X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=cd9a332abe425407cf27b5e35a49db912e3a92b3;hp=9af6f5d04ebf3afb6673fcaf11694e290e52d6df;hb=5c4a21ab03a428373e7659b9606facf85dcbe17b;hpb=1a24a8708494668c07e5c02284bfc2ef3b09603b

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index 9af6f5d0..cd9a332a 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -1,7 +1,7 @@
- /* 
+ /*
    Trivial Database 2: free list/block handling
    Copyright (C) Rusty Russell 2010
-   
+
    This library is free software; you can redistribute it and/or
    modify it under the terms of the GNU Lesser General Public
    License as published by the Free Software Foundation; either
@@ -17,425 +17,633 @@
 */
 #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))
-
 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(unsigned int zone_bits, 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 > BUCKETS_FOR_ZONE(zone_bits)))
-		bucket = BUCKETS_FOR_ZONE(zone_bits);
+	if (unlikely(bucket >= TDB_FREE_BUCKETS))
+		bucket = TDB_FREE_BUCKETS - 1;
 	return bucket;
 }
 
-/* Subtract 1-byte tailer and header.  Then round up to next power of 2. */
-static unsigned max_zone_bits(struct tdb_context *tdb)
+tdb_off_t first_ftable(struct tdb_context *tdb)
 {
-	return fls64(tdb->map_size-1-sizeof(struct tdb_header)-1) + 1;
+	return tdb_read_off(tdb, offsetof(struct tdb_header, free_table));
 }
 
-/* Start by using a random zone to spread the load: returns the offset. */
-static uint64_t random_zone(struct tdb_context *tdb)
+tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable)
 {
-	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;
+	return tdb_read_off(tdb, ftable + offsetof(struct tdb_freetable,next));
+}
 
-	/* 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;
+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;
+	tdb_off_t off;
 
-		if (tdb_read_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1) 
-			return TDB_OFF_ERR;
+	tdb->ftable_off = off = first_ftable(tdb);
+	tdb->ftable = 0;
 
-		if (zhdr.zone_bits == half_bits)
+	while (off) {
+		if (TDB_OFF_IS_ERR(off)) {
 			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;
-}
+		rnd = random();
+		if (rnd >= max) {
+			tdb->ftable_off = off;
+			tdb->ftable = count;
+			max = rnd;
+		}
 
-int tdb_zone_init(struct tdb_context *tdb)
-{
-	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;
+		off = next_ftable(tdb, off);
+		count++;
+	}
+	return TDB_SUCCESS;
 }
 
-/* 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)
+/* Offset of a given bucket. */
+tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
 {
-	off -= sizeof(struct tdb_header);
-	return (off & ~((1ULL << zone_bits) - 1)) + sizeof(struct tdb_header);
+	return ftable_off + offsetof(struct tdb_freetable, buckets)
+		+ bucket * sizeof(tdb_off_t);
 }
 
-/* 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)
+/* 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)
 {
-	return zone_off
-		+ sizeof(struct free_zone_header)
-		+ bucket * sizeof(tdb_off_t);
+	/* Speculatively search for a non-zero bucket. */
+	return tdb_find_nonzero_off(tdb, bucket_off(ftable_off, 0),
+				    bucket, TDB_FREE_BUCKETS);
 }
 
-/* Returns free_buckets + 1, or list number to search. */
-static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
+static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
 {
-	tdb_off_t b;
+#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;
+	}
 
-	/* Speculatively search for a non-zero bucket. */
-	b = tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, bucket),
-				 BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1
-				 - bucket);
-	return bucket + b;
+	if (first) {
+		tdb_read_convert(tdb, first, &r, sizeof(r));
+		if (frec_prev(&r) != prev)
+			abort();
+	}
+#endif
 }
 
 /* Remove from free bucket. */
-static int remove_from_list(struct tdb_context *tdb,
-			    tdb_off_t b_off, struct tdb_free_record *r)
+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;
-
-	/* Front of list? */
-	if (r->prev == 0) {
-		off = b_off;
+	tdb_off_t off, prev_next, head;
+	enum TDB_ERROR ecode;
+
+	/* 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, 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 {
-		off = r->prev + offsetof(struct tdb_free_record, next);
+		/* r->prev->next = r->next */
+		ecode = tdb_write_off(tdb, off, r->next);
+		if (ecode != TDB_SUCCESS)
+			return ecode;
 	}
-	/* r->prev->next = r->next */
-	if (tdb_write_off(tdb, off, r->next)) {
-		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);
 	}
 
-	if (r->next != 0) {
-		off = r->next + offsetof(struct tdb_free_record, prev);
-		/* r->next->prev = r->prev */
-		if (tdb_write_off(tdb, off, r->prev)) {
-			return -1;
-		}
+#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);
 	}
-	return 0;
+#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,
-			   struct tdb_free_record *new)
+static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
+				      tdb_off_t b_off,
+				      tdb_off_t off,
+				      tdb_len_t len)
 {
-	new->prev = 0;
+	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;
+
 	/* new->next = head. */
-	new->next = tdb_read_off(tdb, b_off);
-	if (new->next == TDB_OFF_ERR)
-		return -1;
+	new.next = tdb_read_off(tdb, b_off);
+	if (TDB_OFF_IS_ERR(new.next)) {
+		return new.next;
+	}
 
-	if (new->next) {
+	/* 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);
+		}
 		/* next->prev = new. */
-		if (tdb_write_off(tdb, new->next
-				  + offsetof(struct tdb_free_record, prev),
-				  off) != 0)
-			return -1;
+		ecode = tdb_write_off(tdb, new.next
+				      + offsetof(struct tdb_free_record,
+						 magic_and_prev),
+				      off | magic);
+		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 */
-	if (tdb_write_off(tdb, b_off, off) != 0)
-		return -1;
+	ecode = tdb_write_off(tdb, b_off, off);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
-	return tdb_write_convert(tdb, off, new, sizeof(*new));
+	return tdb_write_convert(tdb, off, &new, sizeof(new));
 }
 
 /* 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)
+enum TDB_ERROR add_free_record(struct tdb_context *tdb,
+			       tdb_off_t off, tdb_len_t len_with_header,
+			       enum tdb_lock_flags waitflag)
 {
-	struct tdb_free_record new;
 	tdb_off_t b_off;
-	int ret;
+	tdb_len_t len;
+	enum TDB_ERROR ecode;
 
-	assert(len_with_header >= sizeof(new));
-	assert(zone_bits < (1 << 6));
+	assert(len_with_header >= sizeof(struct tdb_free_record));
 
-	new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
-	new.data_len = len_with_header - sizeof(struct tdb_used_record);
+	len = len_with_header - sizeof(struct tdb_used_record);
 
-	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;
+	b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
+	ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
-	ret = enqueue_in_free(tdb, b_off, off, &new);
+	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)
+{
+	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,
-			  unsigned int zone_bits,
-			  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;
+
+	if (want_extra)
+		datalen += datalen / 2;
+	leftover = total_len - adjust_size(keylen, datalen);
 
-	leftover = total_len - needed;
-	if (leftover < sizeof(struct tdb_free_record))
-		leftover = 0;
+	if (leftover < (ssize_t)sizeof(struct tdb_free_record))
+		return 0;
 
-	*actual = total_len - leftover;
+	return leftover;
+}
 
-	if (leftover) {
-		if (add_free_record(tdb, zone_bits,
-				    off + sizeof(used) + *actual,
-				    total_len - needed))
-			return -1;
+static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
+{
+	tdb_off_t off;
+	unsigned int i;
+
+	if (likely(tdb->ftable == ftable))
+		return tdb->ftable_off;
+
+	off = first_ftable(tdb);
+	for (i = 0; i < ftable; i++) {
+		if (TDB_OFF_IS_ERR(off)) {
+			break;
+		}
+		off = next_ftable(tdb, off);
 	}
-	return 0;
+	return off;
 }
 
-/* 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)
+/* 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)
 {
-	struct tdb_free_record pad, *r;
-	tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
+	tdb_off_t end;
+	struct tdb_free_record rec;
+	enum TDB_ERROR ecode;
+
+	tdb->stats.alloc_coalesce_tried++;
+	end = off + sizeof(struct tdb_used_record) + data_len;
 
-	while (end < (zone_off + (1ULL << zone_bits))) {
+	while (end < tdb->file->map_size) {
+		const struct tdb_free_record *r;
 		tdb_off_t nb_off;
+		unsigned ftable, bucket;
 
-		/* FIXME: do tdb_get here and below really win? */
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
-		if (!r)
+		r = tdb_access_read(tdb, end, sizeof(*r), true);
+		if (TDB_PTR_IS_ERR(r)) {
+			ecode = TDB_PTR_ERR(r);
 			goto err;
+		}
 
-		if (frec_magic(r) != TDB_FREE_MAGIC)
+		if (frec_magic(r) != TDB_FREE_MAGIC
+		    || frec_ftable(r) == TDB_FTABLE_NONE) {
+			tdb_access_release(tdb, r);
 			break;
+		}
 
-		nb_off = bucket_off(zone_off,
-				    size_to_bucket(zone_bits, r->data_len));
+		ftable = frec_ftable(r);
+		bucket = size_to_bucket(frec_len(r));
+		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) == -1)
+		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
+		    != TDB_SUCCESS) {
+			tdb->stats.alloc_coalesce_lockfail++;
 			break;
+		}
 
 		/* Now we have lock, re-check. */
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
-		if (!r) {
+		ecode = tdb_read_convert(tdb, end, &rec, sizeof(rec));
+		if (ecode != TDB_SUCCESS) {
 			tdb_unlock_free_bucket(tdb, nb_off);
 			goto err;
 		}
 
-		if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
+		if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
+			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)) {
+			tdb->stats.alloc_coalesce_race++;
 			tdb_unlock_free_bucket(tdb, nb_off);
 			break;
 		}
 
-		if (remove_from_list(tdb, nb_off, r) == -1) {
+		/* 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) + r->data_len;
+		end += sizeof(struct tdb_used_record) + frec_len(&rec);
 		tdb_unlock_free_bucket(tdb, nb_off);
+		tdb->stats.alloc_coalesce_num_merged++;
 	}
 
 	/* 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)
+	/* 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));
+	if (ecode != TDB_SUCCESS) {
 		goto err;
+	}
 
-	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);
+	if (frec_len(&rec) != data_len) {
+		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, r) == -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. */
-	tdb_unlock_free_bucket(tdb, b_off);
+	/* Try locking violation first... */
+	ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT);
+	if (ecode != TDB_SUCCESS) {
+		/* 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;
+		}
+
+		tdb->stats.alloc_coalesce_succeeded++;
+		tdb_unlock_free_bucket(tdb, b_off);
+
+		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);
+	}
 
-	if (add_free_record(tdb, zone_bits, off, end - off) == -1)
-		return -1;
-	return 1;
+	/* Return usable length. */
+	return end - off - sizeof(struct tdb_used_record);
 
 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. */
 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
-				tdb_off_t zone_off,
-				unsigned zone_bits,
+				tdb_off_t ftable_off,
 				tdb_off_t bucket,
-				size_t size,
-				tdb_len_t *actual)
+				size_t keylen, size_t datalen,
+				bool want_extra,
+				unsigned magic,
+				unsigned hashlow)
 {
 	tdb_off_t off, b_off,best_off;
-	struct tdb_free_record pad, best = { 0 }, *r;
+	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;
 
+	tdb->stats.allocs++;
 again:
-	b_off = bucket_off(zone_off, bucket);
+	b_off = bucket_off(ftable_off, bucket);
 
+	/* 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;
+	ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
 	}
 
-	best.data_len = -1ULL;
+	best.ftable_and_len = -1ULL;
 	best_off = 0;
-	/* FIXME: Start with larger multiplier if we're growing. */
-	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))
+	if (TDB_OFF_IS_ERR(off)) {
+		ecode = off;
 		goto unlock_err;
+	}
 
 	while (off) {
-		/* FIXME: Does tdb_get win anything here? */
-		r = tdb_get(tdb, off, &pad, sizeof(*r));
-		if (!r)
+		const struct tdb_free_record *r;
+		tdb_len_t len, coal;
+		tdb_off_t next;
+
+		r = tdb_access_read(tdb, off, sizeof(*r), true);
+		if (TDB_PTR_IS_ERR(r)) {
+			ecode = TDB_PTR_ERR(r);
 			goto unlock_err;
+		}
 
 		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_and_meta);
+			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 (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;
+			coalesce_after_best = false;
 		}
 
-		if (best.data_len < size * multiplier && best_off)
-			goto use_best;
+		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, zone_off, zone_bits, off, b_off,
-				 r->data_len)) {
-		case -1:
+		coal = coalesce(tdb, off, b_off, len, &best_off, &next);
+		if (TDB_OFF_IS_ERR(coal)) {
 			/* This has already unlocked on error. */
-			return -1;
-		case 1:
+			return coal;
+		}
+		if (TDB_OFF_IS_ERR(best_off)) {
 			/* This has unlocked list, restart. */
 			goto again;
 		}
-		off = r->next;
+		if (coal > 0)
+			coalesce_after_best = true;
+		off = next;
 	}
 
 	/* If we found anything at all, use it. */
 	if (best_off) {
-	use_best:
+		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. */
-		if (remove_from_list(tdb, b_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;
-		tdb_unlock_free_bucket(tdb, b_off);
+		}
+
+		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! */
+		ecode = set_header(tdb, &rec, magic, keylen, datalen,
+				   frec_len(&best) - leftover, hashlow);
+		if (ecode != TDB_SUCCESS) {
+			goto unlock_err;
+		}
 
-		if (to_used_record(tdb, zone_bits, best_off, size,
-				   best.data_len, actual)) {
-			return -1;
+		ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
+		if (ecode != TDB_SUCCESS) {
+			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) {
+			tdb->stats.alloc_leftover++;
+			ecode = add_free_record(tdb,
+						best_off + sizeof(rec)
+						+ frec_len(&best) - leftover,
+						leftover, TDB_LOCK_WAIT);
+			if (ecode != TDB_SUCCESS) {
+				best_off = ecode;
+			}
+		}
+		tdb_unlock_free_bucket(tdb, b_off);
+
 		return best_off;
 	}
 
@@ -444,234 +652,188 @@ again:
 
 unlock_err:
 	tdb_unlock_free_bucket(tdb, b_off);
-	return TDB_OFF_ERR;
+	return ecode;
 }
 
-static bool next_zone(struct tdb_context *tdb)
+/* 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 next = tdb->zone_off + (1ULL << tdb->zhdr.zone_bits);
-
-	/* We must have a header. */
-	if (tdb->methods->oob(tdb, next + sizeof(tdb->zhdr), true))
-		return false;
-
-	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 start_zone = tdb->zone_off, off;
+	tdb_off_t off, ftable_off;
+	tdb_off_t start_b, b, ftable;
 	bool wrapped = false;
 
-	while (!wrapped || tdb->zone_off != start_zone) {
-		tdb_off_t b;
-
-		/* Shortcut for really huge allocations... */
-		if ((size >> tdb->zhdr.zone_bits) != 0)
-			continue;
+	/* 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));
 
+	ftable_off = tdb->ftable_off;
+	ftable = tdb->ftable;
+	while (!wrapped || ftable_off != tdb->ftable_off) {
 		/* Start at exact size bucket, and search up... */
-		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)) {
+		for (b = find_free_head(tdb, ftable_off, start_b);
+		     b < TDB_FREE_BUCKETS;
+		     b = find_free_head(tdb, ftable_off, 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)
+			off = lock_and_alloc(tdb, ftable_off,
+					     b, keylen, datalen, want_extra,
+					     magic, hashlow);
+			if (TDB_OFF_IS_ERR(off))
+				return off;
+			if (off != 0) {
+				if (b == start_b)
+					tdb->stats.alloc_bucket_exact++;
+				if (b == TDB_FREE_BUCKETS - 1)
+					tdb->stats.alloc_bucket_max++;
+				/* Worked?  Stay using this list. */
+				tdb->ftable_off = ftable_off;
+				tdb->ftable = ftable;
 				return off;
+			}
 			/* Didn't work.  Try next bucket. */
 		}
 
-		/* Didn't work, try next zone, if it exists. */
-		if (!next_zone(tdb)) {
+		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;
-			tdb->zone_off = sizeof(struct tdb_header);
-			if (tdb_read_convert(tdb, tdb->zone_off,
-					     &tdb->zhdr, sizeof(tdb->zhdr))) {
-				return TDB_OFF_ERR;
+			ftable_off = first_ftable(tdb);
+			if (TDB_OFF_IS_ERR(ftable_off)) {
+				return ftable_off;
 			}
+			ftable = 0;
 		}
 	}
+
 	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,
-	       unsigned int zone_bits)
+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;
 
-	/* Use top bits of hash, so it's independent of hash table size. */
-	rec->magic_and_meta
-		= zone_bits
-		| ((hash >> 59) << 6)
+	/* 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);
+		| ((uint64_t)magic << 48);
 	rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
 
 	/* Encoding can fail on big values. */
 	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",
-			 (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;
 }
 
-static tdb_len_t adjust_size(size_t keylen, size_t datalen, bool growing)
+/* Expand the database. */
+static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
-	tdb_len_t size = keylen + datalen;
+	uint64_t old_size, rec_size, map_size;
+	tdb_len_t wanted;
+	enum TDB_ERROR ecode;
 
-	if (size < MIN_DATA_LEN)
-		size = MIN_DATA_LEN;
+	/* 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)) {
+		return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
+				  "tdb_expand: must hold lock during expand");
+	}
 
-	/* Overallocate if this is coming from an enlarging store. */
-	if (growing)
-		size += datalen / 2;
+	/* Only one person can expand file at a time. */
+	ecode = tdb_lock_expand(tdb, F_WRLCK);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
-	/* Round to next uint64_t boundary. */
-	return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
-}
+	/* Someone else may have expanded the file, so retry. */
+	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;
+	}
 
-/* 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;
+	/* 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;
+	}
 
-	/* We don't want header to change during this! */
-	assert(tdb->header_uptodate);
+	/* 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;
+	}
 
-	size = adjust_size(keylen, datalen, growing);
+	if (map_size > rec_size) {
+		wanted = map_size;
+	} else {
+		wanted = rec_size;
+	}
 
-	off = get_free(tdb, size, &actual);
-	if (unlikely(off == TDB_OFF_ERR || off == 0))
-		return off;
+	/* We need room for the record header too. */
+	wanted = adjust_size(0, sizeof(struct tdb_used_record) + wanted);
 
-	/* 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;
+	ecode = tdb->methods->expand_file(tdb, wanted);
+	if (ecode != TDB_SUCCESS) {
+		tdb_unlock_expand(tdb, F_WRLCK);
+		return ecode;
 	}
 
-	if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
-		return TDB_OFF_ERR;
-	
-	return off;
-}
+	/* 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;
+	tdb->stats.expands++;
+	return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT);
 }
 
-/* Assume we want buckets up to the comfort factor. */
-static tdb_len_t overhead(unsigned int zone_bits)
-{
-	return sizeof(struct free_zone_header)
-		+ (BUCKETS_FOR_ZONE(zone_bits) + 1) * sizeof(tdb_off_t);
-}
-
-/* Expand the database (by adding a zone). */
-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, unsigned magic, bool growing)
 {
-	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. */
-	wanted = sizeof(struct tdb_used_record)
-		+ (adjust_size(klen, dlen, growing)<<TDB_COMFORT_FACTOR_BITS);
-
-	/* Only one person can expand file at a time. */
-	if (tdb_lock_expand(tdb, F_WRLCK) != 0)
-		return -1;
-
-	/* 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;
 
-	/* zone bits tailer char is protected by EXPAND lock. */
-	if (tdb->methods->read(tdb, old_size - 1, &zone_bits, 1) == -1)
-		goto fail;
+	/* We can't hold pointers during this: we could unmap! */
+	assert(!tdb->direct_access);
 
-	/* 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))
+		enum TDB_ERROR ecode;
+		off = get_free(tdb, keylen, datalen, growing, magic, hash);
+		if (likely(off != 0))
 			break;
-		zone_bits++;
-	}
-
-	zhdr.zone_bits = zone_bits;
-	num_buckets = BUCKETS_FOR_ZONE(zone_bits);
 
-	if (tdb->methods->expand_file(tdb, 1ULL << zone_bits) == -1)
-		goto fail;
-
-	/* Write new tailer. */
-	if (tdb->methods->write(tdb, tdb->map_size - 1, &zone_bits, 1) == -1)
-		goto fail;
-
-	/* Write new zone header (just before old tailer). */
-	off = old_size - 1;
-	if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
-		goto fail;
-
-	/* 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);
-
-	/* Now add the rest as our free record. */
-	if (add_free_record(tdb, zone_bits, off, tdb->map_size-1-off) == -1)
-		goto fail;
-
-success:
-	tdb_unlock_expand(tdb, F_WRLCK);
-	return 0;
+		ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
+		if (ecode != TDB_SUCCESS) {
+			return ecode;
+		}
+	}
 
-fail:
-	tdb_unlock_expand(tdb, F_WRLCK);
-	return -1;
+	return off;
 }