Whamcloud - gitweb
LU-13004 ptlrpc: Allow BULK_BUF_KIOV to accept a kvec
[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, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/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
53 /**
54  * lockdep markers for nested struct llog_handle::lgh_lock locking.
55  */
56 enum {
57         LLOGH_CAT,
58         LLOGH_LOG,
59 };
60
61 /* Create a new log handle and add it to the open list.
62  * This log handle will be closed when all of the records in it are removed.
63  *
64  * Assumes caller has already pushed us into the kernel context and is locking.
65  */
66 static int llog_cat_new_log(const struct lu_env *env,
67                             struct llog_handle *cathandle,
68                             struct llog_handle *loghandle,
69                             struct thandle *th)
70 {
71         struct llog_thread_info *lgi = llog_info(env);
72         struct llog_logid_rec   *rec = &lgi->lgi_logid;
73         struct thandle *handle = NULL;
74         struct dt_device *dt = NULL;
75         struct llog_log_hdr     *llh = cathandle->lgh_hdr;
76         int                      rc, index;
77
78         ENTRY;
79
80         index = (cathandle->lgh_last_idx + 1) %
81                 (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ? (cfs_fail_val + 1) :
82                                                 LLOG_HDR_BITMAP_SIZE(llh));
83
84         /* check that new llog index will not overlap with the first one.
85          * - llh_cat_idx is the index just before the first/oldest still in-use
86          *      index in catalog
87          * - lgh_last_idx is the last/newest used index in catalog
88          *
89          * When catalog is not wrapped yet then lgh_last_idx is always larger
90          * than llh_cat_idx. After the wrap around lgh_last_idx re-starts
91          * from 0 and llh_cat_idx becomes the upper limit for it
92          *
93          * Check if catalog has already wrapped around or not by comparing
94          * last_idx and cat_idx */
95         if ((index == llh->llh_cat_idx + 1 && llh->llh_count > 1) ||
96             (index == 0 && llh->llh_cat_idx == 0)) {
97                 if (cathandle->lgh_name == NULL) {
98                         CWARN("%s: there are no more free slots in catalog "
99                               DFID":%x\n",
100                               loghandle->lgh_ctxt->loc_obd->obd_name,
101                               PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
102                               cathandle->lgh_id.lgl_ogen);
103                 } else {
104                         CWARN("%s: there are no more free slots in "
105                               "catalog %s\n",
106                               loghandle->lgh_ctxt->loc_obd->obd_name,
107                               cathandle->lgh_name);
108                 }
109                 RETURN(-ENOSPC);
110         }
111
112         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
113                 RETURN(-ENOSPC);
114
115         if (loghandle->lgh_hdr != NULL) {
116                 /* If llog object is remote and creation is failed, lgh_hdr
117                  * might be left over here, free it first */
118                 LASSERT(!llog_exist(loghandle));
119                 OBD_FREE_LARGE(loghandle->lgh_hdr, loghandle->lgh_hdr_size);
120                 loghandle->lgh_hdr = NULL;
121         }
122
123         if (th == NULL) {
124                 dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
125
126                 handle = dt_trans_create(env, dt);
127                 if (IS_ERR(handle))
128                         RETURN(PTR_ERR(handle));
129
130                 /* Create update llog object synchronously, which
131                  * happens during inialization process see
132                  * lod_sub_prep_llog(), to make sure the update
133                  * llog object is created before corss-MDT writing
134                  * updates into the llog object */
135                 if (cathandle->lgh_ctxt->loc_flags & LLOG_CTXT_FLAG_NORMAL_FID)
136                         handle->th_sync = 1;
137
138                 handle->th_wait_submit = 1;
139
140                 rc = llog_declare_create(env, loghandle, handle);
141                 if (rc != 0)
142                         GOTO(out, rc);
143
144                 rec->lid_hdr.lrh_len = sizeof(*rec);
145                 rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
146                 rec->lid_id = loghandle->lgh_id;
147                 rc = llog_declare_write_rec(env, cathandle, &rec->lid_hdr, -1,
148                                             handle);
149                 if (rc != 0)
150                         GOTO(out, rc);
151
152                 rc = dt_trans_start_local(env, dt, handle);
153                 if (rc != 0)
154                         GOTO(out, rc);
155
156                 th = handle;
157         }
158
159         rc = llog_create(env, loghandle, th);
160         /* if llog is already created, no need to initialize it */
161         if (rc == -EEXIST) {
162                 GOTO(out, rc = 0);
163         } else if (rc != 0) {
164                 CERROR("%s: can't create new plain llog in catalog: rc = %d\n",
165                        loghandle->lgh_ctxt->loc_obd->obd_name, rc);
166                 GOTO(out, rc);
167         }
168
169         rc = llog_init_handle(env, loghandle,
170                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
171                               &cathandle->lgh_hdr->llh_tgtuuid);
172         if (rc < 0)
173                 GOTO(out, rc);
174
175         /* build the record for this log in the catalog */
176         rec->lid_hdr.lrh_len = sizeof(*rec);
177         rec->lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
178         rec->lid_id = loghandle->lgh_id;
179
180         /* append the new record into catalog. The new index will be
181          * assigned to the record and updated in rec header */
182         rc = llog_write_rec(env, cathandle, &rec->lid_hdr,
183                             &loghandle->u.phd.phd_cookie, LLOG_NEXT_IDX, th);
184         if (rc < 0)
185                 GOTO(out_destroy, rc);
186
187         CDEBUG(D_OTHER, "new plain log "DFID".%u of catalog "DFID"\n",
188                PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rec->lid_hdr.lrh_index,
189                PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
190
191         loghandle->lgh_hdr->llh_cat_idx = rec->lid_hdr.lrh_index;
192
193         /* limit max size of plain llog so that space can be
194          * released sooner, especially on small filesystems */
195         /* 2MB for the cases when free space hasn't been learned yet */
196         loghandle->lgh_max_size = 2 << 20;
197         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
198         rc = dt_statfs(env, dt, &lgi->lgi_statfs);
199         if (rc == 0 && lgi->lgi_statfs.os_bfree > 0) {
200                 __u64 freespace = (lgi->lgi_statfs.os_bfree *
201                                   lgi->lgi_statfs.os_bsize) >> 6;
202                 if (freespace < loghandle->lgh_max_size)
203                         loghandle->lgh_max_size = freespace;
204                 /* shouldn't be > 128MB in any case?
205                  * it's 256K records of 512 bytes each */
206                 if (freespace > (128 << 20))
207                         loghandle->lgh_max_size = 128 << 20;
208         }
209         if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_PLAIN_RECORDS) ||
210                      OBD_FAIL_PRECHECK(OBD_FAIL_CATALOG_FULL_CHECK))) {
211                 // limit the numer of plain records for test
212                 loghandle->lgh_max_size = loghandle->lgh_hdr_size +
213                        cfs_fail_val * 64;
214         }
215
216         rc = 0;
217
218 out:
219         if (handle != NULL) {
220                 handle->th_result = rc >= 0 ? 0 : rc;
221                 dt_trans_stop(env, dt, handle);
222         }
223         RETURN(rc);
224
225 out_destroy:
226         /* to signal llog_cat_close() it shouldn't try to destroy the llog,
227          * we want to destroy it in this transaction, otherwise the object
228          * becomes an orphan */
229         loghandle->lgh_hdr->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
230         /* this is to mimic full log, so another llog_cat_current_log()
231          * can skip it and ask for another onet */
232         loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(loghandle->lgh_hdr) + 1;
233         llog_trans_destroy(env, loghandle, th);
234         if (handle != NULL)
235                 dt_trans_stop(env, dt, handle);
236         RETURN(rc);
237 }
238
239 static int llog_cat_refresh(const struct lu_env *env,
240                             struct llog_handle *cathandle)
241 {
242         struct llog_handle *loghandle;
243         int rc;
244
245         down_write(&cathandle->lgh_lock);
246         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
247                             u.phd.phd_entry) {
248                 if (!llog_exist(loghandle))
249                         continue;
250
251                 rc = llog_read_header(env, loghandle, NULL);
252                 if (rc)
253                         goto unlock;
254         }
255
256         rc = llog_read_header(env, cathandle, NULL);
257 unlock:
258         up_write(&loghandle->lgh_lock);
259
260         return rc;
261 }
262
263 /*
264  * prepare current/next log for catalog.
265  *
266  * if \a *ploghandle is NULL, open it, and declare create, NB, if \a
267  * *ploghandle is remote, create it synchronously here, see comments
268  * below.
269  *
270  * \a cathandle->lgh_lock is down_read-ed, it gets down_write-ed if \a
271  * *ploghandle has to be opened.
272  */
273 static int llog_cat_prep_log(const struct lu_env *env,
274                              struct llog_handle *cathandle,
275                              struct llog_handle **ploghandle,
276                              struct thandle *th)
277 {
278         int rc;
279         int sem_upgraded;
280
281 start:
282         rc = 0;
283         sem_upgraded = 0;
284         if (IS_ERR_OR_NULL(*ploghandle)) {
285                 up_read(&cathandle->lgh_lock);
286                 down_write(&cathandle->lgh_lock);
287                 sem_upgraded = 1;
288                 if (IS_ERR_OR_NULL(*ploghandle)) {
289                         struct llog_handle *loghandle;
290
291                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
292                                        NULL, NULL, LLOG_OPEN_NEW);
293                         if (!rc) {
294                                 *ploghandle = loghandle;
295                                 list_add_tail(&loghandle->u.phd.phd_entry,
296                                               &cathandle->u.chd.chd_head);
297                         }
298                 }
299                 if (rc)
300                         GOTO(out, rc);
301         }
302
303         rc = llog_exist(*ploghandle);
304         if (rc < 0)
305                 GOTO(out, rc);
306         if (rc)
307                 GOTO(out, rc = 0);
308
309         if (dt_object_remote(cathandle->lgh_obj)) {
310                 down_write_nested(&(*ploghandle)->lgh_lock, LLOGH_LOG);
311                 if (!llog_exist(*ploghandle)) {
312                         /* For remote operation, if we put the llog object
313                          * creation in the current transaction, then the
314                          * llog object will not be created on the remote
315                          * target until the transaction stop, if other
316                          * operations start before the transaction stop,
317                          * and use the same llog object, will be dependent
318                          * on the success of this transaction. So let's
319                          * create the llog object synchronously here to
320                          * remove the dependency. */
321                         rc = llog_cat_new_log(env, cathandle, *ploghandle,
322                                               NULL);
323                         if (rc == -ESTALE) {
324                                 up_write(&(*ploghandle)->lgh_lock);
325                                 if (sem_upgraded)
326                                         up_write(&cathandle->lgh_lock);
327                                 else
328                                         up_read(&cathandle->lgh_lock);
329
330                                 rc = llog_cat_refresh(env, cathandle);
331                                 down_read_nested(&cathandle->lgh_lock,
332                                                  LLOGH_CAT);
333                                 if (rc)
334                                         return rc;
335                                 /* *ploghandle might become NULL, restart */
336                                 goto start;
337                         }
338                 }
339                 up_write(&(*ploghandle)->lgh_lock);
340         } else {
341                 struct llog_thread_info *lgi = llog_info(env);
342                 struct llog_logid_rec *lirec = &lgi->lgi_logid;
343
344                 rc = llog_declare_create(env, *ploghandle, th);
345                 if (rc)
346                         GOTO(out, rc);
347
348                 lirec->lid_hdr.lrh_len = sizeof(*lirec);
349                 rc = llog_declare_write_rec(env, cathandle, &lirec->lid_hdr, -1,
350                                             th);
351         }
352
353 out:
354         if (sem_upgraded) {
355                 up_write(&cathandle->lgh_lock);
356                 down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
357                 if (rc == 0)
358                         goto start;
359         }
360         return rc;
361 }
362
363 /* Open an existent log handle and add it to the open list.
364  * This log handle will be closed when all of the records in it are removed.
365  *
366  * Assumes caller has already pushed us into the kernel context and is locking.
367  * We return a lock on the handle to ensure nobody yanks it from us.
368  *
369  * This takes extra reference on llog_handle via llog_handle_get() and require
370  * this reference to be put by caller using llog_handle_put()
371  */
372 int llog_cat_id2handle(const struct lu_env *env, struct llog_handle *cathandle,
373                        struct llog_handle **res, struct llog_logid *logid)
374 {
375         struct llog_handle      *loghandle;
376         enum llog_flag           fmt;
377         int                      rc = 0;
378
379         ENTRY;
380
381         if (cathandle == NULL)
382                 RETURN(-EBADF);
383
384         fmt = cathandle->lgh_hdr->llh_flags & LLOG_F_EXT_MASK;
385         down_write(&cathandle->lgh_lock);
386         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
387                             u.phd.phd_entry) {
388                 struct llog_logid *cgl = &loghandle->lgh_id;
389
390                 if (ostid_id(&cgl->lgl_oi) == ostid_id(&logid->lgl_oi) &&
391                     ostid_seq(&cgl->lgl_oi) == ostid_seq(&logid->lgl_oi)) {
392                         if (cgl->lgl_ogen != logid->lgl_ogen) {
393                                 CWARN("%s: log "DFID" generation %x != %x\n",
394                                       loghandle->lgh_ctxt->loc_obd->obd_name,
395                                       PFID(&logid->lgl_oi.oi_fid),
396                                       cgl->lgl_ogen, logid->lgl_ogen);
397                                 continue;
398                         }
399                         loghandle->u.phd.phd_cat_handle = cathandle;
400                         up_write(&cathandle->lgh_lock);
401                         GOTO(out, rc = 0);
402                 }
403         }
404         up_write(&cathandle->lgh_lock);
405
406         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle, logid, NULL,
407                        LLOG_OPEN_EXISTS);
408         if (rc < 0) {
409                 CERROR("%s: error opening log id "DFID":%x: rc = %d\n",
410                        cathandle->lgh_ctxt->loc_obd->obd_name,
411                        PFID(&logid->lgl_oi.oi_fid), logid->lgl_ogen, rc);
412                 RETURN(rc);
413         }
414
415         rc = llog_init_handle(env, loghandle, LLOG_F_IS_PLAIN | fmt, NULL);
416         if (rc < 0) {
417                 llog_close(env, loghandle);
418                 loghandle = NULL;
419                 RETURN(rc);
420         }
421
422         down_write(&cathandle->lgh_lock);
423         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
424         up_write(&cathandle->lgh_lock);
425
426         loghandle->u.phd.phd_cat_handle = cathandle;
427         loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
428         loghandle->u.phd.phd_cookie.lgc_index =
429                                 loghandle->lgh_hdr->llh_cat_idx;
430         EXIT;
431 out:
432         llog_handle_get(loghandle);
433         *res = loghandle;
434         return 0;
435 }
436
437 int llog_cat_close(const struct lu_env *env, struct llog_handle *cathandle)
438 {
439         struct llog_handle      *loghandle, *n;
440         int                      rc;
441
442         ENTRY;
443
444         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
445                                  u.phd.phd_entry) {
446                 struct llog_log_hdr     *llh = loghandle->lgh_hdr;
447                 int                      index;
448
449                 /* unlink open-not-created llogs */
450                 list_del_init(&loghandle->u.phd.phd_entry);
451                 llh = loghandle->lgh_hdr;
452                 if (loghandle->lgh_obj != NULL && llh != NULL &&
453                     (llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
454                     (llh->llh_count == 1)) {
455                         rc = llog_destroy(env, loghandle);
456                         if (rc)
457                                 CERROR("%s: failure destroying log during "
458                                        "cleanup: rc = %d\n",
459                                        loghandle->lgh_ctxt->loc_obd->obd_name,
460                                        rc);
461
462                         index = loghandle->u.phd.phd_cookie.lgc_index;
463                         llog_cat_cleanup(env, cathandle, NULL, index);
464                 }
465                 llog_close(env, loghandle);
466         }
467         /* if handle was stored in ctxt, remove it too */
468         if (cathandle->lgh_ctxt->loc_handle == cathandle)
469                 cathandle->lgh_ctxt->loc_handle = NULL;
470         rc = llog_close(env, cathandle);
471         RETURN(rc);
472 }
473 EXPORT_SYMBOL(llog_cat_close);
474
475 /** Return the currently active log handle.  If the current log handle doesn't
476  * have enough space left for the current record, start a new one.
477  *
478  * If reclen is 0, we only want to know what the currently active log is,
479  * otherwise we get a lock on this log so nobody can steal our space.
480  *
481  * Assumes caller has already pushed us into the kernel context and is locking.
482  *
483  * NOTE: loghandle is write-locked upon successful return
484  */
485 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
486                                                 struct thandle *th)
487 {
488         struct llog_handle *loghandle = NULL;
489         ENTRY;
490
491
492         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2)) {
493                 down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
494                 GOTO(next, loghandle);
495         }
496
497         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
498         loghandle = cathandle->u.chd.chd_current_log;
499         if (loghandle) {
500                 struct llog_log_hdr *llh;
501
502                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
503                 llh = loghandle->lgh_hdr;
504                 if (llh == NULL || !llog_is_full(loghandle)) {
505                         up_read(&cathandle->lgh_lock);
506                         RETURN(loghandle);
507                 } else {
508                         up_write(&loghandle->lgh_lock);
509                 }
510         }
511         up_read(&cathandle->lgh_lock);
512
513         /* time to use next log */
514
515         /* first, we have to make sure the state hasn't changed */
516         down_write_nested(&cathandle->lgh_lock, LLOGH_CAT);
517         loghandle = cathandle->u.chd.chd_current_log;
518         if (loghandle) {
519                 struct llog_log_hdr *llh;
520
521                 down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
522                 llh = loghandle->lgh_hdr;
523                 if (llh == NULL || !llog_is_full(loghandle))
524                         GOTO(out_unlock, loghandle);
525                 else
526                         up_write(&loghandle->lgh_lock);
527         }
528
529 next:
530         /* Sigh, the chd_next_log and chd_current_log is initialized
531          * in declare phase, and we do not serialize the catlog
532          * accessing, so it might be possible the llog creation
533          * thread (see llog_cat_declare_add_rec()) did not create
534          * llog successfully, then the following thread might
535          * meet this situation. */
536         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
537                 CERROR("%s: next log does not exist!\n",
538                        cathandle->lgh_ctxt->loc_obd->obd_name);
539                 loghandle = ERR_PTR(-EIO);
540                 if (cathandle->u.chd.chd_next_log == NULL) {
541                         /* Store the error in chd_next_log, so
542                          * the following process can get correct
543                          * failure value */
544                         cathandle->u.chd.chd_next_log = loghandle;
545                 }
546                 GOTO(out_unlock, loghandle);
547         }
548
549         CDEBUG(D_INODE, "use next log\n");
550
551         loghandle = cathandle->u.chd.chd_next_log;
552         cathandle->u.chd.chd_current_log = loghandle;
553         cathandle->u.chd.chd_next_log = NULL;
554         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
555
556 out_unlock:
557         up_write(&cathandle->lgh_lock);
558         LASSERT(loghandle);
559         RETURN(loghandle);
560 }
561
562 /* Add a single record to the recovery log(s) using a catalog
563  * Returns as llog_write_record
564  *
565  * Assumes caller has already pushed us into the kernel context.
566  */
567 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
568                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
569                      struct thandle *th)
570 {
571         struct llog_handle *loghandle;
572         int rc, retried = 0;
573         ENTRY;
574
575         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
576
577 retry:
578         loghandle = llog_cat_current_log(cathandle, th);
579         if (IS_ERR(loghandle))
580                 RETURN(PTR_ERR(loghandle));
581
582         /* loghandle is already locked by llog_cat_current_log() for us */
583         if (!llog_exist(loghandle)) {
584                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
585                 if (rc < 0) {
586                         up_write(&loghandle->lgh_lock);
587                         /* nobody should be trying to use this llog */
588                         down_write(&cathandle->lgh_lock);
589                         if (cathandle->u.chd.chd_current_log == loghandle)
590                                 cathandle->u.chd.chd_current_log = NULL;
591                         up_write(&cathandle->lgh_lock);
592                         RETURN(rc);
593                 }
594         }
595         /* now let's try to add the record */
596         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
597         if (rc < 0) {
598                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
599                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
600                 /* -ENOSPC is returned if no empty records left
601                  * and when it's lack of space on the stogage.
602                  * there is no point to try again if it's the second
603                  * case. many callers (like llog test) expect ENOSPC,
604                  * so we preserve this error code, but look for the
605                  * actual cause here */
606                 if (rc == -ENOSPC && llog_is_full(loghandle))
607                         rc = -ENOBUFS;
608         }
609         up_write(&loghandle->lgh_lock);
610
611         if (rc == -ENOBUFS) {
612                 if (retried++ == 0)
613                         GOTO(retry, rc);
614                 CERROR("%s: error on 2nd llog: rc = %d\n",
615                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
616         }
617
618         RETURN(rc);
619 }
620 EXPORT_SYMBOL(llog_cat_add_rec);
621
622 int llog_cat_declare_add_rec(const struct lu_env *env,
623                              struct llog_handle *cathandle,
624                              struct llog_rec_hdr *rec, struct thandle *th)
625 {
626         int rc;
627
628         ENTRY;
629
630 start:
631         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
632         rc = llog_cat_prep_log(env, cathandle,
633                                &cathandle->u.chd.chd_current_log, th);
634         if (rc)
635                 GOTO(unlock, rc);
636
637         rc = llog_cat_prep_log(env, cathandle, &cathandle->u.chd.chd_next_log,
638                                th);
639         if (rc)
640                 GOTO(unlock, rc);
641
642         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
643                                     rec, -1, th);
644         if (rc == -ESTALE && dt_object_remote(cathandle->lgh_obj)) {
645                 up_read(&cathandle->lgh_lock);
646                 rc = llog_cat_refresh(env, cathandle);
647                 if (rc)
648                         RETURN(rc);
649                 goto start;
650         }
651
652 #if 0
653         /*
654          * XXX: we hope for declarations made for existing llog this might be
655          * not correct with some backends where declarations are expected
656          * against specific object like ZFS with full debugging enabled.
657          */
658         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_next_log, rec, -1,
659                                     th);
660 #endif
661 unlock:
662         up_read(&cathandle->lgh_lock);
663         RETURN(rc);
664 }
665 EXPORT_SYMBOL(llog_cat_declare_add_rec);
666
667 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
668                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
669 {
670         struct llog_ctxt        *ctxt;
671         struct dt_device        *dt;
672         struct thandle          *th = NULL;
673         int                      rc;
674
675         ctxt = cathandle->lgh_ctxt;
676         LASSERT(ctxt);
677         LASSERT(ctxt->loc_exp);
678
679         LASSERT(cathandle->lgh_obj != NULL);
680         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
681
682         th = dt_trans_create(env, dt);
683         if (IS_ERR(th))
684                 RETURN(PTR_ERR(th));
685
686         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
687         if (rc)
688                 GOTO(out_trans, rc);
689
690         rc = dt_trans_start_local(env, dt, th);
691         if (rc)
692                 GOTO(out_trans, rc);
693         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
694 out_trans:
695         dt_trans_stop(env, dt, th);
696         RETURN(rc);
697 }
698 EXPORT_SYMBOL(llog_cat_add);
699
700 int llog_cat_cancel_arr_rec(const struct lu_env *env,
701                             struct llog_handle *cathandle,
702                             struct llog_logid *lgl, int count, int *index)
703 {
704         struct llog_handle *loghandle;
705         int  rc;
706
707         ENTRY;
708         rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
709         if (rc) {
710                 CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
711                        ": rc = %d\n",
712                        cathandle->lgh_ctxt->loc_obd->obd_name,
713                        PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
714                 RETURN(rc);
715         }
716
717         if ((cathandle->lgh_ctxt->loc_flags &
718              LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
719                 /* For update log, some of loghandles of cathandle
720                  * might not exist because remote llog creation might
721                  * be failed, so let's skip the record cancellation
722                  * for these non-exist llogs.
723                  */
724                 rc = -ENOENT;
725                 CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
726                        ": rc = %d\n",
727                        cathandle->lgh_ctxt->loc_obd->obd_name,
728                        PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
729
730                 llog_handle_put(loghandle);
731                 RETURN(rc);
732         }
733
734         rc = llog_cancel_arr_rec(env, loghandle, count, index);
735         if (rc == LLOG_DEL_PLAIN) { /* log has been destroyed */
736                 int cat_index;
737
738                 cat_index = loghandle->u.phd.phd_cookie.lgc_index;
739                 rc = llog_cat_cleanup(env, cathandle, loghandle, cat_index);
740                 if (rc)
741                         CERROR("%s: fail to cancel catalog record: rc = %d\n",
742                                cathandle->lgh_ctxt->loc_obd->obd_name, rc);
743                 rc = 0;
744
745         }
746         llog_handle_put(loghandle);
747
748         if (rc)
749                 CERROR("%s: fail to cancel %d llog-records: rc = %d\n",
750                        cathandle->lgh_ctxt->loc_obd->obd_name, count,
751                        rc);
752
753         RETURN(rc);
754 }
755 EXPORT_SYMBOL(llog_cat_cancel_arr_rec);
756
757 /* For each cookie in the cookie array, we clear the log in-use bit and either:
758  * - the log is empty, so mark it free in the catalog header and delete it
759  * - the log is not empty, just write out the log header
760  *
761  * The cookies may be in different log files, so we need to get new logs
762  * each time.
763  *
764  * Assumes caller has already pushed us into the kernel context.
765  */
766 int llog_cat_cancel_records(const struct lu_env *env,
767                             struct llog_handle *cathandle, int count,
768                             struct llog_cookie *cookies)
769 {
770         int i, rc = 0, failed = 0;
771
772         ENTRY;
773
774         for (i = 0; i < count; i++, cookies++) {
775                 int lrc;
776
777                 lrc = llog_cat_cancel_arr_rec(env, cathandle, &cookies->lgc_lgl,
778                                              1, &cookies->lgc_index);
779                 if (lrc) {
780                         failed++;
781                         if (!rc)
782                                 rc = lrc;
783                 }
784         }
785         if (failed)
786                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
787                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
788                        rc);
789         RETURN(rc);
790 }
791 EXPORT_SYMBOL(llog_cat_cancel_records);
792
793 static int llog_cat_process_common(const struct lu_env *env,
794                                    struct llog_handle *cat_llh,
795                                    struct llog_rec_hdr *rec,
796                                    struct llog_handle **llhp)
797 {
798         struct llog_logid_rec *lir = container_of(rec, typeof(*lir), lid_hdr);
799         struct llog_log_hdr *hdr;
800         int rc;
801
802         ENTRY;
803         if (rec->lrh_type != le32_to_cpu(LLOG_LOGID_MAGIC)) {
804                 rc = -EINVAL;
805                 CWARN("%s: invalid record in catalog "DFID":%x: rc = %d\n",
806                       cat_llh->lgh_ctxt->loc_obd->obd_name,
807                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid),
808                       cat_llh->lgh_id.lgl_ogen, rc);
809                 RETURN(rc);
810         }
811         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "DFID"\n",
812                PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
813                le32_to_cpu(rec->lrh_index),
814                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
815
816         rc = llog_cat_id2handle(env, cat_llh, llhp, &lir->lid_id);
817         if (rc) {
818                 /* After a server crash, a stub of index record in catlog could
819                  * be kept, because plain log destroy + catlog index record
820                  * deletion are not atomic. So we end up with an index but no
821                  * actual record. Destroy the index and move on. */
822                 if (rc == -ENOENT || rc == -ESTALE)
823                         rc = LLOG_DEL_RECORD;
824                 else if (rc)
825                         CWARN("%s: can't find llog handle "DFID":%x: rc = %d\n",
826                               cat_llh->lgh_ctxt->loc_obd->obd_name,
827                               PFID(&lir->lid_id.lgl_oi.oi_fid),
828                               lir->lid_id.lgl_ogen, rc);
829
830                 RETURN(rc);
831         }
832
833         /* clean old empty llogs, do not consider current llog in use */
834         /* ignore remote (lgh_obj == NULL) llogs */
835         hdr = (*llhp)->lgh_hdr;
836         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
837             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
838             *llhp != cat_llh->u.chd.chd_current_log) {
839                 rc = llog_destroy(env, *llhp);
840                 if (rc)
841                         CWARN("%s: can't destroy empty log "DFID": rc = %d\n",
842                               (*llhp)->lgh_ctxt->loc_obd->obd_name,
843                               PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
844                 rc = LLOG_DEL_PLAIN;
845         }
846
847         RETURN(rc);
848 }
849
850 static int llog_cat_process_cb(const struct lu_env *env,
851                                struct llog_handle *cat_llh,
852                                struct llog_rec_hdr *rec, void *data)
853 {
854         struct llog_process_data *d = data;
855         struct llog_handle *llh = NULL;
856         int rc;
857
858         ENTRY;
859         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
860         if (rc)
861                 GOTO(out, rc);
862
863         if (rec->lrh_index < d->lpd_startcat) {
864                 /* Skip processing of the logs until startcat */
865                 rc = 0;
866         } else if (d->lpd_startidx > 0) {
867                 struct llog_process_cat_data cd;
868
869                 cd.lpcd_first_idx = d->lpd_startidx;
870                 cd.lpcd_last_idx = 0;
871                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
872                                           &cd, false);
873                 /* Continue processing the next log from idx 0 */
874                 d->lpd_startidx = 0;
875         } else {
876                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
877                                           NULL, false);
878         }
879
880 out:
881         /* The empty plain log was destroyed while processing */
882         if (rc == LLOG_DEL_PLAIN) {
883                 rc = llog_cat_cleanup(env, cat_llh, llh,
884                                       llh->u.phd.phd_cookie.lgc_index);
885         } else if (rc == LLOG_DEL_RECORD) {
886                 /* clear wrong catalog entry */
887                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
888         }
889
890         if (llh)
891                 llog_handle_put(llh);
892
893         RETURN(rc);
894 }
895
896 int llog_cat_process_or_fork(const struct lu_env *env,
897                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
898                              llog_cb_t cb, void *data, int startcat,
899                              int startidx, bool fork)
900 {
901         struct llog_process_data d;
902         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
903         int rc;
904
905         ENTRY;
906
907         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
908         d.lpd_data = data;
909         d.lpd_cb = cb;
910         d.lpd_startcat = (startcat == LLOG_CAT_FIRST ? 0 : startcat);
911         d.lpd_startidx = startidx;
912
913         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
914             llh->llh_count > 1) {
915                 struct llog_process_cat_data cd;
916
917                 CWARN("%s: catlog "DFID" crosses index zero\n",
918                       cat_llh->lgh_ctxt->loc_obd->obd_name,
919                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
920                 /*startcat = 0 is default value for general processing */
921                 if ((startcat != LLOG_CAT_FIRST &&
922                     startcat >= llh->llh_cat_idx) || !startcat) {
923                         /* processing the catalog part at the end */
924                         cd.lpcd_first_idx = (startcat ? startcat :
925                                              llh->llh_cat_idx);
926                         if (OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS))
927                                 cd.lpcd_last_idx = cfs_fail_val;
928                         else
929                                 cd.lpcd_last_idx = 0;
930                         rc = llog_process_or_fork(env, cat_llh, cat_cb,
931                                                   &d, &cd, fork);
932                         /* Reset the startcat becasue it has already reached
933                          * catalog bottom.
934                          */
935                         startcat = 0;
936                         d.lpd_startcat = 0;
937                         if (rc != 0)
938                                 RETURN(rc);
939                 }
940                 /* processing the catalog part at the begining */
941                 cd.lpcd_first_idx = (startcat == LLOG_CAT_FIRST) ? 0 : startcat;
942                 /* Note, the processing will stop at the lgh_last_idx value,
943                  * and it could be increased during processing. So records
944                  * between current lgh_last_idx and lgh_last_idx in future
945                  * would left unprocessed.
946                  */
947                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
948                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
949                                           &d, &cd, fork);
950         } else {
951                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
952                                           &d, NULL, fork);
953         }
954
955         RETURN(rc);
956 }
957 EXPORT_SYMBOL(llog_cat_process_or_fork);
958
959 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
960                      llog_cb_t cb, void *data, int startcat, int startidx)
961 {
962         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
963                                         cb, data, startcat, startidx, false);
964 }
965 EXPORT_SYMBOL(llog_cat_process);
966
967 static int llog_cat_size_cb(const struct lu_env *env,
968                              struct llog_handle *cat_llh,
969                              struct llog_rec_hdr *rec, void *data)
970 {
971         struct llog_process_data *d = data;
972         struct llog_handle *llh = NULL;
973         __u64 *cum_size = d->lpd_data;
974         __u64 size;
975         int rc;
976
977         ENTRY;
978         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
979
980         if (rc == LLOG_DEL_PLAIN) {
981                 /* empty log was deleted, don't count it */
982                 rc = llog_cat_cleanup(env, cat_llh, llh,
983                                       llh->u.phd.phd_cookie.lgc_index);
984         } else if (rc == LLOG_DEL_RECORD) {
985                 /* clear wrong catalog entry */
986                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
987         } else {
988                 size = llog_size(env, llh);
989                 *cum_size += size;
990
991                 CDEBUG(D_INFO, "Add llog entry "DFID" size=%llu, tot=%llu\n",
992                        PFID(&llh->lgh_id.lgl_oi.oi_fid), size, *cum_size);
993         }
994
995         if (llh != NULL)
996                 llog_handle_put(llh);
997
998         RETURN(0);
999 }
1000
1001 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
1002 {
1003         __u64 size = llog_size(env, cat_llh);
1004
1005         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
1006                                  NULL, &size, 0, 0, false);
1007
1008         return size;
1009 }
1010 EXPORT_SYMBOL(llog_cat_size);
1011
1012 /* currently returns the number of "free" entries in catalog,
1013  * ie the available entries for a new plain LLOG file creation,
1014  * even if catalog has wrapped
1015  */
1016 __u32 llog_cat_free_space(struct llog_handle *cat_llh)
1017 {
1018         /* simulate almost full Catalog */
1019         if (OBD_FAIL_CHECK(OBD_FAIL_CAT_FREE_RECORDS))
1020                 return cfs_fail_val;
1021
1022         if (cat_llh->lgh_hdr->llh_count == 1)
1023                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1;
1024
1025         if (cat_llh->lgh_last_idx > cat_llh->lgh_hdr->llh_cat_idx)
1026                 return LLOG_HDR_BITMAP_SIZE(cat_llh->lgh_hdr) - 1 +
1027                        cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
1028
1029         /* catalog is presently wrapped */
1030         return cat_llh->lgh_hdr->llh_cat_idx - cat_llh->lgh_last_idx;
1031 }
1032 EXPORT_SYMBOL(llog_cat_free_space);
1033
1034 static int llog_cat_reverse_process_cb(const struct lu_env *env,
1035                                        struct llog_handle *cat_llh,
1036                                        struct llog_rec_hdr *rec, void *data)
1037 {
1038         struct llog_process_data *d = data;
1039         struct llog_handle *llh;
1040         int rc;
1041
1042         ENTRY;
1043         rc = llog_cat_process_common(env, cat_llh, rec, &llh);
1044
1045         /* The empty plain log was destroyed while processing */
1046         if (rc == LLOG_DEL_PLAIN) {
1047                 rc = llog_cat_cleanup(env, cat_llh, llh,
1048                                       llh->u.phd.phd_cookie.lgc_index);
1049         } else if (rc == LLOG_DEL_RECORD) {
1050                 /* clear wrong catalog entry */
1051                 rc = llog_cat_cleanup(env, cat_llh, NULL, rec->lrh_index);
1052         }
1053         if (rc)
1054                 RETURN(rc);
1055
1056         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
1057
1058         /* The empty plain was destroyed while processing */
1059         if (rc == LLOG_DEL_PLAIN)
1060                 rc = llog_cat_cleanup(env, cat_llh, llh,
1061                                       llh->u.phd.phd_cookie.lgc_index);
1062
1063         llog_handle_put(llh);
1064         RETURN(rc);
1065 }
1066
1067 int llog_cat_reverse_process(const struct lu_env *env,
1068                              struct llog_handle *cat_llh,
1069                              llog_cb_t cb, void *data)
1070 {
1071         struct llog_process_data d;
1072         struct llog_process_cat_data cd;
1073         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1074         int rc;
1075         ENTRY;
1076
1077         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1078         d.lpd_data = data;
1079         d.lpd_cb = cb;
1080
1081         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1082             llh->llh_count > 1) {
1083                 CWARN("%s: catalog "DFID" crosses index zero\n",
1084                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1085                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1086
1087                 cd.lpcd_first_idx = 0;
1088                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1089                 rc = llog_reverse_process(env, cat_llh,
1090                                           llog_cat_reverse_process_cb,
1091                                           &d, &cd);
1092                 if (rc != 0)
1093                         RETURN(rc);
1094
1095                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1096                 cd.lpcd_last_idx = 0;
1097                 rc = llog_reverse_process(env, cat_llh,
1098                                           llog_cat_reverse_process_cb,
1099                                           &d, &cd);
1100         } else {
1101                 rc = llog_reverse_process(env, cat_llh,
1102                                           llog_cat_reverse_process_cb,
1103                                           &d, NULL);
1104         }
1105
1106         RETURN(rc);
1107 }
1108 EXPORT_SYMBOL(llog_cat_reverse_process);
1109
1110 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1111 {
1112         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1113         int bitmap_size;
1114
1115         ENTRY;
1116
1117         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1118         /*
1119          * The llh_cat_idx equals to the first used index minus 1
1120          * so if we canceled the first index then llh_cat_idx
1121          * must be renewed.
1122          */
1123         if (llh->llh_cat_idx == (idx - 1)) {
1124                 llh->llh_cat_idx = idx;
1125
1126                 while (idx != cathandle->lgh_last_idx) {
1127                         idx = (idx + 1) % bitmap_size;
1128                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1129                                 /* update llh_cat_idx for each unset bit,
1130                                  * expecting the next one is set */
1131                                 llh->llh_cat_idx = idx;
1132                         } else if (idx == 0) {
1133                                 /* skip header bit */
1134                                 llh->llh_cat_idx = 0;
1135                                 continue;
1136                         } else {
1137                                 /* the first index is found */
1138                                 break;
1139                         }
1140                 }
1141
1142                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1143                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1144                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1145         }
1146
1147         RETURN(0);
1148 }
1149
1150 /* Cleanup deleted plain llog traces from catalog */
1151 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1152                      struct llog_handle *loghandle, int index)
1153 {
1154         int rc;
1155
1156         LASSERT(index);
1157         if (loghandle != NULL) {
1158                 /* remove destroyed llog from catalog list and
1159                  * chd_current_log variable */
1160                 down_write(&cathandle->lgh_lock);
1161                 if (cathandle->u.chd.chd_current_log == loghandle)
1162                         cathandle->u.chd.chd_current_log = NULL;
1163                 list_del_init(&loghandle->u.phd.phd_entry);
1164                 up_write(&cathandle->lgh_lock);
1165                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1166                 /* llog was opened and keep in a list, close it now */
1167                 llog_close(env, loghandle);
1168         }
1169
1170         /* do not attempt to cleanup on-disk llog if on client side */
1171         if (cathandle->lgh_obj == NULL)
1172                 return 0;
1173
1174         /* remove plain llog entry from catalog by index */
1175         llog_cat_set_first_idx(cathandle, index);
1176         rc = llog_cancel_rec(env, cathandle, index);
1177         if (rc == 0)
1178                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1179                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1180         return rc;
1181 }