Whamcloud - gitweb
248eab51a91e91189c5bef979189e6c1d274714e
[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                 if (llh == NULL || !llog_is_full(loghandle))
390                         GOTO(out_unlock, loghandle);
391                 else
392                         up_write(&loghandle->lgh_lock);
393         }
394
395 next:
396         /* Sigh, the chd_next_log and chd_current_log is initialized
397          * in declare phase, and we do not serialize the catlog
398          * accessing, so it might be possible the llog creation
399          * thread (see llog_cat_declare_add_rec()) did not create
400          * llog successfully, then the following thread might
401          * meet this situation. */
402         if (IS_ERR_OR_NULL(cathandle->u.chd.chd_next_log)) {
403                 CERROR("%s: next log does not exist!\n",
404                        cathandle->lgh_ctxt->loc_obd->obd_name);
405                 loghandle = ERR_PTR(-EIO);
406                 if (cathandle->u.chd.chd_next_log == NULL) {
407                         /* Store the error in chd_next_log, so
408                          * the following process can get correct
409                          * failure value */
410                         cathandle->u.chd.chd_next_log = loghandle;
411                 }
412                 GOTO(out_unlock, loghandle);
413         }
414
415         CDEBUG(D_INODE, "use next log\n");
416
417         loghandle = cathandle->u.chd.chd_next_log;
418         cathandle->u.chd.chd_current_log = loghandle;
419         cathandle->u.chd.chd_next_log = NULL;
420         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
421
422 out_unlock:
423         up_write(&cathandle->lgh_lock);
424         LASSERT(loghandle);
425         RETURN(loghandle);
426 }
427
428 static int llog_cat_update_header(const struct lu_env *env,
429                            struct llog_handle *cathandle)
430 {
431         struct llog_handle *loghandle;
432         int rc;
433         ENTRY;
434
435         /* refresh llog */
436         down_write(&cathandle->lgh_lock);
437         if (!cathandle->lgh_stale) {
438                 up_write(&cathandle->lgh_lock);
439                 RETURN(0);
440         }
441         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
442                             u.phd.phd_entry) {
443                 if (!llog_exist(loghandle))
444                         continue;
445
446                 rc = llog_read_header(env, loghandle, NULL);
447                 if (rc != 0) {
448                         up_write(&cathandle->lgh_lock);
449                         GOTO(out, rc);
450                 }
451         }
452         rc = llog_read_header(env, cathandle, NULL);
453         if (rc == 0)
454                 cathandle->lgh_stale = 0;
455         up_write(&cathandle->lgh_lock);
456         if (rc != 0)
457                 GOTO(out, rc);
458 out:
459         RETURN(rc);
460 }
461
462 /* Add a single record to the recovery log(s) using a catalog
463  * Returns as llog_write_record
464  *
465  * Assumes caller has already pushed us into the kernel context.
466  */
467 int llog_cat_add_rec(const struct lu_env *env, struct llog_handle *cathandle,
468                      struct llog_rec_hdr *rec, struct llog_cookie *reccookie,
469                      struct thandle *th)
470 {
471         struct llog_handle *loghandle;
472         int rc, retried = 0;
473         ENTRY;
474
475         LASSERT(rec->lrh_len <= cathandle->lgh_ctxt->loc_chunk_size);
476
477 retry:
478         loghandle = llog_cat_current_log(cathandle, th);
479         if (IS_ERR(loghandle))
480                 RETURN(PTR_ERR(loghandle));
481
482         /* loghandle is already locked by llog_cat_current_log() for us */
483         if (!llog_exist(loghandle)) {
484                 rc = llog_cat_new_log(env, cathandle, loghandle, th);
485                 if (rc < 0) {
486                         up_write(&loghandle->lgh_lock);
487                         /* nobody should be trying to use this llog */
488                         down_write(&cathandle->lgh_lock);
489                         /* only reset current log if still room in catalog, to
490                          * avoid unnecessarily and racy creation of new and
491                          * partially initialized llog_handle
492                          */
493                         if ((cathandle->u.chd.chd_current_log == loghandle) &&
494                             rc != -ENOSPC)
495                                 cathandle->u.chd.chd_current_log = NULL;
496                         up_write(&cathandle->lgh_lock);
497                         RETURN(rc);
498                 }
499         }
500         /* now let's try to add the record */
501         rc = llog_write_rec(env, loghandle, rec, reccookie, LLOG_NEXT_IDX, th);
502         if (rc < 0) {
503                 CDEBUG_LIMIT(rc == -ENOSPC ? D_HA : D_ERROR,
504                              "llog_write_rec %d: lh=%p\n", rc, loghandle);
505                 /* -ENOSPC is returned if no empty records left
506                  * and when it's lack of space on the stogage.
507                  * there is no point to try again if it's the second
508                  * case. many callers (like llog test) expect ENOSPC,
509                  * so we preserve this error code, but look for the
510                  * actual cause here */
511                 if (rc == -ENOSPC && llog_is_full(loghandle))
512                         rc = -ENOBUFS;
513         }
514         up_write(&loghandle->lgh_lock);
515
516         if (rc == -ENOBUFS) {
517                 if (retried++ == 0)
518                         GOTO(retry, rc);
519                 CERROR("%s: error on 2nd llog: rc = %d\n",
520                        cathandle->lgh_ctxt->loc_obd->obd_name, rc);
521         }
522
523         RETURN(rc);
524 }
525 EXPORT_SYMBOL(llog_cat_add_rec);
526
527 int llog_cat_declare_add_rec(const struct lu_env *env,
528                              struct llog_handle *cathandle,
529                              struct llog_rec_hdr *rec, struct thandle *th)
530 {
531         struct llog_thread_info *lgi = llog_info(env);
532         struct llog_logid_rec   *lirec = &lgi->lgi_logid;
533         struct llog_handle      *loghandle, *next;
534         int                      rc = 0;
535
536         ENTRY;
537
538         if (cathandle->u.chd.chd_current_log == NULL) {
539                 /* declare new plain llog */
540                 down_write(&cathandle->lgh_lock);
541                 if (cathandle->u.chd.chd_current_log == NULL) {
542                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
543                                        NULL, NULL, LLOG_OPEN_NEW);
544                         if (rc == 0) {
545                                 cathandle->u.chd.chd_current_log = loghandle;
546                                 list_add_tail(&loghandle->u.phd.phd_entry,
547                                               &cathandle->u.chd.chd_head);
548                         }
549                 }
550                 up_write(&cathandle->lgh_lock);
551         } else if (cathandle->u.chd.chd_next_log == NULL ||
552                    IS_ERR(cathandle->u.chd.chd_next_log)) {
553                 /* declare next plain llog */
554                 down_write(&cathandle->lgh_lock);
555                 if (cathandle->u.chd.chd_next_log == NULL ||
556                     IS_ERR(cathandle->u.chd.chd_next_log)) {
557                         rc = llog_open(env, cathandle->lgh_ctxt, &loghandle,
558                                        NULL, NULL, LLOG_OPEN_NEW);
559                         if (rc == 0) {
560                                 cathandle->u.chd.chd_next_log = loghandle;
561                                 list_add_tail(&loghandle->u.phd.phd_entry,
562                                               &cathandle->u.chd.chd_head);
563                         }
564                 }
565                 up_write(&cathandle->lgh_lock);
566         }
567         if (rc)
568                 GOTO(out, rc);
569
570         lirec->lid_hdr.lrh_len = sizeof(*lirec);
571
572         if (!llog_exist(cathandle->u.chd.chd_current_log)) {
573                 if (dt_object_remote(cathandle->lgh_obj)) {
574                         /* For remote operation, if we put the llog object
575                          * creation in the current transaction, then the
576                          * llog object will not be created on the remote
577                          * target until the transaction stop, if other
578                          * operations start before the transaction stop,
579                          * and use the same llog object, will be dependent
580                          * on the success of this transaction. So let's
581                          * create the llog object synchronously here to
582                          * remove the dependency. */
583 create_again:
584                         down_read_nested(&cathandle->lgh_lock, LLOGH_CAT);
585                         loghandle = cathandle->u.chd.chd_current_log;
586                         down_write_nested(&loghandle->lgh_lock, LLOGH_LOG);
587                         if (cathandle->lgh_stale) {
588                                 up_write(&loghandle->lgh_lock);
589                                 up_read(&cathandle->lgh_lock);
590                                 GOTO(out, rc = -EIO);
591                         }
592                         if (!llog_exist(loghandle)) {
593                                 rc = llog_cat_new_log(env, cathandle, loghandle,
594                                                       NULL);
595                                 if (rc == -ESTALE)
596                                         cathandle->lgh_stale = 1;
597                         }
598                         up_write(&loghandle->lgh_lock);
599                         up_read(&cathandle->lgh_lock);
600                         if (rc == -ESTALE) {
601                                 rc = llog_cat_update_header(env, cathandle);
602                                 if (rc != 0)
603                                         GOTO(out, rc);
604                                 goto create_again;
605                         } else if (rc < 0) {
606                                 GOTO(out, rc);
607                         }
608                 } else {
609                         rc = llog_declare_create(env,
610                                         cathandle->u.chd.chd_current_log, th);
611                         if (rc)
612                                 GOTO(out, rc);
613                         llog_declare_write_rec(env, cathandle,
614                                                &lirec->lid_hdr, -1, th);
615                 }
616         }
617
618 write_again:
619         /* declare records in the llogs */
620         rc = llog_declare_write_rec(env, cathandle->u.chd.chd_current_log,
621                                     rec, -1, th);
622         if (rc == -ESTALE) {
623                 down_write(&cathandle->lgh_lock);
624                 if (cathandle->lgh_stale) {
625                         up_write(&cathandle->lgh_lock);
626                         GOTO(out, rc = -EIO);
627                 }
628
629                 cathandle->lgh_stale = 1;
630                 up_write(&cathandle->lgh_lock);
631                 rc = llog_cat_update_header(env, cathandle);
632                 if (rc != 0)
633                         GOTO(out, rc);
634                 goto write_again;
635         } else if (rc < 0) {
636                 GOTO(out, rc);
637         }
638
639         next = cathandle->u.chd.chd_next_log;
640         if (!IS_ERR_OR_NULL(next)) {
641                 if (!llog_exist(next)) {
642                         if (dt_object_remote(cathandle->lgh_obj)) {
643                                 /* For remote operation, if we put the llog
644                                  * object creation in the current transaction,
645                                  * then the llog object will not be created on
646                                  * the remote target until the transaction stop,
647                                  * if other operations start before the
648                                  * transaction stop, and use the same llog
649                                  * object, will be dependent on the success of
650                                  * this transaction. So let's create the llog
651                                  * object synchronously here to remove the
652                                  * dependency. */
653                                 down_write_nested(&cathandle->lgh_lock,
654                                                  LLOGH_CAT);
655                                 next = cathandle->u.chd.chd_next_log;
656                                 if (IS_ERR_OR_NULL(next)) {
657                                         /* Sigh, another thread just tried,
658                                          * let's fail as well */
659                                         up_write(&cathandle->lgh_lock);
660                                         if (next == NULL)
661                                                 rc = -EIO;
662                                         else
663                                                 rc = PTR_ERR(next);
664                                         GOTO(out, rc);
665                                 }
666
667                                 down_write_nested(&next->lgh_lock, LLOGH_LOG);
668                                 if (!llog_exist(next)) {
669                                         rc = llog_cat_new_log(env, cathandle,
670                                                               next, NULL);
671                                         if (rc < 0)
672                                                 cathandle->u.chd.chd_next_log =
673                                                                 ERR_PTR(rc);
674                                 }
675                                 up_write(&next->lgh_lock);
676                                 up_write(&cathandle->lgh_lock);
677                                 if (rc < 0)
678                                         GOTO(out, rc);
679                         } else {
680                                 rc = llog_declare_create(env, next, th);
681                                 llog_declare_write_rec(env, cathandle,
682                                                 &lirec->lid_hdr, -1, th);
683                         }
684                 }
685                 /* XXX: we hope for declarations made for existing llog
686                  *      this might be not correct with some backends
687                  *      where declarations are expected against specific
688                  *      object like ZFS with full debugging enabled */
689                 /*llog_declare_write_rec(env, next, rec, -1, th);*/
690         }
691 out:
692         RETURN(rc);
693 }
694 EXPORT_SYMBOL(llog_cat_declare_add_rec);
695
696 int llog_cat_add(const struct lu_env *env, struct llog_handle *cathandle,
697                  struct llog_rec_hdr *rec, struct llog_cookie *reccookie)
698 {
699         struct llog_ctxt        *ctxt;
700         struct dt_device        *dt;
701         struct thandle          *th = NULL;
702         int                      rc;
703
704         ctxt = cathandle->lgh_ctxt;
705         LASSERT(ctxt);
706         LASSERT(ctxt->loc_exp);
707
708         LASSERT(cathandle->lgh_obj != NULL);
709         dt = lu2dt_dev(cathandle->lgh_obj->do_lu.lo_dev);
710
711         th = dt_trans_create(env, dt);
712         if (IS_ERR(th))
713                 RETURN(PTR_ERR(th));
714
715         rc = llog_cat_declare_add_rec(env, cathandle, rec, th);
716         if (rc)
717                 GOTO(out_trans, rc);
718
719         rc = dt_trans_start_local(env, dt, th);
720         if (rc)
721                 GOTO(out_trans, rc);
722         rc = llog_cat_add_rec(env, cathandle, rec, reccookie, th);
723 out_trans:
724         dt_trans_stop(env, dt, th);
725         RETURN(rc);
726 }
727 EXPORT_SYMBOL(llog_cat_add);
728
729 /* For each cookie in the cookie array, we clear the log in-use bit and either:
730  * - the log is empty, so mark it free in the catalog header and delete it
731  * - the log is not empty, just write out the log header
732  *
733  * The cookies may be in different log files, so we need to get new logs
734  * each time.
735  *
736  * Assumes caller has already pushed us into the kernel context.
737  */
738 int llog_cat_cancel_records(const struct lu_env *env,
739                             struct llog_handle *cathandle, int count,
740                             struct llog_cookie *cookies)
741 {
742         int i, index, rc = 0, failed = 0;
743
744         ENTRY;
745
746         for (i = 0; i < count; i++, cookies++) {
747                 struct llog_handle *loghandle;
748                 struct llog_logid *lgl = &cookies->lgc_lgl;
749                 int  lrc;
750
751                 rc = llog_cat_id2handle(env, cathandle, &loghandle, lgl);
752                 if (rc) {
753                         CDEBUG(D_HA, "%s: cannot find llog for handle "DFID":%x"
754                                ": rc = %d\n",
755                                cathandle->lgh_ctxt->loc_obd->obd_name,
756                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, rc);
757                         failed++;
758                         continue;
759                 }
760
761                 if ((cathandle->lgh_ctxt->loc_flags &
762                      LLOG_CTXT_FLAG_NORMAL_FID) && !llog_exist(loghandle)) {
763                         /* For update log, some of loghandles of cathandle
764                          * might not exist because remote llog creation might
765                          * be failed, so let's skip the record cancellation
766                          * for these non-exist llogs.
767                          */
768                         lrc = -ENOENT;
769                         CDEBUG(D_HA, "%s: llog "DFID":%x does not exist"
770                                ": rc = %d\n",
771                                cathandle->lgh_ctxt->loc_obd->obd_name,
772                                PFID(&lgl->lgl_oi.oi_fid), lgl->lgl_ogen, lrc);
773                         failed++;
774                         if (rc == 0)
775                                 rc = lrc;
776                         continue;
777                 }
778
779                 lrc = llog_cancel_rec(env, loghandle, cookies->lgc_index);
780                 if (lrc == LLOG_DEL_PLAIN) { /* log has been destroyed */
781                         index = loghandle->u.phd.phd_cookie.lgc_index;
782                         lrc = llog_cat_cleanup(env, cathandle, loghandle,
783                                                index);
784                         if (rc == 0)
785                                 rc = lrc;
786                 } else if (lrc == -ENOENT) {
787                         if (rc == 0) /* ENOENT shouldn't rewrite any error */
788                                 rc = lrc;
789                 } else if (lrc < 0) {
790                         failed++;
791                         if (rc == 0)
792                                 rc = lrc;
793                 }
794                 llog_handle_put(loghandle);
795         }
796         if (rc)
797                 CERROR("%s: fail to cancel %d of %d llog-records: rc = %d\n",
798                        cathandle->lgh_ctxt->loc_obd->obd_name, failed, count,
799                        rc);
800
801         RETURN(rc);
802 }
803 EXPORT_SYMBOL(llog_cat_cancel_records);
804
805 static int llog_cat_process_cb(const struct lu_env *env,
806                                struct llog_handle *cat_llh,
807                                struct llog_rec_hdr *rec, void *data)
808 {
809         struct llog_process_data *d = data;
810         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
811         struct llog_handle *llh;
812         struct llog_log_hdr *hdr;
813         int rc;
814
815         ENTRY;
816         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
817                 CERROR("invalid record in catalog\n");
818                 RETURN(-EINVAL);
819         }
820         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
821                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
822                rec->lrh_index, PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
823
824         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
825         if (rc) {
826                 CERROR("%s: cannot find handle for llog "DFID": rc = %d\n",
827                        cat_llh->lgh_ctxt->loc_obd->obd_name,
828                        PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
829                 if (rc == -ENOENT || rc == -ESTALE) {
830                         /* After a server crash, a stub of index
831                          * record in catlog could be kept, because
832                          * plain log destroy + catlog index record
833                          * deletion are not atomic. So we end up with
834                          * an index but no actual record. Destroy the
835                          * index and move on. */
836                         rc = llog_cat_cleanup(env, cat_llh, NULL,
837                                               rec->lrh_index);
838                 }
839
840                 RETURN(rc);
841         }
842
843         /* clean old empty llogs, do not consider current llog in use */
844         /* ignore remote (lgh_obj=NULL) llogs */
845         hdr = llh->lgh_hdr;
846         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
847             hdr->llh_count == 1 && cat_llh->lgh_obj != NULL &&
848             llh != cat_llh->u.chd.chd_current_log) {
849                 rc = llog_destroy(env, llh);
850                 if (rc)
851                         CERROR("%s: fail to destroy empty log: rc = %d\n",
852                                llh->lgh_ctxt->loc_obd->obd_name, rc);
853                 GOTO(out, rc = LLOG_DEL_PLAIN);
854         }
855
856         if (rec->lrh_index < d->lpd_startcat) {
857                 /* Skip processing of the logs until startcat */
858                 rc = 0;
859         } else if (d->lpd_startidx > 0) {
860                 struct llog_process_cat_data cd;
861
862                 cd.lpcd_first_idx = d->lpd_startidx;
863                 cd.lpcd_last_idx = 0;
864                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
865                                           &cd, false);
866                 /* Continue processing the next log from idx 0 */
867                 d->lpd_startidx = 0;
868         } else {
869                 rc = llog_process_or_fork(env, llh, d->lpd_cb, d->lpd_data,
870                                           NULL, false);
871         }
872
873 out:
874         /* The empty plain log was destroyed while processing */
875         if (rc == LLOG_DEL_PLAIN)
876                 rc = llog_cat_cleanup(env, cat_llh, llh,
877                                       llh->u.phd.phd_cookie.lgc_index);
878         llog_handle_put(llh);
879
880         RETURN(rc);
881 }
882
883 int llog_cat_process_or_fork(const struct lu_env *env,
884                              struct llog_handle *cat_llh, llog_cb_t cat_cb,
885                              llog_cb_t cb, void *data, int startcat,
886                              int startidx, bool fork)
887 {
888         struct llog_process_data d;
889         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
890         int rc;
891         ENTRY;
892
893         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
894         d.lpd_data = data;
895         d.lpd_cb = cb;
896         d.lpd_startcat = startcat;
897         d.lpd_startidx = startidx;
898
899         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
900             llh->llh_count > 1) {
901                 struct llog_process_cat_data cd;
902
903                 CWARN("%s: catlog "DFID" crosses index zero\n",
904                       cat_llh->lgh_ctxt->loc_obd->obd_name,
905                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
906
907                 cd.lpcd_first_idx = llh->llh_cat_idx;
908                 cd.lpcd_last_idx = 0;
909                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
910                                           &d, &cd, fork);
911                 if (rc != 0)
912                         RETURN(rc);
913
914                 cd.lpcd_first_idx = 0;
915                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
916                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
917                                           &d, &cd, fork);
918         } else {
919                 rc = llog_process_or_fork(env, cat_llh, cat_cb,
920                                           &d, NULL, fork);
921         }
922
923         RETURN(rc);
924 }
925
926 int llog_cat_process(const struct lu_env *env, struct llog_handle *cat_llh,
927                      llog_cb_t cb, void *data, int startcat, int startidx)
928 {
929         return llog_cat_process_or_fork(env, cat_llh, llog_cat_process_cb,
930                                         cb, data, startcat, startidx, false);
931 }
932 EXPORT_SYMBOL(llog_cat_process);
933
934 static int llog_cat_size_cb(const struct lu_env *env,
935                              struct llog_handle *cat_llh,
936                              struct llog_rec_hdr *rec, void *data)
937 {
938         struct llog_process_data *d = data;
939         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
940         struct llog_handle *llh;
941         int rc;
942         __u64 *cum_size = d->lpd_data;
943         __u64 size;
944
945         ENTRY;
946         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
947                 CERROR("%s: invalid record in catalog, rc = %d\n",
948                        cat_llh->lgh_ctxt->loc_obd->obd_name, -EINVAL);
949                 RETURN(-EINVAL);
950         }
951         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
952                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
953                rec->lrh_index, PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
954
955         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
956         if (rc) {
957                 CWARN("%s: cannot find handle for llog "DFID": rc = %d\n",
958                       cat_llh->lgh_ctxt->loc_obd->obd_name,
959                       PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
960                 RETURN(0);
961         }
962         size = llog_size(env, llh);
963         *cum_size += size;
964
965         CDEBUG(D_INFO, "Add llog entry "DFID" size %llu\n",
966                PFID(&llh->lgh_id.lgl_oi.oi_fid), size);
967
968         llog_handle_put(llh);
969
970         RETURN(0);
971
972 }
973
974 __u64 llog_cat_size(const struct lu_env *env, struct llog_handle *cat_llh)
975 {
976         __u64 size = llog_size(env, cat_llh);
977
978         llog_cat_process_or_fork(env, cat_llh, llog_cat_size_cb,
979                                  NULL, &size, 0, 0, false);
980
981         return size;
982 }
983 EXPORT_SYMBOL(llog_cat_size);
984
985 static int llog_cat_reverse_process_cb(const struct lu_env *env,
986                                        struct llog_handle *cat_llh,
987                                        struct llog_rec_hdr *rec, void *data)
988 {
989         struct llog_process_data *d = data;
990         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
991         struct llog_handle *llh;
992         struct llog_log_hdr *hdr;
993         int rc;
994
995         if (le32_to_cpu(rec->lrh_type) != LLOG_LOGID_MAGIC) {
996                 CERROR("invalid record in catalog\n");
997                 RETURN(-EINVAL);
998         }
999         CDEBUG(D_HA, "processing log "DFID":%x at index %u of catalog "
1000                DFID"\n", PFID(&lir->lid_id.lgl_oi.oi_fid), lir->lid_id.lgl_ogen,
1001                le32_to_cpu(rec->lrh_index),
1002                PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1003
1004         rc = llog_cat_id2handle(env, cat_llh, &llh, &lir->lid_id);
1005         if (rc) {
1006                 CERROR("%s: cannot find handle for llog "DFID": rc = %d\n",
1007                        cat_llh->lgh_ctxt->loc_obd->obd_name,
1008                        PFID(&lir->lid_id.lgl_oi.oi_fid), rc);
1009                 if (rc == -ENOENT || rc == -ESTALE) {
1010                         /* After a server crash, a stub of index
1011                          * record in catlog could be kept, because
1012                          * plain log destroy + catlog index record
1013                          * deletion are not atomic. So we end up with
1014                          * an index but no actual record. Destroy the
1015                          * index and move on. */
1016                         rc = llog_cat_cleanup(env, cat_llh, NULL,
1017                                               rec->lrh_index);
1018                 }
1019
1020                 RETURN(rc);
1021         }
1022
1023         /* clean old empty llogs, do not consider current llog in use */
1024         hdr = llh->lgh_hdr;
1025         if ((hdr->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
1026             hdr->llh_count == 1 &&
1027             llh != cat_llh->u.chd.chd_current_log) {
1028                 rc = llog_destroy(env, llh);
1029                 if (rc)
1030                         CERROR("%s: fail to destroy empty log: rc = %d\n",
1031                                llh->lgh_ctxt->loc_obd->obd_name, rc);
1032                 GOTO(out, rc = LLOG_DEL_PLAIN);
1033         }
1034
1035         rc = llog_reverse_process(env, llh, d->lpd_cb, d->lpd_data, NULL);
1036
1037 out:
1038         /* The empty plain was destroyed while processing */
1039         if (rc == LLOG_DEL_PLAIN)
1040                 rc = llog_cat_cleanup(env, cat_llh, llh,
1041                                       llh->u.phd.phd_cookie.lgc_index);
1042
1043         llog_handle_put(llh);
1044         RETURN(rc);
1045 }
1046
1047 int llog_cat_reverse_process(const struct lu_env *env,
1048                              struct llog_handle *cat_llh,
1049                              llog_cb_t cb, void *data)
1050 {
1051         struct llog_process_data d;
1052         struct llog_process_cat_data cd;
1053         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
1054         int rc;
1055         ENTRY;
1056
1057         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
1058         d.lpd_data = data;
1059         d.lpd_cb = cb;
1060
1061         if (llh->llh_cat_idx >= cat_llh->lgh_last_idx &&
1062             llh->llh_count > 1) {
1063                 CWARN("%s: catalog "DFID" crosses index zero\n",
1064                       cat_llh->lgh_ctxt->loc_obd->obd_name,
1065                       PFID(&cat_llh->lgh_id.lgl_oi.oi_fid));
1066
1067                 cd.lpcd_first_idx = 0;
1068                 cd.lpcd_last_idx = cat_llh->lgh_last_idx;
1069                 rc = llog_reverse_process(env, cat_llh,
1070                                           llog_cat_reverse_process_cb,
1071                                           &d, &cd);
1072                 if (rc != 0)
1073                         RETURN(rc);
1074
1075                 cd.lpcd_first_idx = le32_to_cpu(llh->llh_cat_idx);
1076                 cd.lpcd_last_idx = 0;
1077                 rc = llog_reverse_process(env, cat_llh,
1078                                           llog_cat_reverse_process_cb,
1079                                           &d, &cd);
1080         } else {
1081                 rc = llog_reverse_process(env, cat_llh,
1082                                           llog_cat_reverse_process_cb,
1083                                           &d, NULL);
1084         }
1085
1086         RETURN(rc);
1087 }
1088 EXPORT_SYMBOL(llog_cat_reverse_process);
1089
1090 static int llog_cat_set_first_idx(struct llog_handle *cathandle, int idx)
1091 {
1092         struct llog_log_hdr *llh = cathandle->lgh_hdr;
1093         int bitmap_size;
1094
1095         ENTRY;
1096
1097         bitmap_size = LLOG_HDR_BITMAP_SIZE(llh);
1098         /*
1099          * The llh_cat_idx equals to the first used index minus 1
1100          * so if we canceled the first index then llh_cat_idx
1101          * must be renewed.
1102          */
1103         if (llh->llh_cat_idx == (idx - 1)) {
1104                 llh->llh_cat_idx = idx;
1105
1106                 while (idx != cathandle->lgh_last_idx) {
1107                         idx = (idx + 1) % bitmap_size;
1108                         if (!ext2_test_bit(idx, LLOG_HDR_BITMAP(llh))) {
1109                                 /* update llh_cat_idx for each unset bit,
1110                                  * expecting the next one is set */
1111                                 llh->llh_cat_idx = idx;
1112                         } else if (idx == 0) {
1113                                 /* skip header bit */
1114                                 llh->llh_cat_idx = 0;
1115                                 continue;
1116                         } else {
1117                                 /* the first index is found */
1118                                 break;
1119                         }
1120                 }
1121
1122                 CDEBUG(D_RPCTRACE, "catlog "DFID" first idx %u, last_idx %u\n",
1123                        PFID(&cathandle->lgh_id.lgl_oi.oi_fid),
1124                        llh->llh_cat_idx, cathandle->lgh_last_idx);
1125         }
1126
1127         RETURN(0);
1128 }
1129
1130 /* Cleanup deleted plain llog traces from catalog */
1131 int llog_cat_cleanup(const struct lu_env *env, struct llog_handle *cathandle,
1132                      struct llog_handle *loghandle, int index)
1133 {
1134         int rc;
1135
1136         LASSERT(index);
1137         if (loghandle != NULL) {
1138                 /* remove destroyed llog from catalog list and
1139                  * chd_current_log variable */
1140                 down_write(&cathandle->lgh_lock);
1141                 if (cathandle->u.chd.chd_current_log == loghandle)
1142                         cathandle->u.chd.chd_current_log = NULL;
1143                 list_del_init(&loghandle->u.phd.phd_entry);
1144                 up_write(&cathandle->lgh_lock);
1145                 LASSERT(index == loghandle->u.phd.phd_cookie.lgc_index);
1146                 /* llog was opened and keep in a list, close it now */
1147                 llog_close(env, loghandle);
1148         }
1149
1150         /* do not attempt to cleanup on-disk llog if on client side */
1151         if (cathandle->lgh_obj == NULL)
1152                 return 0;
1153
1154         /* remove plain llog entry from catalog by index */
1155         llog_cat_set_first_idx(cathandle, index);
1156         rc = llog_cancel_rec(env, cathandle, index);
1157         if (rc == 0)
1158                 CDEBUG(D_HA, "cancel plain log at index %u of catalog "DFID"\n",
1159                        index, PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
1160         return rc;
1161 }