2bb085a72e8ee1b1f243de701e4ddd3caf807a4a
[releases.git] / recovery.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
4  * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
5  */
6
7 #include <linux/module.h>
8 #include <linux/slab.h>
9 #include <linux/spinlock.h>
10 #include <linux/completion.h>
11 #include <linux/buffer_head.h>
12 #include <linux/gfs2_ondisk.h>
13 #include <linux/crc32.h>
14 #include <linux/crc32c.h>
15 #include <linux/ktime.h>
16
17 #include "gfs2.h"
18 #include "incore.h"
19 #include "bmap.h"
20 #include "glock.h"
21 #include "glops.h"
22 #include "log.h"
23 #include "lops.h"
24 #include "meta_io.h"
25 #include "recovery.h"
26 #include "super.h"
27 #include "util.h"
28 #include "dir.h"
29
30 struct workqueue_struct *gfs_recovery_wq;
31
32 int gfs2_replay_read_block(struct gfs2_jdesc *jd, unsigned int blk,
33                            struct buffer_head **bh)
34 {
35         struct gfs2_inode *ip = GFS2_I(jd->jd_inode);
36         struct gfs2_glock *gl = ip->i_gl;
37         u64 dblock;
38         u32 extlen;
39         int error;
40
41         extlen = 32;
42         error = gfs2_get_extent(&ip->i_inode, blk, &dblock, &extlen);
43         if (error)
44                 return error;
45         if (!dblock) {
46                 gfs2_consist_inode(ip);
47                 return -EIO;
48         }
49
50         *bh = gfs2_meta_ra(gl, dblock, extlen);
51
52         return error;
53 }
54
55 int gfs2_revoke_add(struct gfs2_jdesc *jd, u64 blkno, unsigned int where)
56 {
57         struct list_head *head = &jd->jd_revoke_list;
58         struct gfs2_revoke_replay *rr = NULL, *iter;
59
60         list_for_each_entry(iter, head, rr_list) {
61                 if (iter->rr_blkno == blkno) {
62                         rr = iter;
63                         break;
64                 }
65         }
66
67         if (rr) {
68                 rr->rr_where = where;
69                 return 0;
70         }
71
72         rr = kmalloc(sizeof(struct gfs2_revoke_replay), GFP_NOFS);
73         if (!rr)
74                 return -ENOMEM;
75
76         rr->rr_blkno = blkno;
77         rr->rr_where = where;
78         list_add(&rr->rr_list, head);
79
80         return 1;
81 }
82
83 int gfs2_revoke_check(struct gfs2_jdesc *jd, u64 blkno, unsigned int where)
84 {
85         struct gfs2_revoke_replay *rr = NULL, *iter;
86         int wrap, a, b, revoke;
87
88         list_for_each_entry(iter, &jd->jd_revoke_list, rr_list) {
89                 if (iter->rr_blkno == blkno) {
90                         rr = iter;
91                         break;
92                 }
93         }
94
95         if (!rr)
96                 return 0;
97
98         wrap = (rr->rr_where < jd->jd_replay_tail);
99         a = (jd->jd_replay_tail < where);
100         b = (where < rr->rr_where);
101         revoke = (wrap) ? (a || b) : (a && b);
102
103         return revoke;
104 }
105
106 void gfs2_revoke_clean(struct gfs2_jdesc *jd)
107 {
108         struct list_head *head = &jd->jd_revoke_list;
109         struct gfs2_revoke_replay *rr;
110
111         while (!list_empty(head)) {
112                 rr = list_first_entry(head, struct gfs2_revoke_replay, rr_list);
113                 list_del(&rr->rr_list);
114                 kfree(rr);
115         }
116 }
117
118 int __get_log_header(struct gfs2_sbd *sdp, const struct gfs2_log_header *lh,
119                      unsigned int blkno, struct gfs2_log_header_host *head)
120 {
121         u32 hash, crc;
122
123         if (lh->lh_header.mh_magic != cpu_to_be32(GFS2_MAGIC) ||
124             lh->lh_header.mh_type != cpu_to_be32(GFS2_METATYPE_LH) ||
125             (blkno && be32_to_cpu(lh->lh_blkno) != blkno))
126                 return 1;
127
128         hash = crc32(~0, lh, LH_V1_SIZE - 4);
129         hash = ~crc32_le_shift(hash, 4); /* assume lh_hash is zero */
130
131         if (be32_to_cpu(lh->lh_hash) != hash)
132                 return 1;
133
134         crc = crc32c(~0, (void *)lh + LH_V1_SIZE + 4,
135                      sdp->sd_sb.sb_bsize - LH_V1_SIZE - 4);
136
137         if ((lh->lh_crc != 0 && be32_to_cpu(lh->lh_crc) != crc))
138                 return 1;
139
140         head->lh_sequence = be64_to_cpu(lh->lh_sequence);
141         head->lh_flags = be32_to_cpu(lh->lh_flags);
142         head->lh_tail = be32_to_cpu(lh->lh_tail);
143         head->lh_blkno = be32_to_cpu(lh->lh_blkno);
144
145         head->lh_local_total = be64_to_cpu(lh->lh_local_total);
146         head->lh_local_free = be64_to_cpu(lh->lh_local_free);
147         head->lh_local_dinodes = be64_to_cpu(lh->lh_local_dinodes);
148
149         return 0;
150 }
151 /**
152  * get_log_header - read the log header for a given segment
153  * @jd: the journal
154  * @blk: the block to look at
155  * @head: the log header to return
156  *
157  * Read the log header for a given segement in a given journal.  Do a few
158  * sanity checks on it.
159  *
160  * Returns: 0 on success,
161  *          1 if the header was invalid or incomplete,
162  *          errno on error
163  */
164
165 static int get_log_header(struct gfs2_jdesc *jd, unsigned int blk,
166                           struct gfs2_log_header_host *head)
167 {
168         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
169         struct buffer_head *bh;
170         int error;
171
172         error = gfs2_replay_read_block(jd, blk, &bh);
173         if (error)
174                 return error;
175
176         error = __get_log_header(sdp, (const struct gfs2_log_header *)bh->b_data,
177                                  blk, head);
178         brelse(bh);
179
180         return error;
181 }
182
183 /**
184  * foreach_descriptor - go through the active part of the log
185  * @jd: the journal
186  * @start: the first log header in the active region
187  * @end: the last log header (don't process the contents of this entry))
188  * @pass: iteration number (foreach_descriptor() is called in a for() loop)
189  *
190  * Call a given function once for every log descriptor in the active
191  * portion of the log.
192  *
193  * Returns: errno
194  */
195
196 static int foreach_descriptor(struct gfs2_jdesc *jd, u32 start,
197                               unsigned int end, int pass)
198 {
199         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
200         struct buffer_head *bh;
201         struct gfs2_log_descriptor *ld;
202         int error = 0;
203         u32 length;
204         __be64 *ptr;
205         unsigned int offset = sizeof(struct gfs2_log_descriptor);
206         offset += sizeof(__be64) - 1;
207         offset &= ~(sizeof(__be64) - 1);
208
209         while (start != end) {
210                 error = gfs2_replay_read_block(jd, start, &bh);
211                 if (error)
212                         return error;
213                 if (gfs2_meta_check(sdp, bh)) {
214                         brelse(bh);
215                         return -EIO;
216                 }
217                 ld = (struct gfs2_log_descriptor *)bh->b_data;
218                 length = be32_to_cpu(ld->ld_length);
219
220                 if (be32_to_cpu(ld->ld_header.mh_type) == GFS2_METATYPE_LH) {
221                         struct gfs2_log_header_host lh;
222                         error = get_log_header(jd, start, &lh);
223                         if (!error) {
224                                 gfs2_replay_incr_blk(jd, &start);
225                                 brelse(bh);
226                                 continue;
227                         }
228                         if (error == 1) {
229                                 gfs2_consist_inode(GFS2_I(jd->jd_inode));
230                                 error = -EIO;
231                         }
232                         brelse(bh);
233                         return error;
234                 } else if (gfs2_metatype_check(sdp, bh, GFS2_METATYPE_LD)) {
235                         brelse(bh);
236                         return -EIO;
237                 }
238                 ptr = (__be64 *)(bh->b_data + offset);
239                 error = lops_scan_elements(jd, start, ld, ptr, pass);
240                 if (error) {
241                         brelse(bh);
242                         return error;
243                 }
244
245                 while (length--)
246                         gfs2_replay_incr_blk(jd, &start);
247
248                 brelse(bh);
249         }
250
251         return 0;
252 }
253
254 /**
255  * clean_journal - mark a dirty journal as being clean
256  * @jd: the journal
257  * @head: the head journal to start from
258  *
259  * Returns: errno
260  */
261
262 static void clean_journal(struct gfs2_jdesc *jd,
263                           struct gfs2_log_header_host *head)
264 {
265         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
266         u32 lblock = head->lh_blkno;
267
268         gfs2_replay_incr_blk(jd, &lblock);
269         gfs2_write_log_header(sdp, jd, head->lh_sequence + 1, 0, lblock,
270                               GFS2_LOG_HEAD_UNMOUNT | GFS2_LOG_HEAD_RECOVERY,
271                               REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC);
272         if (jd->jd_jid == sdp->sd_lockstruct.ls_jid) {
273                 sdp->sd_log_flush_head = lblock;
274                 gfs2_log_incr_head(sdp);
275         }
276 }
277
278
279 static void gfs2_recovery_done(struct gfs2_sbd *sdp, unsigned int jid,
280                                unsigned int message)
281 {
282         char env_jid[20];
283         char env_status[20];
284         char *envp[] = { env_jid, env_status, NULL };
285         struct lm_lockstruct *ls = &sdp->sd_lockstruct;
286
287         ls->ls_recover_jid_done = jid;
288         ls->ls_recover_jid_status = message;
289         sprintf(env_jid, "JID=%u", jid);
290         sprintf(env_status, "RECOVERY=%s",
291                 message == LM_RD_SUCCESS ? "Done" : "Failed");
292         kobject_uevent_env(&sdp->sd_kobj, KOBJ_CHANGE, envp);
293
294         if (sdp->sd_lockstruct.ls_ops->lm_recovery_result)
295                 sdp->sd_lockstruct.ls_ops->lm_recovery_result(sdp, jid, message);
296 }
297
298 /**
299  * update_statfs_inode - Update the master statfs inode or zero out the local
300  *                       statfs inode for a given journal.
301  * @jd: The journal
302  * @head: If NULL, @inode is the local statfs inode and we need to zero it out.
303  *        Otherwise, it @head contains the statfs change info that needs to be
304  *        synced to the master statfs inode (pointed to by @inode).
305  * @inode: statfs inode to update.
306  */
307 static int update_statfs_inode(struct gfs2_jdesc *jd,
308                                struct gfs2_log_header_host *head,
309                                struct inode *inode)
310 {
311         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
312         struct gfs2_inode *ip;
313         struct buffer_head *bh;
314         struct gfs2_statfs_change_host sc;
315         int error = 0;
316
317         BUG_ON(!inode);
318         ip = GFS2_I(inode);
319
320         error = gfs2_meta_inode_buffer(ip, &bh);
321         if (error)
322                 goto out;
323
324         spin_lock(&sdp->sd_statfs_spin);
325
326         if (head) { /* Update the master statfs inode */
327                 gfs2_statfs_change_in(&sc, bh->b_data + sizeof(struct gfs2_dinode));
328                 sc.sc_total += head->lh_local_total;
329                 sc.sc_free += head->lh_local_free;
330                 sc.sc_dinodes += head->lh_local_dinodes;
331                 gfs2_statfs_change_out(&sc, bh->b_data + sizeof(struct gfs2_dinode));
332
333                 fs_info(sdp, "jid=%u: Updated master statfs Total:%lld, "
334                         "Free:%lld, Dinodes:%lld after change "
335                         "[%+lld,%+lld,%+lld]\n", jd->jd_jid, sc.sc_total,
336                         sc.sc_free, sc.sc_dinodes, head->lh_local_total,
337                         head->lh_local_free, head->lh_local_dinodes);
338         } else { /* Zero out the local statfs inode */
339                 memset(bh->b_data + sizeof(struct gfs2_dinode), 0,
340                        sizeof(struct gfs2_statfs_change));
341                 /* If it's our own journal, reset any in-memory changes too */
342                 if (jd->jd_jid == sdp->sd_lockstruct.ls_jid) {
343                         memset(&sdp->sd_statfs_local, 0,
344                                sizeof(struct gfs2_statfs_change_host));
345                 }
346         }
347         spin_unlock(&sdp->sd_statfs_spin);
348
349         mark_buffer_dirty(bh);
350         brelse(bh);
351         gfs2_inode_metasync(ip->i_gl);
352
353 out:
354         return error;
355 }
356
357 /**
358  * recover_local_statfs - Update the master and local statfs changes for this
359  *                        journal.
360  *
361  * Previously, statfs updates would be read in from the local statfs inode and
362  * synced to the master statfs inode during recovery.
363  *
364  * We now use the statfs updates in the journal head to update the master statfs
365  * inode instead of reading in from the local statfs inode. To preserve backward
366  * compatibility with kernels that can't do this, we still need to keep the
367  * local statfs inode up to date by writing changes to it. At some point in the
368  * future, we can do away with the local statfs inodes altogether and keep the
369  * statfs changes solely in the journal.
370  *
371  * @jd: the journal
372  * @head: the journal head
373  *
374  * Returns: errno
375  */
376 static void recover_local_statfs(struct gfs2_jdesc *jd,
377                                  struct gfs2_log_header_host *head)
378 {
379         int error;
380         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
381
382         if (!head->lh_local_total && !head->lh_local_free
383             && !head->lh_local_dinodes) /* No change */
384                 goto zero_local;
385
386          /* First update the master statfs inode with the changes we
387           * found in the journal. */
388         error = update_statfs_inode(jd, head, sdp->sd_statfs_inode);
389         if (error)
390                 goto out;
391
392 zero_local:
393         /* Zero out the local statfs inode so any changes in there
394          * are not re-recovered. */
395         error = update_statfs_inode(jd, NULL,
396                                     find_local_statfs_inode(sdp, jd->jd_jid));
397 out:
398         return;
399 }
400
401 void gfs2_recover_func(struct work_struct *work)
402 {
403         struct gfs2_jdesc *jd = container_of(work, struct gfs2_jdesc, jd_work);
404         struct gfs2_inode *ip = GFS2_I(jd->jd_inode);
405         struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
406         struct gfs2_log_header_host head;
407         struct gfs2_holder j_gh, ji_gh, thaw_gh;
408         ktime_t t_start, t_jlck, t_jhd, t_tlck, t_rep;
409         int ro = 0;
410         unsigned int pass;
411         int error = 0;
412         int jlocked = 0;
413
414         if (gfs2_withdrawn(sdp)) {
415                 fs_err(sdp, "jid=%u: Recovery not attempted due to withdraw.\n",
416                        jd->jd_jid);
417                 goto fail;
418         }
419         t_start = ktime_get();
420         if (sdp->sd_args.ar_spectator)
421                 goto fail;
422         if (jd->jd_jid != sdp->sd_lockstruct.ls_jid) {
423                 fs_info(sdp, "jid=%u: Trying to acquire journal lock...\n",
424                         jd->jd_jid);
425                 jlocked = 1;
426                 /* Acquire the journal lock so we can do recovery */
427
428                 error = gfs2_glock_nq_num(sdp, jd->jd_jid, &gfs2_journal_glops,
429                                           LM_ST_EXCLUSIVE,
430                                           LM_FLAG_NOEXP | LM_FLAG_TRY | GL_NOCACHE,
431                                           &j_gh);
432                 switch (error) {
433                 case 0:
434                         break;
435
436                 case GLR_TRYFAILED:
437                         fs_info(sdp, "jid=%u: Busy\n", jd->jd_jid);
438                         error = 0;
439                         goto fail;
440
441                 default:
442                         goto fail;
443                 }
444
445                 error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED,
446                                            LM_FLAG_NOEXP | GL_NOCACHE, &ji_gh);
447                 if (error)
448                         goto fail_gunlock_j;
449         } else {
450                 fs_info(sdp, "jid=%u, already locked for use\n", jd->jd_jid);
451         }
452
453         t_jlck = ktime_get();
454         fs_info(sdp, "jid=%u: Looking at journal...\n", jd->jd_jid);
455
456         error = gfs2_jdesc_check(jd);
457         if (error)
458                 goto fail_gunlock_ji;
459
460         error = gfs2_find_jhead(jd, &head, true);
461         if (error)
462                 goto fail_gunlock_ji;
463         t_jhd = ktime_get();
464         fs_info(sdp, "jid=%u: Journal head lookup took %lldms\n", jd->jd_jid,
465                 ktime_ms_delta(t_jhd, t_jlck));
466
467         if (!(head.lh_flags & GFS2_LOG_HEAD_UNMOUNT)) {
468                 fs_info(sdp, "jid=%u: Acquiring the transaction lock...\n",
469                         jd->jd_jid);
470
471                 /* Acquire a shared hold on the freeze lock */
472
473                 error = gfs2_freeze_lock(sdp, &thaw_gh, LM_FLAG_PRIORITY);
474                 if (error)
475                         goto fail_gunlock_ji;
476
477                 if (test_bit(SDF_RORECOVERY, &sdp->sd_flags)) {
478                         ro = 1;
479                 } else if (test_bit(SDF_JOURNAL_CHECKED, &sdp->sd_flags)) {
480                         if (!test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))
481                                 ro = 1;
482                 } else {
483                         if (sb_rdonly(sdp->sd_vfs)) {
484                                 /* check if device itself is read-only */
485                                 ro = bdev_read_only(sdp->sd_vfs->s_bdev);
486                                 if (!ro) {
487                                         fs_info(sdp, "recovery required on "
488                                                 "read-only filesystem.\n");
489                                         fs_info(sdp, "write access will be "
490                                                 "enabled during recovery.\n");
491                                 }
492                         }
493                 }
494
495                 if (ro) {
496                         fs_warn(sdp, "jid=%u: Can't replay: read-only block "
497                                 "device\n", jd->jd_jid);
498                         error = -EROFS;
499                         goto fail_gunlock_thaw;
500                 }
501
502                 t_tlck = ktime_get();
503                 fs_info(sdp, "jid=%u: Replaying journal...0x%x to 0x%x\n",
504                         jd->jd_jid, head.lh_tail, head.lh_blkno);
505
506                 /* We take the sd_log_flush_lock here primarily to prevent log
507                  * flushes and simultaneous journal replays from stomping on
508                  * each other wrt jd_log_bio. */
509                 down_read(&sdp->sd_log_flush_lock);
510                 for (pass = 0; pass < 2; pass++) {
511                         lops_before_scan(jd, &head, pass);
512                         error = foreach_descriptor(jd, head.lh_tail,
513                                                    head.lh_blkno, pass);
514                         lops_after_scan(jd, error, pass);
515                         if (error) {
516                                 up_read(&sdp->sd_log_flush_lock);
517                                 goto fail_gunlock_thaw;
518                         }
519                 }
520
521                 recover_local_statfs(jd, &head);
522                 clean_journal(jd, &head);
523                 up_read(&sdp->sd_log_flush_lock);
524
525                 gfs2_freeze_unlock(&thaw_gh);
526                 t_rep = ktime_get();
527                 fs_info(sdp, "jid=%u: Journal replayed in %lldms [jlck:%lldms, "
528                         "jhead:%lldms, tlck:%lldms, replay:%lldms]\n",
529                         jd->jd_jid, ktime_ms_delta(t_rep, t_start),
530                         ktime_ms_delta(t_jlck, t_start),
531                         ktime_ms_delta(t_jhd, t_jlck),
532                         ktime_ms_delta(t_tlck, t_jhd),
533                         ktime_ms_delta(t_rep, t_tlck));
534         }
535
536         gfs2_recovery_done(sdp, jd->jd_jid, LM_RD_SUCCESS);
537
538         if (jlocked) {
539                 gfs2_glock_dq_uninit(&ji_gh);
540                 gfs2_glock_dq_uninit(&j_gh);
541         }
542
543         fs_info(sdp, "jid=%u: Done\n", jd->jd_jid);
544         goto done;
545
546 fail_gunlock_thaw:
547         gfs2_freeze_unlock(&thaw_gh);
548 fail_gunlock_ji:
549         if (jlocked) {
550                 gfs2_glock_dq_uninit(&ji_gh);
551 fail_gunlock_j:
552                 gfs2_glock_dq_uninit(&j_gh);
553         }
554
555         fs_info(sdp, "jid=%u: %s\n", jd->jd_jid, (error) ? "Failed" : "Done");
556 fail:
557         jd->jd_recover_error = error;
558         gfs2_recovery_done(sdp, jd->jd_jid, LM_RD_GAVEUP);
559 done:
560         clear_bit(JDF_RECOVERY, &jd->jd_flags);
561         smp_mb__after_atomic();
562         wake_up_bit(&jd->jd_flags, JDF_RECOVERY);
563 }
564
565 int gfs2_recover_journal(struct gfs2_jdesc *jd, bool wait)
566 {
567         int rv;
568
569         if (test_and_set_bit(JDF_RECOVERY, &jd->jd_flags))
570                 return -EBUSY;
571
572         /* we have JDF_RECOVERY, queue should always succeed */
573         rv = queue_work(gfs_recovery_wq, &jd->jd_work);
574         BUG_ON(!rv);
575
576         if (wait)
577                 wait_on_bit(&jd->jd_flags, JDF_RECOVERY,
578                             TASK_UNINTERRUPTIBLE);
579
580         return wait ? jd->jd_recover_error : 0;
581 }
582