Whamcloud - gitweb
01bead5fea228c3cadaff62828fc71b711c3cca9
[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) 2012, 2013, 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_linkea.h>
38 #include <lustre_fid.h>
39 #include <lustre_lib.h>
40 #include <lustre_net.h>
41 #include <lustre/lustre_user.h>
42
43 #include "lfsck_internal.h"
44
45 #define LFSCK_NAMESPACE_MAGIC   0xA0629D03
46
47 static const char lfsck_namespace_name[] = "lfsck_namespace";
48
49 static void lfsck_namespace_le_to_cpu(struct lfsck_namespace *des,
50                                       struct lfsck_namespace *src)
51 {
52         des->ln_magic = le32_to_cpu(src->ln_magic);
53         des->ln_status = le32_to_cpu(src->ln_status);
54         des->ln_flags = le32_to_cpu(src->ln_flags);
55         des->ln_success_count = le32_to_cpu(src->ln_success_count);
56         des->ln_run_time_phase1 = le32_to_cpu(src->ln_run_time_phase1);
57         des->ln_run_time_phase2 = le32_to_cpu(src->ln_run_time_phase2);
58         des->ln_time_last_complete = le64_to_cpu(src->ln_time_last_complete);
59         des->ln_time_latest_start = le64_to_cpu(src->ln_time_latest_start);
60         des->ln_time_last_checkpoint =
61                                 le64_to_cpu(src->ln_time_last_checkpoint);
62         lfsck_position_le_to_cpu(&des->ln_pos_latest_start,
63                                  &src->ln_pos_latest_start);
64         lfsck_position_le_to_cpu(&des->ln_pos_last_checkpoint,
65                                  &src->ln_pos_last_checkpoint);
66         lfsck_position_le_to_cpu(&des->ln_pos_first_inconsistent,
67                                  &src->ln_pos_first_inconsistent);
68         des->ln_items_checked = le64_to_cpu(src->ln_items_checked);
69         des->ln_items_repaired = le64_to_cpu(src->ln_items_repaired);
70         des->ln_items_failed = le64_to_cpu(src->ln_items_failed);
71         des->ln_dirs_checked = le64_to_cpu(src->ln_dirs_checked);
72         des->ln_mlinked_checked = le64_to_cpu(src->ln_mlinked_checked);
73         des->ln_objs_checked_phase2 = le64_to_cpu(src->ln_objs_checked_phase2);
74         des->ln_objs_repaired_phase2 =
75                                 le64_to_cpu(src->ln_objs_repaired_phase2);
76         des->ln_objs_failed_phase2 = le64_to_cpu(src->ln_objs_failed_phase2);
77         des->ln_objs_nlink_repaired = le64_to_cpu(src->ln_objs_nlink_repaired);
78         des->ln_objs_lost_found = le64_to_cpu(src->ln_objs_lost_found);
79         fid_le_to_cpu(&des->ln_fid_latest_scanned_phase2,
80                       &src->ln_fid_latest_scanned_phase2);
81 }
82
83 static void lfsck_namespace_cpu_to_le(struct lfsck_namespace *des,
84                                       struct lfsck_namespace *src)
85 {
86         des->ln_magic = cpu_to_le32(src->ln_magic);
87         des->ln_status = cpu_to_le32(src->ln_status);
88         des->ln_flags = cpu_to_le32(src->ln_flags);
89         des->ln_success_count = cpu_to_le32(src->ln_success_count);
90         des->ln_run_time_phase1 = cpu_to_le32(src->ln_run_time_phase1);
91         des->ln_run_time_phase2 = cpu_to_le32(src->ln_run_time_phase2);
92         des->ln_time_last_complete = cpu_to_le64(src->ln_time_last_complete);
93         des->ln_time_latest_start = cpu_to_le64(src->ln_time_latest_start);
94         des->ln_time_last_checkpoint =
95                                 cpu_to_le64(src->ln_time_last_checkpoint);
96         lfsck_position_cpu_to_le(&des->ln_pos_latest_start,
97                                  &src->ln_pos_latest_start);
98         lfsck_position_cpu_to_le(&des->ln_pos_last_checkpoint,
99                                  &src->ln_pos_last_checkpoint);
100         lfsck_position_cpu_to_le(&des->ln_pos_first_inconsistent,
101                                  &src->ln_pos_first_inconsistent);
102         des->ln_items_checked = cpu_to_le64(src->ln_items_checked);
103         des->ln_items_repaired = cpu_to_le64(src->ln_items_repaired);
104         des->ln_items_failed = cpu_to_le64(src->ln_items_failed);
105         des->ln_dirs_checked = cpu_to_le64(src->ln_dirs_checked);
106         des->ln_mlinked_checked = cpu_to_le64(src->ln_mlinked_checked);
107         des->ln_objs_checked_phase2 = cpu_to_le64(src->ln_objs_checked_phase2);
108         des->ln_objs_repaired_phase2 =
109                                 cpu_to_le64(src->ln_objs_repaired_phase2);
110         des->ln_objs_failed_phase2 = cpu_to_le64(src->ln_objs_failed_phase2);
111         des->ln_objs_nlink_repaired = cpu_to_le64(src->ln_objs_nlink_repaired);
112         des->ln_objs_lost_found = cpu_to_le64(src->ln_objs_lost_found);
113         fid_cpu_to_le(&des->ln_fid_latest_scanned_phase2,
114                       &src->ln_fid_latest_scanned_phase2);
115 }
116
117 /**
118  * \retval +ve: the lfsck_namespace is broken, the caller should reset it.
119  * \retval 0: succeed.
120  * \retval -ve: failed cases.
121  */
122 static int lfsck_namespace_load(const struct lu_env *env,
123                                 struct lfsck_component *com)
124 {
125         int len = com->lc_file_size;
126         int rc;
127
128         rc = dt_xattr_get(env, com->lc_obj,
129                           lfsck_buf_get(env, com->lc_file_disk, len),
130                           XATTR_NAME_LFSCK_NAMESPACE, BYPASS_CAPA);
131         if (rc == len) {
132                 struct lfsck_namespace *ns = com->lc_file_ram;
133
134                 lfsck_namespace_le_to_cpu(ns,
135                                 (struct lfsck_namespace *)com->lc_file_disk);
136                 if (ns->ln_magic != LFSCK_NAMESPACE_MAGIC) {
137                         CWARN("%.16s: invalid lfsck_namespace magic "
138                               "0x%x != 0x%x\n",
139                               lfsck_lfsck2name(com->lc_lfsck),
140                               ns->ln_magic, LFSCK_NAMESPACE_MAGIC);
141                         rc = 1;
142                 } else {
143                         rc = 0;
144                 }
145         } else if (rc != -ENODATA) {
146                 CERROR("%.16s: fail to load lfsck_namespace, expected = %d, "
147                        "rc = %d\n", lfsck_lfsck2name(com->lc_lfsck), len, rc);
148                 if (rc >= 0)
149                         rc = 1;
150         }
151         return rc;
152 }
153
154 static int lfsck_namespace_store(const struct lu_env *env,
155                                  struct lfsck_component *com, bool init)
156 {
157         struct dt_object        *obj    = com->lc_obj;
158         struct lfsck_instance   *lfsck  = com->lc_lfsck;
159         struct thandle          *handle;
160         int                      len    = com->lc_file_size;
161         int                      rc;
162         ENTRY;
163
164         lfsck_namespace_cpu_to_le((struct lfsck_namespace *)com->lc_file_disk,
165                                   (struct lfsck_namespace *)com->lc_file_ram);
166         handle = dt_trans_create(env, lfsck->li_bottom);
167         if (IS_ERR(handle)) {
168                 rc = PTR_ERR(handle);
169                 CERROR("%.16s: fail to create trans for storing "
170                        "lfsck_namespace: %d\n,", lfsck_lfsck2name(lfsck), rc);
171                 RETURN(rc);
172         }
173
174         rc = dt_declare_xattr_set(env, obj,
175                                   lfsck_buf_get(env, com->lc_file_disk, len),
176                                   XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
177         if (rc != 0) {
178                 CERROR("%.16s: fail to declare trans for storing "
179                        "lfsck_namespace: %d\n,", lfsck_lfsck2name(lfsck), rc);
180                 GOTO(out, rc);
181         }
182
183         rc = dt_trans_start_local(env, lfsck->li_bottom, handle);
184         if (rc != 0) {
185                 CERROR("%.16s: fail to start trans for storing "
186                        "lfsck_namespace: %d\n,", lfsck_lfsck2name(lfsck), rc);
187                 GOTO(out, rc);
188         }
189
190         rc = dt_xattr_set(env, obj,
191                           lfsck_buf_get(env, com->lc_file_disk, len),
192                           XATTR_NAME_LFSCK_NAMESPACE,
193                           init ? LU_XATTR_CREATE : LU_XATTR_REPLACE,
194                           handle, BYPASS_CAPA);
195         if (rc != 0)
196                 CERROR("%.16s: fail to store lfsck_namespace, len = %d, "
197                        "rc = %d\n", lfsck_lfsck2name(lfsck), len, rc);
198
199         GOTO(out, rc);
200
201 out:
202         dt_trans_stop(env, lfsck->li_bottom, handle);
203         return rc;
204 }
205
206 static int lfsck_namespace_init(const struct lu_env *env,
207                                 struct lfsck_component *com)
208 {
209         struct lfsck_namespace *ns = (struct lfsck_namespace *)com->lc_file_ram;
210         int rc;
211
212         memset(ns, 0, sizeof(*ns));
213         ns->ln_magic = LFSCK_NAMESPACE_MAGIC;
214         ns->ln_status = LS_INIT;
215         down_write(&com->lc_sem);
216         rc = lfsck_namespace_store(env, com, true);
217         up_write(&com->lc_sem);
218         return rc;
219 }
220
221 static int lfsck_namespace_lookup(const struct lu_env *env,
222                                   struct lfsck_component *com,
223                                   const struct lu_fid *fid, __u8 *flags)
224 {
225         struct lu_fid *key = &lfsck_env_info(env)->lti_fid;
226         int            rc;
227
228         fid_cpu_to_be(key, fid);
229         rc = dt_lookup(env, com->lc_obj, (struct dt_rec *)flags,
230                        (const struct dt_key *)key, BYPASS_CAPA);
231         return rc;
232 }
233
234 static int lfsck_namespace_delete(const struct lu_env *env,
235                                   struct lfsck_component *com,
236                                   const struct lu_fid *fid)
237 {
238         struct lfsck_instance   *lfsck  = com->lc_lfsck;
239         struct lu_fid           *key    = &lfsck_env_info(env)->lti_fid;
240         struct thandle          *handle;
241         struct dt_object        *obj    = com->lc_obj;
242         int                      rc;
243         ENTRY;
244
245         handle = dt_trans_create(env, lfsck->li_bottom);
246         if (IS_ERR(handle))
247                 RETURN(PTR_ERR(handle));
248
249         rc = dt_declare_delete(env, obj, (const struct dt_key *)fid, handle);
250         if (rc != 0)
251                 GOTO(out, rc);
252
253         rc = dt_trans_start_local(env, lfsck->li_bottom, handle);
254         if (rc != 0)
255                 GOTO(out, rc);
256
257         fid_cpu_to_be(key, fid);
258         rc = dt_delete(env, obj, (const struct dt_key *)key, handle,
259                        BYPASS_CAPA);
260
261         GOTO(out, rc);
262
263 out:
264         dt_trans_stop(env, lfsck->li_bottom, handle);
265         return rc;
266 }
267
268 static int lfsck_namespace_update(const struct lu_env *env,
269                                   struct lfsck_component *com,
270                                   const struct lu_fid *fid,
271                                   __u8 flags, bool force)
272 {
273         struct lfsck_instance   *lfsck  = com->lc_lfsck;
274         struct lu_fid           *key    = &lfsck_env_info(env)->lti_fid;
275         struct thandle          *handle;
276         struct dt_object        *obj    = com->lc_obj;
277         int                      rc;
278         bool                     exist  = false;
279         __u8                     tf;
280         ENTRY;
281
282         rc = lfsck_namespace_lookup(env, com, fid, &tf);
283         if (rc != 0 && rc != -ENOENT)
284                 RETURN(rc);
285
286         if (rc == 0) {
287                 if (!force || flags == tf)
288                         RETURN(0);
289
290                 exist = true;
291                 handle = dt_trans_create(env, lfsck->li_bottom);
292                 if (IS_ERR(handle))
293                         RETURN(PTR_ERR(handle));
294
295                 rc = dt_declare_delete(env, obj, (const struct dt_key *)fid,
296                                        handle);
297                 if (rc != 0)
298                         GOTO(out, rc);
299         } else {
300                 handle = dt_trans_create(env, lfsck->li_bottom);
301                 if (IS_ERR(handle))
302                         RETURN(PTR_ERR(handle));
303         }
304
305         rc = dt_declare_insert(env, obj, (const struct dt_rec *)&flags,
306                                (const struct dt_key *)fid, handle);
307         if (rc != 0)
308                 GOTO(out, rc);
309
310         rc = dt_trans_start_local(env, lfsck->li_bottom, handle);
311         if (rc != 0)
312                 GOTO(out, rc);
313
314         fid_cpu_to_be(key, fid);
315         if (exist) {
316                 rc = dt_delete(env, obj, (const struct dt_key *)key, handle,
317                                BYPASS_CAPA);
318                 if (rc != 0) {
319                         CERROR("%s: fail to insert "DFID", rc = %d\n",
320                                lfsck_lfsck2name(com->lc_lfsck), PFID(fid), rc);
321                         GOTO(out, rc);
322                 }
323         }
324
325         rc = dt_insert(env, obj, (const struct dt_rec *)&flags,
326                        (const struct dt_key *)key, handle, BYPASS_CAPA, 1);
327
328         GOTO(out, rc);
329
330 out:
331         dt_trans_stop(env, lfsck->li_bottom, handle);
332         return rc;
333 }
334
335 static int lfsck_namespace_check_exist(const struct lu_env *env,
336                                        struct lfsck_instance *lfsck,
337                                        struct dt_object *obj, const char *name)
338 {
339         struct dt_object *dir = lfsck->li_obj_dir;
340         struct lu_fid    *fid = &lfsck_env_info(env)->lti_fid;
341         int               rc;
342         ENTRY;
343
344         if (unlikely(lfsck_is_dead_obj(obj)))
345                 RETURN(LFSCK_NAMEENTRY_DEAD);
346
347         rc = dt_lookup(env, dir, (struct dt_rec *)fid,
348                        (const struct dt_key *)name, BYPASS_CAPA);
349         if (rc == -ENOENT)
350                 RETURN(LFSCK_NAMEENTRY_REMOVED);
351
352         if (rc < 0)
353                 RETURN(rc);
354
355         if (!lu_fid_eq(fid, lfsck_dto2fid(obj)))
356                 RETURN(LFSCK_NAMEENTRY_RECREATED);
357
358         RETURN(0);
359 }
360
361 static int lfsck_declare_namespace_exec_dir(const struct lu_env *env,
362                                             struct dt_object *obj,
363                                             struct thandle *handle)
364 {
365         int rc;
366
367         /* For destroying all invalid linkEA entries. */
368         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, handle);
369         if (rc != 0)
370                 return rc;
371
372         /* For insert new linkEA entry. */
373         rc = dt_declare_xattr_set(env, obj,
374                         lfsck_buf_get_const(env, NULL, DEFAULT_LINKEA_SIZE),
375                         XATTR_NAME_LINK, 0, handle);
376         return rc;
377 }
378
379 static int lfsck_links_read(const struct lu_env *env, struct dt_object *obj,
380                             struct linkea_data *ldata)
381 {
382         int rc;
383
384         ldata->ld_buf =
385                 lu_buf_check_and_alloc(&lfsck_env_info(env)->lti_linkea_buf,
386                                        PAGE_CACHE_SIZE);
387         if (ldata->ld_buf->lb_buf == NULL)
388                 return -ENOMEM;
389
390         if (!dt_object_exists(obj))
391                 return -ENODATA;
392
393         rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK, BYPASS_CAPA);
394         if (rc == -ERANGE) {
395                 /* Buf was too small, figure out what we need. */
396                 lu_buf_free(ldata->ld_buf);
397                 rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK,
398                                   BYPASS_CAPA);
399                 if (rc < 0)
400                         return rc;
401
402                 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
403                 if (ldata->ld_buf->lb_buf == NULL)
404                         return -ENOMEM;
405
406                 rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK,
407                                   BYPASS_CAPA);
408         }
409         if (rc < 0)
410                 return rc;
411
412         linkea_init(ldata);
413
414         return 0;
415 }
416
417 static int lfsck_links_write(const struct lu_env *env, struct dt_object *obj,
418                              struct linkea_data *ldata, struct thandle *handle)
419 {
420         const struct lu_buf *buf = lfsck_buf_get_const(env,
421                                                        ldata->ld_buf->lb_buf,
422                                                        ldata->ld_leh->leh_len);
423
424         return dt_xattr_set(env, obj, buf, XATTR_NAME_LINK, 0, handle,
425                             BYPASS_CAPA);
426 }
427
428 /**
429  * \retval ve: removed entries
430  */
431 static int lfsck_linkea_entry_unpack(struct lfsck_instance *lfsck,
432                                      struct linkea_data *ldata,
433                                      struct lu_name *cname,
434                                      struct lu_fid *pfid)
435 {
436         struct link_ea_entry    *oldlee;
437         int                      oldlen;
438         int                      removed = 0;
439
440         linkea_entry_unpack(ldata->ld_lee, &ldata->ld_reclen, cname, pfid);
441         oldlee = ldata->ld_lee;
442         oldlen = ldata->ld_reclen;
443         linkea_next_entry(ldata);
444         while (ldata->ld_lee != NULL) {
445                 ldata->ld_reclen = (ldata->ld_lee->lee_reclen[0] << 8) |
446                                    ldata->ld_lee->lee_reclen[1];
447                 if (unlikely(ldata->ld_reclen == oldlen &&
448                              memcmp(ldata->ld_lee, oldlee, oldlen) == 0)) {
449                         linkea_del_buf(ldata, cname);
450                         removed++;
451                 } else {
452                         linkea_next_entry(ldata);
453                 }
454         }
455         ldata->ld_lee = oldlee;
456         ldata->ld_reclen = oldlen;
457         return removed;
458 }
459
460 /**
461  * \retval +ve  repaired
462  * \retval 0    no need to repair
463  * \retval -ve  error cases
464  */
465 static int lfsck_namespace_double_scan_one(const struct lu_env *env,
466                                            struct lfsck_component *com,
467                                            struct dt_object *child, __u8 flags)
468 {
469         struct lfsck_thread_info *info    = lfsck_env_info(env);
470         struct lu_attr           *la      = &info->lti_la;
471         struct lu_name           *cname   = &info->lti_name;
472         struct lu_fid            *pfid    = &info->lti_fid;
473         struct lu_fid            *cfid    = &info->lti_fid2;
474         struct lfsck_instance   *lfsck    = com->lc_lfsck;
475         struct lfsck_bookmark   *bk       = &lfsck->li_bookmark_ram;
476         struct lfsck_namespace  *ns       =
477                                 (struct lfsck_namespace *)com->lc_file_ram;
478         struct linkea_data       ldata    = { 0 };
479         struct thandle          *handle   = NULL;
480         bool                     locked   = false;
481         bool                     update   = false;
482         int                      rc;
483         ENTRY;
484
485         if (com->lc_journal) {
486
487 again:
488                 LASSERT(!locked);
489
490                 update = false;
491                 com->lc_journal = 1;
492                 handle = dt_trans_create(env, lfsck->li_next);
493                 if (IS_ERR(handle))
494                         RETURN(rc = PTR_ERR(handle));
495
496                 rc = dt_declare_xattr_set(env, child,
497                         lfsck_buf_get_const(env, NULL, DEFAULT_LINKEA_SIZE),
498                         XATTR_NAME_LINK, 0, handle);
499                 if (rc != 0)
500                         GOTO(stop, rc);
501
502                 rc = dt_trans_start(env, lfsck->li_next, handle);
503                 if (rc != 0)
504                         GOTO(stop, rc);
505
506                 dt_write_lock(env, child, MOR_TGT_CHILD);
507                 locked = true;
508         }
509
510         if (unlikely(lfsck_is_dead_obj(child)))
511                 GOTO(stop, rc = 0);
512
513         rc = dt_attr_get(env, child, la, BYPASS_CAPA);
514         if (rc == 0)
515                 rc = lfsck_links_read(env, child, &ldata);
516         if (rc != 0) {
517                 if ((bk->lb_param & LPF_DRYRUN) &&
518                     (rc == -EINVAL || rc == -ENODATA))
519                         rc = 1;
520
521                 GOTO(stop, rc);
522         }
523
524         linkea_first_entry(&ldata);
525         while (ldata.ld_lee != NULL) {
526                 struct dt_object *parent = NULL;
527
528                 rc = lfsck_linkea_entry_unpack(lfsck, &ldata, cname, pfid);
529                 if (rc > 0)
530                         update = true;
531
532                 if (!fid_is_sane(pfid))
533                         goto shrink;
534
535                 parent = lfsck_object_find(env, lfsck, pfid);
536                 if (parent == NULL)
537                         goto shrink;
538                 else if (IS_ERR(parent))
539                         GOTO(stop, rc = PTR_ERR(parent));
540
541                 if (!dt_object_exists(parent))
542                         goto shrink;
543
544                 /* XXX: Currently, skip remote object, the consistency for
545                  *      remote object will be processed in LFSCK phase III. */
546                 if (dt_object_remote(parent)) {
547                         lfsck_object_put(env, parent);
548                         linkea_next_entry(&ldata);
549                         continue;
550                 }
551
552                 if (unlikely(!dt_try_as_dir(env, parent)))
553                         goto shrink;
554
555                 /* To guarantee the 'name' is terminated with '0'. */
556                 memcpy(info->lti_key, cname->ln_name, cname->ln_namelen);
557                 info->lti_key[cname->ln_namelen] = 0;
558                 cname->ln_name = info->lti_key;
559                 rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
560                                (const struct dt_key *)cname->ln_name,
561                                BYPASS_CAPA);
562                 if (rc != 0 && rc != -ENOENT) {
563                         lfsck_object_put(env, parent);
564                         GOTO(stop, rc);
565                 }
566
567                 if (rc == 0) {
568                         if (lu_fid_eq(cfid, lfsck_dto2fid(child))) {
569                                 lfsck_object_put(env, parent);
570                                 linkea_next_entry(&ldata);
571                                 continue;
572                         }
573
574                         goto shrink;
575                 }
576
577                 /* If there is no name entry in the parent dir and the object
578                  * link count is less than the linkea entries count, then the
579                  * linkea entry should be removed. */
580                 if (ldata.ld_leh->leh_reccount > la->la_nlink)
581                         goto shrink;
582
583                 /* XXX: For the case of there is a linkea entry, but without
584                  *      name entry pointing to the object and its hard links
585                  *      count is not less than the object name entries count,
586                  *      then seems we should add the 'missed' name entry back
587                  *      to namespace, but before LFSCK phase III finished, we
588                  *      do not know whether the object has some inconsistency
589                  *      on other MDTs. So now, do NOT add the name entry back
590                  *      to the namespace, but keep the linkEA entry. LU-2914 */
591                 lfsck_object_put(env, parent);
592                 linkea_next_entry(&ldata);
593                 continue;
594
595 shrink:
596                 if (parent != NULL)
597                         lfsck_object_put(env, parent);
598                 if (bk->lb_param & LPF_DRYRUN)
599                         RETURN(1);
600
601                 CDEBUG(D_LFSCK, "Remove linkEA: "DFID"[%.*s], "DFID"\n",
602                        PFID(lfsck_dto2fid(child)), cname->ln_namelen, cname->ln_name,
603                        PFID(pfid));
604                 linkea_del_buf(&ldata, cname);
605                 update = true;
606         }
607
608         if (update) {
609                 if (!com->lc_journal) {
610                         com->lc_journal = 1;
611                         goto again;
612                 }
613
614                 rc = lfsck_links_write(env, child, &ldata, handle);
615         }
616
617         GOTO(stop, rc);
618
619 stop:
620         if (locked) {
621         /* XXX: For the case linkea entries count does not match the object hard
622          *      links count, we cannot update the later one simply. Before LFSCK
623          *      phase III finished, we cannot know whether there are some remote
624          *      name entries to be repaired or not. LU-2914 */
625                 if (rc == 0 && !lfsck_is_dead_obj(child) &&
626                     ldata.ld_leh != NULL &&
627                     ldata.ld_leh->leh_reccount != la->la_nlink)
628                         CWARN("%.16s: the object "DFID" linkEA entry count %u "
629                               "may not match its hardlink count %u\n",
630                               lfsck_lfsck2name(lfsck), PFID(cfid),
631                               ldata.ld_leh->leh_reccount, la->la_nlink);
632
633                 dt_write_unlock(env, child);
634         }
635
636         if (handle != NULL)
637                 dt_trans_stop(env, lfsck->li_next, handle);
638
639         if (rc == 0 && update) {
640                 ns->ln_objs_nlink_repaired++;
641                 rc = 1;
642         }
643
644         return rc;
645 }
646
647 /* namespace APIs */
648
649 static int lfsck_namespace_reset(const struct lu_env *env,
650                                  struct lfsck_component *com, bool init)
651 {
652         struct lfsck_instance   *lfsck = com->lc_lfsck;
653         struct lfsck_namespace  *ns    =
654                                 (struct lfsck_namespace *)com->lc_file_ram;
655         struct dt_object        *dto;
656         int                      rc;
657         ENTRY;
658
659         down_write(&com->lc_sem);
660         if (init) {
661                 memset(ns, 0, sizeof(*ns));
662         } else {
663                 __u32 count = ns->ln_success_count;
664                 __u64 last_time = ns->ln_time_last_complete;
665
666                 memset(ns, 0, sizeof(*ns));
667                 ns->ln_success_count = count;
668                 ns->ln_time_last_complete = last_time;
669         }
670         ns->ln_magic = LFSCK_NAMESPACE_MAGIC;
671         ns->ln_status = LS_INIT;
672
673         rc = local_object_unlink(env, lfsck->li_bottom, lfsck->li_local_root,
674                                  lfsck_namespace_name);
675         if (rc != 0)
676                 GOTO(out, rc);
677
678         dto = local_index_find_or_create(env, lfsck->li_los, lfsck->li_local_root,
679                                          lfsck_namespace_name,
680                                          S_IFREG | S_IRUGO | S_IWUSR,
681                                          &dt_lfsck_features);
682         if (IS_ERR(dto))
683                 GOTO(out, rc = PTR_ERR(dto));
684
685         com->lc_obj = dto;
686         rc = dto->do_ops->do_index_try(env, dto, &dt_lfsck_features);
687         if (rc != 0)
688                 GOTO(out, rc);
689
690         rc = lfsck_namespace_store(env, com, true);
691
692         GOTO(out, rc);
693
694 out:
695         up_write(&com->lc_sem);
696         return rc;
697 }
698
699 static void
700 lfsck_namespace_fail(const struct lu_env *env, struct lfsck_component *com,
701                      bool new_checked)
702 {
703         struct lfsck_namespace *ns = (struct lfsck_namespace *)com->lc_file_ram;
704
705         down_write(&com->lc_sem);
706         if (new_checked)
707                 com->lc_new_checked++;
708         ns->ln_items_failed++;
709         if (lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent))
710                 lfsck_pos_fill(env, com->lc_lfsck,
711                                &ns->ln_pos_first_inconsistent, false);
712         up_write(&com->lc_sem);
713 }
714
715 static int lfsck_namespace_checkpoint(const struct lu_env *env,
716                                       struct lfsck_component *com, bool init)
717 {
718         struct lfsck_instance   *lfsck = com->lc_lfsck;
719         struct lfsck_namespace  *ns    =
720                                 (struct lfsck_namespace *)com->lc_file_ram;
721         int                      rc;
722
723         if (com->lc_new_checked == 0 && !init)
724                 return 0;
725
726         down_write(&com->lc_sem);
727
728         if (init) {
729                 ns->ln_pos_latest_start = lfsck->li_pos_current;
730         } else {
731                 ns->ln_pos_last_checkpoint = lfsck->li_pos_current;
732                 ns->ln_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
733                                 HALF_SEC - lfsck->li_time_last_checkpoint);
734                 ns->ln_time_last_checkpoint = cfs_time_current_sec();
735                 ns->ln_items_checked += com->lc_new_checked;
736                 com->lc_new_checked = 0;
737         }
738
739         rc = lfsck_namespace_store(env, com, false);
740
741         up_write(&com->lc_sem);
742         return rc;
743 }
744
745 static int lfsck_namespace_prep(const struct lu_env *env,
746                                 struct lfsck_component *com)
747 {
748         struct lfsck_instance   *lfsck  = com->lc_lfsck;
749         struct lfsck_namespace  *ns     =
750                                 (struct lfsck_namespace *)com->lc_file_ram;
751         struct lfsck_position   *pos    = &com->lc_pos_start;
752
753         if (ns->ln_status == LS_COMPLETED) {
754                 int rc;
755
756                 rc = lfsck_namespace_reset(env, com, false);
757                 if (rc != 0)
758                         return rc;
759         }
760
761         down_write(&com->lc_sem);
762
763         ns->ln_time_latest_start = cfs_time_current_sec();
764
765         spin_lock(&lfsck->li_lock);
766         if (ns->ln_flags & LF_SCANNED_ONCE) {
767                 if (!lfsck->li_drop_dryrun ||
768                     lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent)) {
769                         ns->ln_status = LS_SCANNING_PHASE2;
770                         cfs_list_del_init(&com->lc_link);
771                         cfs_list_add_tail(&com->lc_link,
772                                           &lfsck->li_list_double_scan);
773                         if (!cfs_list_empty(&com->lc_link_dir))
774                                 cfs_list_del_init(&com->lc_link_dir);
775                         lfsck_pos_set_zero(pos);
776                 } else {
777                         ns->ln_status = LS_SCANNING_PHASE1;
778                         ns->ln_run_time_phase1 = 0;
779                         ns->ln_run_time_phase2 = 0;
780                         ns->ln_items_checked = 0;
781                         ns->ln_items_repaired = 0;
782                         ns->ln_items_failed = 0;
783                         ns->ln_dirs_checked = 0;
784                         ns->ln_mlinked_checked = 0;
785                         ns->ln_objs_checked_phase2 = 0;
786                         ns->ln_objs_repaired_phase2 = 0;
787                         ns->ln_objs_failed_phase2 = 0;
788                         ns->ln_objs_nlink_repaired = 0;
789                         ns->ln_objs_lost_found = 0;
790                         fid_zero(&ns->ln_fid_latest_scanned_phase2);
791                         if (cfs_list_empty(&com->lc_link_dir))
792                                 cfs_list_add_tail(&com->lc_link_dir,
793                                                   &lfsck->li_list_dir);
794                         *pos = ns->ln_pos_first_inconsistent;
795                 }
796         } else {
797                 ns->ln_status = LS_SCANNING_PHASE1;
798                 if (cfs_list_empty(&com->lc_link_dir))
799                         cfs_list_add_tail(&com->lc_link_dir,
800                                           &lfsck->li_list_dir);
801                 if (!lfsck->li_drop_dryrun ||
802                     lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent)) {
803                         *pos = ns->ln_pos_last_checkpoint;
804                         pos->lp_oit_cookie++;
805                 } else {
806                         *pos = ns->ln_pos_first_inconsistent;
807                 }
808         }
809         spin_unlock(&lfsck->li_lock);
810
811         up_write(&com->lc_sem);
812         return 0;
813 }
814
815 static int lfsck_namespace_exec_oit(const struct lu_env *env,
816                                     struct lfsck_component *com,
817                                     struct dt_object *obj)
818 {
819         down_write(&com->lc_sem);
820         com->lc_new_checked++;
821         if (S_ISDIR(lfsck_object_type(obj)))
822                 ((struct lfsck_namespace *)com->lc_file_ram)->ln_dirs_checked++;
823         up_write(&com->lc_sem);
824         return 0;
825 }
826
827 static int lfsck_namespace_exec_dir(const struct lu_env *env,
828                                     struct lfsck_component *com,
829                                     struct dt_object *obj,
830                                     struct lu_dirent *ent)
831 {
832         struct lfsck_thread_info   *info     = lfsck_env_info(env);
833         struct lu_attr             *la       = &info->lti_la;
834         struct lfsck_instance      *lfsck    = com->lc_lfsck;
835         struct lfsck_bookmark      *bk       = &lfsck->li_bookmark_ram;
836         struct lfsck_namespace     *ns       =
837                                 (struct lfsck_namespace *)com->lc_file_ram;
838         struct linkea_data          ldata    = { 0 };
839         const struct lu_fid        *pfid     =
840                                 lu_object_fid(&lfsck->li_obj_dir->do_lu);
841         const struct lu_fid        *cfid     = lfsck_dto2fid(obj);
842         const struct lu_name       *cname;
843         struct thandle             *handle   = NULL;
844         bool                        repaired = false;
845         bool                        locked   = false;
846         bool                        remove;
847         bool                        newdata;
848         int                         count    = 0;
849         int                         rc;
850         ENTRY;
851
852         cname = lfsck_name_get_const(env, ent->lde_name, ent->lde_namelen);
853         down_write(&com->lc_sem);
854         com->lc_new_checked++;
855
856         if (ent->lde_attrs & LUDA_UPGRADE) {
857                 ns->ln_flags |= LF_UPGRADE;
858                 repaired = true;
859         } else if (ent->lde_attrs & LUDA_REPAIR) {
860                 ns->ln_flags |= LF_INCONSISTENT;
861                 repaired = true;
862         }
863
864         if (ent->lde_name[0] == '.' &&
865             (ent->lde_namelen == 1 ||
866              (ent->lde_namelen == 2 && ent->lde_name[1] == '.') ||
867              fid_is_dot_lustre(&ent->lde_fid)))
868                 GOTO(out, rc = 0);
869
870         if (!(bk->lb_param & LPF_DRYRUN) &&
871             (com->lc_journal || repaired)) {
872
873 again:
874                 LASSERT(!locked);
875
876                 com->lc_journal = 1;
877                 handle = dt_trans_create(env, lfsck->li_next);
878                 if (IS_ERR(handle))
879                         GOTO(out, rc = PTR_ERR(handle));
880
881                 rc = lfsck_declare_namespace_exec_dir(env, obj, handle);
882                 if (rc != 0)
883                         GOTO(stop, rc);
884
885                 rc = dt_trans_start(env, lfsck->li_next, handle);
886                 if (rc != 0)
887                         GOTO(stop, rc);
888
889                 dt_write_lock(env, obj, MOR_TGT_CHILD);
890                 locked = true;
891         }
892
893         rc = lfsck_namespace_check_exist(env, lfsck, obj, ent->lde_name);
894         if (rc != 0)
895                 GOTO(stop, rc);
896
897         rc = lfsck_links_read(env, obj, &ldata);
898         if (rc == 0) {
899                 count = ldata.ld_leh->leh_reccount;
900                 rc = linkea_links_find(&ldata, cname, pfid);
901                 if ((rc == 0) &&
902                     (count == 1 || !S_ISDIR(lfsck_object_type(obj))))
903                         goto record;
904
905                 ns->ln_flags |= LF_INCONSISTENT;
906                 /* For dir, if there are more than one linkea entries, or the
907                  * linkea entry does not match the name entry, then remove all
908                  * and add the correct one. */
909                 if (S_ISDIR(lfsck_object_type(obj))) {
910                         remove = true;
911                         newdata = true;
912                 } else {
913                         remove = false;
914                         newdata = false;
915                 }
916                 goto nodata;
917         } else if (unlikely(rc == -EINVAL)) {
918                 count = 1;
919                 ns->ln_flags |= LF_INCONSISTENT;
920                 /* The magic crashed, we are not sure whether there are more
921                  * corrupt data in the linkea, so remove all linkea entries. */
922                 remove = true;
923                 newdata = true;
924                 goto nodata;
925         } else if (rc == -ENODATA) {
926                 count = 1;
927                 ns->ln_flags |= LF_UPGRADE;
928                 remove = false;
929                 newdata = true;
930
931 nodata:
932                 if (bk->lb_param & LPF_DRYRUN) {
933                         repaired = true;
934                         goto record;
935                 }
936
937                 if (!com->lc_journal)
938                         goto again;
939
940                 if (remove) {
941                         LASSERT(newdata);
942
943                         rc = dt_xattr_del(env, obj, XATTR_NAME_LINK, handle,
944                                           BYPASS_CAPA);
945                         if (rc != 0)
946                                 GOTO(stop, rc);
947                 }
948
949                 if (newdata) {
950                         rc = linkea_data_new(&ldata,
951                                         &lfsck_env_info(env)->lti_linkea_buf);
952                         if (rc != 0)
953                                 GOTO(stop, rc);
954                 }
955
956                 rc = linkea_add_buf(&ldata, cname, pfid);
957                 if (rc != 0)
958                         GOTO(stop, rc);
959
960                 rc = lfsck_links_write(env, obj, &ldata, handle);
961                 if (rc != 0)
962                         GOTO(stop, rc);
963
964                 count = ldata.ld_leh->leh_reccount;
965                 repaired = true;
966         } else {
967                 GOTO(stop, rc);
968         }
969
970 record:
971         LASSERT(count > 0);
972
973         rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
974         if (rc != 0)
975                 GOTO(stop, rc);
976
977         if ((count == 1) &&
978             (la->la_nlink == 1 || S_ISDIR(lfsck_object_type(obj))))
979                 /* Usually, it is for single linked object or dir, do nothing.*/
980                 GOTO(stop, rc);
981
982         /* Following modification will be in another transaction.  */
983         if (handle != NULL) {
984                 LASSERT(dt_write_locked(env, obj));
985
986                 dt_write_unlock(env, obj);
987                 locked = false;
988
989                 dt_trans_stop(env, lfsck->li_next, handle);
990                 handle = NULL;
991         }
992
993         ns->ln_mlinked_checked++;
994         rc = lfsck_namespace_update(env, com, cfid,
995                         count != la->la_nlink ? LLF_UNMATCH_NLINKS : 0, false);
996
997         GOTO(out, rc);
998
999 stop:
1000         if (locked)
1001                 dt_write_unlock(env, obj);
1002
1003         if (handle != NULL)
1004                 dt_trans_stop(env, lfsck->li_next, handle);
1005
1006 out:
1007         if (rc < 0) {
1008                 ns->ln_items_failed++;
1009                 if (lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent))
1010                         lfsck_pos_fill(env, lfsck,
1011                                        &ns->ln_pos_first_inconsistent, false);
1012                 if (!(bk->lb_param & LPF_FAILOUT))
1013                         rc = 0;
1014         } else {
1015                 if (repaired)
1016                         ns->ln_items_repaired++;
1017                 else
1018                         com->lc_journal = 0;
1019                 rc = 0;
1020         }
1021         up_write(&com->lc_sem);
1022         return rc;
1023 }
1024
1025 static int lfsck_namespace_post(const struct lu_env *env,
1026                                 struct lfsck_component *com,
1027                                 int result, bool init)
1028 {
1029         struct lfsck_instance   *lfsck = com->lc_lfsck;
1030         struct lfsck_namespace  *ns    =
1031                                 (struct lfsck_namespace *)com->lc_file_ram;
1032         int                      rc;
1033
1034         down_write(&com->lc_sem);
1035
1036         spin_lock(&lfsck->li_lock);
1037         if (!init)
1038                 ns->ln_pos_last_checkpoint = lfsck->li_pos_current;
1039         if (result > 0) {
1040                 ns->ln_status = LS_SCANNING_PHASE2;
1041                 ns->ln_flags |= LF_SCANNED_ONCE;
1042                 ns->ln_flags &= ~LF_UPGRADE;
1043                 cfs_list_del_init(&com->lc_link);
1044                 cfs_list_del_init(&com->lc_link_dir);
1045                 cfs_list_add_tail(&com->lc_link, &lfsck->li_list_double_scan);
1046         } else if (result == 0) {
1047                 if (lfsck->li_paused) {
1048                         ns->ln_status = LS_PAUSED;
1049                 } else {
1050                         ns->ln_status = LS_STOPPED;
1051                         cfs_list_del_init(&com->lc_link);
1052                         cfs_list_del_init(&com->lc_link_dir);
1053                         cfs_list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1054                 }
1055         } else {
1056                 ns->ln_status = LS_FAILED;
1057                 cfs_list_del_init(&com->lc_link);
1058                 cfs_list_del_init(&com->lc_link_dir);
1059                 cfs_list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1060         }
1061         spin_unlock(&lfsck->li_lock);
1062
1063         if (!init) {
1064                 ns->ln_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1065                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1066                 ns->ln_time_last_checkpoint = cfs_time_current_sec();
1067                 ns->ln_items_checked += com->lc_new_checked;
1068                 com->lc_new_checked = 0;
1069         }
1070
1071         rc = lfsck_namespace_store(env, com, false);
1072
1073         up_write(&com->lc_sem);
1074         return rc;
1075 }
1076
1077 static int
1078 lfsck_namespace_dump(const struct lu_env *env, struct lfsck_component *com,
1079                      char *buf, int len)
1080 {
1081         struct lfsck_instance   *lfsck = com->lc_lfsck;
1082         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
1083         struct lfsck_namespace  *ns    =
1084                                 (struct lfsck_namespace *)com->lc_file_ram;
1085         int                      save  = len;
1086         int                      ret   = -ENOSPC;
1087         int                      rc;
1088
1089         down_read(&com->lc_sem);
1090         rc = snprintf(buf, len,
1091                       "name: lfsck_namespace\n"
1092                       "magic: 0x%x\n"
1093                       "version: %d\n"
1094                       "status: %s\n",
1095                       ns->ln_magic,
1096                       bk->lb_version,
1097                       lfsck_status_names[ns->ln_status]);
1098         if (rc <= 0)
1099                 goto out;
1100
1101         buf += rc;
1102         len -= rc;
1103         rc = lfsck_bits_dump(&buf, &len, ns->ln_flags, lfsck_flags_names,
1104                              "flags");
1105         if (rc < 0)
1106                 goto out;
1107
1108         rc = lfsck_bits_dump(&buf, &len, bk->lb_param, lfsck_param_names,
1109                              "param");
1110         if (rc < 0)
1111                 goto out;
1112
1113         rc = lfsck_time_dump(&buf, &len, ns->ln_time_last_complete,
1114                              "time_since_last_completed");
1115         if (rc < 0)
1116                 goto out;
1117
1118         rc = lfsck_time_dump(&buf, &len, ns->ln_time_latest_start,
1119                              "time_since_latest_start");
1120         if (rc < 0)
1121                 goto out;
1122
1123         rc = lfsck_time_dump(&buf, &len, ns->ln_time_last_checkpoint,
1124                              "time_since_last_checkpoint");
1125         if (rc < 0)
1126                 goto out;
1127
1128         rc = lfsck_pos_dump(&buf, &len, &ns->ln_pos_latest_start,
1129                             "latest_start_position");
1130         if (rc < 0)
1131                 goto out;
1132
1133         rc = lfsck_pos_dump(&buf, &len, &ns->ln_pos_last_checkpoint,
1134                             "last_checkpoint_position");
1135         if (rc < 0)
1136                 goto out;
1137
1138         rc = lfsck_pos_dump(&buf, &len, &ns->ln_pos_first_inconsistent,
1139                             "first_failure_position");
1140         if (rc < 0)
1141                 goto out;
1142
1143         if (ns->ln_status == LS_SCANNING_PHASE1) {
1144                 struct lfsck_position pos;
1145                 const struct dt_it_ops *iops;
1146                 cfs_duration_t duration = cfs_time_current() -
1147                                           lfsck->li_time_last_checkpoint;
1148                 __u64 checked = ns->ln_items_checked + com->lc_new_checked;
1149                 __u64 speed = checked;
1150                 __u64 new_checked = com->lc_new_checked * HZ;
1151                 __u32 rtime = ns->ln_run_time_phase1 +
1152                               cfs_duration_sec(duration + HALF_SEC);
1153
1154                 if (duration != 0)
1155                         do_div(new_checked, duration);
1156                 if (rtime != 0)
1157                         do_div(speed, rtime);
1158                 rc = snprintf(buf, len,
1159                               "checked_phase1: "LPU64"\n"
1160                               "checked_phase2: "LPU64"\n"
1161                               "updated_phase1: "LPU64"\n"
1162                               "updated_phase2: "LPU64"\n"
1163                               "failed_phase1: "LPU64"\n"
1164                               "failed_phase2: "LPU64"\n"
1165                               "dirs: "LPU64"\n"
1166                               "M-linked: "LPU64"\n"
1167                               "nlinks_repaired: "LPU64"\n"
1168                               "lost_found: "LPU64"\n"
1169                               "success_count: %u\n"
1170                               "run_time_phase1: %u seconds\n"
1171                               "run_time_phase2: %u seconds\n"
1172                               "average_speed_phase1: "LPU64" items/sec\n"
1173                               "average_speed_phase2: N/A\n"
1174                               "real-time_speed_phase1: "LPU64" items/sec\n"
1175                               "real-time_speed_phase2: N/A\n",
1176                               checked,
1177                               ns->ln_objs_checked_phase2,
1178                               ns->ln_items_repaired,
1179                               ns->ln_objs_repaired_phase2,
1180                               ns->ln_items_failed,
1181                               ns->ln_objs_failed_phase2,
1182                               ns->ln_dirs_checked,
1183                               ns->ln_mlinked_checked,
1184                               ns->ln_objs_nlink_repaired,
1185                               ns->ln_objs_lost_found,
1186                               ns->ln_success_count,
1187                               rtime,
1188                               ns->ln_run_time_phase2,
1189                               speed,
1190                               new_checked);
1191                 if (rc <= 0)
1192                         goto out;
1193
1194                 buf += rc;
1195                 len -= rc;
1196
1197                 LASSERT(lfsck->li_di_oit != NULL);
1198
1199                 iops = &lfsck->li_obj_oit->do_index_ops->dio_it;
1200
1201                 /* The low layer otable-based iteration position may NOT
1202                  * exactly match the namespace-based directory traversal
1203                  * cookie. Generally, it is not a serious issue. But the
1204                  * caller should NOT make assumption on that. */
1205                 pos.lp_oit_cookie = iops->store(env, lfsck->li_di_oit);
1206                 if (!lfsck->li_current_oit_processed)
1207                         pos.lp_oit_cookie--;
1208
1209                 spin_lock(&lfsck->li_lock);
1210                 if (lfsck->li_di_dir != NULL) {
1211                         pos.lp_dir_cookie = lfsck->li_cookie_dir;
1212                         if (pos.lp_dir_cookie >= MDS_DIR_END_OFF) {
1213                                 fid_zero(&pos.lp_dir_parent);
1214                                 pos.lp_dir_cookie = 0;
1215                         } else {
1216                                 pos.lp_dir_parent =
1217                                 *lu_object_fid(&lfsck->li_obj_dir->do_lu);
1218                         }
1219                 } else {
1220                         fid_zero(&pos.lp_dir_parent);
1221                         pos.lp_dir_cookie = 0;
1222                 }
1223                 spin_unlock(&lfsck->li_lock);
1224                 rc = lfsck_pos_dump(&buf, &len, &pos, "current_position");
1225                 if (rc <= 0)
1226                         goto out;
1227         } else if (ns->ln_status == LS_SCANNING_PHASE2) {
1228                 cfs_duration_t duration = cfs_time_current() -
1229                                           lfsck->li_time_last_checkpoint;
1230                 __u64 checked = ns->ln_objs_checked_phase2 +
1231                                 com->lc_new_checked;
1232                 __u64 speed1 = ns->ln_items_checked;
1233                 __u64 speed2 = checked;
1234                 __u64 new_checked = com->lc_new_checked * HZ;
1235                 __u32 rtime = ns->ln_run_time_phase2 +
1236                               cfs_duration_sec(duration + HALF_SEC);
1237
1238                 if (duration != 0)
1239                         do_div(new_checked, duration);
1240                 if (ns->ln_run_time_phase1 != 0)
1241                         do_div(speed1, ns->ln_run_time_phase1);
1242                 if (rtime != 0)
1243                         do_div(speed2, rtime);
1244                 rc = snprintf(buf, len,
1245                               "checked_phase1: "LPU64"\n"
1246                               "checked_phase2: "LPU64"\n"
1247                               "updated_phase1: "LPU64"\n"
1248                               "updated_phase2: "LPU64"\n"
1249                               "failed_phase1: "LPU64"\n"
1250                               "failed_phase2: "LPU64"\n"
1251                               "dirs: "LPU64"\n"
1252                               "M-linked: "LPU64"\n"
1253                               "nlinks_repaired: "LPU64"\n"
1254                               "lost_found: "LPU64"\n"
1255                               "success_count: %u\n"
1256                               "run_time_phase1: %u seconds\n"
1257                               "run_time_phase2: %u seconds\n"
1258                               "average_speed_phase1: "LPU64" items/sec\n"
1259                               "average_speed_phase2: "LPU64" objs/sec\n"
1260                               "real-time_speed_phase1: N/A\n"
1261                               "real-time_speed_phase2: "LPU64" objs/sec\n"
1262                               "current_position: "DFID"\n",
1263                               ns->ln_items_checked,
1264                               checked,
1265                               ns->ln_items_repaired,
1266                               ns->ln_objs_repaired_phase2,
1267                               ns->ln_items_failed,
1268                               ns->ln_objs_failed_phase2,
1269                               ns->ln_dirs_checked,
1270                               ns->ln_mlinked_checked,
1271                               ns->ln_objs_nlink_repaired,
1272                               ns->ln_objs_lost_found,
1273                               ns->ln_success_count,
1274                               ns->ln_run_time_phase1,
1275                               rtime,
1276                               speed1,
1277                               speed2,
1278                               new_checked,
1279                               PFID(&ns->ln_fid_latest_scanned_phase2));
1280                 if (rc <= 0)
1281                         goto out;
1282
1283                 buf += rc;
1284                 len -= rc;
1285         } else {
1286                 __u64 speed1 = ns->ln_items_checked;
1287                 __u64 speed2 = ns->ln_objs_checked_phase2;
1288
1289                 if (ns->ln_run_time_phase1 != 0)
1290                         do_div(speed1, ns->ln_run_time_phase1);
1291                 if (ns->ln_run_time_phase2 != 0)
1292                         do_div(speed2, ns->ln_run_time_phase2);
1293                 rc = snprintf(buf, len,
1294                               "checked_phase1: "LPU64"\n"
1295                               "checked_phase2: "LPU64"\n"
1296                               "updated_phase1: "LPU64"\n"
1297                               "updated_phase2: "LPU64"\n"
1298                               "failed_phase1: "LPU64"\n"
1299                               "failed_phase2: "LPU64"\n"
1300                               "dirs: "LPU64"\n"
1301                               "M-linked: "LPU64"\n"
1302                               "nlinks_repaired: "LPU64"\n"
1303                               "lost_found: "LPU64"\n"
1304                               "success_count: %u\n"
1305                               "run_time_phase1: %u seconds\n"
1306                               "run_time_phase2: %u seconds\n"
1307                               "average_speed_phase1: "LPU64" items/sec\n"
1308                               "average_speed_phase2: "LPU64" objs/sec\n"
1309                               "real-time_speed_phase1: N/A\n"
1310                               "real-time_speed_phase2: N/A\n"
1311                               "current_position: N/A\n",
1312                               ns->ln_items_checked,
1313                               ns->ln_objs_checked_phase2,
1314                               ns->ln_items_repaired,
1315                               ns->ln_objs_repaired_phase2,
1316                               ns->ln_items_failed,
1317                               ns->ln_objs_failed_phase2,
1318                               ns->ln_dirs_checked,
1319                               ns->ln_mlinked_checked,
1320                               ns->ln_objs_nlink_repaired,
1321                               ns->ln_objs_lost_found,
1322                               ns->ln_success_count,
1323                               ns->ln_run_time_phase1,
1324                               ns->ln_run_time_phase2,
1325                               speed1,
1326                               speed2);
1327                 if (rc <= 0)
1328                         goto out;
1329
1330                 buf += rc;
1331                 len -= rc;
1332         }
1333         ret = save - len;
1334
1335 out:
1336         up_read(&com->lc_sem);
1337         return ret;
1338 }
1339
1340 static int lfsck_namespace_double_scan(const struct lu_env *env,
1341                                        struct lfsck_component *com)
1342 {
1343         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1344         struct ptlrpc_thread    *thread = &lfsck->li_thread;
1345         struct lfsck_bookmark   *bk     = &lfsck->li_bookmark_ram;
1346         struct lfsck_namespace  *ns     =
1347                                 (struct lfsck_namespace *)com->lc_file_ram;
1348         struct dt_object        *obj    = com->lc_obj;
1349         const struct dt_it_ops  *iops   = &obj->do_index_ops->dio_it;
1350         struct dt_object        *target;
1351         struct dt_it            *di;
1352         struct dt_key           *key;
1353         struct lu_fid            fid;
1354         int                      rc;
1355         __u8                     flags;
1356         ENTRY;
1357
1358         lfsck->li_new_scanned = 0;
1359         lfsck->li_time_last_checkpoint = cfs_time_current();
1360         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
1361                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1362
1363         di = iops->init(env, obj, 0, BYPASS_CAPA);
1364         if (IS_ERR(di))
1365                 RETURN(PTR_ERR(di));
1366
1367         fid_cpu_to_be(&fid, &ns->ln_fid_latest_scanned_phase2);
1368         rc = iops->get(env, di, (const struct dt_key *)&fid);
1369         if (rc < 0)
1370                 GOTO(fini, rc);
1371
1372         /* Skip the start one, which either has been processed or non-exist. */
1373         rc = iops->next(env, di);
1374         if (rc != 0)
1375                 GOTO(put, rc);
1376
1377         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_DOUBLESCAN))
1378                 GOTO(put, rc = 0);
1379
1380         do {
1381                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY3) &&
1382                     cfs_fail_val > 0) {
1383                         struct l_wait_info lwi;
1384
1385                         lwi = LWI_TIMEOUT(cfs_time_seconds(cfs_fail_val),
1386                                           NULL, NULL);
1387                         l_wait_event(thread->t_ctl_waitq,
1388                                      !thread_is_running(thread),
1389                                      &lwi);
1390                 }
1391
1392                 key = iops->key(env, di);
1393                 fid_be_to_cpu(&fid, (const struct lu_fid *)key);
1394                 target = lfsck_object_find(env, lfsck, &fid);
1395                 down_write(&com->lc_sem);
1396                 if (target == NULL) {
1397                         rc = 0;
1398                         goto checkpoint;
1399                 } else if (IS_ERR(target)) {
1400                         rc = PTR_ERR(target);
1401                         goto checkpoint;
1402                 }
1403
1404                 /* XXX: Currently, skip remote object, the consistency for
1405                  *      remote object will be processed in LFSCK phase III. */
1406                 if (dt_object_exists(target) && !dt_object_remote(target)) {
1407                         rc = iops->rec(env, di, (struct dt_rec *)&flags, 0);
1408                         if (rc == 0)
1409                                 rc = lfsck_namespace_double_scan_one(env, com,
1410                                                                 target, flags);
1411                 }
1412
1413                 lfsck_object_put(env, target);
1414
1415 checkpoint:
1416                 lfsck->li_new_scanned++;
1417                 com->lc_new_checked++;
1418                 ns->ln_fid_latest_scanned_phase2 = fid;
1419                 if (rc > 0)
1420                         ns->ln_objs_repaired_phase2++;
1421                 else if (rc < 0)
1422                         ns->ln_objs_failed_phase2++;
1423                 up_write(&com->lc_sem);
1424
1425                 if ((rc == 0) || ((rc > 0) && !(bk->lb_param & LPF_DRYRUN))) {
1426                         lfsck_namespace_delete(env, com, &fid);
1427                 } else if (rc < 0) {
1428                         flags |= LLF_REPAIR_FAILED;
1429                         lfsck_namespace_update(env, com, &fid, flags, true);
1430                 }
1431
1432                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
1433                         GOTO(put, rc);
1434
1435                 if (unlikely(cfs_time_beforeq(lfsck->li_time_next_checkpoint,
1436                                               cfs_time_current())) &&
1437                     com->lc_new_checked != 0) {
1438                         down_write(&com->lc_sem);
1439                         ns->ln_run_time_phase2 +=
1440                                 cfs_duration_sec(cfs_time_current() +
1441                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1442                         ns->ln_time_last_checkpoint = cfs_time_current_sec();
1443                         ns->ln_objs_checked_phase2 += com->lc_new_checked;
1444                         com->lc_new_checked = 0;
1445                         rc = lfsck_namespace_store(env, com, false);
1446                         up_write(&com->lc_sem);
1447                         if (rc != 0)
1448                                 GOTO(put, rc);
1449
1450                         lfsck->li_time_last_checkpoint = cfs_time_current();
1451                         lfsck->li_time_next_checkpoint =
1452                                 lfsck->li_time_last_checkpoint +
1453                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1454                 }
1455
1456                 lfsck_control_speed(lfsck);
1457                 if (unlikely(!thread_is_running(thread)))
1458                         GOTO(put, rc = 0);
1459
1460                 rc = iops->next(env, di);
1461         } while (rc == 0);
1462
1463         GOTO(put, rc);
1464
1465 put:
1466         iops->put(env, di);
1467
1468 fini:
1469         iops->fini(env, di);
1470         down_write(&com->lc_sem);
1471
1472         ns->ln_run_time_phase2 += cfs_duration_sec(cfs_time_current() +
1473                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1474         ns->ln_time_last_checkpoint = cfs_time_current_sec();
1475         ns->ln_objs_checked_phase2 += com->lc_new_checked;
1476         com->lc_new_checked = 0;
1477
1478         if (rc > 0) {
1479                 com->lc_journal = 0;
1480                 ns->ln_status = LS_COMPLETED;
1481                 if (!(bk->lb_param & LPF_DRYRUN))
1482                         ns->ln_flags &=
1483                         ~(LF_SCANNED_ONCE | LF_INCONSISTENT | LF_UPGRADE);
1484                 ns->ln_time_last_complete = ns->ln_time_last_checkpoint;
1485                 ns->ln_success_count++;
1486         } else if (rc == 0) {
1487                 if (lfsck->li_paused)
1488                         ns->ln_status = LS_PAUSED;
1489                 else
1490                         ns->ln_status = LS_STOPPED;
1491         } else {
1492                 ns->ln_status = LS_FAILED;
1493         }
1494
1495         if (ns->ln_status != LS_PAUSED) {
1496                 spin_lock(&lfsck->li_lock);
1497                 cfs_list_del_init(&com->lc_link);
1498                 cfs_list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1499                 spin_unlock(&lfsck->li_lock);
1500         }
1501
1502         rc = lfsck_namespace_store(env, com, false);
1503
1504         up_write(&com->lc_sem);
1505         return rc;
1506 }
1507
1508 static struct lfsck_operations lfsck_namespace_ops = {
1509         .lfsck_reset            = lfsck_namespace_reset,
1510         .lfsck_fail             = lfsck_namespace_fail,
1511         .lfsck_checkpoint       = lfsck_namespace_checkpoint,
1512         .lfsck_prep             = lfsck_namespace_prep,
1513         .lfsck_exec_oit         = lfsck_namespace_exec_oit,
1514         .lfsck_exec_dir         = lfsck_namespace_exec_dir,
1515         .lfsck_post             = lfsck_namespace_post,
1516         .lfsck_dump             = lfsck_namespace_dump,
1517         .lfsck_double_scan      = lfsck_namespace_double_scan,
1518 };
1519
1520 int lfsck_namespace_setup(const struct lu_env *env,
1521                           struct lfsck_instance *lfsck)
1522 {
1523         struct lfsck_component  *com;
1524         struct lfsck_namespace  *ns;
1525         struct dt_object        *obj;
1526         int                      rc;
1527         ENTRY;
1528
1529         LASSERT(lfsck->li_master);
1530
1531         OBD_ALLOC_PTR(com);
1532         if (com == NULL)
1533                 RETURN(-ENOMEM);
1534
1535         CFS_INIT_LIST_HEAD(&com->lc_link);
1536         CFS_INIT_LIST_HEAD(&com->lc_link_dir);
1537         init_rwsem(&com->lc_sem);
1538         atomic_set(&com->lc_ref, 1);
1539         com->lc_lfsck = lfsck;
1540         com->lc_type = LT_NAMESPACE;
1541         com->lc_ops = &lfsck_namespace_ops;
1542         com->lc_file_size = sizeof(struct lfsck_namespace);
1543         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
1544         if (com->lc_file_ram == NULL)
1545                 GOTO(out, rc = -ENOMEM);
1546
1547         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
1548         if (com->lc_file_disk == NULL)
1549                 GOTO(out, rc = -ENOMEM);
1550
1551         obj = local_index_find_or_create(env, lfsck->li_los,
1552                                          lfsck->li_local_root,
1553                                          lfsck_namespace_name,
1554                                          S_IFREG | S_IRUGO | S_IWUSR,
1555                                          &dt_lfsck_features);
1556         if (IS_ERR(obj))
1557                 GOTO(out, rc = PTR_ERR(obj));
1558
1559         com->lc_obj = obj;
1560         rc = obj->do_ops->do_index_try(env, obj, &dt_lfsck_features);
1561         if (rc != 0)
1562                 GOTO(out, rc);
1563
1564         rc = lfsck_namespace_load(env, com);
1565         if (rc > 0)
1566                 rc = lfsck_namespace_reset(env, com, true);
1567         else if (rc == -ENODATA)
1568                 rc = lfsck_namespace_init(env, com);
1569         if (rc != 0)
1570                 GOTO(out, rc);
1571
1572         ns = (struct lfsck_namespace *)com->lc_file_ram;
1573         switch (ns->ln_status) {
1574         case LS_INIT:
1575         case LS_COMPLETED:
1576         case LS_FAILED:
1577         case LS_STOPPED:
1578                 cfs_list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1579                 break;
1580         default:
1581                 CERROR("%s: unknown lfsck_namespace status: %u\n",
1582                        lfsck_lfsck2name(lfsck), ns->ln_status);
1583                 /* fall through */
1584         case LS_SCANNING_PHASE1:
1585         case LS_SCANNING_PHASE2:
1586                 /* No need to store the status to disk right now.
1587                  * If the system crashed before the status stored,
1588                  * it will be loaded back when next time. */
1589                 ns->ln_status = LS_CRASHED;
1590                 /* fall through */
1591         case LS_PAUSED:
1592         case LS_CRASHED:
1593                 cfs_list_add_tail(&com->lc_link, &lfsck->li_list_scan);
1594                 cfs_list_add_tail(&com->lc_link_dir, &lfsck->li_list_dir);
1595                 break;
1596         }
1597
1598         GOTO(out, rc = 0);
1599
1600 out:
1601         if (rc != 0)
1602                 lfsck_component_cleanup(env, com);
1603         return rc;
1604 }