bucket, TDB_FREE_BUCKETS);
}
+static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
+{
+#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;
+ }
+
+ if (first) {
+ tdb_read_convert(tdb, first, &r, sizeof(r));
+ if (frec_prev(&r) != prev)
+ abort();
+ }
+#endif
+}
+
/* 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;
+ tdb_off_t off, prev_next, head;
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
- if (tdb_read_off(tdb, off) != 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
+ /* 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);
- /* r->prev->next = r->next */
- ecode = tdb_write_off(tdb, off, r->next);
- if (ecode != TDB_SUCCESS) {
- return ecode;
- }
+ /* off = &r->prev->next */
+ off = frec_prev(r) + offsetof(struct tdb_free_record, next);
- if (r->next != 0) {
- off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
- /* r->next->prev = r->prev */
+ /* 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) {
+ if (tdb_read_off(tdb, b_off) != r_off) {
return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
"remove_from_list:"
- " %llu bad list %llu",
- (long long)r_off, (long long)b_off);
+ " %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, off, r->magic_and_prev);
- if (ecode != TDB_SUCCESS) {
+ 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;
- }
}
- return TDB_SUCCESS;
+
+ /* 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);
+ }
+
+#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. */
{
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);
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) {
+ /* 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)b_off);
+ (long long)prev);
}
-#endif
/* next->prev = new. */
ecode = tdb_write_off(tdb, new.next
+ offsetof(struct tdb_free_record,
if (ecode != TDB_SUCCESS) {
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);
/* List need not be locked. */
enum TDB_ERROR add_free_record(struct tdb_context *tdb,
- tdb_off_t off, tdb_len_t len_with_header)
+ tdb_off_t off, tdb_len_t len_with_header,
+ enum tdb_lock_flags waitflag)
{
tdb_off_t b_off;
tdb_len_t len;
len = len_with_header - sizeof(struct tdb_used_record);
b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
- ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
+ ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
if (ecode != TDB_SUCCESS) {
return ecode;
}
ecode = enqueue_in_free(tdb, b_off, off, len);
+ check_list(tdb, b_off);
tdb_unlock_free_bucket(tdb, b_off);
return ecode;
}
return off;
}
-/* Note: we unlock the current bucket if we coalesce or fail. */
-static tdb_bool_err coalesce(struct tdb_context *tdb,
- tdb_off_t off, tdb_off_t b_off,
- tdb_len_t data_len)
+/* Note: we unlock the current bucket if fail (-ve), or coalesce (-ve) and
+ * need to blatt either of the *protect records (which is set to an error). */
+static tdb_len_t coalesce(struct tdb_context *tdb,
+ tdb_off_t off, tdb_off_t b_off,
+ tdb_len_t data_len,
+ tdb_off_t *protect1,
+ tdb_off_t *protect2)
{
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;
/* 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;
}
}
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;
}
+ /* Did we just mess up a record you were hoping to use? */
+ if (end == *protect1 || end == *protect2)
+ *protect1 = TDB_ERR_NOEXIST;
+
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 false;
+ return 0;
+
+ /* Before we expand, check this isn't one you wanted protected? */
+ if (off == *protect1 || off == *protect2)
+ *protect1 = TDB_ERR_EXISTS;
/* OK, expand initial record */
ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
}
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! */
- rec.ftable_and_len = (TDB_FTABLE_NONE << (64 - TDB_OFF_UPPER_STEAL))
- | (end - off - sizeof(struct tdb_used_record));
- ecode = tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
- ftable_and_len),
- rec.ftable_and_len);
+ /* Try locking violation first... */
+ ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT);
if (ecode != TDB_SUCCESS) {
- goto err;
- }
+ /* Need to drop lock. Can't rely on anything stable. */
+ *protect1 = TDB_ERR_CORRUPT;
+
+ /* We have to drop this to avoid deadlocks, so make sure record
+ * doesn't get coalesced by someone else! */
+ rec.ftable_and_len = (TDB_FTABLE_NONE
+ << (64 - TDB_OFF_UPPER_STEAL))
+ | (end - off - sizeof(struct tdb_used_record));
+ ecode = tdb_write_off(tdb,
+ off + offsetof(struct tdb_free_record,
+ ftable_and_len),
+ rec.ftable_and_len);
+ if (ecode != TDB_SUCCESS) {
+ goto err;
+ }
- add_stat(tdb, alloc_coalesce_succeeded, 1);
- tdb_unlock_free_bucket(tdb, b_off);
+ tdb->stats.alloc_coalesce_succeeded++;
+ tdb_unlock_free_bucket(tdb, b_off);
- ecode = add_free_record(tdb, off, end - off);
- if (ecode != TDB_SUCCESS) {
- return ecode;
+ ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT);
+ if (ecode != TDB_SUCCESS) {
+ return ecode;
+ }
+ } else if (TDB_OFF_IS_ERR(*protect1)) {
+ /* For simplicity, we always drop lock if they can't continue */
+ tdb_unlock_free_bucket(tdb, b_off);
}
- return true;
+
+ /* Return usable length. */
+ return end - off - sizeof(struct tdb_used_record);
err:
/* To unify error paths, we *always* unlock bucket on error. */
tdb_off_t off, b_off,best_off;
struct tdb_free_record best = { 0 };
double multiplier;
+ bool coalesce_after_best = false; /* Damn GCC warning! */
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);
while (off) {
const struct tdb_free_record *r;
- tdb_len_t len;
+ tdb_len_t len, coal;
tdb_off_t next;
- int coal;
r = tdb_access_read(tdb, off, sizeof(*r), true);
if (TDB_PTR_IS_ERR(r)) {
}
if (frec_magic(r) != TDB_FREE_MAGIC) {
- tdb_access_release(tdb, r);
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);
goto unlock_err;
}
if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
best_off = off;
best = *r;
+ coalesce_after_best = false;
}
- if (frec_len(&best) < size * multiplier && best_off) {
+ if (frec_len(&best) <= size * multiplier && best_off) {
tdb_access_release(tdb, r);
break;
}
tdb_access_release(tdb, r);
/* Since we're going slow anyway, try coalescing here. */
- coal = coalesce(tdb, off, b_off, len);
- if (coal == 1) {
- /* This has unlocked list, restart. */
- goto again;
- }
- if (coal < 0) {
+ coal = coalesce(tdb, off, b_off, len, &best_off, &next);
+ if (TDB_OFF_IS_ERR(coal)) {
/* This has already unlocked on error. */
return coal;
}
+ if (TDB_OFF_IS_ERR(best_off)) {
+ /* This has unlocked list, restart. */
+ goto again;
+ }
+ if (coal > 0)
+ coalesce_after_best = true;
off = next;
}
struct tdb_used_record rec;
size_t leftover;
+ /* If we coalesced, we might have change prev/next ptrs. */
+ if (coalesce_after_best) {
+ ecode = tdb_read_convert(tdb, best_off, &best,
+ sizeof(best));
+ if (ecode != TDB_SUCCESS)
+ goto unlock_err;
+ }
+
/* We're happy with this size: take it. */
ecode = remove_from_list(tdb, b_off, best_off, &best);
+ check_list(tdb, b_off);
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);
+ tdb->stats.alloc_leftover++;
ecode = add_free_record(tdb,
best_off + sizeof(rec)
+ frec_len(&best) - leftover,
- leftover);
+ leftover, TDB_LOCK_WAIT);
if (ecode != TDB_SUCCESS) {
best_off = ecode;
}
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;
/* Expand the database. */
static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
{
- uint64_t old_size;
+ uint64_t old_size, rec_size, map_size;
tdb_len_t wanted;
enum TDB_ERROR ecode;
- /* We need room for the record header too. */
- wanted = sizeof(struct tdb_used_record) + size;
-
/* Need to hold a hash lock to expand DB: transactions rely on it. */
if (!(tdb->flags & TDB_NOLOCK)
&& !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
"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)
- wanted = size * TDB_EXTENSION_FACTOR;
- else
- wanted = tdb->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) {
}
/* 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 TDB_SUCCESS;
}
+ /* limit size in order to avoid using up huge amounts of memory for
+ * in memory tdbs if an oddball huge record creeps in */
+ if (size > 100 * 1024) {
+ rec_size = size * 2;
+ } else {
+ rec_size = size * 100;
+ }
+
+ /* always make room for at least rec_size more records, and at
+ least 25% more space. if the DB is smaller than 100MiB,
+ otherwise grow it by 10% only. */
+ if (old_size > 100 * 1024 * 1024) {
+ map_size = old_size / 10;
+ } else {
+ map_size = old_size / 4;
+ }
+
+ if (map_size > rec_size) {
+ wanted = map_size;
+ } else {
+ wanted = rec_size;
+ }
+
+ /* We need room for the record header too. */
+ wanted = adjust_size(0, sizeof(struct tdb_used_record) + wanted);
+
ecode = tdb->methods->expand_file(tdb, wanted);
if (ecode != TDB_SUCCESS) {
tdb_unlock_expand(tdb, F_WRLCK);
/* We need to drop this lock before adding free record. */
tdb_unlock_expand(tdb, F_WRLCK);
- add_stat(tdb, expands, 1);
- return add_free_record(tdb, old_size, wanted);
+ tdb->stats.expands++;
+ return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT);
}
/* This won't fail: it will expand the database if it has to. */
projects / ccan / blobdiff