fs/bcachefs/journal_seq_blacklist.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_iter.h"
   5 #include "eytzinger.h"
   6 #include "journal_seq_blacklist.h"
   7 #include "super-io.h"
   8
   9 /*
  10  * journal_seq_blacklist machinery:
  11  *
  12  * To guarantee order of btree updates after a crash, we need to detect when a
  13  * btree node entry (bset) is newer than the newest journal entry that was
  14  * successfully written, and ignore it - effectively ignoring any btree updates
  15  * that didn't make it into the journal.
  16  *
  17  * If we didn't do this, we might have two btree nodes, a and b, both with
  18  * updates that weren't written to the journal yet: if b was updated after a,
  19  * but b was flushed and not a - oops; on recovery we'll find that the updates
  20  * to b happened, but not the updates to a that happened before it.
  21  *
  22  * Ignoring bsets that are newer than the newest journal entry is always safe,
  23  * because everything they contain will also have been journalled - and must
  24  * still be present in the journal on disk until a journal entry has been
  25  * written _after_ that bset was written.
  26  *
  27  * To accomplish this, bsets record the newest journal sequence number they
  28  * contain updates for; then, on startup, the btree code queries the journal
  29  * code to ask "Is this sequence number newer than the newest journal entry? If
  30  * so, ignore it."
  31  *
  32  * When this happens, we must blacklist that journal sequence number: the
  33  * journal must not write any entries with that sequence number, and it must
  34  * record that it was blacklisted so that a) on recovery we don't think we have
  35  * missing journal entries and b) so that the btree code continues to ignore
  36  * that bset, until that btree node is rewritten.
  37  */
  38
  39 static unsigned sb_blacklist_u64s(unsigned nr)
  40 {
  41         struct bch_sb_field_journal_seq_blacklist *bl;
  42
  43         return (sizeof(*bl) + sizeof(bl->start[0]) * nr) / sizeof(u64);
  44 }
  45
  46 int bch2_journal_seq_blacklist_add(struct bch_fs *c, u64 start, u64 end)
  47 {
  48         struct bch_sb_field_journal_seq_blacklist *bl;
  49         unsigned i = 0, nr;
  50         int ret = 0;
  51
  52         mutex_lock(&c->sb_lock);
  53         bl = bch2_sb_field_get(c->disk_sb.sb, journal_seq_blacklist);
  54         nr = blacklist_nr_entries(bl);
  55
  56         while (i < nr) {
  57                 struct journal_seq_blacklist_entry *e =
  58                         bl->start + i;
  59
  60                 if (end < le64_to_cpu(e->start))
  61                         break;
  62
  63                 if (start > le64_to_cpu(e->end)) {
  64                         i++;
  65                         continue;
  66                 }
  67
  68                 /*
  69                  * Entry is contiguous or overlapping with new entry: merge it
  70                  * with new entry, and delete:
  71                  */
  72
  73                 start   = min(start,    le64_to_cpu(e->start));
  74                 end     = max(end,      le64_to_cpu(e->end));
  75                 array_remove_item(bl->start, nr, i);
  76         }
  77
  78         bl = bch2_sb_field_resize(&c->disk_sb, journal_seq_blacklist,
  79                                   sb_blacklist_u64s(nr + 1));
  80         if (!bl) {
  81                 ret = -BCH_ERR_ENOSPC_sb_journal_seq_blacklist;
  82                 goto out;
  83         }
  84
  85         array_insert_item(bl->start, nr, i, ((struct journal_seq_blacklist_entry) {
  86                 .start  = cpu_to_le64(start),
  87                 .end    = cpu_to_le64(end),
  88         }));
  89         c->disk_sb.sb->features[0] |= cpu_to_le64(1ULL << BCH_FEATURE_journal_seq_blacklist_v3);
  90
  91         ret = bch2_write_super(c);
  92 out:
  93         mutex_unlock(&c->sb_lock);
  94
  95         return ret ?: bch2_blacklist_table_initialize(c);
  96 }
  97
  98 static int journal_seq_blacklist_table_cmp(const void *_l, const void *_r)
  99 {
 100         const struct journal_seq_blacklist_table_entry *l = _l;
 101         const struct journal_seq_blacklist_table_entry *r = _r;
 102
 103         return cmp_int(l->start, r->start);
 104 }
 105
 106 bool bch2_journal_seq_is_blacklisted(struct bch_fs *c, u64 seq,
 107                                      bool dirty)
 108 {
 109         struct journal_seq_blacklist_table *t = c->journal_seq_blacklist_table;
 110         struct journal_seq_blacklist_table_entry search = { .start = seq };
 111         int idx;
 112
 113         if (!t)
 114                 return false;
 115
 116         idx = eytzinger0_find_le(t->entries, t->nr,
 117                                  sizeof(t->entries[0]),
 118                                  journal_seq_blacklist_table_cmp,
 119                                  &search);
 120         if (idx < 0)
 121                 return false;
 122
 123         BUG_ON(t->entries[idx].start > seq);
 124
 125         if (seq >= t->entries[idx].end)
 126                 return false;
 127
 128         if (dirty)
 129                 t->entries[idx].dirty = true;
 130         return true;
 131 }
 132
 133 int bch2_blacklist_table_initialize(struct bch_fs *c)
 134 {
 135         struct bch_sb_field_journal_seq_blacklist *bl =
 136                 bch2_sb_field_get(c->disk_sb.sb, journal_seq_blacklist);
 137         struct journal_seq_blacklist_table *t;
 138         unsigned i, nr = blacklist_nr_entries(bl);
 139
 140         if (!bl)
 141                 return 0;
 142
 143         t = kzalloc(struct_size(t, entries, nr), GFP_KERNEL);
 144         if (!t)
 145                 return -BCH_ERR_ENOMEM_blacklist_table_init;
 146
 147         t->nr = nr;
 148
 149         for (i = 0; i < nr; i++) {
 150                 t->entries[i].start     = le64_to_cpu(bl->start[i].start);
 151                 t->entries[i].end       = le64_to_cpu(bl->start[i].end);
 152         }
 153
 154         eytzinger0_sort(t->entries,
 155                         t->nr,
 156                         sizeof(t->entries[0]),
 157                         journal_seq_blacklist_table_cmp,
 158                         NULL);
 159
 160         kfree(c->journal_seq_blacklist_table);
 161         c->journal_seq_blacklist_table = t;
 162         return 0;
 163 }
 164
 165 static int bch2_sb_journal_seq_blacklist_validate(struct bch_sb *sb,
 166                                                   struct bch_sb_field *f,
 167                                                   struct printbuf *err)
 168 {
 169         struct bch_sb_field_journal_seq_blacklist *bl =
 170                 field_to_type(f, journal_seq_blacklist);
 171         unsigned i, nr = blacklist_nr_entries(bl);
 172
 173         for (i = 0; i < nr; i++) {
 174                 struct journal_seq_blacklist_entry *e = bl->start + i;
 175
 176                 if (le64_to_cpu(e->start) >=
 177                     le64_to_cpu(e->end)) {
 178                         prt_printf(err, "entry %u start >= end (%llu >= %llu)",
 179                                i, le64_to_cpu(e->start), le64_to_cpu(e->end));
 180                         return -BCH_ERR_invalid_sb_journal_seq_blacklist;
 181                 }
 182
 183                 if (i + 1 < nr &&
 184                     le64_to_cpu(e[0].end) >
 185                     le64_to_cpu(e[1].start)) {
 186                         prt_printf(err, "entry %u out of order with next entry (%llu > %llu)",
 187                                i + 1, le64_to_cpu(e[0].end), le64_to_cpu(e[1].start));
 188                         return -BCH_ERR_invalid_sb_journal_seq_blacklist;
 189                 }
 190         }
 191
 192         return 0;
 193 }
 194
 195 static void bch2_sb_journal_seq_blacklist_to_text(struct printbuf *out,
 196                                                   struct bch_sb *sb,
 197                                                   struct bch_sb_field *f)
 198 {
 199         struct bch_sb_field_journal_seq_blacklist *bl =
 200                 field_to_type(f, journal_seq_blacklist);
 201         struct journal_seq_blacklist_entry *i;
 202         unsigned nr = blacklist_nr_entries(bl);
 203
 204         for (i = bl->start; i < bl->start + nr; i++) {
 205                 if (i != bl->start)
 206                         prt_printf(out, " ");
 207
 208                 prt_printf(out, "%llu-%llu",
 209                        le64_to_cpu(i->start),
 210                        le64_to_cpu(i->end));
 211         }
 212         prt_newline(out);
 213 }
 214
 215 const struct bch_sb_field_ops bch_sb_field_ops_journal_seq_blacklist = {
 216         .validate       = bch2_sb_journal_seq_blacklist_validate,
 217         .to_text        = bch2_sb_journal_seq_blacklist_to_text
 218 };
 219
 220 void bch2_blacklist_entries_gc(struct work_struct *work)
 221 {
 222         struct bch_fs *c = container_of(work, struct bch_fs,
 223                                         journal_seq_blacklist_gc_work);
 224         struct journal_seq_blacklist_table *t;
 225         struct bch_sb_field_journal_seq_blacklist *bl;
 226         struct journal_seq_blacklist_entry *src, *dst;
 227         struct btree_trans *trans = bch2_trans_get(c);
 228         unsigned i, nr, new_nr;
 229         int ret;
 230
 231         for (i = 0; i < BTREE_ID_NR; i++) {
 232                 struct btree_iter iter;
 233                 struct btree *b;
 234
 235                 bch2_trans_node_iter_init(trans, &iter, i, POS_MIN,
 236                                           0, 0, BTREE_ITER_PREFETCH);
 237 retry:
 238                 bch2_trans_begin(trans);
 239
 240                 b = bch2_btree_iter_peek_node(&iter);
 241
 242                 while (!(ret = PTR_ERR_OR_ZERO(b)) &&
 243                        b &&
 244                        !test_bit(BCH_FS_stopping, &c->flags))
 245                         b = bch2_btree_iter_next_node(&iter);
 246
 247                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 248                         goto retry;
 249
 250                 bch2_trans_iter_exit(trans, &iter);
 251         }
 252
 253         bch2_trans_put(trans);
 254         if (ret)
 255                 return;
 256
 257         mutex_lock(&c->sb_lock);
 258         bl = bch2_sb_field_get(c->disk_sb.sb, journal_seq_blacklist);
 259         if (!bl)
 260                 goto out;
 261
 262         nr = blacklist_nr_entries(bl);
 263         dst = bl->start;
 264
 265         t = c->journal_seq_blacklist_table;
 266         BUG_ON(nr != t->nr);
 267
 268         for (src = bl->start, i = eytzinger0_first(t->nr);
 269              src < bl->start + nr;
 270              src++, i = eytzinger0_next(i, nr)) {
 271                 BUG_ON(t->entries[i].start      != le64_to_cpu(src->start));
 272                 BUG_ON(t->entries[i].end        != le64_to_cpu(src->end));
 273
 274                 if (t->entries[i].dirty)
 275                         *dst++ = *src;
 276         }
 277
 278         new_nr = dst - bl->start;
 279
 280         bch_info(c, "nr blacklist entries was %u, now %u", nr, new_nr);
 281
 282         if (new_nr != nr) {
 283                 bl = bch2_sb_field_resize(&c->disk_sb, journal_seq_blacklist,
 284                                 new_nr ? sb_blacklist_u64s(new_nr) : 0);
 285                 BUG_ON(new_nr && !bl);
 286
 287                 if (!new_nr)
 288                         c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_journal_seq_blacklist_v3));
 289
 290                 bch2_write_super(c);
 291         }
 292 out:
 293         mutex_unlock(&c->sb_lock);
 294 }