid),
1097                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1098
1099         if (rc != 0) {
1100                 struct lfsck_namespace *ns = com->lc_file_ram;
1101
1102                 ns->ln_flags |= LF_INCONSISTENT;
1103         }
1104
1105         return rc;
1106 }
1107
1108 /**
1109  * Add the specified name entry back to namespace.
1110  *
1111  * If there is a linkEA entry that back references a name entry under
1112  * some parent directory, but such parent directory does not have the
1113  * claimed name entry. On the other hand, the linkEA entries count is
1114  * not larger than the MDT-object's hard link count. Under such case,
1115  * it is quite possible that the name entry is lost. Then the LFSCK
1116  * should add the name entry back to the namespace.
1117  *
1118  * \param[in] env       pointer to the thread context
1119  * \param[in] com       pointer to the lfsck component
1120  * \param[in] parent    pointer to the directory under which the name entry
1121  *                      will be inserted into
1122  * \param[in] child     pointer to the object referenced by the name entry
1123  *                      that to be inserted into the parent
1124  * \param[in] name      the name for the child in the parent directory
1125  *
1126  * \retval              positive number for repaired cases
1127  * \retval              0 if nothing to be repaired
1128  * \retval              negative error number on failure
1129  */
1130 static int lfsck_namespace_insert_normal(const struct lu_env *env,
1131                                          struct lfsck_component *com,
1132                                          struct dt_object *parent,
1133                                          struct dt_object *child,
1134                                          const char *name)
1135 {
1136         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1137         struct lu_attr                  *la     = &info->lti_la;
1138         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1139         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1140         /* The child and its name may be on different MDTs. */
1141         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
1142         const struct lu_fid             *cfid   = lfsck_dto2fid(child);
1143         struct dt_device                *dev    = lfsck->li_next;
1144         struct thandle                  *th     = NULL;
1145         struct lfsck_lock_handle        *llh    = &info->lti_llh;
1146         int                              rc     = 0;
1147         ENTRY;
1148
1149         /* @parent/@child may be based on lfsck->li_bottom,
1150          * but here we need the object based on the lfsck->li_next. */
1151
1152         parent = lfsck_object_locate(dev, parent);
1153         if (IS_ERR(parent))
1154                 GOTO(log, rc = PTR_ERR(parent));
1155
1156         if (unlikely(!dt_try_as_dir(env, parent)))
1157                 GOTO(log, rc = -ENOTDIR);
1158
1159         child = lfsck_object_locate(dev, child);
1160         if (IS_ERR(child))
1161                 GOTO(log, rc = PTR_ERR(child));
1162
1163         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1164                 GOTO(log, rc = 1);
1165
1166         rc = lfsck_lock(env, lfsck, parent, name, llh,
1167                         MDS_INODELOCK_UPDATE, LCK_PW);
1168         if (rc != 0)
1169                 GOTO(log, rc);
1170
1171         th = dt_trans_create(env, dev);
1172         if (IS_ERR(th))
1173                 GOTO(unlock, rc = PTR_ERR(th));
1174
1175         rec->rec_type = lfsck_object_type(child) & S_IFMT;
1176         rec->rec_fid = cfid;
1177         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1178                                (const struct dt_key *)name, th);
1179         if (rc != 0)
1180                 GOTO(stop, rc);
1181
1182         if (S_ISDIR(rec->rec_type)) {
1183                 rc = dt_declare_ref_add(env, parent, th);
1184                 if (rc != 0)
1185                         GOTO(stop, rc);
1186         }
1187
1188         memset(la, 0, sizeof(*la));
1189         la->la_ctime = cfs_time_current_sec();
1190         la->la_valid = LA_CTIME;
1191         rc = dt_declare_attr_set(env, parent, la, th);
1192         if (rc != 0)
1193                 GOTO(stop, rc);
1194
1195         rc = dt_declare_attr_set(env, child, la, th);
1196         if (rc != 0)
1197                 GOTO(stop, rc);
1198
1199         rc = dt_trans_start_local(env, dev, th);
1200         if (rc != 0)
1201                 GOTO(stop, rc);
1202
1203         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1204                        (const struct dt_key *)name, th, 1);
1205         if (rc != 0)
1206                 GOTO(stop, rc);
1207
1208         if (S_ISDIR(rec->rec_type)) {
1209                 dt_write_lock(env, parent, 0);
1210                 rc = dt_ref_add(env, parent, th);
1211                 dt_write_unlock(env, parent);
1212                 if (rc != 0)
1213                         GOTO(stop, rc);
1214         }
1215
1216         la->la_ctime = cfs_time_current_sec();
1217         rc = dt_attr_set(env, parent, la, th);
1218         if (rc != 0)
1219                 GOTO(stop, rc);
1220
1221         rc = dt_attr_set(env, child, la, th);
1222
1223         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1224
1225 stop:
1226         dt_trans_stop(env, dev, th);
1227
1228 unlock:
1229         lfsck_unlock(llh);
1230
1231 log:
1232         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert object "DFID" with "
1233                "the name %s and type %o to the parent "DFID": rc = %d\n",
1234                lfsck_lfsck2name(lfsck), PFID(cfid), name,
1235                lfsck_object_type(child) & S_IFMT, PFID(pfid), rc);
1236
1237         if (rc != 0) {
1238                 struct lfsck_namespace *ns = com->lc_file_ram;
1239
1240                 ns->ln_flags |= LF_INCONSISTENT;
1241                 if (rc > 0)
1242                         ns->ln_lost_dirent_repaired++;
1243         }
1244
1245         return rc;
1246 }
1247
1248 /**
1249  * Create the specified orphan directory.
1250  *
1251  * For the case that the parent MDT-object stored in some MDT-object's
1252  * linkEA entry is lost, the LFSCK will re-create the parent object as
1253  * an orphan and insert it into .lustre/lost+found/MDTxxxx/ directory
1254  * with the name ${FID}-P-${conflict_version}.
1255  *
1256  * \param[in] env       pointer to the thread context
1257  * \param[in] com       pointer to the lfsck component
1258  * \param[in] orphan    pointer to the orphan MDT-object to be created
1259  * \param[in] lmv       pointer to master LMV EA that will be set to the orphan
1260  *
1261  * \retval              positive number for repaired cases
1262  * \retval              negative error number on failure
1263  */
1264 static int lfsck_namespace_create_orphan_dir(const struct lu_env *env,
1265                                              struct lfsck_component *com,
1266                                              struct dt_object *orphan,
1267                                              struct lmv_mds_md_v1 *lmv)
1268 {
1269         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1270         struct lu_attr                  *la     = &info->lti_la;
1271         struct dt_allocation_hint       *hint   = &info->lti_hint;
1272         struct dt_object_format         *dof    = &info->lti_dof;
1273         struct lu_name                  *cname  = &info->lti_name2;
1274         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1275         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1276         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
1277         struct lu_fid                    tfid;
1278         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1279         struct lfsck_namespace          *ns     = com->lc_file_ram;
1280         struct dt_device                *dev    = lfsck_obj2dev(orphan);
1281         struct dt_object                *parent = NULL;
1282         struct thandle                  *th     = NULL;
1283         struct lfsck_lock_handle        *llh    = &info->lti_llh;
1284         struct linkea_data               ldata  = { NULL };
1285         struct lu_buf                    linkea_buf;
1286         struct lu_buf                    lmv_buf;
1287         char                             name[32];
1288         int                              namelen;
1289         int                              idx    = 0;
1290         int                              rc     = 0;
1291         int                              rc1    = 0;
1292         ENTRY;
1293
1294         LASSERT(!dt_object_exists(orphan));
1295
1296         cname->ln_name = NULL;
1297         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1298                 GOTO(log, rc = 1);
1299
1300         if (dt_object_remote(orphan)) {
1301                 LASSERT(lfsck->li_lpf_root_obj != NULL);
1302
1303                 idx = lfsck_find_mdt_idx_by_fid(env, lfsck, cfid);
1304                 if (idx < 0)
1305                         GOTO(log, rc = idx);
1306
1307                 snprintf(name, 8, "MDT%04x", idx);
1308                 rc = dt_lookup(env, lfsck->li_lpf_root_obj,
1309                                (struct dt_rec *)&tfid,
1310                                (const struct dt_key *)name);
1311                 if (rc != 0)
1312                         GOTO(log, rc = (rc == -ENOENT ? -ENXIO : rc));
1313
1314                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
1315                 if (IS_ERR(parent))
1316                         GOTO(log, rc = PTR_ERR(parent));
1317
1318                 if (unlikely(!dt_try_as_dir(env, parent)))
1319                         GOTO(log, rc = -ENOTDIR);
1320         } else {
1321                 if (unlikely(lfsck->li_lpf_obj == NULL))
1322                         GOTO(log, rc = -ENXIO);
1323
1324                 parent = lfsck->li_lpf_obj;
1325         }
1326
1327         dev = lfsck_find_dev_by_fid(env, lfsck, cfid);
1328         if (IS_ERR(dev))
1329                 GOTO(log, rc = PTR_ERR(dev));
1330
1331         idx = 0;
1332
1333 again:
1334         do {
1335                 namelen = snprintf(name, 31, DFID"-P-%d",
1336                                    PFID(cfid), idx++);
1337                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1338                                (const struct dt_key *)name);
1339                 if (rc != 0 && rc != -ENOENT)
1340                         GOTO(log, rc);
1341         } while (rc == 0);
1342
1343         rc = lfsck_lock(env, lfsck, parent, name, llh,
1344                         MDS_INODELOCK_UPDATE, LCK_PW);
1345         if (rc != 0)
1346                 GOTO(log, rc);
1347
1348         /* Re-check whether the name conflict with othrs after taken
1349          * the ldlm lock. */
1350         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1351                        (const struct dt_key *)name);
1352         if (unlikely(rc == 0)) {
1353                 lfsck_unlock(llh);
1354                 goto again;
1355         }
1356
1357         if (rc != -ENOENT)
1358                 GOTO(unlock1, rc);
1359
1360         cname->ln_name = name;
1361         cname->ln_namelen = namelen;
1362
1363         memset(la, 0, sizeof(*la));
1364         la->la_mode = S_IFDIR | 0700;
1365         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
1366                        LA_ATIME | LA_MTIME | LA_CTIME;
1367
1368         orphan->do_ops->do_ah_init(env, hint, parent, orphan,
1369                                    la->la_mode & S_IFMT);
1370
1371         memset(dof, 0, sizeof(*dof));
1372         dof->dof_type = dt_mode_to_dft(S_IFDIR);
1373
1374         rc = linkea_data_new(&ldata, &info->lti_linkea_buf2);
1375         if (rc != 0)
1376                 GOTO(unlock1, rc);
1377
1378         rc = linkea_add_buf(&ldata, cname, lfsck_dto2fid(parent));
1379         if (rc != 0)
1380                 GOTO(unlock1, rc);
1381
1382         th = dt_trans_create(env, dev);
1383         if (IS_ERR(th))
1384                 GOTO(unlock1, rc = PTR_ERR(th));
1385
1386         /* Sync the remote transaction to guarantee that the subsequent
1387          * lock against the @orphan can find the @orphan in time. */
1388         if (dt_object_remote(orphan))
1389                 th->th_sync = 1;
1390
1391         rc = dt_declare_create(env, orphan, la, hint, dof, th);
1392         if (rc != 0)
1393                 GOTO(stop, rc);
1394
1395         if (unlikely(!dt_try_as_dir(env, orphan)))
1396                 GOTO(stop, rc = -ENOTDIR);
1397
1398         rc = dt_declare_ref_add(env, orphan, th);
1399         if (rc != 0)
1400                 GOTO(stop, rc);
1401
1402         rec->rec_type = S_IFDIR;
1403         rec->rec_fid = cfid;
1404         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1405                                (const struct dt_key *)dot, th);
1406         if (rc != 0)
1407                 GOTO(stop, rc);
1408
1409         rec->rec_fid = lfsck_dto2fid(parent);
1410         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1411                                (const struct dt_key *)dotdot, th);
1412         if (rc != 0)
1413                 GOTO(stop, rc);
1414
1415         if (lmv != NULL) {
1416                 lmv->lmv_magic = LMV_MAGIC;
1417                 lmv->lmv_master_mdt_index = lfsck_dev_idx(lfsck);
1418                 lfsck_lmv_header_cpu_to_le(lmv2, lmv);
1419                 lfsck_buf_init(&lmv_buf, lmv2, sizeof(*lmv2));
1420                 rc = dt_declare_xattr_set(env, orphan, &lmv_buf,
1421                                           XATTR_NAME_LMV, 0, th);
1422                 if (rc != 0)
1423                         GOTO(stop, rc);
1424         }
1425
1426         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
1427                        ldata.ld_leh->leh_len);
1428         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1429                                   XATTR_NAME_LINK, 0, th);
1430         if (rc != 0)
1431                 GOTO(stop, rc);
1432
1433         rec->rec_fid = cfid;
1434         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1435                                (const struct dt_key *)name, th);
1436         if (rc == 0)
1437                 rc = dt_declare_ref_add(env, parent, th);
1438
1439         if (rc != 0)
1440                 GOTO(stop, rc);
1441
1442         rc = dt_trans_start_local(env, dev, th);
1443         if (rc != 0)
1444                 GOTO(stop, rc);
1445
1446         dt_write_lock(env, orphan, 0);
1447         rc = dt_create(env, orphan, la, hint, dof, th);
1448         if (rc != 0)
1449                 GOTO(unlock2, rc);
1450
1451         rc = dt_ref_add(env, orphan, th);
1452         if (rc != 0)
1453                 GOTO(unlock2, rc);
1454
1455         rec->rec_fid = cfid;
1456         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1457                        (const struct dt_key *)dot, th, 1);
1458         if (rc != 0)
1459                 GOTO(unlock2, rc);
1460
1461         rec->rec_fid = lfsck_dto2fid(parent);
1462         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1463                        (const struct dt_key *)dotdot, th, 1);
1464         if (rc != 0)
1465                 GOTO(unlock2, rc);
1466
1467         if (lmv != NULL) {
1468                 rc = dt_xattr_set(env, orphan, &lmv_buf, XATTR_NAME_LMV, 0, th);
1469                 if (rc != 0)
1470                         GOTO(unlock2, rc);
1471         }
1472
1473         rc = dt_xattr_set(env, orphan, &linkea_buf,
1474                           XATTR_NAME_LINK, 0, th);
1475         dt_write_unlock(env, orphan);
1476         if (rc != 0)
1477                 GOTO(stop, rc);
1478
1479         rec->rec_fid = cfid;
1480         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1481                        (const struct dt_key *)name, th, 1);
1482         if (rc == 0) {
1483                 dt_write_lock(env, parent, 0);
1484                 rc = dt_ref_add(env, parent, th);
1485                 dt_write_unlock(env, parent);
1486         }
1487
1488         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1489
1490 unlock2:
1491         dt_write_unlock(env, orphan);
1492
1493 stop:
1494         rc1 = dt_trans_stop(env, dev, th);
1495         if (rc1 != 0 && rc > 0)
1496                 rc = rc1;
1497
1498 unlock1:
1499         lfsck_unlock(llh);
1500
1501 log:
1502         CDEBUG(D_LFSCK, "%s: namespace LFSCK create orphan dir for "
1503                "the object "DFID", name = %s: rc = %d\n",
1504                lfsck_lfsck2name(lfsck), PFID(cfid),
1505                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1506
1507         if (parent != NULL && !IS_ERR(parent) && parent != lfsck->li_lpf_obj)
1508                 lfsck_object_put(env, parent);
1509
1510         if (rc != 0)
1511                 ns->ln_flags |= LF_INCONSISTENT;
1512
1513         return rc;
1514 }
1515
1516 /**
1517  * Remove the specified entry from the linkEA.
1518  *
1519  * Locate the linkEA entry with the given @cname and @pfid, then
1520  * remove this entry or the other entries those are repeated with
1521  * this entry.
1522  *
1523  * \param[in] env       pointer to the thread context
1524  * \param[in] com       pointer to the lfsck component
1525  * \param[in] obj       pointer to the dt_object to be handled
1526  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1527  * \param[in] cname     the name for the child in the parent directory
1528  * \param[in] pfid      the parent directory's FID for the linkEA
1529  * \param[in] next      if true, then remove the first found linkEA
1530  *                      entry, and move the ldata->ld_lee to next entry
1531  *
1532  * \retval              positive number for repaired cases
1533  * \retval              0 if nothing to be repaired
1534  * \retval              negative error number on failure
1535  */
1536 static int lfsck_namespace_shrink_linkea(const struct lu_env *env,
1537                                          struct lfsck_component *com,
1538                                          struct dt_object *obj,
1539                                          struct linkea_data *ldata,
1540                                          struct lu_name *cname,
1541                                          struct lu_fid *pfid,
1542                                          bool next)
1543 {
1544         struct lfsck_instance           *lfsck     = com->lc_lfsck;
1545         struct dt_device                *dev       = lfsck_obj2dev(obj);
1546         struct lfsck_bookmark           *bk        = &lfsck->li_bookmark_ram;
1547         struct thandle                  *th        = NULL;
1548         struct lustre_handle             lh        = { 0 };
1549         struct linkea_data               ldata_new = { NULL };
1550         struct lu_buf                    linkea_buf;
1551         int                              rc        = 0;
1552         ENTRY;
1553
1554         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1555                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1556                               LCK_EX);
1557         if (rc != 0)
1558                 GOTO(log, rc);
1559
1560         if (next)
1561                 linkea_del_buf(ldata, cname);
1562         else
1563                 lfsck_namespace_filter_linkea_entry(ldata, cname, pfid,
1564                                                     true);
1565         lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
1566                        ldata->ld_leh->leh_len);
1567
1568 again:
1569         th = dt_trans_create(env, dev);
1570         if (IS_ERR(th))
1571                 GOTO(unlock1, rc = PTR_ERR(th));
1572
1573         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
1574                                   XATTR_NAME_LINK, 0, th);
1575         if (rc != 0)
1576                 GOTO(stop, rc);
1577
1578         rc = dt_trans_start_local(env, dev, th);
1579         if (rc != 0)
1580                 GOTO(stop, rc);
1581
1582         dt_write_lock(env, obj, 0);
1583         if (unlikely(lfsck_is_dead_obj(obj)))
1584                 GOTO(unlock2, rc = -ENOENT);
1585
1586         rc = lfsck_links_read2(env, obj, &ldata_new);
1587         if (rc != 0)
1588                 GOTO(unlock2, rc);
1589
1590         /* The specified linkEA entry has been removed by race. */
1591         rc = linkea_links_find(&ldata_new, cname, pfid);
1592         if (rc != 0)
1593                 GOTO(unlock2, rc = 0);
1594
1595         if (bk->lb_param & LPF_DRYRUN)
1596                 GOTO(unlock2, rc = 1);
1597
1598         if (next)
1599                 linkea_del_buf(&ldata_new, cname);
1600         else
1601                 lfsck_namespace_filter_linkea_entry(&ldata_new, cname, pfid,
1602                                                     true);
1603
1604         if (linkea_buf.lb_len < ldata_new.ld_leh->leh_len) {
1605                 dt_write_unlock(env, obj);
1606                 dt_trans_stop(env, dev, th);
1607                 lfsck_buf_init(&linkea_buf, ldata_new.ld_buf->lb_buf,
1608                                ldata_new.ld_leh->leh_len);
1609                 goto again;
1610         }
1611
1612         lfsck_buf_init(&linkea_buf, ldata_new.ld_buf->lb_buf,
1613                        ldata_new.ld_leh->leh_len);
1614         rc = dt_xattr_set(env, obj, &linkea_buf, XATTR_NAME_LINK, 0, th);
1615
1616         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
1617
1618 unlock2:
1619         dt_write_unlock(env, obj);
1620
1621 stop:
1622         dt_trans_stop(env, dev, th);
1623
1624 unlock1:
1625         lfsck_ibits_unlock(&lh, LCK_EX);
1626
1627 log:
1628         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove %s linkEA entry "
1629                "for the object: "DFID", parent "DFID", name %.*s\n",
1630                lfsck_lfsck2name(lfsck), next ? "invalid" : "redundant",
1631                PFID(lfsck_dto2fid(obj)), PFID(pfid), cname->ln_namelen,
1632                cname->ln_name);
1633
1634         if (rc != 0) {
1635                 struct lfsck_namespace *ns = com->lc_file_ram;
1636
1637                 ns->ln_flags |= LF_INCONSISTENT;
1638         }
1639
1640         return rc;
1641 }
1642
1643 /**
1644  * Conditionally remove the specified entry from the linkEA.
1645  *
1646  * Take the parent lock firstly, then check whether the specified
1647  * name entry exists or not: if yes, do nothing; otherwise, call
1648  * lfsck_namespace_shrink_linkea() to remove the linkea entry.
1649  *
1650  * \param[in] env       pointer to the thread context
1651  * \param[in] com       pointer to the lfsck component
1652  * \param[in] parent    pointer to the parent directory
1653  * \param[in] child     pointer to the child object that holds the linkEA
1654  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1655  * \param[in] cname     the name for the child in the parent directory
1656  * \param[in] pfid      the parent directory's FID for the linkEA
1657  *
1658  * \retval              positive number for repaired cases
1659  * \retval              0 if nothing to be repaired
1660  * \retval              negative error number on failure
1661  */
1662 static int lfsck_namespace_shrink_linkea_cond(const struct lu_env *env,
1663                                               struct lfsck_component *com,
1664                                               struct dt_object *parent,
1665                                               struct dt_object *child,
1666                                               struct linkea_data *ldata,
1667                                               struct lu_name *cname,
1668                                               struct lu_fid *pfid)
1669 {
1670         struct lfsck_thread_info *info  = lfsck_env_info(env);
1671         struct lu_fid            *cfid  = &info->lti_fid3;
1672         struct lfsck_lock_handle *llh   = &info->lti_llh;
1673         int                       rc;
1674         ENTRY;
1675
1676         rc = lfsck_lock(env, com->lc_lfsck, parent, cname->ln_name, llh,
1677                         MDS_INODELOCK_UPDATE, LCK_PR);
1678         if (rc != 0)
1679                 RETURN(rc);
1680
1681         dt_read_lock(env, parent, 0);
1682         if (unlikely(lfsck_is_dead_obj(parent))) {
1683                 dt_read_unlock(env, parent);
1684                 lfsck_unlock(llh);
1685                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1686                                                    cname, pfid, true);
1687
1688                 RETURN(rc);
1689         }
1690
1691         rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
1692                        (const struct dt_key *)cname->ln_name);
1693         dt_read_unlock(env, parent);
1694
1695         /* It is safe to release the ldlm lock, because when the logic come
1696          * here, we have got all the needed information above whether the
1697          * linkEA entry is valid or not. It is not important that others
1698          * may add new linkEA entry after the ldlm lock released. If other
1699          * has removed the specified linkEA entry by race, then it is OK,
1700          * because the subsequent lfsck_namespace_shrink_linkea() can handle
1701          * such case. */
1702         lfsck_unlock(llh);
1703         if (rc == -ENOENT) {
1704                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1705                                                    cname, pfid, true);
1706
1707                 RETURN(rc);
1708         }
1709
1710         if (rc != 0)
1711                 RETURN(rc);
1712
1713         /* The LFSCK just found some internal status of cross-MDTs
1714          * create operation. That is normal. */
1715         if (lu_fid_eq(cfid, lfsck_dto2fid(child))) {
1716                 linkea_next_entry(ldata);
1717
1718                 RETURN(0);
1719         }
1720
1721         rc = lfsck_namespace_shrink_linkea(env, com, child, ldata, cname,
1722                                            pfid, true);
1723
1724         RETURN(rc);
1725 }
1726
1727 /**
1728  * Conditionally replace name entry in the parent.
1729  *
1730  * As required, the LFSCK may re-create the lost MDT-object for dangling
1731  * name entry, but such repairing may be wrong because of bad FID in the
1732  * name entry. As the LFSCK processing, the real MDT-object may be found,
1733  * then the LFSCK should check whether the former re-created MDT-object
1734  * has been modified or not, if not, then destroy it and update the name
1735  * entry in the parent to reference the real MDT-object.
1736  *
1737  * \param[in] env       pointer to the thread context
1738  * \param[in] com       pointer to the lfsck component
1739  * \param[in] parent    pointer to the parent directory
1740  * \param[in] child     pointer to the MDT-object that may be the real
1741  *                      MDT-object corresponding to the name entry in parent
1742  * \param[in] cfid      the current FID in the name entry
1743  * \param[in] cname     contains the name of the child in the parent directory
1744  *
1745  * \retval              positive number for repaired cases
1746  * \retval              0 if nothing to be repaired
1747  * \retval              negative error number on failure
1748  */
1749 static int lfsck_namespace_replace_cond(const struct lu_env *env,
1750                                         struct lfsck_component *com,
1751                                         struct dt_object *parent,
1752                                         struct dt_object *child,
1753                                         const struct lu_fid *cfid,
1754                                         const struct lu_name *cname)
1755 {
1756         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1757         struct lu_attr                  *la     = &info->lti_la;
1758         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1759         struct lu_fid                    tfid;
1760         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1761         /* The child and its name may be on different MDTs. */
1762         struct dt_device                *dev    = lfsck->li_next;
1763         const char                      *name   = cname->ln_name;
1764         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
1765         struct dt_object                *cobj   = NULL;
1766         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
1767         struct lustre_handle             clh    = { 0 };
1768         struct linkea_data               ldata  = { NULL };
1769         struct thandle                  *th     = NULL;
1770         bool                             exist  = true;
1771         int                              rc     = 0;
1772         ENTRY;
1773
1774         /* @parent/@child may be based on lfsck->li_bottom,
1775          * but here we need the object based on the lfsck->li_next. */
1776
1777         parent = lfsck_object_locate(dev, parent);
1778         if (IS_ERR(parent))
1779                 GOTO(log, rc = PTR_ERR(parent));
1780
1781         if (unlikely(!dt_try_as_dir(env, parent)))
1782                 GOTO(log, rc = -ENOTDIR);
1783
1784         rc = lfsck_lock(env, lfsck, parent, name, pllh,
1785                         MDS_INODELOCK_UPDATE, LCK_PW);
1786         if (rc != 0)
1787                 GOTO(log, rc);
1788
1789         if (!fid_is_sane(cfid)) {
1790                 exist = false;
1791                 goto replace;
1792         }
1793
1794         cobj = lfsck_object_find_by_dev(env, dev, cfid);
1795         if (IS_ERR(cobj)) {
1796                 rc = PTR_ERR(cobj);
1797                 if (rc == -ENOENT) {
1798                         exist = false;
1799                         goto replace;
1800                 }
1801
1802                 GOTO(log, rc);
1803         }
1804
1805         if (!dt_object_exists(cobj)) {
1806                 exist = false;
1807                 goto replace;
1808         }
1809
1810         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1811                        (const struct dt_key *)name);
1812         if (rc == -ENOENT) {
1813                 exist = false;
1814                 goto replace;
1815         }
1816
1817         if (rc != 0)
1818                 GOTO(log, rc);
1819
1820         /* Someone changed the name entry, cannot replace it. */
1821         if (!lu_fid_eq(cfid, &tfid))
1822                 GOTO(log, rc = 0);
1823
1824         /* lock the object to be destroyed. */
1825         rc = lfsck_ibits_lock(env, lfsck, cobj, &clh,
1826                               MDS_INODELOCK_UPDATE |
1827                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1828                               LCK_EX);
1829         if (rc != 0)
1830                 GOTO(log, rc);
1831
1832         if (unlikely(lfsck_is_dead_obj(cobj))) {
1833                 exist = false;
1834                 goto replace;
1835         }
1836
1837         rc = dt_attr_get(env, cobj, la);
1838         if (rc != 0)
1839                 GOTO(log, rc);
1840
1841         /* The object has been modified by other(s), or it is not created by
1842          * LFSCK, the two cases are indistinguishable. So cannot replace it. */
1843         if (la->la_ctime != 0)
1844                 GOTO(log, rc);
1845
1846         if (S_ISREG(la->la_mode)) {
1847                 rc = dt_xattr_get(env, cobj, &LU_BUF_NULL, XATTR_NAME_LOV);
1848                 /* If someone has created related OST-object(s),
1849                  * then keep it. */
1850                 if ((rc > 0) || (rc < 0 && rc != -ENODATA))
1851                         GOTO(log, rc = (rc > 0 ? 0 : rc));
1852         }
1853
1854 replace:
1855         dt_read_lock(env, child, 0);
1856         rc = lfsck_links_read2(env, child, &ldata);
1857         dt_read_unlock(env, child);
1858
1859         /* Someone changed the child, no need to replace. */
1860         if (rc == -ENODATA)
1861                 GOTO(log, rc = 0);
1862
1863         if (rc != 0)
1864                 GOTO(log, rc);
1865
1866         rc = linkea_links_find(&ldata, cname, pfid);
1867         /* Someone moved the child, no need to replace. */
1868         if (rc != 0)
1869                 GOTO(log, rc = 0);
1870
1871         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1872                 GOTO(log, rc = 1);
1873
1874         th = dt_trans_create(env, dev);
1875         if (IS_ERR(th))
1876                 GOTO(log, rc = PTR_ERR(th));
1877
1878         if (exist) {
1879                 rc = dt_declare_destroy(env, cobj, th);
1880                 if (rc != 0)
1881                         GOTO(stop, rc);
1882         }
1883
1884         rc = dt_declare_delete(env, parent, (const struct dt_key *)name, th);
1885         if (rc != 0)
1886                 GOTO(stop, rc);
1887
1888         rec->rec_type = S_IFDIR;
1889         rec->rec_fid = lfsck_dto2fid(child);
1890         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1891                                (const struct dt_key *)name, th);
1892         if (rc != 0)
1893                 GOTO(stop, rc);
1894
1895         rc = dt_trans_start_local(env, dev, th);
1896         if (rc != 0)
1897                 GOTO(stop, rc);
1898
1899         if (exist) {
1900                 rc = dt_destroy(env, cobj, th);
1901                 if (rc != 0)
1902                         GOTO(stop, rc);
1903         }
1904
1905         /* The old name entry maybe not exist. */
1906         rc = dt_delete(env, parent, (const struct dt_key *)name, th);
1907         if (rc != 0 && rc != -ENOENT)
1908                 GOTO(stop, rc);
1909
1910         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1911                        (const struct dt_key *)name, th, 1);
1912
1913         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1914
1915 stop:
1916         dt_trans_stop(env, dev, th);
1917
1918 log:
1919         lfsck_ibits_unlock(&clh, LCK_EX);
1920         lfsck_unlock(pllh);
1921
1922         if (cobj != NULL && !IS_ERR(cobj))
1923                 lfsck_object_put(env, cobj);
1924
1925         CDEBUG(D_LFSCK, "%s: namespace LFSCK conditionally destroy the "
1926                "object "DFID" because of conflict with the object "DFID
1927                " under the parent "DFID" with name %s: rc = %d\n",
1928                lfsck_lfsck2name(lfsck), PFID(cfid),
1929                PFID(lfsck_dto2fid(child)), PFID(pfid), name, rc);
1930
1931         return rc;
1932 }
1933
1934 /**
1935  * Overwrite the linkEA for the object with the given ldata.
1936  *
1937  * The caller should take the ldlm lock before the calling.
1938  *
1939  * \param[in] env       pointer to the thread context
1940  * \param[in] com       pointer to the lfsck component
1941  * \param[in] obj       pointer to the dt_object to be handled
1942  * \param[in] ldata     pointer to the new linkEA data
1943  *
1944  * \retval              positive number for repaired cases
1945  * \retval              0 if nothing to be repaired
1946  * \retval              negative error number on failure
1947  */
1948 int lfsck_namespace_rebuild_linkea(const struct lu_env *env,
1949                                    struct lfsck_component *com,
1950                                    struct dt_object *obj,
1951                                    struct linkea_data *ldata)
1952 {
1953         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1954         struct dt_device                *dev    = lfsck_obj2dev(obj);
1955         struct thandle                  *th     = NULL;
1956         struct lu_buf                    linkea_buf;
1957         int                              rc     = 0;
1958         ENTRY;
1959
1960         th = dt_trans_create(env, dev);
1961         if (IS_ERR(th))
1962                 GOTO(log, rc = PTR_ERR(th));
1963
1964         lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
1965                        ldata->ld_leh->leh_len);
1966         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
1967                                   XATTR_NAME_LINK, 0, th);
1968         if (rc != 0)
1969                 GOTO(stop, rc);
1970
1971         rc = dt_trans_start_local(env, dev, th);
1972         if (rc != 0)
1973                 GOTO(stop, rc);
1974
1975         dt_write_lock(env, obj, 0);
1976         if (unlikely(lfsck_is_dead_obj(obj)))
1977                 GOTO(unlock, rc = 0);
1978
1979         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1980                 GOTO(unlock, rc = 1);
1981
1982         rc = dt_xattr_set(env, obj, &linkea_buf,
1983                           XATTR_NAME_LINK, 0, th);
1984
1985         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
1986
1987 unlock:
1988         dt_write_unlock(env, obj);
1989
1990 stop:
1991         dt_trans_stop(env, dev, th);
1992
1993 log:
1994         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild linkEA for the "
1995                "object "DFID": rc = %d\n",
1996                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
1997
1998         if (rc != 0) {
1999                 struct lfsck_namespace *ns = com->lc_file_ram;
2000
2001                 ns->ln_flags |= LF_INCONSISTENT;
2002         }
2003
2004         return rc;
2005 }
2006
2007 /**
2008  * Repair invalid name entry.
2009  *
2010  * If the name entry contains invalid information, such as bad file type
2011  * or (and) corrupted object FID, then either remove the name entry or
2012  * udpate the name entry with the given (right) information.
2013  *
2014  * \param[in] env       pointer to the thread context
2015  * \param[in] com       pointer to the lfsck component
2016  * \param[in] parent    pointer to the parent directory
2017  * \param[in] child     pointer to the object referenced by the name entry
2018  * \param[in] name      the old name of the child under the parent directory
2019  * \param[in] name2     the new name of the child under the parent directory
2020  * \param[in] type      the type claimed by the name entry
2021  * \param[in] update    update the name entry if true; otherwise, remove it
2022  * \param[in] dec       decrease the parent nlink count if true
2023  *
2024  * \retval              positive number for repaired successfully
2025  * \retval              0 if nothing to be repaired
2026  * \retval              negative error number on failure
2027  */
2028 int lfsck_namespace_repair_dirent(const struct lu_env *env,
2029                                   struct lfsck_component *com,
2030                                   struct dt_object *parent,
2031                                   struct dt_object *child,
2032                                   const char *name, const char *name2,
2033                                   __u16 type, bool update, bool dec)
2034 {
2035         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2036         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2037         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
2038         const struct lu_fid             *cfid   = lfsck_dto2fid(child);
2039         struct lu_fid                    tfid;
2040         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2041         struct dt_device                *dev    = lfsck->li_next;
2042         struct thandle                  *th     = NULL;
2043         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2044         struct lustre_handle             lh     = { 0 };
2045         int                              rc     = 0;
2046         ENTRY;
2047
2048         parent = lfsck_object_locate(dev, parent);
2049         if (IS_ERR(parent))
2050                 GOTO(log, rc = PTR_ERR(parent));
2051
2052         if (unlikely(!dt_try_as_dir(env, parent)))
2053                 GOTO(log, rc = -ENOTDIR);
2054
2055         if (!update || strcmp(name, name2) == 0)
2056                 rc = lfsck_lock(env, lfsck, parent, name, llh,
2057                                 MDS_INODELOCK_UPDATE, LCK_PW);
2058         else
2059                 rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2060                                       MDS_INODELOCK_UPDATE, LCK_PW);
2061         if (rc != 0)
2062                 GOTO(log, rc);
2063
2064         th = dt_trans_create(env, dev);
2065         if (IS_ERR(th))
2066                 GOTO(unlock1, rc = PTR_ERR(th));
2067
2068         rc = dt_declare_delete(env, parent, (const struct dt_key *)name, th);
2069         if (rc != 0)
2070                 GOTO(stop, rc);
2071
2072         if (update) {
2073                 rec->rec_type = lfsck_object_type(child) & S_IFMT;
2074                 rec->rec_fid = cfid;
2075                 rc = dt_declare_insert(env, parent,
2076                                        (const struct dt_rec *)rec,
2077                                        (const struct dt_key *)name2, th);
2078                 if (rc != 0)
2079                         GOTO(stop, rc);
2080         }
2081
2082         if (dec) {
2083                 rc = dt_declare_ref_del(env, parent, th);
2084                 if (rc != 0)
2085                         GOTO(stop, rc);
2086         }
2087
2088         rc = dt_trans_start_local(env, dev, th);
2089         if (rc != 0)
2090                 GOTO(stop, rc);
2091
2092
2093         dt_write_lock(env, parent, 0);
2094         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2095                        (const struct dt_key *)name);
2096         /* Someone has removed the bad name entry by race. */
2097         if (rc == -ENOENT)
2098                 GOTO(unlock2, rc = 0);
2099
2100         if (rc != 0)
2101                 GOTO(unlock2, rc);
2102
2103         /* Someone has removed the bad name entry and reused it for other
2104          * object by race. */
2105         if (!lu_fid_eq(&tfid, cfid))
2106                 GOTO(unlock2, rc = 0);
2107
2108         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2109                 GOTO(unlock2, rc = 1);
2110
2111         rc = dt_delete(env, parent, (const struct dt_key *)name, th);
2112         if (rc != 0)
2113                 GOTO(unlock2, rc);
2114
2115         if (update) {
2116                 rc = dt_insert(env, parent,
2117                                (const struct dt_rec *)rec,
2118                                (const struct dt_key *)name2, th, 1);
2119                 if (rc != 0)
2120                         GOTO(unlock2, rc);
2121         }
2122
2123         if (dec) {
2124                 rc = dt_ref_del(env, parent, th);
2125                 if (rc != 0)
2126                         GOTO(unlock2, rc);
2127         }
2128
2129         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
2130
2131 unlock2:
2132         dt_write_unlock(env, parent);
2133
2134 stop:
2135         dt_trans_stop(env, dev, th);
2136
2137         /* We are not sure whether the child will become orphan or not.
2138          * Record it in the LFSCK trace file for further checking in
2139          * the second-stage scanning. */
2140         if (!update && !dec && rc == 0)
2141                 lfsck_namespace_trace_update(env, com, cfid,
2142                                              LNTF_CHECK_LINKEA, true);
2143
2144 unlock1:
2145         /* It is harmless even if unlock the unused lock_handle */
2146         lfsck_ibits_unlock(&lh, LCK_PW);
2147         lfsck_unlock(llh);
2148
2149 log:
2150         CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found bad name "
2151                "entry for: parent "DFID", child "DFID", name %s, type "
2152                "in name entry %o, type claimed by child %o. repair it "
2153                "by %s with new name2 %s: rc = %d\n",
2154                lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
2155                name, type, update ? lfsck_object_type(child) : 0,
2156                update ? "updating" : "removing", name2, rc);
2157
2158         if (rc != 0) {
2159                 struct lfsck_namespace *ns = com->lc_file_ram;
2160
2161                 ns->ln_flags |= LF_INCONSISTENT;
2162         }
2163
2164         return rc;
2165 }
2166
2167 /**
2168  * Update the ".." name entry for the given object.
2169  *
2170  * The object's ".." is corrupted, this function will update the ".." name
2171  * entry with the given pfid, and the linkEA with the given ldata.
2172  *
2173  * The caller should take the ldlm lock before the calling.
2174  *
2175  * \param[in] env       pointer to the thread context
2176  * \param[in] com       pointer to the lfsck component
2177  * \param[in] obj       pointer to the dt_object to be handled
2178  * \param[in] pfid      the new fid for the object's ".." name entry
2179  * \param[in] cname     the name for the @obj in the parent directory
2180  *
2181  * \retval              positive number for repaired cases
2182  * \retval              0 if nothing to be repaired
2183  * \retval              negative error number on failure
2184  */
2185 static int lfsck_namespace_repair_unmatched_pairs(const struct lu_env *env,
2186                                                   struct lfsck_component *com,
2187                                                   struct dt_object *obj,
2188                                                   const struct lu_fid *pfid,
2189                                                   struct lu_name *cname)
2190 {
2191         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2192         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2193         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2194         struct dt_device                *dev    = lfsck_obj2dev(obj);
2195         struct thandle                  *th     = NULL;
2196         struct linkea_data               ldata  = { NULL };
2197         struct lu_buf                    linkea_buf;
2198         int                              rc     = 0;
2199         ENTRY;
2200
2201         LASSERT(!dt_object_remote(obj));
2202         LASSERT(S_ISDIR(lfsck_object_type(obj)));
2203
2204         rc = linkea_data_new(&ldata, &info->lti_big_buf);
2205         if (rc != 0)
2206                 GOTO(log, rc);
2207
2208         rc = linkea_add_buf(&ldata, cname, pfid);
2209         if (rc != 0)
2210                 GOTO(log, rc);
2211
2212         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2213                        ldata.ld_leh->leh_len);
2214
2215         th = dt_trans_create(env, dev);
2216         if (IS_ERR(th))
2217                 GOTO(log, rc = PTR_ERR(th));
2218
2219         rc = dt_declare_delete(env, obj, (const struct dt_key *)dotdot, th);
2220         if (rc != 0)
2221                 GOTO(stop, rc);
2222
2223         rec->rec_type = S_IFDIR;
2224         rec->rec_fid = pfid;
2225         rc = dt_declare_insert(env, obj, (const struct dt_rec *)rec,
2226                                (const struct dt_key *)dotdot, th);
2227         if (rc != 0)
2228                 GOTO(stop, rc);
2229
2230         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
2231                                   XATTR_NAME_LINK, 0, th);
2232         if (rc != 0)
2233                 GOTO(stop, rc);
2234
2235         rc = dt_trans_start_local(env, dev, th);
2236         if (rc != 0)
2237                 GOTO(stop, rc);
2238
2239         dt_write_lock(env, obj, 0);
2240         if (unlikely(lfsck_is_dead_obj(obj)))
2241                 GOTO(unlock, rc = 0);
2242
2243         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2244                 GOTO(unlock, rc = 1);
2245
2246         /* The old ".." name entry maybe not exist. */
2247         dt_delete(env, obj, (const struct dt_key *)dotdot, th);
2248
2249         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
2250                        (const struct dt_key *)dotdot, th, 1);
2251         if (rc != 0)
2252                 GOTO(unlock, rc);
2253
2254         rc = dt_xattr_set(env, obj, &linkea_buf,
2255                           XATTR_NAME_LINK, 0, th);
2256
2257         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2258
2259 unlock:
2260         dt_write_unlock(env, obj);
2261
2262 stop:
2263         dt_trans_stop(env, dev, th);
2264
2265 log:
2266         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild dotdot name entry for "
2267                "the object "DFID", new parent "DFID": rc = %d\n",
2268                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)),
2269                PFID(pfid), rc);
2270
2271         if (rc != 0) {
2272                 struct lfsck_namespace *ns = com->lc_file_ram;
2273
2274                 ns->ln_flags |= LF_INCONSISTENT;
2275         }
2276
2277         return rc;
2278 }
2279
2280 /**
2281  * Handle orphan @obj during Double Scan Directory.
2282  *
2283  * Remove the @obj's current (invalid) linkEA entries, and insert
2284  * it in the directory .lustre/lost+found/MDTxxxx/ with the name:
2285  * ${FID}-${PFID}-D-${conflict_version}
2286  *
2287  * The caller should take the ldlm lock before the calling.
2288  *
2289  * \param[in] env       pointer to the thread context
2290  * \param[in] com       pointer to the lfsck component
2291  * \param[in] obj       pointer to the orphan object to be handled
2292  * \param[in] pfid      the new fid for the object's ".." name entry
2293  * \param[in,out] lh    ldlm lock handler for the given @obj
2294  * \param[out] type     to tell the caller what the inconsistency is
2295  *
2296  * \retval              positive number for repaired cases
2297  * \retval              0 if nothing to be repaired
2298  * \retval              negative error number on failure
2299  */
2300 static int
2301 lfsck_namespace_dsd_orphan(const struct lu_env *env,
2302                            struct lfsck_component *com,
2303                            struct dt_object *obj,
2304                            const struct lu_fid *pfid,
2305                            struct lustre_handle *lh,
2306                            enum lfsck_namespace_inconsistency_type *type)
2307 {
2308         struct lfsck_thread_info *info = lfsck_env_info(env);
2309         struct lfsck_namespace   *ns   = com->lc_file_ram;
2310         int                       rc;
2311         ENTRY;
2312
2313         /* Remove the unrecognized linkEA. */
2314         rc = lfsck_namespace_links_remove(env, com, obj);
2315         lfsck_ibits_unlock(lh, LCK_EX);
2316         if (rc < 0 && rc != -ENODATA)
2317                 RETURN(rc);
2318
2319         *type = LNIT_MUL_REF;
2320
2321         /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT has
2322          * ever tried to verify some remote MDT-object that resides on this
2323          * MDT, but this MDT failed to respond such request. So means there
2324          * may be some remote name entry on other MDT that references this
2325          * object with another name, so we cannot know whether this linkEA
2326          * is valid or not. So keep it there and maybe resolved when next
2327          * LFSCK run. */
2328         if (ns->ln_flags & LF_INCOMPLETE)
2329                 RETURN(0);
2330
2331         /* The unique linkEA is invalid, even if the ".." name entry may be
2332          * valid, we still cannot know via which name entry this directory
2333          * will be referenced. Then handle it as pure orphan. */
2334         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2335                  "-"DFID, PFID(pfid));
2336         rc = lfsck_namespace_insert_orphan(env, com, obj,
2337                                            info->lti_tmpbuf, "D", NULL);
2338
2339         RETURN(rc);
2340 }
2341
2342 /**
2343  * Double Scan Directory object for single linkEA entry case.
2344  *
2345  * The given @child has unique linkEA entry. If the linkEA entry is valid,
2346  * then check whether the name is in the namespace or not, if not, add the
2347  * missing name entry back to namespace. If the linkEA entry is invalid,
2348  * then remove it and insert the @child in the .lustre/lost+found/MDTxxxx/
2349  * as an orphan.
2350  *
2351  * \param[in] env       pointer to the thread context
2352  * \param[in] com       pointer to the lfsck component
2353  * \param[in] child     pointer to the directory to be double scanned
2354  * \param[in] pfid      the FID corresponding to the ".." entry
2355  * \param[in] ldata     pointer to the linkEA data for the given @child
2356  * \param[in,out] lh    ldlm lock handler for the given @child
2357  * \param[out] type     to tell the caller what the inconsistency is
2358  * \param[in] retry     if found inconsistency, but the caller does not hold
2359  *                      ldlm lock on the @child, then set @retry as true
2360  *
2361  * \retval              positive number for repaired cases
2362  * \retval              0 if nothing to be repaired
2363  * \retval              negative error number on failure
2364  */
2365 static int
2366 lfsck_namespace_dsd_single(const struct lu_env *env,
2367                            struct lfsck_component *com,
2368                            struct dt_object *child,
2369                            const struct lu_fid *pfid,
2370                            struct linkea_data *ldata,
2371                            struct lustre_handle *lh,
2372                            enum lfsck_namespace_inconsistency_type *type,
2373                            bool *retry)
2374 {
2375         struct lfsck_thread_info *info          = lfsck_env_info(env);
2376         struct lu_name           *cname         = &info->lti_name;
2377         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2378         struct lu_fid             tfid;
2379         struct lfsck_namespace   *ns            = com->lc_file_ram;
2380         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2381         struct dt_object         *parent        = NULL;
2382         struct lmv_mds_md_v1     *lmv;
2383         int                       rc            = 0;
2384         ENTRY;
2385
2386         lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid, info->lti_key);
2387         /* The unique linkEA entry with bad parent will be handled as orphan. */
2388         if (!fid_is_sane(&tfid)) {
2389                 if (!lustre_handle_is_used(lh) && retry != NULL)
2390                         *retry = true;
2391                 else
2392                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2393                                                         pfid, lh, type);
2394
2395                 GOTO(out, rc);
2396         }
2397
2398         parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2399         if (IS_ERR(parent))
2400                 GOTO(out, rc = PTR_ERR(parent));
2401
2402         /* We trust the unique linkEA entry in spite of whether it matches the
2403          * ".." name entry or not. Because even if the linkEA entry is wrong
2404          * and the ".." name entry is right, we still cannot know via which
2405          * name entry the child will be referenced, since all known entries
2406          * have been verified during the first-stage scanning. */
2407         if (!dt_object_exists(parent)) {
2408                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2409                  * has ever tried to verify some remote MDT-object that resides
2410                  * on this MDT, but this MDT failed to respond such request. So
2411                  * means there may be some remote name entry on other MDT that
2412                  * references this object with another name, so we cannot know
2413                  * whether this linkEA is valid or not. So keep it there and
2414                  * maybe resolved when next LFSCK run. */
2415                 if (ns->ln_flags & LF_INCOMPLETE)
2416                         GOTO(out, rc = 0);
2417
2418                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2419                         *retry = true;
2420
2421                         GOTO(out, rc = 0);
2422                 }
2423
2424                 lfsck_ibits_unlock(lh, LCK_EX);
2425
2426 lost_parent:
2427                 lmv = &info->lti_lmv;
2428                 rc = lfsck_read_stripe_lmv(env, child, lmv);
2429                 if (rc != 0 && rc != -ENODATA)
2430                         GOTO(out, rc);
2431
2432                 if (rc == -ENODATA || lmv->lmv_magic != LMV_MAGIC_STRIPE) {
2433                         lmv = NULL;
2434                 } else if (lfsck_shard_name_to_index(env,
2435                                         cname->ln_name, cname->ln_namelen,
2436                                         S_IFDIR, cfid) < 0) {
2437                         /* It is an invalid name entry, we
2438                          * cannot trust the parent also. */
2439                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2440                                                 ldata, cname, &tfid, true);
2441                         if (rc < 0)
2442                                 GOTO(out, rc);
2443
2444                         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2445                                  "-"DFID, PFID(pfid));
2446                         rc = lfsck_namespace_insert_orphan(env, com, child,
2447                                                 info->lti_tmpbuf, "S", NULL);
2448
2449                         GOTO(out, rc);
2450                 }
2451
2452                 /* Create the lost parent as an orphan. */
2453                 rc = lfsck_namespace_create_orphan_dir(env, com, parent, lmv);
2454                 if (rc >= 0) {
2455                         /* Add the missing name entry to the parent. */
2456                         rc = lfsck_namespace_insert_normal(env, com, parent,
2457                                                         child, cname->ln_name);
2458                         if (unlikely(rc == -EEXIST)) {
2459                                 /* Unfortunately, someone reused the name
2460                                  * under the parent by race. So we have
2461                                  * to remove the linkEA entry from
2462                                  * current child object. It means that the
2463                                  * LFSCK cannot recover the system
2464                                  * totally back to its original status,
2465                                  * but it is necessary to make the
2466                                  * current system to be consistent. */
2467                                 rc = lfsck_namespace_shrink_linkea(env,
2468                                                 com, child, ldata,
2469                                                 cname, &tfid, true);
2470                                 if (rc >= 0) {
2471                                         snprintf(info->lti_tmpbuf,
2472                                                  sizeof(info->lti_tmpbuf),
2473                                                  "-"DFID, PFID(pfid));
2474                                         rc = lfsck_namespace_insert_orphan(env,
2475                                                 com, child, info->lti_tmpbuf,
2476                                                 "D", NULL);
2477                                 }
2478                         }
2479                 }
2480
2481                 GOTO(out, rc);
2482         }
2483
2484         /* The unique linkEA entry with bad parent will be handled as orphan. */
2485         if (unlikely(!dt_try_as_dir(env, parent))) {
2486                 if (!lustre_handle_is_used(lh) && retry != NULL)
2487                         *retry = true;
2488                 else
2489                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2490                                                         pfid, lh, type);
2491
2492                 GOTO(out, rc);
2493         }
2494
2495         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2496                        (const struct dt_key *)cname->ln_name);
2497         if (rc == -ENOENT) {
2498                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2499                  * has ever tried to verify some remote MDT-object that resides
2500                  * on this MDT, but this MDT failed to respond such request. So
2501                  * means there may be some remote name entry on other MDT that
2502                  * references this object with another name, so we cannot know
2503                  * whether this linkEA is valid or not. So keep it there and
2504                  * maybe resolved when next LFSCK run. */
2505                 if (ns->ln_flags & LF_INCOMPLETE)
2506                         GOTO(out, rc = 0);
2507
2508                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2509                         *retry = true;
2510
2511                         GOTO(out, rc = 0);
2512                 }
2513
2514                 lfsck_ibits_unlock(lh, LCK_EX);
2515                 rc = lfsck_namespace_check_name(env, parent, child, cname);
2516                 if (rc == -ENOENT)
2517                         goto lost_parent;
2518
2519                 if (rc < 0)
2520                         GOTO(out, rc);
2521
2522                 /* It is an invalid name entry, drop it. */
2523                 if (unlikely(rc > 0)) {
2524                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2525                                                 ldata, cname, &tfid, true);
2526                         if (rc >= 0) {
2527                                 snprintf(info->lti_tmpbuf,
2528                                          sizeof(info->lti_tmpbuf),
2529                                          "-"DFID, PFID(pfid));
2530                                 rc = lfsck_namespace_insert_orphan(env, com,
2531                                         child, info->lti_tmpbuf, "D", NULL);
2532                         }
2533
2534                         GOTO(out, rc);
2535                 }
2536
2537                 /* Add the missing name entry back to the namespace. */
2538                 rc = lfsck_namespace_insert_normal(env, com, parent, child,
2539                                                    cname->ln_name);
2540                 if (unlikely(rc == -ESTALE))
2541                         /* It may happen when the remote object has been
2542                          * removed, but the local MDT is not aware of that. */
2543                         goto lost_parent;
2544
2545                 if (unlikely(rc == -EEXIST)) {
2546                         /* Unfortunately, someone reused the name under the
2547                          * parent by race. So we have to remove the linkEA
2548                          * entry from current child object. It means that the
2549                          * LFSCK cannot recover the system totally back to
2550                          * its original status, but it is necessary to make
2551                          * the current system to be consistent.
2552                          *
2553                          * It also may be because of the LFSCK found some
2554                          * internal status of create operation. Under such
2555                          * case, nothing to be done. */
2556                         rc = lfsck_namespace_shrink_linkea_cond(env, com,
2557                                         parent, child, ldata, cname, &tfid);
2558                         if (rc >= 0) {
2559                                 snprintf(info->lti_tmpbuf,
2560                                          sizeof(info->lti_tmpbuf),
2561                                          "-"DFID, PFID(pfid));
2562                                 rc = lfsck_namespace_insert_orphan(env, com,
2563                                         child, info->lti_tmpbuf, "D", NULL);
2564                         }
2565                 }
2566
2567                 GOTO(out, rc);
2568         }
2569
2570         if (rc != 0)
2571                 GOTO(out, rc);
2572
2573         if (!lu_fid_eq(&tfid, cfid)) {
2574                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2575                         *retry = true;
2576
2577                         GOTO(out, rc = 0);
2578                 }
2579
2580                 lfsck_ibits_unlock(lh, LCK_EX);
2581                 /* The name entry references another MDT-object that
2582                  * may be created by the LFSCK for repairing dangling
2583                  * name entry. Try to replace it. */
2584                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
2585                                                   &tfid, cname);
2586                 if (rc == 0)
2587                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2588                                                         pfid, lh, type);
2589
2590                 GOTO(out, rc);
2591         }
2592
2593         if (fid_is_zero(pfid))
2594                 GOTO(out, rc = 0);
2595
2596         /* The ".." name entry is wrong, update it. */
2597         if (!lu_fid_eq(pfid, lfsck_dto2fid(parent))) {
2598                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2599                         *retry = true;
2600
2601                         GOTO(out, rc = 0);
2602                 }
2603
2604                 *type = LNIT_UNMATCHED_PAIRS;
2605                 rc = lfsck_namespace_repair_unmatched_pairs(env, com, child,
2606                                                 lfsck_dto2fid(parent), cname);
2607         }
2608
2609         GOTO(out, rc);
2610
2611 out:
2612         if (parent != NULL && !IS_ERR(parent))
2613                 lfsck_object_put(env, parent);
2614
2615         return rc;
2616 }
2617
2618 /**
2619  * Double Scan Directory object for multiple linkEA entries case.
2620  *
2621  * The given @child has multiple linkEA entries. There is at most one linkEA
2622  * entry will be valid, all the others will be removed. Firstly, the function
2623  * will try to find out the linkEA entry for which the name entry exists under
2624  * the given parent (@pfid). If there is no linkEA entry that matches the given
2625  * ".." name entry, then tries to find out the first linkEA entry that both the
2626  * parent and the name entry exist to rebuild a new ".." name entry.
2627  *
2628  * \param[in] env       pointer to the thread context
2629  * \param[in] com       pointer to the lfsck component
2630  * \param[in] child     pointer to the directory to be double scanned
2631  * \param[in] pfid      the FID corresponding to the ".." entry
2632  * \param[in] ldata     pointer to the linkEA data for the given @child
2633  * \param[in,out] lh    ldlm lock handler for the given @child
2634  * \param[out] type     to tell the caller what the inconsistency is
2635  * \param[in] lpf       true if the ".." entry is under lost+found/MDTxxxx/
2636  *
2637  * \retval              positive number for repaired cases
2638  * \retval              0 if nothing to be repaired
2639  * \retval              negative error number on failure
2640  */
2641 static int
2642 lfsck_namespace_dsd_multiple(const struct lu_env *env,
2643                              struct lfsck_component *com,
2644                              struct dt_object *child,
2645                              const struct lu_fid *pfid,
2646                              struct linkea_data *ldata,
2647                              struct lustre_handle *lh,
2648                              enum lfsck_namespace_inconsistency_type *type,
2649                              bool lpf)
2650 {
2651         struct lfsck_thread_info *info          = lfsck_env_info(env);
2652         struct lu_name           *cname         = &info->lti_name;
2653         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2654         struct lu_fid            *pfid2         = &info->lti_fid3;
2655         struct lu_fid             tfid;
2656         struct lfsck_namespace   *ns            = com->lc_file_ram;
2657         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2658         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2659         struct dt_object         *parent        = NULL;
2660         struct linkea_data        ldata_new     = { NULL };
2661         int                       dirent_count  = 0;
2662         int                       linkea_count  = 0;
2663         int                       rc            = 0;
2664         bool                      once          = true;
2665         ENTRY;
2666
2667 again:
2668         while (ldata->ld_lee != NULL) {
2669                 lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid,
2670                                                     info->lti_key);
2671                 /* Drop repeated linkEA entries. */
2672                 lfsck_namespace_filter_linkea_entry(ldata, cname, &tfid, true);
2673                 /* Drop invalid linkEA entry. */
2674                 if (!fid_is_sane(&tfid)) {
2675                         linkea_del_buf(ldata, cname);
2676                         linkea_count++;
2677                         continue;
2678                 }
2679
2680                 /* If current dotdot is the .lustre/lost+found/MDTxxxx/,
2681                  * then it is possible that: the directry object has ever
2682                  * been lost, but its name entry was there. In the former
2683                  * LFSCK run, during the first-stage scanning, the LFSCK
2684                  * found the dangling name entry, but it did not recreate
2685                  * the lost object, and when moved to the second-stage
2686                  * scanning, some children objects of the lost directory
2687                  * object were found, then the LFSCK recreated such lost
2688                  * directory object as an orphan.
2689                  *
2690                  * When the LFSCK runs again, if the dangling name is still
2691                  * there, the LFSCK should move the orphan directory object
2692                  * back to the normal namespace. */
2693                 if (!lpf && !lu_fid_eq(pfid, &tfid) && once) {
2694                         linkea_next_entry(ldata);
2695                         continue;
2696                 }
2697
2698                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2699                 if (IS_ERR(parent))
2700                         RETURN(PTR_ERR(parent));
2701
2702                 if (!dt_object_exists(parent)) {
2703                         lfsck_object_put(env, parent);
2704                         if (ldata->ld_leh->leh_reccount > 1) {
2705                                 /* If it is NOT the last linkEA entry, then
2706                                  * there is still other chance to make the
2707                                  * child to be visible via other parent, then
2708                                  * remove this linkEA entry. */
2709                                 linkea_del_buf(ldata, cname);
2710                                 linkea_count++;
2711                                 continue;
2712                         }
2713
2714                         break;
2715                 }
2716
2717                 /* The linkEA entry with bad parent will be removed. */
2718                 if (unlikely(!dt_try_as_dir(env, parent))) {
2719                         lfsck_object_put(env, parent);
2720                         linkea_del_buf(ldata, cname);
2721                         linkea_count++;
2722                         continue;
2723                 }
2724
2725                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2726                                (const struct dt_key *)cname->ln_name);
2727                 *pfid2 = *lfsck_dto2fid(parent);
2728                 if (rc == -ENOENT) {
2729                         lfsck_object_put(env, parent);
2730                         linkea_next_entry(ldata);
2731                         continue;
2732                 }
2733
2734                 if (rc != 0) {
2735                         lfsck_object_put(env, parent);
2736
2737                         RETURN(rc);
2738                 }
2739
2740                 if (lu_fid_eq(&tfid, cfid)) {
2741                         lfsck_object_put(env, parent);
2742                         if (!lu_fid_eq(pfid, pfid2)) {
2743                                 *type = LNIT_UNMATCHED_PAIRS;
2744                                 rc = lfsck_namespace_repair_unmatched_pairs(env,
2745                                                 com, child, pfid2, cname);
2746
2747                                 RETURN(rc);
2748                         }
2749
2750 rebuild:
2751                         /* It is the most common case that we find the
2752                          * name entry corresponding to the linkEA entry
2753                          * that matches the ".." name entry. */
2754                         rc = linkea_data_new(&ldata_new, &info->lti_big_buf);
2755                         if (rc != 0)
2756                                 RETURN(rc);
2757
2758                         rc = linkea_add_buf(&ldata_new, cname, pfid2);
2759                         if (rc != 0)
2760                                 RETURN(rc);
2761
2762                         rc = lfsck_namespace_rebuild_linkea(env, com, child,
2763                                                             &ldata_new);
2764                         if (rc < 0)
2765                                 RETURN(rc);
2766
2767                         linkea_del_buf(ldata, cname);
2768                         linkea_count++;
2769                         linkea_first_entry(ldata);
2770                         /* There may be some invalid dangling name entries under
2771                          * other parent directories, remove all of them. */
2772                         while (ldata->ld_lee != NULL) {
2773                                 lfsck_namespace_unpack_linkea_entry(ldata,
2774                                                 cname, &tfid, info->lti_key);
2775                                 if (!fid_is_sane(&tfid))
2776                                         goto next;
2777
2778                                 parent = lfsck_object_find_bottom(env, lfsck,
2779                                                                   &tfid);
2780                                 if (IS_ERR(parent)) {
2781                                         rc = PTR_ERR(parent);
2782                                         if (rc != -ENOENT &&
2783                                             bk->lb_param & LPF_FAILOUT)
2784                                                 RETURN(rc);
2785
2786                                         goto next;
2787                                 }
2788
2789                                 if (!dt_object_exists(parent)) {
2790                                         lfsck_object_put(env, parent);
2791                                         goto next;
2792                                 }
2793
2794                                 rc = lfsck_namespace_repair_dirent(env, com,
2795                                         parent, child, cname->ln_name,
2796                                         cname->ln_name, S_IFDIR, false, true);
2797                                 lfsck_object_put(env, parent);
2798                                 if (rc < 0) {
2799                                         if (bk->lb_param & LPF_FAILOUT)
2800                                                 RETURN(rc);
2801
2802                                         goto next;
2803                                 }
2804
2805                                 dirent_count += rc;
2806
2807 next:
2808                                 linkea_del_buf(ldata, cname);
2809                         }
2810
2811                         ns->ln_dirent_repaired += dirent_count;
2812
2813                         RETURN(rc);
2814                 }
2815
2816                 lfsck_ibits_unlock(lh, LCK_EX);
2817                 /* The name entry references another MDT-object that may be
2818                  * created by the LFSCK for repairing dangling name entry.
2819                  * Try to replace it. */
2820                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
2821                                                   &tfid, cname);
2822                 lfsck_object_put(env, parent);
2823                 if (rc < 0)
2824                         RETURN(rc);
2825
2826                 if (rc > 0)
2827                         goto rebuild;
2828
2829                 linkea_del_buf(ldata, cname);
2830         }
2831
2832         linkea_first_entry(ldata);
2833         if (ldata->ld_leh->leh_reccount == 1) {
2834                 rc = lfsck_namespace_dsd_single(env, com, child, pfid, ldata,
2835                                                 lh, type, NULL);
2836
2837                 if (rc == 0 && fid_is_zero(pfid) && linkea_count > 0)
2838                         rc = lfsck_namespace_rebuild_linkea(env, com, child,
2839                                                             ldata);
2840
2841                 RETURN(rc);
2842         }
2843
2844         /* All linkEA entries are invalid and removed, then handle the @child
2845          * as an orphan.*/
2846         if (ldata->ld_leh->leh_reccount == 0) {
2847                 rc = lfsck_namespace_dsd_orphan(env, com, child, pfid, lh,
2848                                                 type);
2849
2850                 RETURN(rc);
2851         }
2852
2853         /* If the dangling name entry for the orphan directory object has
2854          * been remvoed, then just check whether the directory object is
2855          * still under the .lustre/lost+found/MDTxxxx/ or not. */
2856         if (lpf) {
2857                 lpf = false;
2858                 goto again;
2859         }
2860
2861         /* There is no linkEA entry that matches the ".." name entry. Find
2862          * the first linkEA entry that both parent and name entry exist to
2863          * rebuild a new ".." name entry. */
2864         if (once) {
2865                 once = false;
2866                 goto again;
2867         }
2868
2869         RETURN(rc);
2870 }
2871
2872 /**
2873  * Repair the object's nlink attribute.
2874  *
2875  * If all the known name entries have been verified, then the object's hard
2876  * link attribute should match the object's linkEA entries count unless the
2877  * object's has too much hard link to be recorded in the linkEA. Such cases
2878  * should have been marked in the LFSCK trace file. Otherwise, trust the
2879  * linkEA to update the object's nlink attribute.
2880  *
2881  * \param[in] env       pointer to the thread context
2882  * \param[in] com       pointer to the lfsck component
2883  * \param[in] obj       pointer to the dt_object to be handled
2884  * \param[in,out] la    pointer to buffer to object's attribute before
2885  *                      and after the repairing
2886  *
2887  * \retval              positive number for repaired cases
2888  * \retval              0 if nothing to be repaired
2889  * \retval              negative error number on failure
2890  */
2891 static int lfsck_namespace_repair_nlink(const struct lu_env *env,
2892                                         struct lfsck_component *com,
2893                                         struct dt_object *obj,
2894                                         struct lu_attr *la)
2895 {
2896         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2897         struct lu_fid                   *tfid   = &info->lti_fid3;
2898         struct lfsck_namespace          *ns     = com->lc_file_ram;
2899         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2900         struct dt_device                *dev    = lfsck_obj2dev(obj);
2901         const struct lu_fid             *cfid   = lfsck_dto2fid(obj);
2902         struct thandle                  *th     = NULL;
2903         struct linkea_data               ldata  = { NULL };
2904         struct lustre_handle             lh     = { 0 };
2905         __u32                            old    = la->la_nlink;
2906         int                              idx;
2907         int                              rc     = 0;
2908         __u8                             flags;
2909         ENTRY;
2910
2911         LASSERT(!dt_object_remote(obj));
2912         LASSERT(S_ISREG(lfsck_object_type(obj)));
2913
2914         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
2915                               MDS_INODELOCK_UPDATE, LCK_PW);
2916         if (rc != 0)
2917                 GOTO(log, rc);
2918
2919         th = dt_trans_create(env, dev);
2920         if (IS_ERR(th))
2921                 GOTO(log, rc = PTR_ERR(th));
2922
2923         la->la_valid = LA_NLINK;
2924         rc = dt_declare_attr_set(env, obj, la, th);
2925         if (rc != 0)
2926                 GOTO(stop, rc);
2927
2928         rc = dt_trans_start_local(env, dev, th);
2929         if (rc != 0)
2930                 GOTO(stop, rc);
2931
2932         dt_write_lock(env, obj, 0);
2933         /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT has
2934          * ever tried to verify some remote MDT-object that resides on this
2935          * MDT, but this MDT failed to respond such request. So means there
2936          * may be some remote name entry on other MDT that references this
2937          * object with another name, so we cannot know whether this linkEA
2938          * is valid or not. So keep it there and maybe resolved when next
2939          * LFSCK run. */
2940         if (ns->ln_flags & LF_INCOMPLETE)
2941                 GOTO(unlock, rc = 0);
2942
2943         fid_cpu_to_be(tfid, cfid);
2944         idx = lfsck_sub_trace_file_fid2idx(cfid);
2945         rc = dt_lookup(env, com->lc_sub_trace_objs[idx].lsto_obj,
2946                        (struct dt_rec *)&flags, (const struct dt_key *)tfid);
2947         if (rc != 0)
2948                 GOTO(unlock, rc);
2949
2950         if (flags & LNTF_SKIP_NLINK)
2951                 GOTO(unlock, rc = 0);
2952
2953         rc = dt_attr_get(env, obj, la);
2954         if (rc != 0)
2955                 GOTO(unlock, rc = (rc == -ENOENT ? 0 : rc));
2956
2957         rc = lfsck_links_read2(env, obj, &ldata);
2958         if (rc != 0)
2959                 GOTO(unlock, rc = (rc == -ENODATA ? 0 : rc));
2960
2961         if (la->la_nlink == ldata.ld_leh->leh_reccount ||
2962             unlikely(la->la_nlink == 0))
2963                 GOTO(unlock, rc = 0);
2964
2965         la->la_nlink = ldata.ld_leh->leh_reccount;
2966         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2967                 GOTO(unlock, rc = 1);
2968
2969         rc = dt_attr_set(env, obj, la, th);
2970
2971         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2972
2973 unlock:
2974         dt_write_unlock(env, obj);
2975
2976 stop:
2977         dt_trans_stop(env, dev, th);
2978
2979 log:
2980         lfsck_ibits_unlock(&lh, LCK_PW);
2981         CDEBUG(D_LFSCK, "%s: namespace LFSCK repaired the object "DFID"'s "
2982                "nlink count from %u to %u: rc = %d\n",
2983                lfsck_lfsck2name(lfsck), PFID(cfid), old, la->la_nlink, rc);
2984
2985         if (rc != 0)
2986                 ns->ln_flags |= LF_INCONSISTENT;
2987
2988         return rc;
2989 }
2990
2991 /**
2992  * Double scan the directory object for namespace LFSCK.
2993  *
2994  * This function will verify the <parent, child> pairs in the namespace tree:
2995  * the parent references the child via some name entry that should be in the
2996  * child's linkEA entry, the child should back references the parent via its
2997  * ".." name entry.
2998  *
2999  * The LFSCK will scan every linkEA entry in turn until find out the first
3000  * matched pairs. If found, then all other linkEA entries will be dropped.
3001  * If all the linkEA entries cannot match the ".." name entry, then there
3002  * are serveral possible cases:
3003  *
3004  * 1) If there is only one linkEA entry, then trust it as long as the PFID
3005  *    in the linkEA entry is valid.
3006  *
3007  * 2) If there are multiple linkEA entries, then try to find the linkEA
3008  *    that matches the ".." name entry. If found, then all other entries
3009  *    are invalid; otherwise, it is quite possible that the ".." name entry
3010  *    is corrupted. Under such case, the LFSCK will rebuild the ".." name
3011  *    entry according to the first valid linkEA entry (both the parent and
3012  *    the name entry should exist).
3013  *
3014  * 3) If the directory object has no (valid) linkEA entry, then the
3015  *    directory object will be handled as pure orphan and inserted
3016  *    in the .lustre/lost+found/MDTxxxx/ with the name:
3017  *    ${self_FID}-${PFID}-D-${conflict_version}
3018  *
3019  * \param[in] env       pointer to the thread context
3020  * \param[in] com       pointer to the lfsck component
3021  * \param[in] child     pointer to the directory object to be handled
3022  * \param[in] flags     to indicate the specical checking on the @child
3023  *
3024  * \retval              positive number for repaired cases
3025  * \retval              0 if nothing to be repaired
3026  * \retval              negative error number on failure
3027  */
3028 static int lfsck_namespace_double_scan_dir(const struct lu_env *env,
3029                                            struct lfsck_component *com,
3030                                            struct dt_object *child, __u8 flags)
3031 {
3032         struct lfsck_thread_info *info          = lfsck_env_info(env);
3033         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
3034         struct lu_fid            *pfid          = &info->lti_fid2;
3035         struct lfsck_namespace   *ns            = com->lc_file_ram;
3036         struct lfsck_instance    *lfsck         = com->lc_lfsck;
3037         struct lustre_handle      lh            = { 0 };
3038         struct linkea_data        ldata         = { NULL };
3039         bool                      unknown       = false;
3040         bool                      lpf           = false;
3041         bool                      retry         = false;
3042         enum lfsck_namespace_inconsistency_type type = LNIT_BAD_LINKEA;
3043         int                       rc            = 0;
3044         ENTRY;
3045
3046         LASSERT(!dt_object_remote(child));
3047
3048         if (flags & LNTF_UNCERTAIN_LMV) {
3049                 if (flags & LNTF_RECHECK_NAME_HASH) {
3050                         rc = lfsck_namespace_scan_shard(env, com, child);
3051                         if (rc < 0)
3052                                 RETURN(rc);
3053
3054                         ns->ln_striped_shards_scanned++;
3055                 } else {
3056                         ns->ln_striped_shards_skipped++;
3057                 }
3058         }
3059
3060         flags &= ~(LNTF_RECHECK_NAME_HASH | LNTF_UNCERTAIN_LMV);
3061         if (flags == 0)
3062                 RETURN(0);
3063
3064         if (flags & (LNTF_CHECK_LINKEA | LNTF_CHECK_PARENT) &&
3065             !(lfsck->li_bookmark_ram.lb_param & LPF_ALL_TGT)) {
3066                 CDEBUG(D_LFSCK, "%s: some MDT(s) maybe NOT take part in the"
3067                        "the namespace LFSCK, then the LFSCK cannot guarantee"
3068                        "all the name entries have been verified in first-stage"
3069                        "scanning. So have to skip orphan related handling for"
3070                        "the directory object "DFID" with remote name entry\n",
3071                        lfsck_lfsck2name(lfsck), PFID(cfid));
3072
3073                 RETURN(0);
3074         }
3075
3076         if (unlikely(!dt_try_as_dir(env, child)))
3077                 GOTO(out, rc = -ENOTDIR);
3078
3079         /* We only take ldlm lock on the @child when required. When the
3080          * logic comes here for the first time, it is always false. */
3081         if (0) {
3082
3083 lock:
3084                 rc = lfsck_ibits_lock(env, lfsck, child, &lh,
3085                                       MDS_INODELOCK_UPDATE |
3086                                       MDS_INODELOCK_XATTR, LCK_EX);
3087                 if (rc != 0)
3088                         GOTO(out, rc);
3089         }
3090
3091         dt_read_lock(env, child, 0);
3092         if (unlikely(lfsck_is_dead_obj(child))) {
3093                 dt_read_unlock(env, child);
3094
3095                 GOTO(out, rc = 0);
3096         }
3097
3098         rc = dt_lookup(env, child, (struct dt_rec *)pfid,
3099                        (const struct dt_key *)dotdot);
3100         if (rc != 0) {
3101                 if (rc != -ENOENT && rc != -ENODATA && rc != -EINVAL) {
3102                         dt_read_unlock(env, child);
3103
3104                         GOTO(out, rc);
3105                 }
3106
3107                 if (!lustre_handle_is_used(&lh)) {
3108                         dt_read_unlock(env, child);
3109                         goto lock;
3110                 }
3111
3112                 fid_zero(pfid);
3113         } else if (lfsck->li_lpf_obj != NULL &&
3114                    lu_fid_eq(pfid, lfsck_dto2fid(lfsck->li_lpf_obj))) {
3115                 lpf = true;
3116         } else if (unlikely(!fid_is_sane(pfid))) {
3117                 fid_zero(pfid);
3118         }
3119
3120         rc = lfsck_links_read(env, child, &ldata);
3121         dt_read_unlock(env, child);
3122         if (rc != 0) {
3123                 if (rc != -ENODATA && rc != -EINVAL)
3124                         GOTO(out, rc);
3125
3126                 if (!lustre_handle_is_used(&lh))
3127                         goto lock;
3128
3129                 if (rc == -EINVAL && !fid_is_zero(pfid)) {
3130                         /* Remove the corrupted linkEA. */
3131                         rc = lfsck_namespace_links_remove(env, com, child);
3132                         if (rc == 0)
3133                                 /* Here, because of the crashed linkEA, we
3134                                  * cannot know whether there is some parent
3135                                  * that references the child directory via
3136                                  * some name entry or not. So keep it there,
3137                                  * when the LFSCK run next time, if there is
3138                                  * some parent that references this object,
3139                                  * then the LFSCK can rebuild the linkEA;
3140                                  * otherwise, this object will be handled
3141                                  * as orphan as above. */
3142                                 unknown = true;
3143                 } else {
3144                         /* 1. If we have neither ".." nor linkEA,
3145                          *    then it is an orphan.
3146                          *
3147                          * 2. If we only have the ".." name entry,
3148                          *    but no parent references this child
3149                          *    directory, then handle it as orphan. */
3150                         lfsck_ibits_unlock(&lh, LCK_EX);
3151                         type = LNIT_MUL_REF;
3152
3153                         /* If the LFSCK is marked as LF_INCOMPLETE,
3154                          * then means some MDT has ever tried to
3155                          * verify some remote MDT-object that resides
3156                          * on this MDT, but this MDT failed to respond
3157                          * such request. So means there may be some
3158                          * remote name entry on other MDT that
3159                          * references this object with another name,
3160                          * so we cannot know whether this linkEA is
3161                          * valid or not. So keep it there and maybe
3162                          * resolved when next LFSCK run. */
3163                         if (ns->ln_flags & LF_INCOMPLETE)
3164                                 GOTO(out, rc = 0);
3165
3166                         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
3167                                  "-"DFID, PFID(pfid));
3168                         rc = lfsck_namespace_insert_orphan(env, com, child,
3169                                                 info->lti_tmpbuf, "D", NULL);
3170                 }
3171
3172                 GOTO(out, rc);
3173         }
3174
3175         linkea_first_entry(&ldata);
3176         /* This is the most common case: the object has unique linkEA entry. */
3177         if (ldata.ld_leh->leh_reccount == 1) {
3178                 rc = lfsck_namespace_dsd_single(env, com, child, pfid, &ldata,
3179                                                 &lh, &type, &retry);
3180                 if (retry) {
3181                         LASSERT(!lustre_handle_is_used(&lh));
3182
3183                         retry = false;
3184                         goto lock;
3185                 }
3186
3187                 GOTO(out, rc);
3188         }
3189
3190         if (!lustre_handle_is_used(&lh))
3191                 goto lock;
3192
3193         if (unlikely(ldata.ld_leh->leh_reccount == 0)) {
3194                 rc = lfsck_namespace_dsd_orphan(env, com, child, pfid, &lh,
3195                                                 &type);
3196
3197                 GOTO(out, rc);
3198         }
3199
3200         /* When we come here, the cases usually like that:
3201          * 1) The directory object has a corrupted linkEA entry. During the
3202          *    first-stage scanning, the LFSCK cannot know such corruption,
3203          *    then it appends the right linkEA entry according to the found
3204          *    name entry after the bad one.
3205          *
3206          * 2) The directory object has a right linkEA entry. During the
3207          *    first-stage scanning, the LFSCK finds some bad name entry,
3208          *    but the LFSCK cannot aware that at that time, then it adds
3209          *    the bad linkEA entry for further processing. */
3210         rc = lfsck_namespace_dsd_multiple(env, com, child, pfid, &ldata,
3211                                           &lh, &type, lpf);
3212
3213         GOTO(out, rc);
3214
3215 out:
3216         lfsck_ibits_unlock(&lh, LCK_EX);
3217         if (rc > 0) {
3218                 switch (type) {
3219                 case LNIT_BAD_LINKEA:
3220                         ns->ln_linkea_repaired++;
3221                         break;
3222                 case LNIT_UNMATCHED_PAIRS:
3223                         ns->ln_unmatched_pairs_repaired++;
3224                         break;
3225                 case LNIT_MUL_REF:
3226                         ns->ln_mul_ref_repaired++;
3227                         break;
3228                 default:
3229                         break;
3230                 }
3231         }
3232
3233         if (unknown)
3234                 ns->ln_unknown_inconsistency++;
3235
3236         return rc;
3237 }
3238
3239 /**
3240  * Double scan the MDT-object for namespace LFSCK.
3241  *
3242  * If the MDT-object contains invalid or repeated linkEA entries, then drop
3243  * those entries from the linkEA; if the linkEA becomes empty or the object
3244  * has no linkEA, then it is an orphan and will be added into the directory
3245  * .lustre/lost+found/MDTxxxx/; if the remote parent is lost, then recreate
3246  * the remote parent; if the name entry corresponding to some linkEA entry
3247  * is lost, then add the name entry back to the namespace.
3248  *
3249  * \param[in] env       pointer to the thread context
3250  * \param[in] com       pointer to the lfsck component
3251  * \param[in] child     pointer to the dt_object to be handled
3252  * \param[in] flags     some hints to indicate how the @child should be handled
3253  *
3254  * \retval              positive number for repaired cases
3255  * \retval              0 if nothing to be repaired
3256  * \retval              negative error number on failure
3257  */
3258 static int lfsck_namespace_double_scan_one(const struct lu_env *env,
3259                                            struct lfsck_component *com,
3260                                            struct dt_object *child, __u8 flags)
3261 {
3262         struct lfsck_thread_info *info     = lfsck_env_info(env);
3263         struct lu_attr           *la       = &info->lti_la;
3264         struct lu_name           *cname    = &info->lti_name;
3265         struct lu_fid            *pfid     = &info->lti_fid;
3266         struct lu_fid            *cfid     = &info->lti_fid2;
3267         struct lfsck_instance    *lfsck    = com->lc_lfsck;
3268         struct lfsck_namespace   *ns       = com->lc_file_ram;
3269         struct dt_object         *parent   = NULL;
3270         struct linkea_data        ldata    = { NULL };
3271         bool                      repaired = false;
3272         int                       count    = 0;
3273         int                       rc;
3274         ENTRY;
3275
3276         dt_read_lock(env, child, 0);
3277         if (unlikely(lfsck_is_dead_obj(child))) {
3278                 dt_read_unlock(env, child);
3279
3280                 RETURN(0);
3281         }
3282
3283         if (S_ISDIR(lfsck_object_type(child))) {
3284                 dt_read_unlock(env, child);
3285                 rc = lfsck_namespace_double_scan_dir(env, com, child, flags);
3286
3287                 RETURN(rc);
3288         }
3289
3290         rc = lfsck_links_read(env, child, &ldata);
3291         dt_read_unlock(env, child);
3292         if (rc != 0)
3293                 GOTO(out, rc);
3294
3295         linkea_first_entry(&ldata);
3296         while (ldata.ld_lee != NULL) {
3297                 lfsck_namespace_unpack_linkea_entry(&ldata, cname, pfid,
3298                                                     info->lti_key);
3299                 rc = lfsck_namespace_filter_linkea_entry(&ldata, cname, pfid,
3300                                                          false);
3301                 /* Found repeated linkEA entries */
3302                 if (rc > 0) {
3303                         rc = lfsck_namespace_shrink_linkea(env, com, child,
3304                                                 &ldata, cname, pfid, false);
3305                         if (rc < 0)
3306                                 GOTO(out, rc);
3307
3308                         if (rc == 0)
3309                                 continue;
3310
3311                         repaired = true;
3312
3313                         /* fall through */
3314                 }
3315
3316                 /* Invalid PFID in the linkEA entry. */
3317                 if (!fid_is_sane(pfid)) {
3318                         rc = lfsck_namespace_shrink_linkea(env, com, child,
3319                                                 &ldata, cname, pfid, true);
3320                         if (rc < 0)
3321                                 GOTO(out, rc);
3322
3323                         if (rc > 0)
3324                                 repaired = true;
3325
3326                         continue;
3327                 }
3328
3329                 parent = lfsck_object_find_bottom(env, lfsck, pfid);
3330                 if (IS_ERR(parent))
3331                         GOTO(out, rc = PTR_ERR(parent));
3332
3333                 if (!dt_object_exists(parent)) {
3334
3335 lost_parent:
3336                         if (ldata.ld_leh->leh_reccount > 1) {
3337                                 /* If it is NOT the last linkEA entry, then
3338                                  * there is still other chance to make the
3339                                  * child to be visible via other parent, then
3340                                  * remove this linkEA entry. */
3341                                 rc = lfsck_namespace_shrink_linkea(env, com,
3342                                         child, &ldata, cname, pfid, true);
3343                         } else {
3344                                 /* If the LFSCK is marked as LF_INCOMPLETE,
3345                                  * then means some MDT has ever tried to
3346                                  * verify some remote MDT-object that resides
3347                                  * on this MDT, but this MDT failed to respond
3348                                  * such request. So means there may be some
3349                                  * remote name entry on other MDT that
3350                                  * references this object with another name,
3351                                  * so we cannot know whether this linkEA is
3352                                  * valid or not. So keep it there and maybe
3353                                  * resolved when next LFSCK run. */
3354                                 if (ns->ln_flags & LF_INCOMPLETE) {
3355                                         lfsck_object_put(env, parent);
3356
3357                                         GOTO(out, rc = 0);
3358                                 }
3359
3360                                 /* Create the lost parent as an orphan. */
3361                                 rc = lfsck_namespace_create_orphan_dir(env, com,
3362                                                                 parent, NULL);
3363                                 if (rc < 0) {
3364                                         lfsck_object_put(env, parent);
3365
3366                                         GOTO(out, rc);
3367                                 }
3368
3369                                 if (rc > 0)
3370                                         repaired = true;
3371
3372                                 /* Add the missing name entry to the parent. */
3373                                 rc = lfsck_namespace_insert_normal(env, com,
3374                                                 parent, child, cname->ln_name);
3375                                 if (unlikely(rc == -EEXIST))
3376                                         /* Unfortunately, someone reused the
3377                                          * name under the parent by race. So we
3378                                          * have to remove the linkEA entry from
3379                                          * current child object. It means that
3380                                          * the LFSCK cannot recover the system
3381                                          * totally back to its original status,
3382                                          * but it is necessary to make the
3383                                          * current system to be consistent. */
3384                                         rc = lfsck_namespace_shrink_linkea(env,
3385                                                         com, child, &ldata,
3386                                                         cname, pfid, true);
3387                                 else
3388                                         linkea_next_entry(&ldata);
3389                         }
3390
3391                         lfsck_object_put(env, parent);
3392                         if (rc < 0)
3393                                 GOTO(out, rc);
3394
3395                         if (rc > 0)
3396                                 repaired = true;
3397
3398                         continue;
3399                 }
3400
3401                 /* The linkEA entry with bad parent will be removed. */
3402                 if (unlikely(!dt_try_as_dir(env, parent))) {
3403                         lfsck_object_put(env, parent);
3404                         rc = lfsck_namespace_shrink_linkea(env, com, child,
3405                                                 &ldata, cname, pfid, true);
3406                         if (rc < 0)
3407                                 GOTO(out, rc);
3408
3409                         if (rc > 0)
3410                                 repaired = true;
3411
3412                         continue;
3413                 }
3414
3415                 rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
3416                                (const struct dt_key *)cname->ln_name);
3417                 if (rc != 0 && rc != -ENOENT) {
3418                         lfsck_object_put(env, parent);
3419
3420                         GOTO(out, rc);
3421                 }
3422
3423                 if (rc == 0) {
3424                         if (lu_fid_eq(cfid, lfsck_dto2fid(child))) {
3425                                 /* It is the most common case that we
3426                                  * find the name entry corresponding
3427                                  * to the linkEA entry. */
3428                                 lfsck_object_put(env, parent);
3429                                 linkea_next_entry(&ldata);
3430                         } else {
3431                                 /* The name entry references another
3432                                  * MDT-object that may be created by
3433                                  * the LFSCK for repairing dangling
3434                                  * name entry. Try to replace it. */
3435                                 rc = lfsck_namespace_replace_cond(env, com,
3436                                                 parent, child, cfid, cname);
3437                                 lfsck_object_put(env, parent);
3438                                 if (rc < 0)
3439                                         GOTO(out, rc);