X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=b41e1d58d2a699a76f74c98cb17f0e5e21ba5a97;hp=83d916fc6986ceec757131b51b44a8c13eaa022a;hb=539f1af037858b905c50c560f2a608555d8457ff;hpb=39f01834db9b6a21d076e67d1e3143ab99aaf43e

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index 83d916fc..b41e1d58 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -17,362 +17,426 @@
 */
 #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-	/* num_zones might be out of date, but can only increase */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(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 >= TDB_FREE_BUCKETS))
+		bucket = TDB_FREE_BUCKETS - 1;
 	return bucket;
 }
 
-/* What zone does a block belong in? */ 
-tdb_off_t zone_of(struct tdb_context *tdb, tdb_off_t off)
+tdb_off_t first_flist(struct tdb_context *tdb)
 {
-	assert(tdb->header_uptodate);
+	return tdb_read_off(tdb, offsetof(struct tdb_header, free_list));
+}
 
-	return off >> tdb->header.v.zone_bits;
+tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist)
+{
+	return tdb_read_off(tdb, flist + offsetof(struct tdb_freelist, next));
 }
 
-/* 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)
+int tdb_flist_init(struct tdb_context *tdb)
 {
-	tdb_off_t first, off;
+	/* Use reservoir sampling algorithm to select a free list at random. */
+	unsigned int rnd, max = 0, count = 0;
+	tdb_off_t off;
+
+	tdb->flist_off = off = first_flist(tdb);
+	tdb->flist = 0;
+
+	while (off) {
+		if (off == TDB_OFF_ERR)
+			return -1;
+
+		rnd = random();
+		if (rnd >= max) {
+			tdb->flist_off = off;
+			tdb->flist = count;
+			max = rnd;
+		}
+
+		off = next_flist(tdb, off);
+		count++;
+	}
+	return 0;
+}
 
+/* Offset of a given bucket. */
+tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket)
+{
+	return flist_off + offsetof(struct tdb_freelist, buckets)
+		+ 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 flist_off,
+				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(flist_off, 0),
+				    bucket, TDB_FREE_BUCKETS);
 }
 
+/* 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,
+			    const struct tdb_free_record *r)
 {
 	tdb_off_t off;
 
 	/* Front of list? */
-	if (r->prev == 0) {
-		off = free_list_off(tdb, list);
+	if (frec_prev(r) == 0) {
+		off = b_off;
 	} else {
-		off = r->prev + offsetof(struct tdb_free_record, next);
+		off = frec_prev(r) + offsetof(struct tdb_free_record, next);
 	}
+
+#ifdef DEBUG
+	if (tdb_read_off(tdb, off) != r_off) {
+		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+			 "remove_from_list: %llu bad prev in list %llu",
+			 (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;
 	}
 
 	if (r->next != 0) {
-		off = r->next + offsetof(struct tdb_free_record, prev);
+		off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
 		/* r->next->prev = r->prev */
-		if (tdb_write_off(tdb, off, r->prev)) {
+
+#ifdef DEBUG
+		if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				   "remove_from_list: %llu bad list %llu",
+				   (long long)r_off, (long long)b_off);
+			return -1;
+		}
+#endif
+
+		if (tdb_write_off(tdb, off, r->magic_and_prev)) {
 			return -1;
 		}
 	}
 	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)
+			   tdb_len_t len)
 {
-	new->prev = 0;
+	struct tdb_free_record new;
+	uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
+
+	/* We only need to set flist_and_len; rest is set in enqueue_in_free */
+	new.flist_and_len = ((uint64_t)tdb->flist << (64 - TDB_OFF_UPPER_STEAL))
+		| len;
+	/* prev = 0. */
+	new.magic_and_prev = magic;
+
 	/* new->next = head. */
-	new->next = tdb_read_off(tdb, free_list_off(tdb, list));
-	if (new->next == TDB_OFF_ERR)
+	new.next = tdb_read_off(tdb, b_off);
+	if (new.next == TDB_OFF_ERR)
 		return -1;
 
-	if (new->next) {
+	if (new.next) {
+#ifdef DEBUG
+		if (tdb_read_off(tdb,
+				 new.next + offsetof(struct tdb_free_record,
+						     magic_and_prev))
+		    != magic) {
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				   "enqueue_in_free: %llu bad head"
+				   " prev %llu",
+				   (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),
-				  off) != 0)
+		if (tdb_write_off(tdb, new.next
+				  + offsetof(struct tdb_free_record,
+					     magic_and_prev),
+				  off | magic) != 0)
 			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));
+
+	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,
 		    tdb_off_t off, tdb_len_t len_with_header)
 {
-	struct tdb_free_record new;
-	tdb_off_t list;
+	tdb_off_t b_off;
+	tdb_len_t len;
 	int ret;
 
-	assert(len_with_header >= sizeof(new));
+	assert(len_with_header >= sizeof(struct tdb_free_record));
 
-	new.magic = TDB_FREE_MAGIC;
-	new.data_len = len_with_header - sizeof(struct tdb_used_record);
+	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(tdb->flist_off, size_to_bucket(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, len);
+	tdb_unlock_free_bucket(tdb, b_off);
 	return ret;
 }
 
+static size_t adjust_size(size_t keylen, size_t datalen)
+{
+	size_t size = keylen + datalen;
+
+	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,
-			  tdb_off_t off,
-			  tdb_len_t needed,
-			  tdb_len_t total_len,
-			  tdb_len_t *actual)
+static size_t record_leftover(size_t keylen, size_t datalen,
+			      bool want_extra, size_t total_len)
 {
-	struct tdb_used_record used;
-	tdb_len_t leftover;
+	ssize_t leftover;
 
-	leftover = total_len - needed;
-	if (leftover < sizeof(struct tdb_free_record))
-		leftover = 0;
+	if (want_extra)
+		datalen += datalen / 2;
+	leftover = total_len - adjust_size(keylen, datalen);
 
-	*actual = total_len - leftover;
+	if (leftover < (ssize_t)sizeof(struct tdb_free_record))
+		return 0;
 
-	if (leftover) {
-		if (add_free_record(tdb, off + sizeof(used) + *actual,
-				    total_len - needed))
-			return -1;
-	}
-	return 0;
+	return leftover;
 }
 
-/* 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)
+static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
 {
-	struct tdb_free_record pad, *r;
-	tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
+	tdb_off_t off;
+	unsigned int i;
 
-	while (!tdb->methods->oob(tdb, end + sizeof(*r), 1)) {
-		tdb_off_t nlist;
+	if (likely(tdb->flist == flist))
+		return tdb->flist_off;
 
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
+	off = first_flist(tdb);
+	for (i = 0; i < flist; i++)
+		off = next_flist(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)
+{
+	tdb_off_t end;
+	struct tdb_free_record rec;
+
+	add_stat(tdb, alloc_coalesce_tried, 1);
+	end = off + sizeof(struct tdb_used_record) + data_len;
+
+	while (end < tdb->map_size) {
+		const struct tdb_free_record *r;
+		tdb_off_t nb_off;
+		unsigned flist, bucket;
+
+		r = tdb_access_read(tdb, end, sizeof(*r), true);
 		if (!r)
 			goto err;
 
-		if (r->magic != TDB_FREE_MAGIC)
+		if (frec_magic(r) != TDB_FREE_MAGIC
+		    || frec_flist(r) == TDB_FLIST_NONE) {
+			tdb_access_release(tdb, r);
 			break;
+		}
 
-		nlist = zone_of(tdb, end) * (tdb->header.v.free_buckets+1)
-			+ size_to_bucket(tdb, r->data_len);
+		flist = frec_flist(r);
+		bucket = size_to_bucket(frec_len(r));
+		nb_off = bucket_off(flist_offset(tdb, flist), bucket);
+		tdb_access_release(tdb, r);
 
 		/* 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) {
+			add_stat(tdb, alloc_coalesce_lockfail, 1);
 			break;
+		}
 
 		/* Now we have lock, re-check. */
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
-		if (!r) {
-			tdb_unlock_free_list(tdb, nlist);
+		if (tdb_read_convert(tdb, end, &rec, sizeof(rec))) {
+			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(&rec) != TDB_FREE_MAGIC)) {
+			add_stat(tdb, alloc_coalesce_race, 1);
+			tdb_unlock_free_bucket(tdb, nb_off);
+			break;
+		}
+
+		if (unlikely(frec_flist(&rec) != flist)
+		    || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
+			add_stat(tdb, alloc_coalesce_race, 1);
+			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, &rec) == -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);
+		end += sizeof(struct tdb_used_record) + frec_len(&rec);
+		tdb_unlock_free_bucket(tdb, nb_off);
+		add_stat(tdb, alloc_coalesce_num_merged, 1);
 	}
 
 	/* Didn't find any adjacent free? */
 	if (end == off + sizeof(struct tdb_used_record) + data_len)
 		return 0;
 
-	/* OK, expand record */
-	r = tdb_get(tdb, off, &pad, sizeof(pad));
-	if (!r)
+	/* OK, expand initial record */
+	if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
+		goto err;
+
+	if (frec_len(&rec) != data_len) {
+		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+			   "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)
 		goto err;
 
-	if (remove_from_list(tdb, list, r) == -1)
+	/* We have to drop this to avoid deadlocks, so make sure record
+	 * doesn't get coalesced by someone else! */
+	rec.flist_and_len = (TDB_FLIST_NONE << (64 - TDB_OFF_UPPER_STEAL))
+		| (end - off - sizeof(struct tdb_used_record));
+	if (tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
+					      flist_and_len),
+			  rec.flist_and_len) != 0)
 		goto err;
 
-	/* We have to drop this to avoid deadlocks. */
-	tdb_unlock_free_list(tdb, list);
+	add_stat(tdb, alloc_coalesce_succeeded, 1);
+	tdb_unlock_free_bucket(tdb, b_off);
 
 	if (add_free_record(tdb, 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_len_t *actual)
+				tdb_off_t flist_off,
+				tdb_off_t bucket,
+				size_t keylen, size_t datalen,
+				bool want_extra,
+				unsigned hashlow)
 {
-	tdb_off_t list;
-	tdb_off_t off, prev, best_off;
-	struct tdb_free_record pad, best = { 0 }, *r;
+	tdb_off_t off, b_off,best_off;
+	struct tdb_free_record best = { 0 };
 	double multiplier;
+	size_t size = adjust_size(keylen, datalen);
 
+	add_stat(tdb, allocs, 1);
 again:
-	list = tdb->last_zone * (tdb->header.v.free_buckets+1) + bucket;
+	b_off = bucket_off(flist_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. */
+	/* 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.flist_and_len = -1ULL;
 	best_off = 0;
-	multiplier = 1.0;
+
+	/* Get slack if we're after extra. */
+	if (want_extra)
+		multiplier = 1.5;
+	else
+		multiplier = 1.0;
 
 	/* 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))
+		goto unlock_err;
+
 	while (off) {
-		prev = off;
-		off = tdb_read_off(tdb, prev);
-		if (unlikely(off == TDB_OFF_ERR))
-			goto unlock_err;
+		const struct tdb_free_record *r;
+		tdb_len_t len;
+		tdb_off_t next;
 
-		r = tdb_get(tdb, off, &pad, sizeof(*r));
+		r = tdb_access_read(tdb, off, sizeof(*r), true);
 		if (!r)
 			goto unlock_err;
-		if (r->magic != 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);
+
+		if (frec_magic(r) != TDB_FREE_MAGIC) {
+			tdb_access_release(tdb, r);
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				 "lock_and_alloc: %llu non-free 0x%llx",
+				 (long long)off, (long long)r->magic_and_prev);
 			goto unlock_err;
 		}
 
-		if (r->data_len >= size && r->data_len < best.data_len) {
+		if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
 			best_off = off;
 			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 (to_used_record(tdb, best_off, size, best.data_len,
-					   actual)) {
-				return -1;
-			}
-			return best_off;
+		if (frec_len(&best) < size * multiplier && best_off) {
+			tdb_access_release(tdb, r);
+			break;
 		}
+
 		multiplier *= 1.01;
 
+		next = r->next;
+		len = frec_len(r);
+		tdb_access_release(tdb, r);
+
 		/* Since we're going slow anyway, try coalescing here. */
-		switch (coalesce(tdb, off, list, r->data_len)) {
+		switch (coalesce(tdb, off, b_off, len)) {
 		case -1:
 			/* This has already unlocked on error. */
 			return -1;
@@ -380,95 +444,121 @@ again:
 			/* This has unlocked list, restart. */
 			goto again;
 		}
+		off = next;
+	}
+
+	/* If we found anything at all, use it. */
+	if (best_off) {
+		struct tdb_used_record rec;
+		size_t leftover;
+
+		/* We're happy with this size: take it. */
+		if (remove_from_list(tdb, b_off, best_off, &best) != 0)
+			goto unlock_err;
+
+		leftover = record_leftover(keylen, datalen, want_extra,
+					   frec_len(&best));
+
+		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_used_header(tdb, &rec, keylen, datalen,
+				    frec_len(&best) - leftover,
+				    hashlow) != 0)
+			goto unlock_err;
+
+		if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
+			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_unlock_free_bucket(tdb, b_off);
+
+		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)
+/* Get a free block from current free list, or 0 if none. */
+static tdb_off_t get_free(struct tdb_context *tdb,
+			  size_t keylen, size_t datalen, bool want_extra,
+			  unsigned hashlow)
 {
-	tdb_off_t i, off;
-
-	do {
-		for (i = 0; i < tdb->header.v.num_zones; i++) {
+	tdb_off_t off, flist_off;
+	unsigned start_b, b, flist;
+	bool wrapped = false;
+
+	/* If they are growing, add 50% to get to higher bucket. */
+	if (want_extra)
+		start_b = size_to_bucket(adjust_size(keylen,
+						     datalen + datalen / 2));
+	else
+		start_b = size_to_bucket(adjust_size(keylen, datalen));
+
+	flist_off = tdb->flist_off;
+	flist = tdb->flist;
+	while (!wrapped || flist_off != tdb->flist_off) {
+		/* Start at exact size bucket, and search up... */
+		for (b = find_free_head(tdb, flist_off, start_b);
+		     b < TDB_FREE_BUCKETS;
+		     b = find_free_head(tdb, flist_off, b + 1)) {
 			/* Try getting one from list. */
-			off = lock_and_alloc(tdb, tdb->header.v.free_buckets,
-					     size, actual);
+			off = lock_and_alloc(tdb, flist_off,
+					     b, keylen, datalen, want_extra,
+					     hashlow);
 			if (off == TDB_OFF_ERR)
 				return TDB_OFF_ERR;
-			if (off != 0)
+			if (off != 0) {
+				if (b == start_b)
+					add_stat(tdb, alloc_bucket_exact, 1);
+				if (b == TDB_FREE_BUCKETS - 1)
+					add_stat(tdb, alloc_bucket_max, 1);
+				/* Worked?  Stay using this list. */
+				tdb->flist_off = flist_off;
+				tdb->flist = flist;
 				return off;
-			/* FIXME: Coalesce! */
+			}
+			/* Didn't work.  Try next bucket. */
 		}
-	} while (tdb_expand(tdb, 0, size, false) == 0);
-
-	return TDB_OFF_ERR;
-}
-
-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 b;
+		/* Hmm, try next list. */
+		flist_off = next_flist(tdb, flist_off);
+		flist++;
 
-		/* 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. */
-			}
+		if (flist_off == 0) {
+			wrapped = true;
+			flist_off = first_flist(tdb);
+			flist = 0;
 		}
-
-		/* 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;
 	}
+
 	return 0;
 }
 
-int set_header(struct tdb_context *tdb,
-	       struct tdb_used_record *rec,
-	       uint64_t keylen, uint64_t datalen,
-	       uint64_t actuallen, uint64_t hash)
+int set_used_header(struct tdb_context *tdb,
+		    struct tdb_used_record *rec,
+		    uint64_t keylen, uint64_t datalen,
+		    uint64_t actuallen, unsigned hashlow)
 {
 	uint64_t keybits = (fls64(keylen) + 1) / 2;
 
-	/* Use top bits of hash, so it's independent of hash table size. */
-	rec->magic_and_meta
-		= (actuallen - (keylen + datalen))
-		| ((hash >> 53) << 32)
+	/* Use bottom bits of hash, so it's independent of hash table size. */
+	rec->magic_and_meta = (hashlow & ((1 << 11)-1))
+		| ((actuallen - (keylen + datalen)) << 11)
 		| (keybits << 43)
 		| (TDB_MAGIC << 48);
 	rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
@@ -477,9 +567,8 @@ 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->ecode = TDB_ERR_IO;
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "Could not encode k=%llu,d=%llu,a=%llu\n",
+		tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+			 "Could not encode k=%llu,d=%llu,a=%llu",
 			 (long long)keylen, (long long)datalen,
 			 (long long)actuallen);
 		return -1;
@@ -487,224 +576,72 @@ int set_header(struct tdb_context *tdb,
 	return 0;
 }
 
-static tdb_len_t adjust_size(size_t keylen, size_t datalen, bool growing)
+/* Expand the database. */
+static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
-	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);
-}
+	uint64_t old_size;
+	tdb_len_t wanted;
 
-/* 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 need room for the record header too. */
+	wanted = sizeof(struct tdb_used_record) + size;
 
-	/* We don't want header to change during this! */
-	assert(tdb->header_uptodate);
+	/* 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_DEBUG_ERROR,
+			   "tdb_expand: must hold lock during expand");
+		return -1;
+	}
 
-	size = adjust_size(keylen, datalen, growing);
+	/* 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);
 
-	off = get_free(tdb, size, &actual);
-	if (unlikely(off == TDB_OFF_ERR || off == 0))
-		return off;
+	/* Only one person can expand file at a time. */
+	if (tdb_lock_expand(tdb, F_WRLCK) != 0)
+		return -1;
 
-	/* 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;
+	/* 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) {
+		tdb_unlock_expand(tdb, F_WRLCK);
+		return 0;
 	}
 
-	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)
-{
-	/* 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;
+	if (tdb->methods->expand_file(tdb, wanted) == -1) {
+		tdb_unlock_expand(tdb, F_WRLCK);
+		return -1;
 	}
 
-	/* FIXME: Put stats in tdb_context or examine db itself! */
-	/* It's fairly cheap to do as we expand database. */
-	return true;
-}
+	/* We need to drop this lock before adding free record. */
+	tdb_unlock_expand(tdb, F_WRLCK);
 
-static bool zones_happy(struct tdb_context *tdb)
-{
-	/* FIXME: look at distribution of zones. */
-	return true;
+	add_stat(tdb, expands, 1);
+	return add_free_record(tdb, old_size, wanted);
 }
 
-/* Expand the database. */
-int tdb_expand(struct tdb_context *tdb, tdb_len_t klen, tdb_len_t dlen,
-	       bool growing)
+/* 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)
 {
-	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;
-	
-	/* We need room for the record header too. */
-	needed = sizeof(struct tdb_used_record)
-		+ adjust_size(klen, dlen, growing);
-
-	/* FIXME: this is overkill.  An expand lock? */
-	if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
-		return -1;
-
-	/* Someone may have expanded for us. */
-	if (update_header(tdb))
-		goto success;
-
-	/* Make sure we have the latest size. */
-	tdb->methods->oob(tdb, tdb->map_size + 1, true);
-
-	/* 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;
-		}
-	}
+	tdb_off_t off;
 
-	/* 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++;
-		}
+	/* We can't hold pointers during this: we could unmap! */
+	assert(!tdb->direct_access);
 
-		/* We expand by enough zones to meet the need. */
-		add = (needed + (1ULL << new_zone_bits)-1)
-			& ~((1ULL << new_zone_bits)-1);
-	}
+	for (;;) {
+		off = get_free(tdb, keylen, datalen, growing, hash);
+		if (likely(off != 0))
+			break;
 
-	/* Updates tdb->map_size. */
-	if (tdb->methods->expand_file(tdb, tdb->map_size, add) == -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)
-		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)
-		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;
-		}
+		if (tdb_expand(tdb, adjust_size(keylen, datalen)))
+			return TDB_OFF_ERR;
 	}
 
-
-	/* 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;
-
-	/* Add the rest as a new free record. */
-	if (add_free_record(tdb, tdb->map_size - add, add) == -1)
-		goto fail_release;
-
-	/* 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);
-	return 0;
-
-fail_release:
-	tdb_access_release(tdb, oldf);
-fail:
-	tdb_allrecord_unlock(tdb, F_WRLCK);
-	return -1;
+	return off;
 }