Whamcloud - gitweb
LU-7340 mdd: changelogs garbage collection
[fs/lustre-release.git] / lustre / obdclass / llog_cat.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2016, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/obdclass/llog_cat.c
33  *
34  * OST<->MDS recovery logging infrastructure.
35  *
36  * Invariants in implementation:
37  * - we do not share logs among different OST<->MDS connections, so that
38  *   if an OST or MDS fails it need only look at log(s) relevant to itself
39  *
40  * Author: Andreas Dilger <adilger@clusterfs.com>
41  * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
42  * Author: Mikhail Pershin <mike.pershin@intel.com>
43  */
44
45 #define DEBUG_SUBSYSTEM S_LOG
46
47
48 #include <obd_class.h>
49
50 #include "llog_internal.h"
51
52 /* Create a new log handle and add it to the open list.
53  * This log handle will be closed when all of the records in it are removed.
54  *
55  * Assumes caller has already pushed us into the kernel context and is locking.
56  */
57 static int llog_cat_new_log(const struct lu_env *env,
58                             struct llog_handle *cathandle,
59                             struct llog_handle *loghandle,
60                             struct thandle *th)
61 {
62         struct llog_thread_info *lgi = llog_info(env);
63         struct llog_logid_rec   *rec = &lgi->lgi_logid;
64         struct thandle *handle = NULL;
65         struct dt_device *dt = NULL;
66         struct llog_log_hdr     *llh = cathandle->lgh_hdr;
67         int                      rc, index;
68
69         ENTRY;
70
71         index = (cathandle->lgh_last_idx + 1) %
72                 (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ? (cfs_fail_val + 1) :
73                                                 LLOG_HDR_BITMAP_SIZE(llh));
74
75         /* check that new llog index will not overlap with the first one.
76          * - llh_cat_idx is the index just before the first/oldest still in-use
77          *      index in catalog
78          * - lgh_last_idx is the last/newest used index in catalog
79          *
80          * When catalog is not wrapped yet then lgh_last_idx is always larger
81          * than llh_cat_idx. After the wrap around lgh_last_idx re-starts
82          * from 0 and llh_cat_idx becomes the upper limit for it
83          *
84          * Check if catalog has already wrapped around or not by comparing
85          * last_idx and cat_idx */
86         if ((index == llh->llh_cat_idx + 1 && llh->llh_count > 1) ||
87             (index == 0 && llh->llh_cat_idx == 0)) {
88                 if (cathandle->lgh_name == NULL) {
89                         CWARN("%s: there are no more free slots in catalog "
90                               DFID":%x\n",
91                               loghandle->lgh_ctxt->loc_obd->obd_name,
92                               PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
93                               cathandle->lgh_id.lgl_ogen);
94                 } else {
95                         CWARN("%s: there are no more free slots in "
96                               "catalog %s\n",
97                               loghandle->lgh_ctxt->loc_obd->obd_name,
98                               cathandle->lgh_name);
99                 }
100                 RETURN(-ENOSPC);
101         }
102
103         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
104                 RETURN(-ENOSPC);
105
106         if (loghandle->lgh_hdr != NULL) {
107                 /* If llog object is remote and creation is failed, lgh_hdr
108                  * might be left over here, free it first */
109                 LASSERT(!llog_exist(loghandle));
110                 OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
111                 loghandle->lgh_hdr = NULL;
112         }
113
114         if (th == NULL) {
115                 dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
116
117                 handle = dt_trans_create(env, dt);
118                 if (IS_ERR(handle))
119                         RETURN(PTR_ERR(handle));
120
121                 /* Create update llog object synchronously, which
122                  * happens during inialization process see
123                  * lod_sub_prep_llog(), to make sure the update
124                  * llog object is created before corss-MDT writing
125                  * updates into the llog object */
126                 if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
127                         handle->th_sync = 1;
128
129                 handle->th_wait_submit = 1;
130
131                 rc = llog_declare_create(env, loghandle, handle);
132                 if (rc != 0)
133                         GOTO(out, rc);
134
135                 rec->lid_hdr.lrh_len = sizeof(*rec);
136                 rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
137                 rec->lid_id = loghandle->lgh_id;
138                 rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
139                                             handle);
140                 if (rc != 0)
141                         GOTO(out, rc);
142
143                 rc = dt_trans_start_local(env, dt, handle);
144                 if (rc != 0)
145                         GOTO(out, rc);
146
147                 th = handle;
148         }
149
150         rc = llog_create(env, loghandle, th);
151         /* if llog is already created, no need to initialize it */
152         if (rc == -EEXIST) {
153                 GOTO(out, rc = 0);
154         } else if (rc != 0) {
155                 CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
156                        loghandle->lgh_ctxt->loc_obd->obd_name, rc);
157                 GOTO(out, rc);
158         }
159
160         rc = llog_init_handle(env, loghandle,
161                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
162                               &cathandle->lgh_hdr->llh_tgtuuid);
163         if (rc < 0)
164                 GOTO(out, rc);
165
166         /* build the record for this log in the catalog */
167         rec->lid_hdr.lrh_len = sizeof(*rec);
168         rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
169         rec->lid_id = loghandle->lgh_id;
170
171         /* append the new record into catalog. The new index will be
172          * assigned to the record and updated in rec header */
173         rc = llog_write_rec(env, cathandle, &rec->lid_hdr,
174                             &loghandle->u.phd.phd_cookie, LLOG_NEXT_IDX, th);
175         if (rc < 0)
176                 GOTO(out_destroy, rc);
177
178         CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
179                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rec->lid_hdr.lrh_index,
180                PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
181
182         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
183
184         /* limit max size of plain llog so that space can be
185          * released sooner, especially on small filesystems */
186         /* 2MB for the cases when free space hasn't been learned yet */
187         loghandle->lgh_max_size = 2 << 20;
188         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
189         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
190         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
191                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
192                                   lgi->lgi_statfs.os_bsize) >> 6;
193                 if (freespace < loghandle->lgh_max_size)
194                         loghandle->lgh_max_size = freespace;
195                 /* shouldn't be > 128MB in any case?
196                  * it's 256K records of 512 bytes each */
197                 if (freespace > (128 << 20))
198                         loghandle->lgh_max_size = 128 << 20;
199         }
200         rc = 0;
201
202 out:
203         if (handle != NULL) {
204                 handle->th_result = rc >= 0 ? 0 : rc;
205                 dt_trans_stop(env, dt, handle);
206         }
207         RETURN(rc);
208
209 out_destroy:
210         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
211          * we want to destroy it in this transaction, otherwise the object
212          * becomes an orphan */
213         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
214         /* this is to mimic full log, so another llog_cat_current_log()
215          * can skip it and ask for another onet */
216         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(llh) + 1;
217         llog_trans_destroy(env, loghandle, th);
218         RETURN(rc);
219 }
220
221 /* Open an existent log handle and add it to the open list.
222  * This log handle will be closed when all of the records in it are removed.
223  *
224  * Assumes caller has already pushed us into the kernel context and is locking.
225  * We return a lock on the handle to ensure nobody yanks it from us.
226  *
227  * This takes extra reference on llog_handle via llog_handle_get() and require
228  * this reference to be put by caller using llog_handle_put()
229  */
230 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
231                        struct llog_handle **res, struct llog_logid *logid)
232 {
233         struct llog_handle      *loghandle;
234         enum llog_flag           fmt;
235         int                      rc = 0;
236
237         ENTRY;
238
239         if (cathandle == NULL)
240                 RETURN(-EBADF);
241
242         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
243         down_write(&cathandle->lgh_lock);
244         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
245                             u.phd.phd_entry) {
246                 struct llog_logid *cgl = &loghandle->lgh_id;
247
248                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
249                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
250                         if (cgl->lgl_ogen != logid->lgl_ogen) {
251                                 CWARN("%s: log "DFID" generation %x != %x\n",
252                                       loghandle->lgh_ctxt->loc_obd->obd_name,
253                                       PFID(&logid->lgl_oi.oi_fid),
254                                       cgl->lgl_ogen, logid->lgl_ogen);
255                                 continue;
256                         }
257                         loghandle->u.phd.phd_cat_handle = cathandle;
258                         up_write(&cathandle->lgh_lock);
259                         GOTO(out, rc = 0);
260                 }
261         }
262         up_write(&cathandle->lgh_lock);
263
264         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
265                        LLOG_OPEN_EXISTS);
266         if (rc < 0) {
267                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
268                        cathandle->lgh_ctxt->loc_obd->obd_name,
269                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
270                 RETURN(rc);
271         }
272
273         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
274         if (rc < 0) {
275                 llog_close(env, loghandle);
276                 loghandle = NULL;
277                 RETURN(rc);
278         }
279
280         down_write(&cathandle->lgh_lock);
281         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
282         up_write(&cathandle->lgh_lock);
283
284         loghandle->u.phd.phd_cat_handle = cathandle;
285         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
286         loghandle->u.phd.phd_cookie.lgc_index =
287                                 loghandle->lgh_hdr->llh_cat_idx;
288         EXIT;
289 out:
290         llog_handle_get(loghandle);
291         *res = loghandle;
292         return 0;
293 }
294
295 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
296 {
297         struct llog_handle      *loghandle, *n;
298         int                      rc;
299
300         ENTRY;
301
302         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
303                                  u.phd.phd_entry) {
304                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
305                 int                      index;
306
307                 /* unlink open-not-created llogs */
308                 list_del_init(&loghandle->u.phd.phd_entry);
309                 llh = loghandle->lgh_hdr;
310                 if (loghandle->lgh_obj != NULL && llh != NULL &&
311                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
312                     (llh->llh_count == 1)) {
313                         rc = llog_destroy(env, loghandle);
314                         if (rc)
315                                 CERROR("%s: failure destroying log during "
316                                        "cleanup: rc = %d\n",
317                                        loghandle->lgh_ctxt->loc_obd->obd_name,
318                                        rc);
319
320                         index = loghandle->u.phd.phd_cookie.lgc_index;
321                         llog_cat_cleanup(env, cathandle, NULL, index);
322                 }
323                 llog_close(env, loghandle);
324         }
325         /* if handle was stored in ctxt, remove it too */
326         if (cathandle->lgh_ctxt->loc_handle == cathandle)
327                 cathandle->lgh_ctxt->loc_handle = NULL;
328         rc = llog_close(env, cathandle);
329         RETURN(rc);
330 }
331 EXPORT_SYMBOL(llog_cat_close);
332
333 /**
334  * lockdep markers for nested struct llog_handle::lgh_lock locking.
335  */
336 enum {
337         LLOGH_CAT,
338         LLOGH_LOG
339 };
340
341 /** Return the currently active log handle.  If the current log handle doesn't
342  * have enough space left for the current record, start a new one.
343  *
344  * If reclen is 0, we only want to know what the currently active log is,
345  * otherwise we get a lock on this log so nobody can steal our space.
346  *
347  * Assumes caller has already pushed us into the kernel context and is locking.
348  *
349  * NOTE: loghandle is write-locked upon successful return
350  */
351 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
352                                                 struct thandle *th)
353 {
354         struct llog_handle *loghandle = NULL;
355         ENTRY;
356
357
358         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
359                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
360                 GOTO(next, loghandle);
361         }
362
363         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
364         loghandle = cathandle->u.chd.chd_current_log;
365         if (loghandle) {
366                 struct llog_log_hdr *llh;
367
368                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
369                 llh = loghandle->lgh_hdr;
370                 if (llh == NULL || !llog_is_full(loghandle)) {
371                         up_read(&cathandle->lgh_lock);
372                         RETURN(loghandle);
373                 } else {
374                         up_write(&loghandle->lgh_lock);
375                 }
376         }
377         up_read(&cathandle->lgh_lock);
378
379         /* time to use next log */
380
381         /* first, we have to make sure the state hasn't changed */
382         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
383         loghandle = cathandle->u.chd.chd_current_log;
384         if (loghandle) {
385                 struct llog_log_hdr *llh;
386
387                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
388                 llh = loghandle->lgh_hdr;
389                 LASSERT(llh);
390                 if (!llog_is_full(loghandle))
391                         GOTO(out_unlock, loghandle);
392                 else
393                         up_write(&loghandle->lgh_lock);
394         }
395
396 next:
397         /* Sigh, the chd_next_log and chd_current_log is initialized
398          * in declare phase, and we do not serialize the catlog
399          * accessing, so it might be possible the llog creation
400          * thread (see llog_cat_declare_add_rec()) did not create
401          * llog successfully, then the following thread might
402          * meet this situation. */
403         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
404                 CERROR("%s: next log does not exist!\n",
405                        cathandle->lgh_ctxt->loc_obd->obd_name);
406                 loghandle = ERR_PTR(-EIO);
407                 if (cathandle->u.chd.chd_next_log == NULL) {
408                         /* Store the error in chd_next_log, so
409                          * the following process can get correct
410                          * failure value */
411                         cathandle->u.chd.chd_next_log = loghandle;
412                 }
413                 GOTO(out_unlock, loghandle);
414         }
415
416         CDEBUG(D_INODE, "use next log\n");
417
418         loghandle = cathandle->u.chd.chd_next_log;
419         cathandle->u.chd.chd_current_log = loghandle;
420         cathandle->u.chd.chd_next_log = NULL;
421         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
422
423 out_unlock:
424         up_write(&cathandle->lgh_lock);
425         LASSERT(loghandle);
426         RETURN(loghandle);
427 }
428
429 static int llog_cat_update_header(const struct lu_env *env,
430                            struct llog_handle *cathandle)
431 {
432         struct llog_handle *loghandle;
433         int rc;
434         ENTRY;
435
436         /* refresh llog */
437         down_write(&cathandle->lgh_lock);
438         if (!cathandle->lgh_stale) {
439                 up_write(&cathandle->lgh_lock);
440                 RETURN(0);
441         }
442         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
443                             u.phd.phd_entry) {
444                 if (!llog_exist(loghandle))
445                         continue;
446
447                 rc = llog_read_header(env, loghandle, NULL);
448                 if (rc != 0) {
449                         up_write(&cathandle->lgh_lock);
450                         GOTO(out, rc);
451                 }
452         }
453         rc = llog_read_header(env, cathandle, NULL);
454         if (rc == 0)
455                 cathandle->lgh_stale = 0;
456         up_write(&cathandle->lgh_lock);
457         if (rc != 0)
458                 GOTO(out, rc);
459 out:
460         RETURN(rc);
461 }
462
463 /* Add a single record to the recovery log(s) using a catalog
464  * Returns as llog_write_record
465  *
466  * Assumes caller has already pushed us into the kernel context.
467  */
468 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
469                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
470                      struct thandle *th)
471 {
472         struct llog_handle *loghandle;
473         int rc, retried = 0;
474         ENTRY;
475
476         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
477
478 retry:
479         loghandle = llog_cat_current_log(cathandle, th);
480         if (IS_ERR(loghandle))
481                 RETURN(PTR_ERR(loghandle));
482
483         /* loghandle is already locked by llog_cat_current_log() for us */
484         if (!llog_exist(loghandle)) {
485                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
486                 if (rc < 0) {
487                         up_write(&loghandle->lgh_lock);
488                         /* nobody should be trying to use this llog */
489                         down_write(&cathandle->lgh_lock);
490                         if (cathandle->u.chd.chd_current_log == loghandle)
491                                 cathandle->u.chd.chd_current_log = NULL;
492                         up_write(&cathandle->lgh_lock);
493                         RETURN(rc);
494                 }
495         }
496         /* now let's try to add the record */
497         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
498         if (rc < 0) {
499                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
500                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
501                 /* -ENOSPC is returned if no empty records left
502                  * and when it's lack of space on the stogage.
503                  * there is no point to try again if it's the second
504                  * case. many callers (like llog test) expect ENOSPC,
505                  * so we preserve this error code, but look for the
506                  * actual cause here */
507                 if (rc == -ENOSPC && llog_is_full(loghandle))
508                         rc = -ENOBUFS;
509         }
510         up_write(&loghandle->lgh_lock);
511
512         if (rc == -ENOBUFS) {
513                 if (retried++ == 0)
514                         GOTO(retry, rc);
515                 CERROR("%s: error on 2nd llog: rc = %d\n",
516                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
517         }
518
519         RETURN(rc);
520 }
521 EXPORT_SYMBOL(llog_cat_add_rec);
522
523 int llog_cat_declare_add_rec(const struct lu_env *env,
524                              struct llog_handle *cathandle,
525                              struct llog_rec_hdr *rec, struct thandle *th)
526 {
527         struct llog_thread_info *lgi = llog_info(env);
528         struct llog_logid_rec   *lirec = &lgi->lgi_logid;
529         struct llog_handle      *loghandle, *next;
530         int                      rc = 0;
531
532         ENTRY;
533
534         if (cathandle->u.chd.chd_current_log == NULL) {
535                 /* declare new plain llog */
536                 down_write(&cathandle->lgh_lock);
537                 if (cathandle->u.chd.chd_current_log == NULL) {
538                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
539                                        NULL, NULL, LLOG_OPEN_NEW);
540                         if (rc == 0) {
541                                 cathandle->u.chd.chd_current_log = loghandle;
542                                 list_add_tail(&loghandle->u.phd.phd_entry,
543                                               &cathandle->u.chd.chd_head);
544                         }
545                 }
546                 up_write(&cathandle->lgh_lock);
547         } else if (cathandle->u.chd.chd_next_log == NULL ||
548                    IS_ERR(cathandle->u.chd.chd_next_log)) {
549                 /* declare next plain llog */
550                 down_write(&cathandle->lgh_lock);
551                 if (cathandle->u.chd.chd_next_log == NULL ||
552                     IS_ERR(cathandle->u.chd.chd_next_log)) {
553                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
554                                        NULL, NULL, LLOG_OPEN_NEW);
555                         if (rc == 0) {
556                                 cathandle->u.chd.chd_next_log = loghandle;
557                                 list_add_tail(&loghandle->u.phd.phd_entry,
558                                               &cathandle->u.chd.chd_head);
559                         }
560                 }
561                 up_write(&cathandle->lgh_lock);
562         }
563         if (rc)
564                 GOTO(out, rc);
565
566         lirec->lid_hdr.lrh_len = sizeof(*lirec);
567
568         if (!llog_exist(cathandle->u.chd.chd_current_log)) {
569                 if (dt_object_remote(cathandle->lgh_obj)) {
570                         /* For remote operation, if we put the llog object
571                          * creation in the current transaction, then the
572                          * llog object will not be created on the remote
573                          * target until the transaction stop, if other
574                          * operations start before the transaction stop,
575                          * and use the same llog object, will be dependent
576                          * on the success of this transaction. So let's
577                          * create the llog object synchronously here to
578                          * remove the dependency. */
579 create_again:
580                         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
581                         loghandle = cathandle->u.chd.chd_current_log;
582                         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
583                         if (cathandle->lgh_stale) {
584                                 up_write(&loghandle->lgh_lock);
585                                 up_read(&cathandle->lgh_lock);
586                                 GOTO(out, rc = -EIO);
587                         }
588                         if (!llog_exist(loghandle)) {
589                                 rc = llog_cat_new_log(env, cathandle, loghandle,
590                                                       NULL);
591                                 if (rc == -ESTALE)
592                                         cathandle->lgh_stale = 1;
593                         }
594                         up_write(&loghandle->lgh_lock);
595                         up_read(&cathandle->lgh_lock);
596                         if (rc == -ESTALE) {
597                                 rc = llog_cat_update_header(env, cathandle);
598                                 if (rc != 0)
599                                         GOTO(out, rc);
600                                 goto create_again;
601                         } else if (rc < 0) {
602                                 GOTO(out, rc);
603                         }
604                 } else {
605                         rc = llog_declare_create(env,
606                                         cathandle->u.chd.chd_current_log, th);
607                         if (rc)
608                                 GOTO(out, rc);
609                         llog_declare_write_rec(env, cathandle,
610                                                &lirec->lid_hdr, -1, th);
611                 }
612         }
613
614 write_again:
615         /* declare records in the llogs */
616         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
617                                     rec, -1, th);
618         if (rc == -ESTALE) {
619                 down_write(&cathandle->lgh_lock);
620                 if (cathandle->lgh_stale) {
621                         up_write(&cathandle->lgh_lock);
622                         GOTO(out, rc = -EIO);
623                 }
624
625                 cathandle->lgh_stale = 1;
626                 up_write(&cathandle->lgh_lock);
627                 rc = llog_cat_update_header(env, cathandle);
628                 if (rc != 0)
629                         GOTO(out, rc);
630                 goto write_again;
631         } else if (rc < 0) {
632                 GOTO(out, rc);
633         }
634
635         next = cathandle->u.chd.chd_next_log;
636         if (!IS_ERR_OR_NULL(next)) {
637                 if (!llog_exist(next)) {
638                         if (dt_object_remote(cathandle->lgh_obj)) {
639                                 /* For remote operation, if we put the llog
640                                  * object creation in the current transaction,
641                                  * then the llog object will not be created on
642                                  * the remote target until the transaction stop,
643                                  * if other operations start before the
644                                  * transaction stop, and use the same llog
645                                  * object, will be dependent on the success of
646                                  * this transaction. So let's create the llog
647                                  * object synchronously here to remove the
648                                  * dependency. */
649                                 down_write_nested(&cathandle->lgh_lock,
650                                                  LLOGH_CAT);
651                                 next = cathandle->u.chd.chd_next_log;
652                                 if (IS_ERR_OR_NULL(next)) {
653                                         /* Sigh, another thread just tried,
654                                          * let's fail as well */
655                                         up_write(&cathandle->lgh_lock);
656                                         if (next == NULL)
657                                                 rc = -EIO;
658                                         else
659                                                 rc = PTR_ERR(next);
660                                         GOTO(out, rc);
661                                 }
662
663                                 down_write_nested(&next->lgh_lock, LLOGH_LOG);
664                                 if (!llog_exist(next)) {
665                                         rc = llog_cat_new_log(env, cathandle,
666                                                               next, NULL);
667                                         if (rc < 0)
668                                                 cathandle->u.chd.chd_next_log =
669                                                                 ERR_PTR(rc);
670                                 }
671                                 up_write(&next->lgh_lock);
672                                 up_write(&cathandle->lgh_lock);
673                                 if (rc < 0)
674                                         GOTO(out, rc);
675                         } else {
676                                 rc = llog_declare_create(env, next, th);
677                                 llog_declare_write_rec(env, cathandle,
678                                                 &lirec->lid_hdr, -1, th);
679                         }
680                 }
681                 /* XXX: we hope for declarations made for existing llog
682                  *      this might be not correct with some backends
683                  *      where declarations are expected against specific
684                  *      object like ZFS with full debugging enabled */
685                 /*llog_declare_write_rec(env, next, rec, -1, th);*/
686         }
687 out:
688         RETURN(rc);
689 }
690 EXPORT_SYMBOL(llog_cat_declare_add_rec);
691
692 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
693                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
694 {
695         struct llog_ctxt        *ctxt;
696         struct dt_device        *dt;
697         struct thandle          *th = NULL;
698         int                      rc;
699
700         ctxt = cathandle->lgh_ctxt;
701         LASSERT(ctxt);
702         LASSERT(ctxt->loc_exp);
703
704         LASSERT(cathandle->lgh_obj != NULL);
705         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
706
707         th = dt_trans_create(env, dt);
708         if (IS_ERR(th))
709                 RETURN(PTR_ERR(th));
710
711         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
712         if (rc)
713                 GOTO(out_trans, rc);
714
715         rc = dt_trans_start_local(env, dt, th);
716         if (rc)
717                 GOTO(out_trans, rc);
718         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
719 out_trans:
720         dt_trans_stop(env, dt, th);
721         RETURN(rc);
722 }
723 EXPORT_SYMBOL(llog_cat_add);
724
725 /* For each cookie in the cookie array, we clear the log in-use bit and either:
726  * - the log is empty, so mark it free in the catalog header and delete it
727  * - the log is not empty, just write out the log header
728  *
729  * The cookies may be in different log files, so we need to get new logs
730  * each time.
731  *
732  * Assumes caller has already pushed us into the kernel context.
733  */
734 int llog_cat_cancel_records(const struct lu_env *env,
735                             struct llog_handle *cathandle, int count,
736                             struct llog_cookie *cookies)
737 {
738         int i, index, rc = 0, failed = 0;
739
740         ENTRY;
741
742         for (i = 0; i < count; i++, cookies++) {
743                 struct llog_handle *loghandle;
744                 struct llog_logid *lgl = &cookies->lgc_lgl;
745                 int  lrc;
746
747                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
748                 if (rc) {
749                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
750                                ": rc = %d\n",
751                                cathandle->lgh_ctxt->loc_obd->obd_name,
752                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
753                         failed++;
754                         continue;
755                 }
756
757                 if ((cathandle->lgh_ctxt->loc_flags &
758                      LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
759                         /* For update log, some of loghandles of cathandle
760                          * might not exist because remote llog creation might
761                          * be failed, so let's skip the record cancellation
762                          * for these non-exist llogs.
763                          */
764                         lrc = -ENOENT;
765                         CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
766                                ": rc = %d\n",
767                                cathandle->lgh_ctxt->loc_obd->obd_name,
768                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, lrc);
769                         failed++;
770                         if (rc == 0)
771                                 rc = lrc;
772                         continue;
773                 }
774
775                 lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
776                 if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
777                         index = loghandle->u.phd.phd_cookie.lgc_index;
778                         lrc = llog_cat_cleanup(env, cathandle, loghandle,
779                                                index);
780                         if (rc == 0)
781                                 rc = lrc;
782                 } else if (lrc == -ENOENT) {
783                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
784                                 rc = lrc;
785                 } else if (lrc < 0) {
786                         failed++;
787                         if (rc == 0)
788                                 rc = lrc;
789                 }
790                 llog_handle_put(loghandle);
791         }
792         if (rc)
793                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
794                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
795                        rc);
796
797         RETURN(rc);
798 }
799 EXPORT_SYMBOL(llog_cat_cancel_records);
800
801 static int llog_cat_process_common(const struct lu_env *env,
802                                    struct llog_handle *cat_llh,
803                                    struct llog_rec_hdr *rec,
804                                    struct llog_handle **llhp)
805 {
806         struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
807         struct llog_log_hdr *hdr;
808         int rc;
809
810         ENTRY;
811         if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
812                 rc = -EINVAL;
813                 CWARN("%s: invalid record in catalog "DFID":%x: rc = %d\n",
814                       cat_llh->lgh_ctxt->loc_obd->obd_name,
815                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid),
816                       cat_llh->lgh_id.lgl_ogen, rc);
817                 RETURN(rc);
818         }
819         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "DFID"\n",
820                PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
821                le32_to_cpu(rec->lrh_index),
822                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
823
824         rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
825         if (rc) {
826                 /* After a server crash, a stub of index record in catlog could
827                  * be kept, because plain log destroy + catlog index record
828                  * deletion are not atomic. So we end up with an index but no
829                  * actual record. Destroy the index and move on. */
830                 if (rc == -ENOENT || rc == -ESTALE)
831                         rc = LLOG_DEL_RECORD;
832                 else if (rc)
833                         CWARN("%s: can't find llog handle "DFID":%x: rc = %d\n",
834                               cat_llh->lgh_ctxt->loc_obd->obd_name,
835                               PFID(&lir->lid_id.lgl_oi.oi_fid),
836                               lir->lid_id.lgl_ogen, rc);
837
838                 RETURN(rc);
839         }
840
841         /* clean old empty llogs, do not consider current llog in use */
842         /* ignore remote (lgh_obj == NULL) llogs */
843         hdr = (*llhp)->lgh_hdr;
844         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
845             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
846             *llhp != cat_llh->u.chd.chd_current_log) {
847                 rc = llog_destroy(env, *llhp);
848                 if (rc)
849                         CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
850                               (*llhp)->lgh_ctxt->loc_obd->obd_name,
851                               PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
852                 rc = LLOG_DEL_PLAIN;
853         }
854
855         RETURN(rc);
856 }
857
858 static int llog_cat_process_cb(const struct lu_env *env,
859                                struct llog_handle *cat_llh,
860                                struct llog_rec_hdr *rec, void *data)
861 {
862         struct llog_process_data *d = data;
863         struct llog_handle *llh = NULL;
864         int rc;
865
866         ENTRY;
867         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
868         if (rc)
869                 GOTO(out, rc);
870
871         if (rec->lrh_index < d->lpd_startcat) {
872                 /* Skip processing of the logs until startcat */
873                 rc = 0;
874         } else if (d->lpd_startidx > 0) {
875                 struct llog_process_cat_data cd;
876
877                 cd.lpcd_first_idx = d->lpd_startidx;
878                 cd.lpcd_last_idx = 0;
879                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
880                                           &cd, false);
881                 /* Continue processing the next log from idx 0 */
882                 d->lpd_startidx = 0;
883         } else {
884                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
885                                           NULL, false);
886         }
887
888 out:
889         /* The empty plain log was destroyed while processing */
890         if (rc == LLOG_DEL_PLAIN) {
891                 rc = llog_cat_cleanup(env, cat_llh, llh,
892                                       llh->u.phd.phd_cookie.lgc_index);
893         } else if (rc == LLOG_DEL_RECORD) {
894                 /* clear wrong catalog entry */
895                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
896         }
897
898         if (llh)
899                 llog_handle_put(llh);
900
901         RETURN(rc);
902 }
903
904 int llog_cat_process_or_fork(const struct lu_env *env,
905                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
906                              llog_cb_t cb, void *data, int startcat,
907                              int startidx, bool fork)
908 {
909         struct llog_process_data d;
910         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
911         int rc;
912         ENTRY;
913
914         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
915         d.lpd_data = data;
916         d.lpd_cb = cb;
917         d.lpd_startcat = startcat;
918         d.lpd_startidx = startidx;
919
920         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
921             llh->llh_count > 1) {
922                 struct llog_process_cat_data cd;
923
924                 CWARN("%s: catlog "DFID" crosses index zero\n",
925                       cat_llh->lgh_ctxt->loc_obd->obd_name,
926                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
927
928                 cd.lpcd_first_idx = llh->llh_cat_idx;
929                 cd.lpcd_last_idx = 0;
930                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
931                                           &d, &cd, fork);
932                 if (rc != 0)
933                         RETURN(rc);
934
935                 cd.lpcd_first_idx = 0;
936                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
937                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
938                                           &d, &cd, fork);
939         } else {
940                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
941                                           &d, NULL, fork);
942         }
943
944         RETURN(rc);
945 }
946
947 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
948                      llog_cb_t cb, void *data, int startcat, int startidx)
949 {
950         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
951                                         cb, data, startcat, startidx, false);
952 }
953 EXPORT_SYMBOL(llog_cat_process);
954
955 static int llog_cat_size_cb(const struct lu_env *env,
956                              struct llog_handle *cat_llh,
957                              struct llog_rec_hdr *rec, void *data)
958 {
959         struct llog_process_data *d = data;
960         struct llog_handle *llh = NULL;
961         __u64 *cum_size = d->lpd_data;
962         __u64 size;
963         int rc;
964
965         ENTRY;
966         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
967
968         if (rc == LLOG_DEL_PLAIN) {
969                 /* empty log was deleted, don't count it */
970                 rc = llog_cat_cleanup(env, cat_llh, llh,
971                                       llh->u.phd.phd_cookie.lgc_index);
972         } else if (rc == LLOG_DEL_RECORD) {
973                 /* clear wrong catalog entry */
974                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
975         } else {
976                 size = llog_size(env, llh);
977                 *cum_size += size;
978
979                 CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
980                        PFID(&llh->lgh_id.lgl_oi.oi_fid), size, *cum_size);
981         }
982
983         if (llh != NULL)
984                 llog_handle_put(llh);
985
986         RETURN(0);
987 }
988
989 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
990 {
991         __u64 size = llog_size(env, cat_llh);
992
993         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
994                                  NULL, &size, 0, 0, false);
995
996         return size;
997 }
998 EXPORT_SYMBOL(llog_cat_size);
999
1000 /* currently returns the number of "free" entries in catalog,
1001  * ie the available entries for a new plain LLOG file creation,
1002  * even if catalog has wrapped
1003  */
1004 __u32 llog_cat_free_space(struct llog_handle *cat_llh)
1005 {
1006         /* simulate almost full Catalog */
1007         if (OBD_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
1008                 return cfs_fail_val;
1009
1010         if (cat_llh->lgh_hdr->llh_count == 1)
1011                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1;
1012
1013         if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
1014                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1 +
1015                        cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
1016
1017         /* catalog is presently wrapped */
1018         return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
1019 }
1020 EXPORT_SYMBOL(llog_cat_free_space);
1021
1022 static int llog_cat_reverse_process_cb(const struct lu_env *env,
1023                                        struct llog_handle *cat_llh,
1024                                        struct llog_rec_hdr *rec, void *data)
1025 {
1026         struct llog_process_data *d = data;
1027         struct llog_handle *llh;
1028         int rc;
1029
1030         ENTRY;
1031         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
1032
1033         /* The empty plain log was destroyed while processing */
1034         if (rc == LLOG_DEL_PLAIN) {
1035                 rc = llog_cat_cleanup(env, cat_llh, llh,
1036                                       llh->u.phd.phd_cookie.lgc_index);
1037         } else if (rc == LLOG_DEL_RECORD) {
1038                 /* clear wrong catalog entry */
1039                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
1040         }
1041         if (rc)
1042                 RETURN(rc);
1043
1044         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
1045
1046         /* The empty plain was destroyed while processing */
1047         if (rc == LLOG_DEL_PLAIN)
1048                 rc = llog_cat_cleanup(env, cat_llh, llh,
1049                                       llh->u.phd.phd_cookie.lgc_index);
1050
1051         llog_handle_put(llh);
1052         RETURN(rc);
1053 }
1054
1055 int llog_cat_reverse_process(const struct lu_env *env,
1056                              struct llog_handle *cat_llh,
1057                              llog_cb_t cb, void *data)
1058 {
1059         struct llog_process_data d;
1060         struct llog_process_cat_data cd;
1061         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1062         int rc;
1063         ENTRY;
1064
1065         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1066         d.lpd_data = data;
1067         d.lpd_cb = cb;
1068
1069         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1070             llh->llh_count > 1) {
1071                 CWARN("%s: catalog "DFID" crosses index zero\n",
1072                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1073                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1074
1075                 cd.lpcd_first_idx = 0;
1076                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1077                 rc = llog_reverse_process(env, cat_llh,
1078                                           llog_cat_reverse_process_cb,
1079                                           &d, &cd);
1080                 if (rc != 0)
1081                         RETURN(rc);
1082
1083                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1084                 cd.lpcd_last_idx = 0;
1085                 rc = llog_reverse_process(env, cat_llh,
1086                                           llog_cat_reverse_process_cb,
1087                                           &d, &cd);
1088         } else {
1089                 rc = llog_reverse_process(env, cat_llh,
1090                                           llog_cat_reverse_process_cb,
1091                                           &d, NULL);
1092         }
1093
1094         RETURN(rc);
1095 }
1096 EXPORT_SYMBOL(llog_cat_reverse_process);
1097
1098 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1099 {
1100         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1101         int bitmap_size;
1102
1103         ENTRY;
1104
1105         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1106         /*
1107          * The llh_cat_idx equals to the first used index minus 1
1108          * so if we canceled the first index then llh_cat_idx
1109          * must be renewed.
1110          */
1111         if (llh->llh_cat_idx == (idx - 1)) {
1112                 llh->llh_cat_idx = idx;
1113
1114                 while (idx != cathandle->lgh_last_idx) {
1115                         idx = (idx + 1) % bitmap_size;
1116                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1117                                 /* update llh_cat_idx for each unset bit,
1118                                  * expecting the next one is set */
1119                                 llh->llh_cat_idx = idx;
1120                         } else if (idx == 0) {
1121                                 /* skip header bit */
1122                                 llh->llh_cat_idx = 0;
1123                                 continue;
1124                         } else {
1125                                 /* the first index is found */
1126                                 break;
1127                         }
1128                 }
1129
1130                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1131                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1132                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1133         }
1134
1135         RETURN(0);
1136 }
1137
1138 /* Cleanup deleted plain llog traces from catalog */
1139 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1140                      struct llog_handle *loghandle, int index)
1141 {
1142         int rc;
1143
1144         LASSERT(index);
1145         if (loghandle != NULL) {
1146                 /* remove destroyed llog from catalog list and
1147                  * chd_current_log variable */
1148                 down_write(&cathandle->lgh_lock);
1149                 if (cathandle->u.chd.chd_current_log == loghandle)
1150                         cathandle->u.chd.chd_current_log = NULL;
1151                 list_del_init(&loghandle->u.phd.phd_entry);
1152                 up_write(&cathandle->lgh_lock);
1153                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1154                 /* llog was opened and keep in a list, close it now */
1155                 llog_close(env, loghandle);
1156         }
1157
1158         /* do not attempt to cleanup on-disk llog if on client side */
1159         if (cathandle->lgh_obj == NULL)
1160                 return 0;
1161
1162         /* remove plain llog entry from catalog by index */
1163         llog_cat_set_first_idx(cathandle, index);
1164         rc = llog_cancel_rec(env, cathandle, index);
1165         if (rc == 0)
1166                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1167                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1168         return rc;
1169 }