]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb2/free.c
tdb2: allow multiple chain locks.
[ccan] / ccan / tdb2 / free.c
index fb45c80c7aecc65e6156180c3dc61e4db2750d17..e2b9cb18c5043a98a173eca52fecab4c22ad3dae 100644 (file)
@@ -59,7 +59,7 @@ tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable)
        return tdb_read_off(tdb, ftable + offsetof(struct tdb_freetable,next));
 }
 
-int tdb_ftable_init(struct tdb_context *tdb)
+enum TDB_ERROR tdb_ftable_init(struct tdb_context *tdb)
 {
        /* Use reservoir sampling algorithm to select a free list at random. */
        unsigned int rnd, max = 0, count = 0;
@@ -69,8 +69,9 @@ int tdb_ftable_init(struct tdb_context *tdb)
        tdb->ftable = 0;
 
        while (off) {
-               if (off == TDB_OFF_ERR)
-                       return -1;
+               if (TDB_OFF_IS_ERR(off)) {
+                       return off;
+               }
 
                rnd = random();
                if (rnd >= max) {
@@ -82,7 +83,7 @@ int tdb_ftable_init(struct tdb_context *tdb)
                off = next_ftable(tdb, off);
                count++;
        }
-       return 0;
+       return TDB_SUCCESS;
 }
 
 /* Offset of a given bucket. */
@@ -92,7 +93,7 @@ tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
                + bucket * sizeof(tdb_off_t);
 }
 
-/* Returns free_buckets + 1, or list number to search. */
+/* Returns free_buckets + 1, or list number to search, or -ve error. */
 static tdb_off_t find_free_head(struct tdb_context *tdb,
                                tdb_off_t ftable_off,
                                tdb_off_t bucket)
@@ -102,125 +103,173 @@ static tdb_off_t find_free_head(struct tdb_context *tdb,
                                    bucket, TDB_FREE_BUCKETS);
 }
 
-/* Remove from free bucket. */
-static int remove_from_list(struct tdb_context *tdb,
-                           tdb_off_t b_off, tdb_off_t r_off,
-                           const struct tdb_free_record *r)
+static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
 {
-       tdb_off_t off;
-       enum TDB_ERROR ecode;
-
-       /* Front of list? */
-       if (frec_prev(r) == 0) {
-               off = b_off;
-       } else {
-               off = frec_prev(r) + offsetof(struct tdb_free_record, next);
+#ifdef CCAN_TDB2_DEBUG
+       tdb_off_t off, prev = 0, first;
+       struct tdb_free_record r;
+
+       first = off = tdb_read_off(tdb, b_off);
+       while (off != 0) {
+               tdb_read_convert(tdb, off, &r, sizeof(r));
+               if (frec_magic(&r) != TDB_FREE_MAGIC)
+                       abort();
+               if (prev && frec_prev(&r) != prev)
+                       abort();
+               prev = off;
+               off = r.next;
        }
 
-#ifdef CCAN_TDB2_DEBUG
-       if (tdb_read_off(tdb, off) != r_off) {
-               tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
-                        "remove_from_list: %llu bad prev in list %llu",
-                        (long long)r_off, (long long)b_off);
-               return -1;
+       if (first) {
+               tdb_read_convert(tdb, first, &r, sizeof(r));
+               if (frec_prev(&r) != prev)
+                       abort();
        }
 #endif
+}
 
-       /* r->prev->next = r->next */
-       ecode = tdb_write_off(tdb, off, r->next);
-       if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
-       }
+/* Remove from free bucket. */
+static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
+                                      tdb_off_t b_off, tdb_off_t r_off,
+                                      const struct tdb_free_record *r)
+{
+       tdb_off_t off, prev_next, head;
+       enum TDB_ERROR ecode;
 
-       if (r->next != 0) {
-               off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
-               /* r->next->prev = r->prev */
+       /* Is this only element in list?  Zero out bucket, and we're done. */
+       if (frec_prev(r) == r_off)
+               return tdb_write_off(tdb, b_off, 0);
 
+       /* off = &r->prev->next */
+       off = frec_prev(r) + offsetof(struct tdb_free_record, next);
+
+       /* Get prev->next */
+       prev_next = tdb_read_off(tdb, off);
+       if (TDB_OFF_IS_ERR(prev_next))
+               return prev_next;
+
+       /* If prev->next == 0, we were head: update bucket to point to next. */
+       if (prev_next == 0) {
 #ifdef CCAN_TDB2_DEBUG
-               if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
-                       tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
-                                  "remove_from_list: %llu bad list %llu",
-                                  (long long)r_off, (long long)b_off);
-                       return -1;
+               if (tdb_read_off(tdb, b_off) != r_off) {
+                       return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                         "remove_from_list:"
+                                         " %llu head %llu on list %llu",
+                                         (long long)r_off,
+                                         (long long)tdb_read_off(tdb, b_off),
+                                         (long long)b_off);
                }
 #endif
+               ecode = tdb_write_off(tdb, b_off, r->next);
+               if (ecode != TDB_SUCCESS)
+                       return ecode;
+       } else {
+               /* r->prev->next = r->next */
+               ecode = tdb_write_off(tdb, off, r->next);
+               if (ecode != TDB_SUCCESS)
+                       return ecode;
+       }
 
-               ecode = tdb_write_off(tdb, off, r->magic_and_prev);
-               if (ecode != TDB_SUCCESS) {
-                       tdb->ecode = ecode;
-                       return -1;
-               }
+       /* If we were the tail, off = &head->prev. */
+       if (r->next == 0) {
+               head = tdb_read_off(tdb, b_off);
+               if (TDB_OFF_IS_ERR(head))
+                       return head;
+               off = head + offsetof(struct tdb_free_record, magic_and_prev);
+       } else {
+               /* off = &r->next->prev */
+               off = r->next + offsetof(struct tdb_free_record,
+                                        magic_and_prev);
        }
-       return 0;
+
+#ifdef CCAN_TDB2_DEBUG
+       /* *off == r */
+       if ((tdb_read_off(tdb, off) & TDB_OFF_MASK) != r_off) {
+               return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                 "remove_from_list:"
+                                 " %llu bad prev in list %llu",
+                                 (long long)r_off, (long long)b_off);
+       }
+#endif
+       /* r->next->prev = r->prev */
+       return tdb_write_off(tdb, off, r->magic_and_prev);
 }
 
 /* Enqueue in this free bucket. */
-static int enqueue_in_free(struct tdb_context *tdb,
-                          tdb_off_t b_off,
-                          tdb_off_t off,
-                          tdb_len_t len)
+static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
+                                     tdb_off_t b_off,
+                                     tdb_off_t off,
+                                     tdb_len_t len)
 {
        struct tdb_free_record new;
        enum TDB_ERROR ecode;
+       tdb_off_t prev;
        uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
 
        /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
        new.ftable_and_len = ((uint64_t)tdb->ftable << (64 - TDB_OFF_UPPER_STEAL))
                | len;
-       /* prev = 0. */
-       new.magic_and_prev = magic;
 
        /* new->next = head. */
        new.next = tdb_read_off(tdb, b_off);
-       if (new.next == TDB_OFF_ERR)
-               return -1;
+       if (TDB_OFF_IS_ERR(new.next)) {
+               return new.next;
+       }
 
-       if (new.next) {
-#ifdef CCAN_TDB2_DEBUG
-               if (tdb_read_off(tdb,
-                                new.next + offsetof(struct tdb_free_record,
-                                                    magic_and_prev))
-                   != magic) {
-                       tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
-                                  "enqueue_in_free: %llu bad head"
-                                  " prev %llu",
-                                  (long long)new.next, (long long)b_off);
-                       return -1;
+       /* First element?  Prev points to ourselves. */
+       if (!new.next) {
+               new.magic_and_prev = (magic | off);
+       } else {
+               /* new->prev = next->prev */
+               prev = tdb_read_off(tdb,
+                                   new.next + offsetof(struct tdb_free_record,
+                                                       magic_and_prev));
+               new.magic_and_prev = prev;
+               if (frec_magic(&new) != TDB_FREE_MAGIC) {
+                       return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                         "enqueue_in_free: %llu bad head"
+                                         " prev %llu",
+                                         (long long)new.next,
+                                         (long long)prev);
                }
-#endif
                /* next->prev = new. */
                ecode = tdb_write_off(tdb, new.next
                                      + offsetof(struct tdb_free_record,
                                                 magic_and_prev),
                                      off | magic);
                if (ecode != TDB_SUCCESS) {
-                       tdb->ecode = ecode;
-                       return -1;
+                       return ecode;
                }
+
+#ifdef CCAN_TDB2_DEBUG
+               prev = tdb_read_off(tdb, frec_prev(&new)
+                                   + offsetof(struct tdb_free_record, next));
+               if (prev != 0) {
+                       return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                         "enqueue_in_free:"
+                                         " %llu bad tail next ptr %llu",
+                                         (long long)frec_prev(&new)
+                                         + offsetof(struct tdb_free_record,
+                                                    next),
+                                         (long long)prev);
+               }
+#endif
        }
        /* head = new */
        ecode = tdb_write_off(tdb, b_off, off);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
+               return ecode;
        }
 
-       ecode = tdb_write_convert(tdb, off, &new, sizeof(new));
-       if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
-       }
-       return 0;
+       return tdb_write_convert(tdb, off, &new, sizeof(new));
 }
 
 /* List need not be locked. */
-int add_free_record(struct tdb_context *tdb,
-                   tdb_off_t off, tdb_len_t len_with_header)
+enum TDB_ERROR add_free_record(struct tdb_context *tdb,
+                              tdb_off_t off, tdb_len_t len_with_header)
 {
        tdb_off_t b_off;
        tdb_len_t len;
-       int ret;
        enum TDB_ERROR ecode;
 
        assert(len_with_header >= sizeof(struct tdb_free_record));
@@ -230,13 +279,13 @@ int add_free_record(struct tdb_context *tdb,
        b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
        ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
+               return ecode;
        }
 
-       ret = enqueue_in_free(tdb, b_off, off, len);
+       ecode = enqueue_in_free(tdb, b_off, off, len);
+       check_list(tdb, b_off);
        tdb_unlock_free_bucket(tdb, b_off);
-       return ret;
+       return ecode;
 }
 
 static size_t adjust_size(size_t keylen, size_t datalen)
@@ -275,30 +324,37 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
                return tdb->ftable_off;
 
        off = first_ftable(tdb);
-       for (i = 0; i < ftable; i++)
+       for (i = 0; i < ftable; i++) {
+               if (TDB_OFF_IS_ERR(off)) {
+                       break;
+               }
                off = next_ftable(tdb, off);
+       }
        return off;
 }
 
 /* Note: we unlock the current bucket if we coalesce or fail. */
-static int coalesce(struct tdb_context *tdb,
-                   tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
+static tdb_bool_err coalesce(struct tdb_context *tdb,
+                            tdb_off_t off, tdb_off_t b_off,
+                            tdb_len_t data_len)
 {
        tdb_off_t end;
        struct tdb_free_record rec;
        enum TDB_ERROR ecode;
 
-       add_stat(tdb, alloc_coalesce_tried, 1);
+       tdb->stats.alloc_coalesce_tried++;
        end = off + sizeof(struct tdb_used_record) + data_len;
 
-       while (end < tdb->map_size) {
+       while (end < tdb->file->map_size) {
                const struct tdb_free_record *r;
                tdb_off_t nb_off;
                unsigned ftable, bucket;
 
                r = tdb_access_read(tdb, end, sizeof(*r), true);
-               if (!r)
+               if (TDB_PTR_IS_ERR(r)) {
+                       ecode = TDB_PTR_ERR(r);
                        goto err;
+               }
 
                if (frec_magic(r) != TDB_FREE_MAGIC
                    || frec_ftable(r) == TDB_FTABLE_NONE) {
@@ -308,67 +364,76 @@ static int coalesce(struct tdb_context *tdb,
 
                ftable = frec_ftable(r);
                bucket = size_to_bucket(frec_len(r));
-               nb_off = bucket_off(ftable_offset(tdb, ftable), bucket);
+               nb_off = ftable_offset(tdb, ftable);
+               if (TDB_OFF_IS_ERR(nb_off)) {
+                       tdb_access_release(tdb, r);
+                       ecode = nb_off;
+                       goto err;
+               }
+               nb_off = bucket_off(nb_off, bucket);
                tdb_access_release(tdb, r);
 
                /* We may be violating lock order here, so best effort. */
                if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
                    != TDB_SUCCESS) {
-                       add_stat(tdb, alloc_coalesce_lockfail, 1);
+                       tdb->stats.alloc_coalesce_lockfail++;
                        break;
                }
 
                /* Now we have lock, re-check. */
                ecode = tdb_read_convert(tdb, end, &rec, sizeof(rec));
                if (ecode != TDB_SUCCESS) {
-                       tdb->ecode = ecode;
                        tdb_unlock_free_bucket(tdb, nb_off);
                        goto err;
                }
 
                if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
-                       add_stat(tdb, alloc_coalesce_race, 1);
+                       tdb->stats.alloc_coalesce_race++;
                        tdb_unlock_free_bucket(tdb, nb_off);
                        break;
                }
 
                if (unlikely(frec_ftable(&rec) != ftable)
                    || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
-                       add_stat(tdb, alloc_coalesce_race, 1);
+                       tdb->stats.alloc_coalesce_race++;
                        tdb_unlock_free_bucket(tdb, nb_off);
                        break;
                }
 
-               if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
+               ecode = remove_from_list(tdb, nb_off, end, &rec);
+               check_list(tdb, nb_off);
+               if (ecode != TDB_SUCCESS) {
                        tdb_unlock_free_bucket(tdb, nb_off);
                        goto err;
                }
 
                end += sizeof(struct tdb_used_record) + frec_len(&rec);
                tdb_unlock_free_bucket(tdb, nb_off);
-               add_stat(tdb, alloc_coalesce_num_merged, 1);
+               tdb->stats.alloc_coalesce_num_merged++;
        }
 
        /* Didn't find any adjacent free? */
        if (end == off + sizeof(struct tdb_used_record) + data_len)
-               return 0;
+               return false;
 
        /* OK, expand initial record */
        ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
                goto err;
        }
 
        if (frec_len(&rec) != data_len) {
-               tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
-                          "coalesce: expected data len %zu not %zu",
-                          (size_t)data_len, (size_t)frec_len(&rec));
+               ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                  "coalesce: expected data len %zu not %zu",
+                                  (size_t)data_len, (size_t)frec_len(&rec));
                goto err;
        }
 
-       if (remove_from_list(tdb, b_off, off, &rec) == -1)
+       ecode = remove_from_list(tdb, b_off, off, &rec);
+       check_list(tdb, b_off);
+       if (ecode != TDB_SUCCESS) {
                goto err;
+       }
 
        /* We have to drop this to avoid deadlocks, so make sure record
         * doesn't get coalesced by someone else! */
@@ -378,21 +443,22 @@ static int coalesce(struct tdb_context *tdb,
                                                  ftable_and_len),
                              rec.ftable_and_len);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
                goto err;
        }
 
-       add_stat(tdb, alloc_coalesce_succeeded, 1);
+       tdb->stats.alloc_coalesce_succeeded++;
        tdb_unlock_free_bucket(tdb, b_off);
 
-       if (add_free_record(tdb, off, end - off) == -1)
-               return -1;
-       return 1;
+       ecode = add_free_record(tdb, off, end - off);
+       if (ecode != TDB_SUCCESS) {
+               return ecode;
+       }
+       return true;
 
 err:
        /* To unify error paths, we *always* unlock bucket on error. */
        tdb_unlock_free_bucket(tdb, b_off);
-       return -1;
+       return ecode;
 }
 
 /* We need size bytes to put our key and data in. */
@@ -410,7 +476,7 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
        size_t size = adjust_size(keylen, datalen);
        enum TDB_ERROR ecode;
 
-       add_stat(tdb, allocs, 1);
+       tdb->stats.allocs++;
 again:
        b_off = bucket_off(ftable_off, bucket);
 
@@ -418,8 +484,7 @@ again:
        /* Lock this bucket. */
        ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return TDB_OFF_ERR;
+               return ecode;
        }
 
        best.ftable_and_len = -1ULL;
@@ -434,23 +499,30 @@ again:
        /* Walk the list to see if any are large enough, getting less fussy
         * as we go. */
        off = tdb_read_off(tdb, b_off);
-       if (unlikely(off == TDB_OFF_ERR))
+       if (TDB_OFF_IS_ERR(off)) {
+               ecode = off;
                goto unlock_err;
+       }
 
        while (off) {
                const struct tdb_free_record *r;
                tdb_len_t len;
                tdb_off_t next;
+               int coal;
 
                r = tdb_access_read(tdb, off, sizeof(*r), true);
-               if (!r)
+               if (TDB_PTR_IS_ERR(r)) {
+                       ecode = TDB_PTR_ERR(r);
                        goto unlock_err;
+               }
 
                if (frec_magic(r) != TDB_FREE_MAGIC) {
+                       ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+                                          "lock_and_alloc:"
+                                          " %llu non-free 0x%llx",
+                                          (long long)off,
+                                          (long long)r->magic_and_prev);
                        tdb_access_release(tdb, r);
-                       tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
-                                "lock_and_alloc: %llu non-free 0x%llx",
-                                (long long)off, (long long)r->magic_and_prev);
                        goto unlock_err;
                }
 
@@ -459,7 +531,7 @@ again:
                        best = *r;
                }
 
-               if (frec_len(&best) < size * multiplier && best_off) {
+               if (frec_len(&best) <= size * multiplier && best_off) {
                        tdb_access_release(tdb, r);
                        break;
                }
@@ -471,14 +543,15 @@ again:
                tdb_access_release(tdb, r);
 
                /* Since we're going slow anyway, try coalescing here. */
-               switch (coalesce(tdb, off, b_off, len)) {
-               case -1:
-                       /* This has already unlocked on error. */
-                       return -1;
-               case 1:
+               coal = coalesce(tdb, off, b_off, len);
+               if (coal == 1) {
                        /* This has unlocked list, restart. */
                        goto again;
                }
+               if (coal < 0) {
+                       /* This has already unlocked on error. */
+                       return coal;
+               }
                off = next;
        }
 
@@ -488,8 +561,11 @@ again:
                size_t leftover;
 
                /* We're happy with this size: take it. */
-               if (remove_from_list(tdb, b_off, best_off, &best) != 0)
+               ecode = remove_from_list(tdb, b_off, best_off, &best);
+               check_list(tdb, b_off);
+               if (ecode != TDB_SUCCESS) {
                        goto unlock_err;
+               }
 
                leftover = record_leftover(keylen, datalen, want_extra,
                                           frec_len(&best));
@@ -497,25 +573,37 @@ again:
                assert(keylen + datalen + leftover <= frec_len(&best));
                /* We need to mark non-free before we drop lock, otherwise
                 * coalesce() could try to merge it! */
-               if (set_header(tdb, &rec, magic, keylen, datalen,
-                              frec_len(&best) - leftover, hashlow) != 0)
+               ecode = set_header(tdb, &rec, magic, keylen, datalen,
+                                  frec_len(&best) - leftover, hashlow);
+               if (ecode != TDB_SUCCESS) {
                        goto unlock_err;
+               }
 
                ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
                if (ecode != TDB_SUCCESS) {
-                       tdb->ecode = ecode;
                        goto unlock_err;
                }
 
+               /* For futureproofing, we put a 0 in any unused space. */
+               if (rec_extra_padding(&rec)) {
+                       ecode = tdb->methods->twrite(tdb, best_off + sizeof(rec)
+                                                    + keylen + datalen, "", 1);
+                       if (ecode != TDB_SUCCESS) {
+                               goto unlock_err;
+                       }
+               }
+
                /* Bucket of leftover will be <= current bucket, so nested
                 * locking is allowed. */
                if (leftover) {
-                       add_stat(tdb, alloc_leftover, 1);
-                       if (add_free_record(tdb,
-                                           best_off + sizeof(rec)
-                                           + frec_len(&best) - leftover,
-                                           leftover))
-                               best_off = TDB_OFF_ERR;
+                       tdb->stats.alloc_leftover++;
+                       ecode = add_free_record(tdb,
+                                               best_off + sizeof(rec)
+                                               + frec_len(&best) - leftover,
+                                               leftover);
+                       if (ecode != TDB_SUCCESS) {
+                               best_off = ecode;
+                       }
                }
                tdb_unlock_free_bucket(tdb, b_off);
 
@@ -527,16 +615,16 @@ again:
 
 unlock_err:
        tdb_unlock_free_bucket(tdb, b_off);
-       return TDB_OFF_ERR;
+       return ecode;
 }
 
-/* Get a free block from current free list, or 0 if none. */
+/* Get a free block from current free list, or 0 if none, -ve on error. */
 static tdb_off_t get_free(struct tdb_context *tdb,
                          size_t keylen, size_t datalen, bool want_extra,
                          unsigned magic, unsigned hashlow)
 {
        tdb_off_t off, ftable_off;
-       unsigned start_b, b, ftable;
+       tdb_off_t start_b, b, ftable;
        bool wrapped = false;
 
        /* If they are growing, add 50% to get to higher bucket. */
@@ -557,13 +645,13 @@ static tdb_off_t get_free(struct tdb_context *tdb,
                        off = lock_and_alloc(tdb, ftable_off,
                                             b, keylen, datalen, want_extra,
                                             magic, hashlow);
-                       if (off == TDB_OFF_ERR)
-                               return TDB_OFF_ERR;
+                       if (TDB_OFF_IS_ERR(off))
+                               return off;
                        if (off != 0) {
                                if (b == start_b)
-                                       add_stat(tdb, alloc_bucket_exact, 1);
+                                       tdb->stats.alloc_bucket_exact++;
                                if (b == TDB_FREE_BUCKETS - 1)
-                                       add_stat(tdb, alloc_bucket_max, 1);
+                                       tdb->stats.alloc_bucket_max++;
                                /* Worked?  Stay using this list. */
                                tdb->ftable_off = ftable_off;
                                tdb->ftable = ftable;
@@ -572,13 +660,23 @@ static tdb_off_t get_free(struct tdb_context *tdb,
                        /* Didn't work.  Try next bucket. */
                }
 
+               if (TDB_OFF_IS_ERR(b)) {
+                       return b;
+               }
+
                /* Hmm, try next table. */
                ftable_off = next_ftable(tdb, ftable_off);
+               if (TDB_OFF_IS_ERR(ftable_off)) {
+                       return ftable_off;
+               }
                ftable++;
 
                if (ftable_off == 0) {
                        wrapped = true;
                        ftable_off = first_ftable(tdb);
+                       if (TDB_OFF_IS_ERR(ftable_off)) {
+                               return ftable_off;
+                       }
                        ftable = 0;
                }
        }
@@ -586,10 +684,10 @@ static tdb_off_t get_free(struct tdb_context *tdb,
        return 0;
 }
 
-int set_header(struct tdb_context *tdb,
-              struct tdb_used_record *rec,
-              unsigned magic, uint64_t keylen, uint64_t datalen,
-              uint64_t actuallen, unsigned hashlow)
+enum TDB_ERROR set_header(struct tdb_context *tdb,
+                         struct tdb_used_record *rec,
+                         unsigned magic, uint64_t keylen, uint64_t datalen,
+                         uint64_t actuallen, unsigned hashlow)
 {
        uint64_t keybits = (fls64(keylen) + 1) / 2;
 
@@ -604,17 +702,16 @@ int set_header(struct tdb_context *tdb,
        if (rec_key_length(rec) != keylen
            || rec_data_length(rec) != datalen
            || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
-               tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
-                        "Could not encode k=%llu,d=%llu,a=%llu",
-                        (long long)keylen, (long long)datalen,
-                        (long long)actuallen);
-               return -1;
+               return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
+                                 "Could not encode k=%llu,d=%llu,a=%llu",
+                                 (long long)keylen, (long long)datalen,
+                                 (long long)actuallen);
        }
-       return 0;
+       return TDB_SUCCESS;
 }
 
 /* Expand the database. */
-static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
+static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
        uint64_t old_size;
        tdb_len_t wanted;
@@ -625,46 +722,43 @@ static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 
        /* Need to hold a hash lock to expand DB: transactions rely on it. */
        if (!(tdb->flags & TDB_NOLOCK)
-           && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
-               tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
-                          "tdb_expand: must hold lock during expand");
-               return -1;
+           && !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
+               return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
+                                 "tdb_expand: must hold lock during expand");
        }
 
        /* always make room for at least 100 more records, and at
            least 25% more space. */
-       if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
+       if (size * TDB_EXTENSION_FACTOR > tdb->file->map_size / 4)
                wanted = size * TDB_EXTENSION_FACTOR;
        else
-               wanted = tdb->map_size / 4;
+               wanted = tdb->file->map_size / 4;
        wanted = adjust_size(0, wanted);
 
        /* Only one person can expand file at a time. */
        ecode = tdb_lock_expand(tdb, F_WRLCK);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
+               return ecode;
        }
 
        /* Someone else may have expanded the file, so retry. */
-       old_size = tdb->map_size;
-       tdb->methods->oob(tdb, tdb->map_size + 1, true);
-       if (tdb->map_size != old_size) {
+       old_size = tdb->file->map_size;
+       tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
+       if (tdb->file->map_size != old_size) {
                tdb_unlock_expand(tdb, F_WRLCK);
-               return 0;
+               return TDB_SUCCESS;
        }
 
        ecode = tdb->methods->expand_file(tdb, wanted);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
                tdb_unlock_expand(tdb, F_WRLCK);
-               return -1;
+               return ecode;
        }
 
        /* We need to drop this lock before adding free record. */
        tdb_unlock_expand(tdb, F_WRLCK);
 
-       add_stat(tdb, expands, 1);
+       tdb->stats.expands++;
        return add_free_record(tdb, old_size, wanted);
 }
 
@@ -678,12 +772,15 @@ tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
        assert(!tdb->direct_access);
 
        for (;;) {
+               enum TDB_ERROR ecode;
                off = get_free(tdb, keylen, datalen, growing, magic, hash);
                if (likely(off != 0))
                        break;
 
-               if (tdb_expand(tdb, adjust_size(keylen, datalen)))
-                       return TDB_OFF_ERR;
+               ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
+               if (ecode != TDB_SUCCESS) {
+                       return ecode;
+               }
        }
 
        return off;