tdb2: minor optimization for set_header
[ccan] / ccan / tdb2 / free.c
index 83d916fc6986ceec757131b51b44a8c13eaa022a..e6e77bf616666deed3940b07707616485bd29812 100644 (file)
 */
 #include "private.h"
 #include <ccan/likely/likely.h>
+#include <ccan/ilog/ilog.h>
 #include <time.h>
 #include <assert.h>
 #include <limits.h>
 
-/* We have to be able to fit a free record here. */
-#define MIN_DATA_LEN   \
-       (sizeof(struct tdb_free_record) - sizeof(struct tdb_used_record))
-
-/* We have a series of free lists, each one covering a "zone" of the file.
- *
- * For each zone we have a series of per-size buckets, and a final bucket for
- * "too big".
- *
- * It's possible to move the free_list_head, but *only* under the allrecord
- * lock. */
-static tdb_off_t free_list_off(struct tdb_context *tdb, unsigned int list)
-{
-       return tdb->header.v.free_off + list * sizeof(tdb_off_t);
-}
-
-/* We're a library: playing with srandom() is unfriendly.  srandom_r
- * probably lacks portability.  We don't need very random here. */
-static unsigned int quick_random(struct tdb_context *tdb)
-{
-       return getpid() + time(NULL) + (unsigned long)tdb;
-}
-
-/* Start by using a random zone to spread the load. */
-uint64_t random_free_zone(struct tdb_context *tdb)
-{
-       /* num_zones might be out of date, but can only increase */
-       return quick_random(tdb) % tdb->header.v.num_zones;
-}
-
 static unsigned fls64(uint64_t val)
 {
-#if HAVE_BUILTIN_CLZL
-       if (val <= ULONG_MAX) {
-               /* This is significantly faster! */
-               return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
-       } else {
-#endif
-       uint64_t r = 64;
-
-       if (!val)
-               return 0;
-       if (!(val & 0xffffffff00000000ull)) {
-               val <<= 32;
-               r -= 32;
-       }
-       if (!(val & 0xffff000000000000ull)) {
-               val <<= 16;
-               r -= 16;
-       }
-       if (!(val & 0xff00000000000000ull)) {
-               val <<= 8;
-               r -= 8;
-       }
-       if (!(val & 0xf000000000000000ull)) {
-               val <<= 4;
-               r -= 4;
-       }
-       if (!(val & 0xc000000000000000ull)) {
-               val <<= 2;
-               r -= 2;
-       }
-       if (!(val & 0x8000000000000000ull)) {
-               val <<= 1;
-               r -= 1;
-       }
-       return r;
-#if HAVE_BUILTIN_CLZL
-       }
-#endif
+       return ilog64(val);
 }
 
 /* In which bucket would we find a particular record size? (ignoring header) */
-unsigned int size_to_bucket(struct tdb_context *tdb, tdb_len_t data_len)
+unsigned int size_to_bucket(unsigned int zone_bits, tdb_len_t data_len)
 {
        unsigned int bucket;
 
        /* We can't have records smaller than this. */
-       assert(data_len >= MIN_DATA_LEN);
+       assert(data_len >= TDB_MIN_DATA_LEN);
 
        /* Ignoring the header... */
-       if (data_len - MIN_DATA_LEN <= 64) {
-               /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 6. */
-               bucket = (data_len - MIN_DATA_LEN) / 8;
+       if (data_len - TDB_MIN_DATA_LEN <= 64) {
+               /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
+               bucket = (data_len - TDB_MIN_DATA_LEN) / 8;
        } else {
                /* After that we go power of 2. */
-               bucket = fls64(data_len - MIN_DATA_LEN) + 2;
+               bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
        }
 
-       if (unlikely(bucket > tdb->header.v.free_buckets))
-               bucket = tdb->header.v.free_buckets;
+       if (unlikely(bucket > BUCKETS_FOR_ZONE(zone_bits)))
+               bucket = BUCKETS_FOR_ZONE(zone_bits);
        return bucket;
 }
 
-/* What zone does a block belong in? */ 
-tdb_off_t zone_of(struct tdb_context *tdb, tdb_off_t off)
+/* Subtract 1-byte tailer and header.  Then round up to next power of 2. */
+static unsigned max_zone_bits(struct tdb_context *tdb)
+{
+       return fls64(tdb->map_size-1-sizeof(struct tdb_header)-1) + 1;
+}
+
+/* Start by using a random zone to spread the load: returns the offset. */
+static uint64_t random_zone(struct tdb_context *tdb)
+{
+       struct free_zone_header zhdr;
+       tdb_off_t off = sizeof(struct tdb_header);
+       tdb_len_t half_bits;
+       uint64_t randbits = 0;
+       unsigned int i;
+
+       for (i = 0; i < 64; i += fls64(RAND_MAX)) 
+               randbits ^= ((uint64_t)random()) << i;
+
+       /* FIXME: Does this work?  Test! */
+       half_bits = max_zone_bits(tdb) - 1;
+       do {
+               /* Pick left or right side (not outside file) */
+               if ((randbits & 1)
+                   && !tdb->methods->oob(tdb, off + (1ULL << half_bits)
+                                         + sizeof(zhdr), true)) {
+                       off += 1ULL << half_bits;
+               }
+               randbits >>= 1;
+
+               if (tdb_read_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1) 
+                       return TDB_OFF_ERR;
+
+               if (zhdr.zone_bits == half_bits)
+                       return off;
+
+               half_bits--;
+       } while (half_bits >= INITIAL_ZONE_BITS);
+
+       tdb->ecode = TDB_ERR_CORRUPT;
+       tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+                "random_zone: zone at %llu smaller than %u bits?",
+                (long long)off, INITIAL_ZONE_BITS);
+       return TDB_OFF_ERR;
+}
+
+int tdb_zone_init(struct tdb_context *tdb)
 {
-       assert(tdb->header_uptodate);
+       tdb->zone_off = random_zone(tdb);
+       if (tdb->zone_off == TDB_OFF_ERR)
+               return -1;
+       if (tdb_read_convert(tdb, tdb->zone_off,
+                            &tdb->zhdr, sizeof(tdb->zhdr)) == -1) 
+               return -1;
+       return 0;
+}
 
-       return off >> tdb->header.v.zone_bits;
+/* Where's the header, given a zone size of 1 << zone_bits? */
+static tdb_off_t zone_off(tdb_off_t off, unsigned int zone_bits)
+{
+       off -= sizeof(struct tdb_header);
+       return (off & ~((1ULL << zone_bits) - 1)) + sizeof(struct tdb_header);
 }
 
-/* Returns fl->max_bucket + 1, or list number to search. */
-static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
+/* Offset of a given bucket. */
+/* FIXME: bucket can be "unsigned" everywhere, or even uint8/16. */
+tdb_off_t bucket_off(tdb_off_t zone_off, tdb_off_t bucket)
 {
-       tdb_off_t first, off;
+       return zone_off
+               + sizeof(struct free_zone_header)
+               + bucket * sizeof(tdb_off_t);
+}
 
+/* Returns free_buckets + 1, or list number to search. */
+static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
+{
        /* Speculatively search for a non-zero bucket. */
-       first = tdb->last_zone * (tdb->header.v.free_buckets+1) + bucket;
-       off = tdb_find_nonzero_off(tdb, free_list_off(tdb, first),
-                                  tdb->header.v.free_buckets - bucket);
-       return bucket + off;
+       return tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, 0),
+                                   bucket,
+                                   BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1);
 }
 
+/* Remove from free bucket. */
 static int remove_from_list(struct tdb_context *tdb,
-                           tdb_off_t list, struct tdb_free_record *r)
+                           tdb_off_t b_off, tdb_off_t r_off,
+                           struct tdb_free_record *r)
 {
        tdb_off_t off;
 
        /* Front of list? */
        if (r->prev == 0) {
-               off = free_list_off(tdb, list);
+               off = b_off;
        } else {
                off = r->prev + offsetof(struct tdb_free_record, next);
        }
+
+#ifdef DEBUG
+       if (tdb_read_off(tdb, off) != r_off) {
+               tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+                        "remove_from_list: %llu bad prev in list %llu\n",
+                        (long long)r_off, (long long)b_off);
+               return -1;
+       }
+#endif
+
        /* r->prev->next = r->next */
        if (tdb_write_off(tdb, off, r->next)) {
                return -1;
@@ -154,6 +161,16 @@ static int remove_from_list(struct tdb_context *tdb,
        if (r->next != 0) {
                off = r->next + offsetof(struct tdb_free_record, prev);
                /* r->next->prev = r->prev */
+
+#ifdef DEBUG
+               if (tdb_read_off(tdb, off) != r_off) {
+                       tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+                                "remove_from_list: %llu bad list %llu\n",
+                                (long long)r_off, (long long)b_off);
+                       return -1;
+               }
+#endif
+
                if (tdb_write_off(tdb, off, r->prev)) {
                        return -1;
                }
@@ -161,19 +178,30 @@ static int remove_from_list(struct tdb_context *tdb,
        return 0;
 }
 
-/* Enqueue in this free list. */
+/* Enqueue in this free bucket. */
 static int enqueue_in_free(struct tdb_context *tdb,
-                          tdb_off_t list,
+                          tdb_off_t b_off,
                           tdb_off_t off,
                           struct tdb_free_record *new)
 {
        new->prev = 0;
        /* new->next = head. */
-       new->next = tdb_read_off(tdb, free_list_off(tdb, list));
+       new->next = tdb_read_off(tdb, b_off);
        if (new->next == TDB_OFF_ERR)
                return -1;
 
        if (new->next) {
+#ifdef DEBUG
+               if (tdb_read_off(tdb,
+                                new->next
+                                + offsetof(struct tdb_free_record, prev))
+                   != 0) {
+                       tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+                                "enqueue_in_free: %llu bad head prev %llu\n",
+                                (long long)new->next, (long long)b_off);
+                       return -1;
+               }
+#endif
                /* next->prev = new. */
                if (tdb_write_off(tdb, new->next
                                  + offsetof(struct tdb_free_record, prev),
@@ -181,39 +209,55 @@ static int enqueue_in_free(struct tdb_context *tdb,
                        return -1;
        }
        /* head = new */
-       if (tdb_write_off(tdb, free_list_off(tdb, list), off) != 0)
+       if (tdb_write_off(tdb, b_off, off) != 0)
                return -1;
-       
+
        return tdb_write_convert(tdb, off, new, sizeof(*new));
 }
 
-/* List isn't locked. */
+/* List need not be locked. */
 int add_free_record(struct tdb_context *tdb,
+                   unsigned int zone_bits,
                    tdb_off_t off, tdb_len_t len_with_header)
 {
        struct tdb_free_record new;
-       tdb_off_t list;
+       tdb_off_t b_off;
        int ret;
 
        assert(len_with_header >= sizeof(new));
+       assert(zone_bits < (1 << 6));
 
-       new.magic = TDB_FREE_MAGIC;
+       new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
        new.data_len = len_with_header - sizeof(struct tdb_used_record);
 
-       tdb->last_zone = zone_of(tdb, off);
-       list = tdb->last_zone * (tdb->header.v.free_buckets+1)
-               + size_to_bucket(tdb, new.data_len);
-               
-       if (tdb_lock_free_list(tdb, list, TDB_LOCK_WAIT) != 0)
+       b_off = bucket_off(zone_off(off, zone_bits),
+                          size_to_bucket(zone_bits, new.data_len));
+       if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
                return -1;
 
-       ret = enqueue_in_free(tdb, list, off, &new);
-       tdb_unlock_free_list(tdb, list);
+       ret = enqueue_in_free(tdb, b_off, off, &new);
+       tdb_unlock_free_bucket(tdb, b_off);
        return ret;
 }
 
+static size_t adjust_size(size_t keylen, size_t datalen, bool want_extra)
+{
+       size_t size = keylen + datalen;
+
+       /* We want at least 50% growth for data. */
+       if (want_extra)
+               size += datalen/2;
+
+       if (size < TDB_MIN_DATA_LEN)
+               size = TDB_MIN_DATA_LEN;
+
+       /* Round to next uint64_t boundary. */
+       return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
+}
+
 /* If we have enough left over to be useful, split that off. */
 static int to_used_record(struct tdb_context *tdb,
+                         unsigned int zone_bits,
                          tdb_off_t off,
                          tdb_len_t needed,
                          tdb_len_t total_len,
@@ -229,56 +273,59 @@ static int to_used_record(struct tdb_context *tdb,
        *actual = total_len - leftover;
 
        if (leftover) {
-               if (add_free_record(tdb, off + sizeof(used) + *actual,
+               if (add_free_record(tdb, zone_bits,
+                                   off + sizeof(used) + *actual,
                                    total_len - needed))
                        return -1;
        }
        return 0;
 }
 
-/* Note: we unlock the current list if we coalesce or fail. */
-static int coalesce(struct tdb_context *tdb, tdb_off_t off,
-                   tdb_off_t list, tdb_len_t data_len)
+/* Note: we unlock the current bucket if we coalesce or fail. */
+static int coalesce(struct tdb_context *tdb,
+                   tdb_off_t zone_off, unsigned zone_bits,
+                   tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
 {
        struct tdb_free_record pad, *r;
        tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
 
-       while (!tdb->methods->oob(tdb, end + sizeof(*r), 1)) {
-               tdb_off_t nlist;
+       while (end < (zone_off + (1ULL << zone_bits))) {
+               tdb_off_t nb_off;
 
+               /* FIXME: do tdb_get here and below really win? */
                r = tdb_get(tdb, end, &pad, sizeof(pad));
                if (!r)
                        goto err;
 
-               if (r->magic != TDB_FREE_MAGIC)
+               if (frec_magic(r) != TDB_FREE_MAGIC)
                        break;
 
-               nlist = zone_of(tdb, end) * (tdb->header.v.free_buckets+1)
-                       + size_to_bucket(tdb, r->data_len);
+               nb_off = bucket_off(zone_off,
+                                   size_to_bucket(zone_bits, r->data_len));
 
                /* We may be violating lock order here, so best effort. */
-               if (tdb_lock_free_list(tdb, nlist, TDB_LOCK_NOWAIT) == -1)
+               if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
                        break;
 
                /* Now we have lock, re-check. */
                r = tdb_get(tdb, end, &pad, sizeof(pad));
                if (!r) {
-                       tdb_unlock_free_list(tdb, nlist);
+                       tdb_unlock_free_bucket(tdb, nb_off);
                        goto err;
                }
 
-               if (unlikely(r->magic != TDB_FREE_MAGIC)) {
-                       tdb_unlock_free_list(tdb, nlist);
+               if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
+                       tdb_unlock_free_bucket(tdb, nb_off);
                        break;
                }
 
-               if (remove_from_list(tdb, list, r) == -1) {
-                       tdb_unlock_free_list(tdb, nlist);
+               if (remove_from_list(tdb, nb_off, end, r) == -1) {
+                       tdb_unlock_free_bucket(tdb, nb_off);
                        goto err;
                }
 
                end += sizeof(struct tdb_used_record) + r->data_len;
-               tdb_unlock_free_list(tdb, nlist);
+               tdb_unlock_free_bucket(tdb, nb_off);
        }
 
        /* Didn't find any adjacent free? */
@@ -290,65 +337,74 @@ static int coalesce(struct tdb_context *tdb, tdb_off_t off,
        if (!r)
                goto err;
 
-       if (remove_from_list(tdb, list, r) == -1)
+       if (r->data_len != data_len) {
+               tdb->ecode = TDB_ERR_CORRUPT;
+               tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+                        "coalesce: expected data len %llu not %llu\n",
+                        (long long)data_len, (long long)r->data_len);
+               goto err;
+       }
+
+       if (remove_from_list(tdb, b_off, off, r) == -1)
                goto err;
 
        /* We have to drop this to avoid deadlocks. */
-       tdb_unlock_free_list(tdb, list);
+       tdb_unlock_free_bucket(tdb, b_off);
 
-       if (add_free_record(tdb, off, end - off) == -1)
+       if (add_free_record(tdb, zone_bits, off, end - off) == -1)
                return -1;
        return 1;
 
 err:
-       /* To unify error paths, we *always* unlock list. */
-       tdb_unlock_free_list(tdb, list);
+       /* To unify error paths, we *always* unlock bucket on error. */
+       tdb_unlock_free_bucket(tdb, b_off);
        return -1;
 }
 
 /* We need size bytes to put our key and data in. */
 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
-                               tdb_off_t bucket, size_t size,
+                               tdb_off_t zone_off,
+                               unsigned zone_bits,
+                               tdb_off_t bucket,
+                               size_t size,
                                tdb_len_t *actual)
 {
-       tdb_off_t list;
-       tdb_off_t off, prev, best_off;
+       tdb_off_t off, b_off,best_off;
        struct tdb_free_record pad, best = { 0 }, *r;
        double multiplier;
 
 again:
-       list = tdb->last_zone * (tdb->header.v.free_buckets+1) + bucket;
+       b_off = bucket_off(zone_off, bucket);
 
-       /* Lock this list. */
-       if (tdb_lock_free_list(tdb, list, TDB_LOCK_WAIT) == -1) {
+       /* FIXME: Try non-blocking wait first, to measure contention.
+        * If we're contented, try switching zones, and don't enlarge zone
+        * next time (we want more zones). */
+       /* Lock this bucket. */
+       if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
                return TDB_OFF_ERR;
        }
 
-       prev = free_list_off(tdb, list);
-       off = tdb_read_off(tdb, prev);
-
-       if (unlikely(off == TDB_OFF_ERR))
-               goto unlock_err;
-
        best.data_len = -1ULL;
        best_off = 0;
+       /* FIXME: Start with larger multiplier if we're growing. */
        multiplier = 1.0;
 
        /* Walk the list to see if any are large enough, getting less fussy
         * as we go. */
-       while (off) {
-               prev = off;
-               off = tdb_read_off(tdb, prev);
-               if (unlikely(off == TDB_OFF_ERR))
-                       goto unlock_err;
+       off = tdb_read_off(tdb, b_off);
+       if (unlikely(off == TDB_OFF_ERR))
+               goto unlock_err;
 
+       while (off) {
+               /* FIXME: Does tdb_get win anything here? */
                r = tdb_get(tdb, off, &pad, sizeof(*r));
                if (!r)
                        goto unlock_err;
-               if (r->magic != TDB_FREE_MAGIC) {
+
+               if (frec_magic(r) != TDB_FREE_MAGIC) {
                        tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
                                 "lock_and_alloc: %llu non-free 0x%llx\n",
-                                (long long)off, (long long)r->magic);
+                                (long long)off, (long long)r->magic_and_meta);
                        goto unlock_err;
                }
 
@@ -357,22 +413,14 @@ again:
                        best = *r;
                }
 
-               if (best.data_len < size * multiplier && best_off) {
-                       /* We're happy with this size: take it. */
-                       if (remove_from_list(tdb, list, &best) != 0)
-                               goto unlock_err;
-                       tdb_unlock_free_list(tdb, list);
+               if (best.data_len < size * multiplier && best_off)
+                       goto use_best;
 
-                       if (to_used_record(tdb, best_off, size, best.data_len,
-                                          actual)) {
-                               return -1;
-                       }
-                       return best_off;
-               }
                multiplier *= 1.01;
 
                /* Since we're going slow anyway, try coalescing here. */
-               switch (coalesce(tdb, off, list, r->data_len)) {
+               switch (coalesce(tdb, zone_off, zone_bits, off, b_off,
+                                r->data_len)) {
                case -1:
                        /* This has already unlocked on error. */
                        return -1;
@@ -380,80 +428,87 @@ again:
                        /* This has unlocked list, restart. */
                        goto again;
                }
+               off = r->next;
+       }
+
+       /* If we found anything at all, use it. */
+       if (best_off) {
+       use_best:
+               /* We're happy with this size: take it. */
+               if (remove_from_list(tdb, b_off, best_off, &best) != 0)
+                       goto unlock_err;
+               tdb_unlock_free_bucket(tdb, b_off);
+
+               if (to_used_record(tdb, zone_bits, best_off, size,
+                                  best.data_len, actual)) {
+                       return -1;
+               }
+               return best_off;
        }
 
-       tdb_unlock_free_list(tdb, list);
+       tdb_unlock_free_bucket(tdb, b_off);
        return 0;
 
 unlock_err:
-       tdb_unlock_free_list(tdb, list);
+       tdb_unlock_free_bucket(tdb, b_off);
        return TDB_OFF_ERR;
 }
 
-/* We want a really big chunk.  Look through every zone's oversize bucket */
-static tdb_off_t huge_alloc(struct tdb_context *tdb, size_t size,
-                           tdb_len_t *actual)
+static bool next_zone(struct tdb_context *tdb)
 {
-       tdb_off_t i, off;
+       tdb_off_t next = tdb->zone_off + (1ULL << tdb->zhdr.zone_bits);
 
-       do {
-               for (i = 0; i < tdb->header.v.num_zones; i++) {
-                       /* Try getting one from list. */
-                       off = lock_and_alloc(tdb, tdb->header.v.free_buckets,
-                                            size, actual);
-                       if (off == TDB_OFF_ERR)
-                               return TDB_OFF_ERR;
-                       if (off != 0)
-                               return off;
-                       /* FIXME: Coalesce! */
-               }
-       } while (tdb_expand(tdb, 0, size, false) == 0);
+       /* We must have a header. */
+       if (tdb->methods->oob(tdb, next + sizeof(tdb->zhdr), true))
+               return false;
 
-       return TDB_OFF_ERR;
+       tdb->zone_off = next;
+       return tdb_read_convert(tdb, next, &tdb->zhdr, sizeof(tdb->zhdr)) == 0;
 }
 
+/* Offset returned is within current zone (which it may alter). */
 static tdb_off_t get_free(struct tdb_context *tdb, size_t size,
                          tdb_len_t *actual)
 {
-       tdb_off_t off, bucket;
-       unsigned int num_empty, step = 0;
-
-       bucket = size_to_bucket(tdb, size);
-
-       /* If we're after something bigger than a single zone, handle
-        * specially. */
-       if (unlikely(sizeof(struct tdb_used_record) + size
-                    >= (1ULL << tdb->header.v.zone_bits))) {
-               return huge_alloc(tdb, size, actual);
-       }
-
-       /* Number of zones we search is proportional to the log of them. */
-       for (num_empty = 0; num_empty < fls64(tdb->header.v.num_zones);
-            num_empty++) {
+       tdb_off_t start_zone = tdb->zone_off, off;
+       bool wrapped = false;
+
+       /* FIXME: If we don't get a hit in the first bucket we want,
+        * try changing zones for next time.  That should help wear
+        * zones evenly, so we don't need to search all of them before
+        * expanding. */
+       while (!wrapped || tdb->zone_off != start_zone) {
                tdb_off_t b;
 
+               /* Shortcut for really huge allocations... */
+               if ((size >> tdb->zhdr.zone_bits) != 0)
+                       continue;
+
                /* Start at exact size bucket, and search up... */
-               for (b = bucket; b <= tdb->header.v.num_zones; b++) {
-                       b = find_free_head(tdb, b);
-
-                       /* Non-empty list?  Try getting block. */
-                       if (b <= tdb->header.v.num_zones) {
-                               /* Try getting one from list. */
-                               off = lock_and_alloc(tdb, b, size, actual);
-                               if (off == TDB_OFF_ERR)
-                                       return TDB_OFF_ERR;
-                               if (off != 0)
-                                       return off;
-                               /* Didn't work.  Try next bucket. */
-                       }
+               b = size_to_bucket(tdb->zhdr.zone_bits, size);
+               for (b = find_free_head(tdb, b);
+                    b <= BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits);
+                    b += find_free_head(tdb, b + 1)) {
+                       /* Try getting one from list. */
+                       off = lock_and_alloc(tdb, tdb->zone_off,
+                                            tdb->zhdr.zone_bits,
+                                            b, size, actual);
+                       if (off == TDB_OFF_ERR)
+                               return TDB_OFF_ERR;
+                       if (off != 0)
+                               return off;
+                       /* Didn't work.  Try next bucket. */
                }
 
-               /* Try another zone, at pseudo random.  Avoid duplicates by
-                  using an odd step. */
-               if (step == 0)
-                       step = ((quick_random(tdb)) % 65536) * 2 + 1;
-               tdb->last_zone = (tdb->last_zone + step)
-                       % tdb->header.v.num_zones;
+               /* Didn't work, try next zone, if it exists. */
+               if (!next_zone(tdb)) {
+                       wrapped = true;
+                       tdb->zone_off = sizeof(struct tdb_header);
+                       if (tdb_read_convert(tdb, tdb->zone_off,
+                                            &tdb->zhdr, sizeof(tdb->zhdr))) {
+                               return TDB_OFF_ERR;
+                       }
+               }
        }
        return 0;
 }
@@ -461,14 +516,16 @@ static tdb_off_t get_free(struct tdb_context *tdb, size_t size,
 int set_header(struct tdb_context *tdb,
               struct tdb_used_record *rec,
               uint64_t keylen, uint64_t datalen,
-              uint64_t actuallen, uint64_t hash)
+              uint64_t actuallen, unsigned hashlow,
+              unsigned int zone_bits)
 {
        uint64_t keybits = (fls64(keylen) + 1) / 2;
 
-       /* Use top bits of hash, so it's independent of hash table size. */
+       /* Use bottom bits of hash, so it's independent of hash table size. */
        rec->magic_and_meta
-               = (actuallen - (keylen + datalen))
-               | ((hash >> 53) << 32)
+               = zone_bits
+               | ((hashlow & ((1 << 5)-1)) << 6)
+               | ((actuallen - (keylen + datalen)) << 11)
                | (keybits << 43)
                | (TDB_MAGIC << 48);
        rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
@@ -487,224 +544,141 @@ int set_header(struct tdb_context *tdb,
        return 0;
 }
 
-static tdb_len_t adjust_size(size_t keylen, size_t datalen, bool growing)
-{
-       tdb_len_t size = keylen + datalen;
-
-       if (size < MIN_DATA_LEN)
-               size = MIN_DATA_LEN;
-
-       /* Overallocate if this is coming from an enlarging store. */
-       if (growing)
-               size += datalen / 2;
-
-       /* Round to next uint64_t boundary. */
-       return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
-}
-
-/* If this fails, try tdb_expand. */
-tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
-               uint64_t hash, bool growing)
-{
-       tdb_off_t off;
-       tdb_len_t size, actual;
-       struct tdb_used_record rec;
-
-       /* We don't want header to change during this! */
-       assert(tdb->header_uptodate);
-
-       size = adjust_size(keylen, datalen, growing);
-
-       off = get_free(tdb, size, &actual);
-       if (unlikely(off == TDB_OFF_ERR || off == 0))
-               return off;
-
-       /* Some supergiant values can't be encoded. */
-       if (set_header(tdb, &rec, keylen, datalen, actual, hash) != 0) {
-               add_free_record(tdb, off, sizeof(rec) + actual);
-               return TDB_OFF_ERR;
-       }
-
-       if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
-               return TDB_OFF_ERR;
-       
-       return off;
-}
-
-static bool larger_buckets_might_help(struct tdb_context *tdb)
+static bool zones_happy(struct tdb_context *tdb)
 {
-       /* If our buckets are already covering 1/8 of a zone, don't
-        * bother (note: might become an 1/16 of a zone if we double
-        * zone size). */
-       tdb_len_t size = (1ULL << tdb->header.v.zone_bits) / 8;
-
-       if (size >= MIN_DATA_LEN
-           && size_to_bucket(tdb, size) < tdb->header.v.free_buckets) {
-               return false;
-       }
-
-       /* FIXME: Put stats in tdb_context or examine db itself! */
-       /* It's fairly cheap to do as we expand database. */
+       /* FIXME: look at distribution of zones. */
        return true;
 }
 
-static bool zones_happy(struct tdb_context *tdb)
+/* Assume we want buckets up to the comfort factor. */
+static tdb_len_t overhead(unsigned int zone_bits)
 {
-       /* FIXME: look at distribution of zones. */
-       return true;
+       return sizeof(struct free_zone_header)
+               + (BUCKETS_FOR_ZONE(zone_bits) + 1) * sizeof(tdb_off_t);
 }
 
-/* Expand the database. */
-int tdb_expand(struct tdb_context *tdb, tdb_len_t klen, tdb_len_t dlen,
-              bool growing)
+/* Expand the database (by adding a zone). */
+static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
-       uint64_t new_num_buckets, new_num_zones, new_zone_bits;
-       uint64_t old_num_total, i;
-       tdb_len_t add, freebucket_size, needed;
-       tdb_off_t off, old_free_off;
-       const tdb_off_t *oldf;
-       struct tdb_used_record fhdr;
-       
+       uint64_t old_size;
+       tdb_off_t off;
+       uint8_t zone_bits;
+       unsigned int num_buckets;
+       tdb_len_t wanted;
+       struct free_zone_header zhdr;
+       bool enlarge_zone;
+
        /* We need room for the record header too. */
-       needed = sizeof(struct tdb_used_record)
-               + adjust_size(klen, dlen, growing);
+       wanted = sizeof(struct tdb_used_record) + size;
 
-       /* FIXME: this is overkill.  An expand lock? */
-       if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
+       /* Only one person can expand file at a time. */
+       if (tdb_lock_expand(tdb, F_WRLCK) != 0)
                return -1;
 
-       /* Someone may have expanded for us. */
-       if (update_header(tdb))
+       /* 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)
                goto success;
 
-       /* Make sure we have the latest size. */
-       tdb->methods->oob(tdb, tdb->map_size + 1, true);
+       /* FIXME: Tailer is a bogus optimization, remove it. */
+       /* zone bits tailer char is protected by EXPAND lock. */
+       if (tdb->methods->read(tdb, old_size - 1, &zone_bits, 1) == -1)
+               goto fail;
 
-       /* Did we enlarge zones without enlarging file? */
-       if (tdb->map_size < tdb->header.v.num_zones<<tdb->header.v.zone_bits) {
-               add = (tdb->header.v.num_zones<<tdb->header.v.zone_bits)
-                       - tdb->map_size;
-               /* Updates tdb->map_size. */
-               if (tdb->methods->expand_file(tdb, tdb->map_size, add) == -1)
-                       goto fail;
-               if (add_free_record(tdb, tdb->map_size - add, add) == -1)
-                       goto fail;
-               if (add >= needed) {
-                       /* Allocate from this zone. */
-                       tdb->last_zone = zone_of(tdb, tdb->map_size - add);
-                       goto success;
+       /* If zones aren't working well, add larger zone if possible. */
+       enlarge_zone = !zones_happy(tdb);
+
+       /* New zone can be between zone_bits or larger if we're on the right
+        * boundary. */
+       for (;;) {
+               /* Does this fit the allocation comfortably? */
+               if ((1ULL << zone_bits) >= overhead(zone_bits) + wanted) {
+                       /* Only let enlarge_zone enlarge us once. */
+                       if (!enlarge_zone)
+                               break;
+                       enlarge_zone = false;
                }
+               if ((old_size - 1 - sizeof(struct tdb_header))
+                   & (1 << zone_bits))
+                       break;
+               zone_bits++;
        }
 
-       /* Slow path.  Should we increase the number of buckets? */
-       new_num_buckets = tdb->header.v.free_buckets;
-       if (larger_buckets_might_help(tdb))
-               new_num_buckets++;
-
-       /* Now we'll need room for the new free buckets, too.  Assume
-        * worst case (zones expand). */
-       needed += sizeof(fhdr)
-               + ((tdb->header.v.num_zones+1)
-                  * (new_num_buckets+1) * sizeof(tdb_off_t));
-
-       /* If we need less that one zone, and they're working well, just add
-        * another one. */
-       if (needed < (1UL<<tdb->header.v.zone_bits) && zones_happy(tdb)) {
-               new_num_zones = tdb->header.v.num_zones+1;
-               new_zone_bits = tdb->header.v.zone_bits;
-               add = 1ULL << tdb->header.v.zone_bits;
-       } else {
-               /* Increase the zone size. */
-               new_num_zones = tdb->header.v.num_zones;
-               new_zone_bits = tdb->header.v.zone_bits+1;
-               while ((new_num_zones << new_zone_bits) - tdb->map_size
-                      < needed) {
-                       new_zone_bits++;
-               }
+       zhdr.zone_bits = zone_bits;
+       num_buckets = BUCKETS_FOR_ZONE(zone_bits);
 
-               /* We expand by enough zones to meet the need. */
-               add = (needed + (1ULL << new_zone_bits)-1)
-                       & ~((1ULL << new_zone_bits)-1);
-       }
-
-       /* Updates tdb->map_size. */
-       if (tdb->methods->expand_file(tdb, tdb->map_size, add) == -1)
+       /* FIXME: I don't think we need to expand to full zone, do we? */
+       if (tdb->methods->expand_file(tdb, 1ULL << zone_bits) == -1)
                goto fail;
 
-       /* Use first part as new free bucket array. */
-       off = tdb->map_size - add;
-       freebucket_size = new_num_zones
-               * (new_num_buckets + 1) * sizeof(tdb_off_t);
-
-       /* Write header. */
-       if (set_header(tdb, &fhdr, 0, freebucket_size, freebucket_size, 0))
-               goto fail;
-       if (tdb_write_convert(tdb, off, &fhdr, sizeof(fhdr)) == -1)
+       /* Write new tailer. */
+       if (tdb->methods->write(tdb, tdb->map_size - 1, &zone_bits, 1) == -1)
                goto fail;
 
-       /* Adjust off to point to start of buckets, add to be remainder. */
-       add -= freebucket_size + sizeof(fhdr);
-       off += sizeof(fhdr);
-
-       /* Access the old zones. */
-       old_num_total = tdb->header.v.num_zones*(tdb->header.v.free_buckets+1);
-       old_free_off = tdb->header.v.free_off;
-       oldf = tdb_access_read(tdb, old_free_off,
-                              old_num_total * sizeof(tdb_off_t));
-       if (!oldf)
+       /* Write new zone header (just before old tailer). */
+       off = old_size - 1;
+       if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
                goto fail;
 
-       /* Switch to using our new zone. */
-       if (zero_out(tdb, off, new_num_zones * (new_num_buckets + 1)) == -1)
-               goto fail_release;
-       tdb->header.v.free_off = off;
-       tdb->header.v.num_zones = new_num_zones;
-       tdb->header.v.free_buckets = new_num_buckets;
-
-       /* FIXME: If zone size hasn't changed, can simply copy pointers. */
-       /* FIXME: Coalesce? */
-       for (i = 0; i < old_num_total; i++) {
-               tdb_off_t next;
-               struct tdb_free_record rec;
-               tdb_off_t list;
-
-               for (off = oldf[i]; off; off = next) {
-                       if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
-                               goto fail_release;
-
-                       list = zone_of(tdb, off)
-                               * (tdb->header.v.free_buckets+1)
-                               + size_to_bucket(tdb, rec.data_len);
-                       next = rec.next;
-               
-                       if (enqueue_in_free(tdb, list, off, &rec) == -1)
-                               goto fail_release;
-               }
-       }
-
+       /* Now write empty buckets. */
+       off += sizeof(zhdr);
+       if (zero_out(tdb, off, (num_buckets+1) * sizeof(tdb_off_t)) == -1)
+               goto fail;
+       off += (num_buckets+1) * sizeof(tdb_off_t);
 
-       /* Free up the old free buckets. */
-       old_free_off -= sizeof(fhdr);
-       if (tdb_read_convert(tdb, old_free_off, &fhdr, sizeof(fhdr)) == -1)
-               goto fail_release;
-       if (add_free_record(tdb, old_free_off,
-                           rec_data_length(&fhdr)+rec_extra_padding(&fhdr)))
-               goto fail_release;
+       /* Now add the rest as our free record. */
+       if (add_free_record(tdb, zone_bits, off, tdb->map_size-1-off) == -1)
+               goto fail;
 
-       /* Add the rest as a new free record. */
-       if (add_free_record(tdb, tdb->map_size - add, add) == -1)
-               goto fail_release;
+       /* Try allocating from this zone now. */
+       tdb->zone_off = old_size - 1;
+       tdb->zhdr = zhdr;
 
-       /* Start allocating from where the new space is. */
-       tdb->last_zone = zone_of(tdb, tdb->map_size - add);
-       tdb_access_release(tdb, oldf);
 success:
-       tdb_allrecord_unlock(tdb, F_WRLCK);
+       tdb_unlock_expand(tdb, F_WRLCK);
        return 0;
 
-fail_release:
-       tdb_access_release(tdb, oldf);
 fail:
-       tdb_allrecord_unlock(tdb, F_WRLCK);
+       tdb_unlock_expand(tdb, F_WRLCK);
        return -1;
 }
+
+/* This won't fail: it will expand the database if it has to. */
+tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
+               uint64_t hash, bool growing)
+{
+       tdb_off_t off;
+       tdb_len_t size, actual;
+       struct tdb_used_record rec;
+
+       /* We can't hold pointers during this: we could unmap! */
+       assert(!tdb->direct_access);
+
+       size = adjust_size(keylen, datalen, growing);
+
+again:
+       off = get_free(tdb, size, &actual);
+       if (unlikely(off == TDB_OFF_ERR))
+               return off;
+
+       if (unlikely(off == 0)) {
+               if (tdb_expand(tdb, size) == -1)
+                       return TDB_OFF_ERR;
+               goto again;
+       }
+
+       /* Some supergiant values can't be encoded. */
+       /* FIXME: Check before, and limit actual in get_free. */
+       if (set_header(tdb, &rec, keylen, datalen, actual, hash,
+                      tdb->zhdr.zone_bits) != 0) {
+               add_free_record(tdb, tdb->zhdr.zone_bits, off,
+                               sizeof(rec) + actual);
+               return TDB_OFF_ERR;
+       }
+
+       if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
+               return TDB_OFF_ERR;
+       
+       return off;
+}