alloc: speed up tiny allocator
[ccan] / ccan / alloc / alloc.c
index d5accb40a581308b60ca76e233ee73968bcf9a2b..8e0b72aa0432227990f7b053f6620020fd53dcac 100644 (file)
 #include <assert.h>
 #include <stdlib.h>
 #include "alloc.h"
+#include "bitops.h"
+#include "tiny.h"
 #include <ccan/build_assert/build_assert.h>
+#include <ccan/likely/likely.h>
 #include <ccan/alignof/alignof.h>
+#include <ccan/short_types/short_types.h>
 #include "config.h"
 
-/* FIXME: We assume getpagesize() doesnt change.  Remapping file with
- * different pagesize should still work. */
+/*
+   Inspired by (and parts taken from) Andrew Tridgell's alloc_mmap:
+   http://samba.org/~tridge/junkcode/alloc_mmap/
+
+   Copyright (C) Andrew Tridgell 2007
+   
+   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
+   version 2 of the License, or (at your option) any later version.
+
+   This library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with this library; if not, write to the Free Software
+   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
+/* We divide the pool into this many large pages (nearest power of 2) */
+#define MAX_LARGE_PAGES (256UL)
+
+/* 32 small pages == 1 large page. */
+#define BITS_FROM_SMALL_TO_LARGE_PAGE 5
+
+#define MAX_SMALL_PAGES (MAX_LARGE_PAGES << BITS_FROM_SMALL_TO_LARGE_PAGE)
+
+/* Smallest pool size for this scheme: 128-byte small pages.  That's
+ * 9/13% overhead for 32/64 bit. */
+#define MIN_USEFUL_SIZE (MAX_SMALL_PAGES * 128)
+
+/* Every 4 buckets, we jump up a power of 2. ...8 10 12 14 16 20 24 28 32... */
+#define INTER_BUCKET_SPACE 4
+
+#define SMALL_PAGES_PER_LARGE_PAGE (1 << BITS_FROM_SMALL_TO_LARGE_PAGE)
+
+/* FIXME: Figure this out properly. */
+#define MAX_SIZE (1 << 30)
+
+/* How few object to fit in a page before using a larger one? (8) */
+#define MAX_PAGE_OBJECT_ORDER  3
 
-/* FIXME: Doesn't handle non-page-aligned poolsize. */
+#define BITS_PER_LONG (sizeof(long) * CHAR_BIT)
 
-/* FIXME: Reduce. */
-#define MIN_SIZE (getpagesize() * 2)
+struct bucket_state {
+       u32 elements_per_page;
+       u16 page_list;
+       u16 full_list;
+};
+
+struct header {
+       /* Bitmap of which pages are large. */
+       unsigned long pagesize[MAX_LARGE_PAGES / BITS_PER_LONG];
+
+       /* List of unused small/large pages. */
+       u16 small_free_list;
+       u16 large_free_list;
 
-/* What's the granularity of sub-page allocs? */
-#define BITMAP_GRANULARITY 4
+       /* List of huge allocs. */
+       unsigned long huge;
+
+       /* This is less defined: we have two buckets for each power of 2 */
+       struct bucket_state bs[1];
+};
+
+struct huge_alloc {
+       unsigned long next, prev;
+       unsigned long off, len;
+};
+
+struct page_header {
+       u16 next, prev;
+       /* FIXME: We can just count all-0 and all-1 used[] elements. */
+       unsigned elements_used : 25;
+       unsigned bucket : 7;
+       unsigned long used[1]; /* One bit per element. */
+};
 
-/* File layout:
+/*
+ * Every 4 buckets, the size doubles.
+ * Between buckets, sizes increase linearly.
  *
- *  file := pagestates pad uniform-cache metadata
- *  pagestates := pages * 2-bits-per-page
- *  pad := pad to next ALIGNOF(metaheader)
+ * eg. bucket 40 = 2^10                        = 1024
+ *     bucket 41 = 2^10 + 2^10*4       = 1024 + 256
+ *     bucket 42 = 2^10 + 2^10*4       = 1024 + 512
+ *     bucket 43 = 2^10 + 2^10*4       = 1024 + 768
+ *     bucket 45 = 2^11                        = 2048
  *
- *  metadata := metalen next-ptr metabits
- *  metabits := freeblock | bitblock | uniformblock
- *  freeblock := FREE +
- *  bitblock := BITMAP + 2-bits-per-bit-in-page + pad-to-byte
- *  uniformblock := UNIFORM + 14-bit-byte-len + bits + pad-to-byte
+ * Care is taken to handle low numbered buckets, at cost of overflow.
  */
-#define UNIFORM_CACHE_NUM 16
-struct uniform_cache
+static unsigned long bucket_to_size(unsigned int bucket)
 {
-       uint16_t size[UNIFORM_CACHE_NUM];
-       /* These could be u32 if we're prepared to limit size. */
-       unsigned long page[UNIFORM_CACHE_NUM];
-};
+       unsigned long base = 1 << (bucket / INTER_BUCKET_SPACE);
+       return base + ((bucket % INTER_BUCKET_SPACE)
+                      << (bucket / INTER_BUCKET_SPACE))
+               / INTER_BUCKET_SPACE;
+}
 
-struct metaheader
+/*
+ * Say size is 10.
+ *   fls(size/2) == 3.  1 << 3 == 8, so we're 2 too large, out of a possible
+ * 8 too large.  That's 1/4 of the way to the next power of 2 == 1 bucket.
+ *
+ * We make sure we round up.  Note that this fails on 32 bit at size
+ * 1879048193 (around bucket 120).
+ */
+static unsigned int size_to_bucket(unsigned long size)
 {
-       /* Next meta header, or 0 */
-       unsigned long next;
-       /* Bits start here. */
-};
+       unsigned int base = fls(size/2);
+       unsigned long overshoot;
 
-/* Assumes a is a power of two. */
-static unsigned long align_up(unsigned long x, unsigned long a)
-{
-       return (x + a - 1) & ~(a - 1);
+       overshoot = size - (1 << base);
+       return base * INTER_BUCKET_SPACE
+               + ((overshoot * INTER_BUCKET_SPACE + (1 << base)-1) >> base);
 }
 
-static unsigned long align_down(unsigned long x, unsigned long a)
+static unsigned int small_page_bits(unsigned long poolsize)
 {
-       return x & ~(a - 1);
+       return fls(poolsize / MAX_SMALL_PAGES / 2);
 }
 
-static unsigned long div_up(unsigned long x, unsigned long a)
+static struct page_header *from_pgnum(struct header *head,
+                                     unsigned long pgnum,
+                                     unsigned sp_bits)
 {
-       return (x + a - 1) / a;
+       return (struct page_header *)((char *)head + (pgnum << sp_bits));
 }
 
-/* It turns out that we spend a lot of time dealing with bit pairs.
- * These routines manipulate them.
- */
-static uint8_t get_bit_pair(const uint8_t *bits, unsigned long index)
+static u16 to_pgnum(struct header *head, void *p, unsigned sp_bits)
 {
-       return bits[index * 2 / CHAR_BIT] >> (index * 2 % CHAR_BIT) & 3;
+       return ((char *)p - (char *)head) >> sp_bits;
 }
 
-static void set_bit_pair(uint8_t *bits, unsigned long index, uint8_t val)
+static size_t used_size(unsigned int num_elements)
 {
-       bits[index * 2 / CHAR_BIT] &= ~(3 << (index * 2 % CHAR_BIT));
-       bits[index * 2 / CHAR_BIT] |= (val << (index * 2 % CHAR_BIT));
+       return align_up(num_elements, BITS_PER_LONG) / CHAR_BIT;
 }
 
-/* This is used for page states and subpage allocations */
-enum alloc_state
+/*
+ * We always align the first entry to the lower power of 2.
+ * eg. the 12-byte bucket gets 8-byte aligned.  The 4096-byte bucket
+ * gets 4096-byte aligned.
+ */
+static unsigned long page_header_size(unsigned int align_bits,
+                                     unsigned long num_elements)
 {
-       FREE,
-       TAKEN,
-       TAKEN_START,
-       SPECIAL,        /* Sub-page allocation for page states. */
-};
+       unsigned long size;
 
-/* The types for subpage metadata. */
-enum sub_metadata_type
-{
-       /* FREE is same as alloc state */
-       BITMAP = 1, /* bitmap allocated page */
-       UNIFORM, /* uniform size allocated page */
-};
+       size = sizeof(struct page_header)
+               - sizeof(((struct page_header *)0)->used)
+               + used_size(num_elements);
+       return align_up(size, 1 << align_bits);
+}
 
-/* Page states are represented by bitpairs, at the start of the pool. */
-#define BITS_PER_PAGE 2
+static void add_to_list(struct header *head,
+                       u16 *list, struct page_header *ph, unsigned sp_bits)
+{
+       unsigned long h = *list, offset = to_pgnum(head, ph, sp_bits);
 
-/* How much metadata info per byte? */
-#define METADATA_PER_BYTE (CHAR_BIT / 2)
+       ph->next = h;
+       if (h) {
+               struct page_header *prev = from_pgnum(head, h, sp_bits);
+               assert(prev->prev == 0);
+               prev->prev = offset;
+       }
+       *list = offset;
+       ph->prev = 0;
+}
 
-static uint8_t *get_page_statebits(const void *pool)
+static void del_from_list(struct header *head,
+                         u16 *list, struct page_header *ph, unsigned sp_bits)
 {
-       return (uint8_t *)pool + sizeof(struct uniform_cache);
+       /* Front of list? */
+       if (ph->prev == 0) {
+               *list = ph->next;
+       } else {
+               struct page_header *prev = from_pgnum(head, ph->prev, sp_bits);
+               prev->next = ph->next;
+       }
+       if (ph->next != 0) {
+               struct page_header *next = from_pgnum(head, ph->next, sp_bits);
+               next->prev = ph->prev;
+       }
 }
 
-static enum alloc_state get_page_state(const void *pool, unsigned long page)
+static u16 pop_from_list(struct header *head,
+                                  u16 *list,
+                                  unsigned int sp_bits)
 {
-       return get_bit_pair(get_page_statebits(pool), page);
+       u16 h = *list;
+       struct page_header *ph = from_pgnum(head, h, sp_bits);
+
+       if (likely(h)) {
+               *list = ph->next;
+               if (*list)
+                       from_pgnum(head, *list, sp_bits)->prev = 0;
+       }
+       return h;
 }
 
-static void set_page_state(void *pool, unsigned long page, enum alloc_state s)
+static void add_to_huge_list(struct header *head, struct huge_alloc *ha)
 {
-       set_bit_pair(get_page_statebits(pool), page, s);
+       unsigned long h = head->huge;
+       unsigned long offset = (char *)ha - (char *)head;
+
+       ha->next = h;
+       if (h) {
+               struct huge_alloc *prev = (void *)((char *)head + h);
+               assert(prev->prev == 0);
+               prev->prev = offset;
+       }
+       head->huge = offset;
+       ha->prev = 0;
 }
 
-/* The offset of metadata for a subpage allocation is found at the end
- * of the subpage */
-#define SUBPAGE_METAOFF (getpagesize() - sizeof(unsigned long))
-
-/* This is the length of metadata in bits.  It consists of two bits
- * for every BITMAP_GRANULARITY of usable bytes in the page, then two
- * bits for the tailer.. */
-#define BITMAP_METABITLEN                                              \
-    ((div_up(SUBPAGE_METAOFF, BITMAP_GRANULARITY) + 1) * BITS_PER_PAGE)
-
-/* This is the length in bytes. */
-#define BITMAP_METALEN (div_up(BITMAP_METABITLEN, CHAR_BIT))
-
-static struct metaheader *first_mheader(void *pool, unsigned long poolsize)
+static void del_from_huge(struct header *head, struct huge_alloc *ha)
 {
-       unsigned int pagestatelen;
-
-       pagestatelen = align_up(div_up(poolsize/getpagesize() * BITS_PER_PAGE,
-                                      CHAR_BIT),
-                               ALIGNOF(struct metaheader));
-       return (struct metaheader *)(get_page_statebits(pool) + pagestatelen);
+       /* Front of list? */
+       if (ha->prev == 0) {
+               head->huge = ha->next;
+       } else {
+               struct huge_alloc *prev = (void *)((char *)head + ha->prev);
+               prev->next = ha->next;
+       }
+       if (ha->next != 0) {
+               struct huge_alloc *next = (void *)((char *)head + ha->next);
+               next->prev = ha->prev;
+       }
 }
 
-static struct metaheader *next_mheader(void *pool, struct metaheader *mh)
+static void add_small_page_to_freelist(struct header *head,
+                                      struct page_header *ph,
+                                      unsigned int sp_bits)
 {
-       if (!mh->next)
-               return NULL;
-
-       return (struct metaheader *)((char *)pool + mh->next);
+       add_to_list(head, &head->small_free_list, ph, sp_bits);
 }
 
-static unsigned long pool_offset(void *pool, void *p)
+static void add_large_page_to_freelist(struct header *head,
+                                      struct page_header *ph,
+                                      unsigned int sp_bits)
 {
-       return (char *)p - (char *)pool;
+       add_to_list(head, &head->large_free_list, ph, sp_bits);
 }
 
-void alloc_init(void *pool, unsigned long poolsize)
+static void add_to_bucket_list(struct header *head,
+                              struct bucket_state *bs,
+                              struct page_header *ph,
+                              unsigned int sp_bits)
 {
-       /* FIXME: Alignment assumptions about pool. */
-       unsigned long len, i;
-       struct metaheader *mh;
-
-       if (poolsize < MIN_SIZE)
-               return;
-
-       mh = first_mheader(pool, poolsize);
-
-       /* Mark all page states FREE, all uniform caches zero, and all of
-        * metaheader bitmap which takes rest of first page. */
-       len = align_up(pool_offset(pool, mh + 1), getpagesize());
-       BUILD_ASSERT(FREE == 0);
-       memset(pool, 0, len);
-
-       /* Mark the pagestate and metadata page(s) allocated. */
-       set_page_state(pool, 0, TAKEN_START);
-       for (i = 1; i < div_up(len, getpagesize()); i++)
-               set_page_state(pool, i, TAKEN);
+       add_to_list(head, &bs->page_list, ph, sp_bits);
 }
 
-/* Two bits per element, representing page states.  Returns 0 on fail.
- * off is used to allocate from subpage bitmaps, which use the first 2
- * bits as the type, so the real bitmap is offset by 1. */
-static unsigned long alloc_from_bitmap(uint8_t *bits, unsigned long off,
-                                      unsigned long elems,
-                                      unsigned long want, unsigned long align)
+static void del_from_bucket_list(struct header *head,
+                                struct bucket_state *bs,
+                                struct page_header *ph,
+                                unsigned int sp_bits)
 {
-       long i;
-       unsigned long free;
-
-       free = 0;
-       /* We allocate from far end, to increase ability to expand metadata. */
-       for (i = elems - 1; i >= 0; i--) {
-               switch (get_bit_pair(bits, off+i)) {
-               case FREE:
-                       if (++free >= want) {
-                               unsigned long j;
-
-                               /* They might ask for large alignment. */
-                               if (align && i % align)
-                                       continue;
-
-                               set_bit_pair(bits, off+i, TAKEN_START);
-                               for (j = i+1; j < i + want; j++)
-                                       set_bit_pair(bits, off+j, TAKEN);
-                               return off+i;
-                       }
-                       break;
-               case SPECIAL:
-               case TAKEN_START:
-               case TAKEN:
-                       free = 0;
-                       break;
-               }
-       }
-
-       return 0;
+       del_from_list(head, &bs->page_list, ph, sp_bits);
 }
 
-static unsigned long alloc_get_pages(void *pool, unsigned long poolsize,
-                                    unsigned long pages, unsigned long align)
+static void del_from_bucket_full_list(struct header *head,
+                                     struct bucket_state *bs,
+                                     struct page_header *ph,
+                                     unsigned int sp_bits)
 {
-       return alloc_from_bitmap(get_page_statebits(pool),
-                                0, poolsize / getpagesize(), pages,
-                                align / getpagesize());
+       del_from_list(head, &bs->full_list, ph, sp_bits);
 }
 
-/* Offset to metadata is at end of page. */
-static unsigned long *metadata_off(void *pool, unsigned long page)
+static void add_to_bucket_full_list(struct header *head,
+                                   struct bucket_state *bs,
+                                   struct page_header *ph,
+                                   unsigned int sp_bits)
 {
-       return (unsigned long *)
-               ((char *)pool + (page+1)*getpagesize() - sizeof(unsigned long));
+       add_to_list(head, &bs->full_list, ph, sp_bits);
 }
 
-static uint8_t *get_page_metadata(void *pool, unsigned long page)
+static void clear_bit(unsigned long bitmap[], unsigned int off)
 {
-       return (uint8_t *)pool + *metadata_off(pool, page);
+       bitmap[off / BITS_PER_LONG] &= ~(1 << (off % BITS_PER_LONG));
 }
 
-static void set_page_metadata(void *pool, unsigned long page, uint8_t *meta)
+static bool test_bit(const unsigned long bitmap[], unsigned int off)
 {
-       *metadata_off(pool, page) = meta - (uint8_t *)pool;
+       return bitmap[off / BITS_PER_LONG] & (1 << (off % BITS_PER_LONG));
 }
 
-static unsigned long sub_page_alloc(void *pool, unsigned long page,
-                                   unsigned long size, unsigned long align)
+static void set_bit(unsigned long bitmap[], unsigned int off)
 {
-       uint8_t *bits = get_page_metadata(pool, page);
-       unsigned long i;
-       enum sub_metadata_type type;
-
-       type = get_bit_pair(bits, 0);
-
-       /* If this is a uniform page, we can't allocate from it. */
-       if (type == UNIFORM)
-               return 0;
-
-       assert(type == BITMAP);
-
-       /* We use a standart bitmap, but offset because of that BITMAP
-        * header. */
-       i = alloc_from_bitmap(bits, 1, SUBPAGE_METAOFF/BITMAP_GRANULARITY,
-                             div_up(size, BITMAP_GRANULARITY),
-                             align / BITMAP_GRANULARITY);
+       bitmap[off / BITS_PER_LONG] |= (1 << (off % BITS_PER_LONG));
+}
 
-       /* Can't allocate? */
-       if (i == 0)
-               return 0;
+/* There must be a bit to be found. */
+static unsigned int find_free_bit(const unsigned long bitmap[])
+{
+       unsigned int i;
 
-       /* i-1 because of the header. */
-       return page*getpagesize() + (i-1)*BITMAP_GRANULARITY;
+       for (i = 0; bitmap[i] == -1UL; i++);
+       return (i*BITS_PER_LONG) + ffsl(~bitmap[i]) - 1;
 }
 
-/* We look at the page states to figure out where the allocation for this
- * metadata ends. */
-static unsigned long get_metalen(void *pool, unsigned long poolsize,
-                                struct metaheader *mh)
+/* How many elements can we fit in a page? */
+static unsigned long elements_per_page(unsigned long align_bits,
+                                      unsigned long esize,
+                                      unsigned long psize)
 {
-       unsigned long i, first, pages = poolsize / getpagesize();
-
-       first = pool_offset(pool, mh + 1)/getpagesize();
+       unsigned long num, overhead;
 
-       for (i = first + 1; i < pages && get_page_state(pool,i) == TAKEN; i++);
+       /* First approximation: no extra room for bitmap. */
+       overhead = align_up(sizeof(struct page_header), 1 << align_bits);
+       num = (psize - overhead) / esize;
 
-       return i * getpagesize() - pool_offset(pool, mh + 1);
+       while (page_header_size(align_bits, num) + esize * num > psize)
+               num--;
+       return num;
 }
 
-static unsigned int uniform_metalen(unsigned int usize)
+static bool large_page_bucket(unsigned int bucket, unsigned int sp_bits)
 {
-       unsigned int metalen;
-
-       assert(usize < (1 << 14));
+       unsigned long max_smallsize;
 
-       /* Two bits for the header, 14 bits for size, then one bit for each
-        * element the page can hold.  Round up to number of bytes. */
-       metalen = div_up(2 + 14 + SUBPAGE_METAOFF / usize, CHAR_BIT);
+       /* Note: this doesn't take into account page header. */
+       max_smallsize = (1UL << sp_bits) >> MAX_PAGE_OBJECT_ORDER;
 
-       /* To ensure metaheader is always aligned, round bytes up. */
-       metalen = align_up(metalen, ALIGNOF(struct metaheader));
-
-       return metalen;
+       return bucket_to_size(bucket) > max_smallsize;
 }
 
-static unsigned int decode_usize(uint8_t *meta)
+static unsigned int max_bucket(unsigned int lp_bits)
 {
-       return ((unsigned)meta[1] << (CHAR_BIT-2)) | (meta[0] >> 2);
+       return (lp_bits - MAX_PAGE_OBJECT_ORDER) * INTER_BUCKET_SPACE;
 }
 
-static void encode_usize(uint8_t *meta, unsigned int usize)
+void alloc_init(void *pool, unsigned long poolsize)
 {
-       meta[0] = (UNIFORM | (usize << 2));
-       meta[1] = (usize >> (CHAR_BIT - 2));
-}
+       struct header *head = pool;
+       struct page_header *ph;
+       unsigned int lp_bits, sp_bits, num_buckets;
+       unsigned long header_size, i;
 
-static uint8_t *alloc_metaspace(void *pool, unsigned long poolsize,
-                               struct metaheader *mh, unsigned long bytes,
-                               enum sub_metadata_type type)
-{
-       uint8_t *meta = (uint8_t *)(mh + 1);
-       unsigned long free = 0, len, i, metalen;
-
-       metalen = get_metalen(pool, poolsize, mh);
-
-       /* Walk through metadata looking for free. */
-       for (i = 0; i < metalen * METADATA_PER_BYTE; i += len) {
-               switch (get_bit_pair(meta, i)) {
-               case FREE:
-                       len = 1;
-                       free++;
-                       if (free == bytes * METADATA_PER_BYTE) {
-                               /* Mark this as a bitmap. */
-                               set_bit_pair(meta, i - free + 1, type);
-                               return meta + (i - free + 1)/METADATA_PER_BYTE;
-                       }
-                       break;
-               case BITMAP:
-                       /* Skip over this allocated part. */
-                       len = BITMAP_METALEN * METADATA_PER_BYTE;
-                       free = 0;
-                       break;
-               case UNIFORM:
-                       /* Figure metalen given usize. */
-                       len = decode_usize(meta + i / METADATA_PER_BYTE);
-                       len = uniform_metalen(len) * METADATA_PER_BYTE;
-                       free = 0;
-                       break;
-               default:
-                       assert(0);
-                       return NULL;
-               }
+       if (poolsize < MIN_USEFUL_SIZE) {
+               tiny_alloc_init(pool, poolsize);
+               return;
        }
-       return NULL;
-}
 
-/* We need this many bytes of metadata. */
-static uint8_t *new_metadata(void *pool, unsigned long poolsize,
-                            unsigned long bytes, enum sub_metadata_type type)
-{
-       struct metaheader *mh, *newmh;
-       unsigned long page;
-       uint8_t *meta;
+       /* We rely on page numbers fitting in 16 bit. */
+       BUILD_ASSERT(MAX_SMALL_PAGES < 65536);
+       
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
 
-       for (mh = first_mheader(pool,poolsize); mh; mh = next_mheader(pool,mh))
-               if ((meta = alloc_metaspace(pool, poolsize, mh, bytes, type)))
-                       return meta;
+       num_buckets = max_bucket(lp_bits);
 
-       /* No room for metadata?  Can we expand an existing one? */
-       for (mh = first_mheader(pool,poolsize); mh; mh = next_mheader(pool,mh)){
-               unsigned long nextpage;
+       head = pool;
+       header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
 
-               /* We start on this page. */
-               nextpage = pool_offset(pool, (char *)(mh+1))/getpagesize();
-               /* Iterate through any other pages we own. */
-               while (get_page_state(pool, ++nextpage) == TAKEN);
+       memset(head, 0, header_size);
+       for (i = 0; i < num_buckets; i++) {
+               unsigned long pagesize;
 
-               /* Now, can we grab that page? */
-               if (get_page_state(pool, nextpage) != FREE)
-                       continue;
+               if (large_page_bucket(i, sp_bits))
+                       pagesize = 1UL << lp_bits;
+               else
+                       pagesize = 1UL << sp_bits;
 
-               /* OK, expand metadata, do it again. */
-               set_page_state(pool, nextpage, TAKEN);
-               BUILD_ASSERT(FREE == 0);
-               memset((char *)pool + nextpage*getpagesize(), 0, getpagesize());
-               return alloc_metaspace(pool, poolsize, mh, bytes, type);
+               head->bs[i].elements_per_page
+                       = elements_per_page(i / INTER_BUCKET_SPACE,
+                                           bucket_to_size(i),
+                                           pagesize);
        }
 
-       /* No metadata left at all? */
-       page = alloc_get_pages(pool, poolsize, div_up(bytes, getpagesize()), 1);
-       if (!page)
-               return NULL;
+       /* They start as all large pages. */
+       memset(head->pagesize, 0xFF, sizeof(head->pagesize));
+       /* FIXME: small pages for last bit? */
+
+       /* Split first page into small pages. */
+       assert(header_size < (1UL << lp_bits));
+       clear_bit(head->pagesize, 0);
+
+       /* Skip over page(s) used by header, add rest to free list */
+       for (i = align_up(header_size, (1 << sp_bits)) >> sp_bits;
+            i < SMALL_PAGES_PER_LARGE_PAGE;
+            i++) {
+               ph = from_pgnum(head, i, sp_bits);
+               ph->elements_used = 0;
+               add_small_page_to_freelist(head, ph, sp_bits);
+       }
 
-       newmh = (struct metaheader *)((char *)pool + page * getpagesize());
-       BUILD_ASSERT(FREE == 0);
-       memset(newmh + 1, 0, getpagesize() - sizeof(*mh));
+       /* Add the rest of the pages as large pages. */
+       i = SMALL_PAGES_PER_LARGE_PAGE;
+       while ((i << sp_bits) + (1 << lp_bits) <= poolsize) {
+               ph = from_pgnum(head, i, sp_bits);
+               ph->elements_used = 0;
+               add_large_page_to_freelist(head, ph, sp_bits);
+               i += SMALL_PAGES_PER_LARGE_PAGE;
+       }
+}
 
-       /* Sew it into linked list */
-       mh = first_mheader(pool,poolsize);
-       newmh->next = mh->next;
-       mh->next = pool_offset(pool, newmh);
+/* A large page worth of small pages are free: delete them from free list. */
+static void del_large_from_small_free_list(struct header *head,
+                                          struct page_header *ph,
+                                          unsigned int sp_bits)
+{
+       unsigned long i;
 
-       return alloc_metaspace(pool, poolsize, newmh, bytes, type);
+       for (i = 0; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
+               del_from_list(head, &head->small_free_list,
+                             (void *)ph + (i << sp_bits),
+                             sp_bits);
+       }
 }
 
-static void alloc_free_pages(void *pool, unsigned long pagenum)
+static bool all_empty(struct header *head,
+                     unsigned long pgnum,
+                     unsigned sp_bits)
 {
-       assert(get_page_state(pool, pagenum) == TAKEN_START);
-       set_page_state(pool, pagenum, FREE);
-       while (get_page_state(pool, ++pagenum) == TAKEN)
-               set_page_state(pool, pagenum, FREE);
+       unsigned long i;
+
+       for (i = 0; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
+               struct page_header *ph = from_pgnum(head, pgnum + i, sp_bits);
+               if (ph->elements_used)
+                       return false;
+       }
+       return true;
 }
 
-static void maybe_transform_uniform_page(void *pool, unsigned long offset)
+static void recombine_small_pages(struct header *head, unsigned long poolsize,
+                                 unsigned int sp_bits)
 {
-       /* FIXME: If possible and page isn't full, change to a bitmap */
+       unsigned long i;
+       unsigned int lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
+
+       /* Look for small pages to coalesce, after first large page. */
+       for (i = SMALL_PAGES_PER_LARGE_PAGE;
+            i < (poolsize >> lp_bits) << BITS_FROM_SMALL_TO_LARGE_PAGE;
+            i += SMALL_PAGES_PER_LARGE_PAGE) {
+               /* Already a large page? */
+               if (test_bit(head->pagesize, i / SMALL_PAGES_PER_LARGE_PAGE))
+                       continue;
+               if (all_empty(head, i, sp_bits)) {
+                       struct page_header *ph = from_pgnum(head, i, sp_bits);
+                       set_bit(head->pagesize,
+                               i / SMALL_PAGES_PER_LARGE_PAGE);
+                       del_large_from_small_free_list(head, ph, sp_bits);
+                       add_large_page_to_freelist(head, ph, sp_bits);
+               }
+       }
 }
 
-/* Returns 0 or the size of the uniform alloc to use */
-static unsigned long suitable_for_uc(unsigned long size, unsigned long align)
+static u16 get_large_page(struct header *head, unsigned long poolsize,
+                         unsigned int sp_bits)
 {
-       unsigned long num_elems, wastage, usize;
-       unsigned long bitmap_cost;
+       unsigned int lp_bits, page;
 
-       if (size == 0)
-               size = 1;
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
 
-       /* Fix up silly alignments. */
-       usize = align_up(size, align);
+       page = pop_from_list(head, &head->large_free_list, sp_bits);
+       if (likely(page))
+               return page;
 
-       /* How many can fit in this page? */
-       num_elems = SUBPAGE_METAOFF / usize;
+       recombine_small_pages(head, poolsize, sp_bits);
 
-       /* Can happen with bigger alignments. */
-       if (!num_elems)
-               return 0;
-
-       /* Usize maxes out at 14 bits. */
-       if (usize >= (1 << 14))
-               return 0;
-
-       /* How many bytes would be left at the end? */
-       wastage = SUBPAGE_METAOFF % usize;
+       return pop_from_list(head, &head->large_free_list, sp_bits);
+}
 
-       /* If we can get a larger allocation within alignment constraints, we
-        * should do it, otherwise might as well leave wastage at the end. */
-       usize += align_down(wastage / num_elems, align);
+/* Returns small page. */
+static unsigned long break_up_large_page(struct header *head,
+                                        unsigned int sp_bits,
+                                        u16 lpage)
+{
+       unsigned int i;
 
-       /* Bitmap allocation costs 2 bits per BITMAP_GRANULARITY bytes, plus
-        * however much we waste in rounding up to BITMAP_GRANULARITY. */
-       bitmap_cost = 2 * div_up(size, BITMAP_GRANULARITY)
-               + CHAR_BIT * (align_up(size, BITMAP_GRANULARITY) - size);
+       clear_bit(head->pagesize, lpage >> BITS_FROM_SMALL_TO_LARGE_PAGE);
 
-       /* Our cost is 1 bit, plus usize overhead */
-       if (bitmap_cost < 1 + (usize - size) * CHAR_BIT)
-               return 0;
+       for (i = 1; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
+               struct page_header *ph = from_pgnum(head, lpage + i, sp_bits);
+               add_small_page_to_freelist(head, ph, sp_bits);
+       }
 
-       return usize;
+       return lpage;
 }
 
-static unsigned long uniform_alloc(void *pool, unsigned long poolsize,
-                                  struct uniform_cache *uc,
-                                  unsigned long ucnum)
+static u16 get_small_page(struct header *head, unsigned long poolsize,
+                         unsigned int sp_bits)
 {
-       uint8_t *metadata = get_page_metadata(pool, uc->page[ucnum]) + 2;
-       unsigned long i, max;
-
-       /* Simple one-bit-per-object bitmap. */
-       max = SUBPAGE_METAOFF / uc->size[ucnum];
-       for (i = 0; i < max; i++) {
-               if (!(metadata[i / CHAR_BIT] & (1 << (i % CHAR_BIT)))) {
-                       metadata[i / CHAR_BIT] |= (1 << (i % CHAR_BIT));
-                       return uc->page[ucnum] * getpagesize()
-                               + i * uc->size[ucnum];
-               }
-       }
-
-       return 0;
+       u16 ret;
+
+       ret = pop_from_list(head, &head->small_free_list, sp_bits);
+       if (likely(ret))
+               return ret;
+       ret = get_large_page(head, poolsize, sp_bits);
+       if (likely(ret))
+               ret = break_up_large_page(head, sp_bits, ret);
+       return ret;
 }
 
-static unsigned long new_uniform_page(void *pool, unsigned long poolsize,
-                                     unsigned long usize)
+static bool huge_allocated(struct header *head, unsigned long offset)
 {
-       unsigned long page, metalen;
-       uint8_t *metadata;
-
-       page = alloc_get_pages(pool, poolsize, 1, 1);
-       if (page == 0)
-               return 0;
-
-       metalen = uniform_metalen(usize);
+       unsigned long i;
+       struct huge_alloc *ha;
 
-       /* Get metadata for page. */
-       metadata = new_metadata(pool, poolsize, metalen, UNIFORM);
-       if (!metadata) {
-               alloc_free_pages(pool, page);
-               return 0;
+       for (i = head->huge; i; i = ha->next) {
+               ha = (void *)((char *)head + i);
+               if (ha->off <= offset && ha->off + ha->len > offset)
+                       return true;
        }
+       return false;
+}
 
-       encode_usize(metadata, usize);
+/* They want something really big.  Aim for contiguous pages (slow). */
+static void *unlikely_func huge_alloc(void *pool, unsigned long poolsize,
+                                     unsigned long size, unsigned long align)
+{
+       struct header *head = pool;
+       struct huge_alloc *ha;
+       unsigned long i, sp_bits, lp_bits, num, header_size;
 
-       BUILD_ASSERT(FREE == 0);
-       memset(metadata + 2, 0, metalen - 2);
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
 
-       /* Actually, this is a subpage page now. */
-       set_page_state(pool, page, SPECIAL);
+       /* Allocate tracking structure optimistically. */
+       ha = alloc_get(pool, poolsize, sizeof(*ha), ALIGNOF(*ha));
+       if (!ha)
+               return NULL;
 
-       /* Set metadata pointer for page. */
-       set_page_metadata(pool, page, metadata);
+       /* First search for contiguous small pages... */
+       header_size = sizeof(*head) + sizeof(head->bs) * (max_bucket(lp_bits)-1);
 
-       return page;
-}
+       num = 0;
+       for (i = (header_size + (1 << sp_bits) - 1) >> sp_bits;
+            i << sp_bits < poolsize;
+            i++) {
+               struct page_header *pg;
+               unsigned long off = (i << sp_bits);
 
-static unsigned long alloc_sub_page(void *pool, unsigned long poolsize,
-                                   unsigned long size, unsigned long align)
-{
-       unsigned long i, usize;
-       uint8_t *metadata;
-       struct uniform_cache *uc = pool;
-
-       usize = suitable_for_uc(size, align);
-       if (usize) {
-               /* Look for a uniform page. */
-               for (i = 0; i < UNIFORM_CACHE_NUM; i++) {
-                       if (uc->size[i] == usize) {
-                               unsigned long ret;
-                               ret = uniform_alloc(pool, poolsize, uc, i);
-                               if (ret != 0)
-                                       return ret;
-                               /* OK, that one is full, remove from cache. */
-                               uc->size[i] = 0;
-                               break;
-                       }
+               /* Skip over large pages. */
+               if (test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
+                       i += (1 << BITS_FROM_SMALL_TO_LARGE_PAGE)-1;
+                       continue;
                }
 
-               /* OK, try a new uniform page.  Use random discard for now. */
-               i = random() % UNIFORM_CACHE_NUM;
-               maybe_transform_uniform_page(pool, uc->page[i]);
+               /* Does this page meet alignment requirements? */
+               if (!num && off % align != 0)
+                       continue;
 
-               uc->page[i] = new_uniform_page(pool, poolsize, usize);
-               if (uc->page[i]) {
-                       uc->size[i] = usize;
-                       return uniform_alloc(pool, poolsize, uc, i);
+               /* FIXME: This makes us O(n^2). */
+               if (huge_allocated(head, off)) {
+                       num = 0;
+                       continue;
                }
-               uc->size[i] = 0;
-       }
 
-       /* Look for partial page. */
-       for (i = 0; i < poolsize / getpagesize(); i++) {
-               unsigned long ret;
-               if (get_page_state(pool, i) != SPECIAL)
+               pg = (struct page_header *)((char *)head + off);
+               if (pg->elements_used) {
+                       num = 0;
                        continue;
+               }
 
-               ret = sub_page_alloc(pool, i, size, align);
-               if (ret)
-                       return ret;
-       }
-
-       /* Create new SUBPAGE page. */
-       i = alloc_get_pages(pool, poolsize, 1, 1);
-       if (i == 0)
-               return 0;
+               num++;
+               if (num << sp_bits >= size) {
+                       unsigned long pgnum;
 
-       /* Get metadata for page. */
-       metadata = new_metadata(pool, poolsize, BITMAP_METALEN, BITMAP);
-       if (!metadata) {
-               alloc_free_pages(pool, i);
-               return 0;
+                       /* Remove from free list. */
+                       for (pgnum = i; pgnum > i - num; pgnum--) {
+                               pg = from_pgnum(head, pgnum, sp_bits);
+                               del_from_list(head,
+                                             &head->small_free_list,
+                                             pg, sp_bits);
+                       }
+                       ha->off = (i - num + 1) << sp_bits;
+                       ha->len = num << sp_bits;
+                       goto done;
+               }
        }
 
-       /* Actually, this is a subpage page now. */
-       set_page_state(pool, i, SPECIAL);
+       /* Now search for large pages... */
+       recombine_small_pages(head, poolsize, sp_bits);
 
-       /* Set metadata pointer for page. */
-       set_page_metadata(pool, i, metadata);
+       num = 0;
+       for (i = (header_size + (1 << lp_bits) - 1) >> lp_bits;
+            (i << lp_bits) < poolsize; i++) {
+               struct page_header *pg;
+               unsigned long off = (i << lp_bits);
 
-       /* Do allocation like normal */
-       return sub_page_alloc(pool, i, size, align);
-}
-
-static bool bitmap_page_is_empty(uint8_t *meta)
-{
-       unsigned int i;
+               /* Ignore small pages. */
+               if (!test_bit(head->pagesize, i))
+                       continue;
 
-       /* Skip the header (first bit of metadata). */
-       for (i = 1; i < SUBPAGE_METAOFF/BITMAP_GRANULARITY+1; i++)
-               if (get_bit_pair(meta, i) != FREE)
-                       return false;
+               /* Does this page meet alignment requirements? */
+               if (!num && off % align != 0)
+                       continue;
 
-       return true;
-}
+               /* FIXME: This makes us O(n^2). */
+               if (huge_allocated(head, off)) {
+                       num = 0;
+                       continue;
+               }
 
-static bool uniform_page_is_empty(uint8_t *meta)
-{
-       unsigned int i, metalen;
+               pg = (struct page_header *)((char *)head + off);
+               if (pg->elements_used) {
+                       num = 0;
+                       continue;
+               }
 
-       metalen = uniform_metalen(decode_usize(meta));
+               num++;
+               if (num << lp_bits >= size) {
+                       unsigned long pgnum;
 
-       /* Skip the header (first two bytes of metadata). */
-       for (i = 2; i < metalen + 2; i++) {
-               BUILD_ASSERT(FREE == 0);
-               if (meta[i])
-                       return false;
+                       /* Remove from free list. */
+                       for (pgnum = i; pgnum > i - num; pgnum--) {
+                               pg = from_pgnum(head, pgnum, lp_bits);
+                               del_from_list(head,
+                                             &head->large_free_list,
+                                             pg, sp_bits);
+                       }
+                       ha->off = (i - num + 1) << lp_bits;
+                       ha->len = num << lp_bits;
+                       goto done;
+               }
        }
-       return true;
-}
 
-static bool special_page_is_empty(void *pool, unsigned long page)
-{
-       uint8_t *meta;
-       enum sub_metadata_type type;
-
-       meta = get_page_metadata(pool, page);
-       type = get_bit_pair(meta, 0);
-
-       switch (type) {
-       case UNIFORM:
-               return uniform_page_is_empty(meta);
-       case BITMAP:
-               return bitmap_page_is_empty(meta);
-       default:
-               assert(0);
-       }
-}
+       /* Unable to satisfy: free huge alloc structure. */
+       alloc_free(pool, poolsize, ha);
+       return NULL;
 
-static void clear_special_metadata(void *pool, unsigned long page)
-{
-       uint8_t *meta;
-       enum sub_metadata_type type;
-
-       meta = get_page_metadata(pool, page);
-       type = get_bit_pair(meta, 0);
-
-       switch (type) {
-       case UNIFORM:
-               /* First two bytes are the header, rest is already FREE */
-               BUILD_ASSERT(FREE == 0);
-               memset(meta, 0, 2);
-               break;
-       case BITMAP:
-               /* First two bits is the header. */
-               BUILD_ASSERT(BITMAP_METALEN > 1);
-               meta[0] = 0;
-               break;
-       default:
-               assert(0);
-       }
+done:
+       add_to_huge_list(pool, ha);
+       return (char *)pool + ha->off;
 }
 
-/* Returns true if we cleaned any pages. */
-static bool clean_empty_subpages(void *pool, unsigned long poolsize)
+static void unlikely_func huge_free(struct header *head,
+                                   unsigned long poolsize, void *free)
 {
-       unsigned long i;
-       bool progress = false;
-
-       for (i = 0; i < poolsize/getpagesize(); i++) {
-               if (get_page_state(pool, i) != SPECIAL)
-                       continue;
+       unsigned long i, off, pgnum, free_off = (char *)free - (char *)head;
+       unsigned int sp_bits, lp_bits;
+       struct huge_alloc *ha;
 
-               if (special_page_is_empty(pool, i)) {
-                       clear_special_metadata(pool, i);
-                       set_page_state(pool, i, FREE);
-                       progress = true;
+       for (i = head->huge; i; i = ha->next) {
+               ha = (void *)((char *)head + i);
+               if (free_off == ha->off)
+                       break;
+       }
+       assert(i);
+
+       /* Free up all the pages, delete and free ha */
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
+       pgnum = free_off >> sp_bits;
+
+       if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
+               for (off = ha->off; off < ha->off + ha->len; off += 1 << lp_bits) {
+                       add_large_page_to_freelist(head,
+                                                  (void *)((char *)head + off),
+                                                  sp_bits);
+               }
+       } else {
+               for (off = ha->off; off < ha->off + ha->len; off += 1 << sp_bits) {
+                       add_small_page_to_freelist(head,
+                                                  (void *)((char *)head + off),
+                                                  sp_bits);
                }
        }
-       return progress;
+       del_from_huge(head, ha);
+       alloc_free(head, poolsize, ha);
 }
 
-/* Returns true if we cleaned any pages. */
-static bool clean_metadata(void *pool, unsigned long poolsize)
+static unsigned long unlikely_func huge_size(struct header *head, void *p)
 {
-       struct metaheader *mh, *prev_mh = NULL;
-       bool progress = false;
-
-       for (mh = first_mheader(pool,poolsize); mh; mh = next_mheader(pool,mh)){
-               uint8_t *meta;
-               long i;
-               unsigned long metalen = get_metalen(pool, poolsize, mh);
-
-               meta = (uint8_t *)(mh + 1);
-               BUILD_ASSERT(FREE == 0);
-               for (i = metalen - 1; i > 0; i--)
-                       if (meta[i] != 0)
-                               break;
-
-               /* Completely empty? */
-               if (prev_mh && i == metalen) {
-                       alloc_free_pages(pool,
-                                        pool_offset(pool, mh)/getpagesize());
-                       prev_mh->next = mh->next;
-                       mh = prev_mh;
-                       progress = true;
-               } else {
-                       uint8_t *p;
-
-                       /* Some pages at end are free? */
-                       for (p = (uint8_t *)(mh+1) + metalen - getpagesize();
-                            p > meta + i;
-                            p -= getpagesize()) {
-                               set_page_state(pool,
-                                              pool_offset(pool, p)
-                                              / getpagesize(),
-                                              FREE);
-                               progress = true;
-                       }
+       unsigned long i, off = (char *)p - (char *)head;
+       struct huge_alloc *ha;
+
+       for (i = head->huge; i; i = ha->next) {
+               ha = (void *)((char *)head + i);
+               if (off == ha->off) {
+                       return ha->len;
                }
        }
-
-       return progress;
+       abort();
 }
 
 void *alloc_get(void *pool, unsigned long poolsize,
                unsigned long size, unsigned long align)
 {
-       bool subpage_clean = false, metadata_clean = false;
-       unsigned long ret;
-
-       if (poolsize < MIN_SIZE)
-               return NULL;
-
-again:
-       /* Sub-page allocations have an overhead of ~12%. */
-       if (size + size/8 >= getpagesize() || align >= getpagesize()) {
-               unsigned long pages = div_up(size, getpagesize());
-
-               ret = alloc_get_pages(pool, poolsize, pages, align)
-                       * getpagesize();
-       } else
-               ret = alloc_sub_page(pool, poolsize, size, align);
-
-       if (ret != 0)
-               return (char *)pool + ret;
-
-       /* Allocation failed: garbage collection. */
-       if (!subpage_clean) {
-               subpage_clean = true;
-               if (clean_empty_subpages(pool, poolsize))
-                       goto again;
-       }
+       struct header *head = pool;
+       unsigned int bucket;
+       unsigned long i;
+       struct bucket_state *bs;
+       struct page_header *ph;
+       unsigned int sp_bits;
 
-       if (!metadata_clean) {
-               metadata_clean = true;
-               if (clean_metadata(pool, poolsize))
-                       goto again;
+       if (poolsize < MIN_USEFUL_SIZE) {
+               return tiny_alloc_get(pool, poolsize, size, align);
        }
 
-       /* FIXME: Compact metadata? */
-       return NULL;
-}
-
-static void bitmap_free(void *pool, unsigned long pagenum, unsigned long off,
-                       uint8_t *metadata)
-{
-       assert(off % BITMAP_GRANULARITY == 0);
+       size = align_up(size, align);
+       if (unlikely(!size))
+               size = 1;
+       bucket = size_to_bucket(size);
 
-       off /= BITMAP_GRANULARITY;
+       sp_bits = small_page_bits(poolsize);
 
-       /* Offset by one because first bit is used for header. */
-       off++;
+       if (bucket >= max_bucket(sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE)) {
+               return huge_alloc(pool, poolsize, size, align);
+       }
 
-       set_bit_pair(metadata, off++, FREE);
-       while (off < SUBPAGE_METAOFF / BITMAP_GRANULARITY
-              && get_bit_pair(metadata, off) == TAKEN)
-               set_bit_pair(metadata, off++, FREE);
-}
+       bs = &head->bs[bucket];
 
-static void uniform_free(void *pool, unsigned long pagenum, unsigned long off,
-                        uint8_t *metadata)
-{
-       unsigned int usize, bit;
+       if (!bs->page_list) {
+               struct page_header *ph;
 
-       usize = decode_usize(metadata);
-       /* Must have been this size. */
-       assert(off % usize == 0);
-       bit = off / usize;
+               if (large_page_bucket(bucket, sp_bits))
+                       bs->page_list = get_large_page(head, poolsize,
+                                                      sp_bits);
+               else
+                       bs->page_list = get_small_page(head, poolsize,
+                                                      sp_bits);
+               /* FIXME: Try large-aligned alloc?  Header stuffing? */
+               if (unlikely(!bs->page_list))
+                       return NULL;
+               ph = from_pgnum(head, bs->page_list, sp_bits);
+               ph->bucket = bucket;
+               ph->elements_used = 0;
+               ph->next = 0;
+               memset(ph->used, 0, used_size(bs->elements_per_page));
+       }
 
-       /* Skip header. */
-       metadata += 2;
+       ph = from_pgnum(head, bs->page_list, sp_bits);
 
-       /* Must have been allocated. */
-       assert(metadata[bit / CHAR_BIT] & (1 << (bit % CHAR_BIT)));
-       metadata[bit / CHAR_BIT] &= ~(1 << (bit % CHAR_BIT));
-}
+       i = find_free_bit(ph->used);
+       set_bit(ph->used, i);
+       ph->elements_used++;
 
-static void subpage_free(void *pool, unsigned long pagenum, void *free)
-{
-       unsigned long off = (unsigned long)free % getpagesize();
-       uint8_t *metadata = get_page_metadata(pool, pagenum);
-       enum sub_metadata_type type;
-
-       type = get_bit_pair(metadata, 0);
-
-       assert(off < SUBPAGE_METAOFF);
-
-       switch (type) {
-       case BITMAP:
-               bitmap_free(pool, pagenum, off, metadata);
-               break;
-       case UNIFORM:
-               uniform_free(pool, pagenum, off, metadata);
-               break;
-       default:
-               assert(0);
+       /* check if this page is now full */
+       if (unlikely(ph->elements_used == bs->elements_per_page)) {
+               del_from_bucket_list(head, bs, ph, sp_bits);
+               add_to_bucket_full_list(head, bs, ph, sp_bits);
        }
+
+       return (char *)ph + page_header_size(ph->bucket / INTER_BUCKET_SPACE,
+                                            bs->elements_per_page)
+              + i * bucket_to_size(bucket);
 }
 
 void alloc_free(void *pool, unsigned long poolsize, void *free)
 {
-       unsigned long pagenum;
-       struct metaheader *mh;
+       struct header *head = pool;
+       struct bucket_state *bs;
+       unsigned int sp_bits;
+       unsigned long i, pgnum, pgoffset, offset = (char *)free - (char *)pool;
+       bool smallpage;
+       struct page_header *ph;
+
+       if (poolsize < MIN_USEFUL_SIZE) {
+               return tiny_alloc_free(pool, poolsize, free);
+       }
+       
+       /* Get page header. */
+       sp_bits = small_page_bits(poolsize);
+       pgnum = offset >> sp_bits;
+
+       /* Big page? Round down further. */
+       if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
+               smallpage = false;
+               pgnum &= ~(SMALL_PAGES_PER_LARGE_PAGE - 1);
+       } else
+               smallpage = true;
 
-       if (!free)
+       /* Step back to page header. */
+       ph = from_pgnum(head, pgnum, sp_bits);
+       if ((void *)ph == free) {
+               huge_free(head, poolsize, free);
                return;
+       }
 
-       assert(poolsize >= MIN_SIZE);
-
-       mh = first_mheader(pool, poolsize);
-       assert((char *)free >= (char *)(mh + 1));
-       assert((char *)pool + poolsize > (char *)free);
+       bs = &head->bs[ph->bucket];
+       pgoffset = offset - (pgnum << sp_bits)
+               - page_header_size(ph->bucket / INTER_BUCKET_SPACE,
+                                  bs->elements_per_page);
 
-       pagenum = pool_offset(pool, free) / getpagesize();
+       if (unlikely(ph->elements_used == bs->elements_per_page)) {
+               del_from_bucket_full_list(head, bs, ph, sp_bits);
+               add_to_bucket_list(head, bs, ph, sp_bits);
+       }
 
-       if (get_page_state(pool, pagenum) == SPECIAL)
-               subpage_free(pool, pagenum, free);
-       else {
-               assert((unsigned long)free % getpagesize() == 0);
-               alloc_free_pages(pool, pagenum);
+       /* Which element are we? */
+       i = pgoffset / bucket_to_size(ph->bucket);
+       clear_bit(ph->used, i);
+       ph->elements_used--;
+
+       if (unlikely(ph->elements_used == 0)) {
+               bs = &head->bs[ph->bucket];
+               del_from_bucket_list(head, bs, ph, sp_bits);
+               if (smallpage)
+                       add_small_page_to_freelist(head, ph, sp_bits);
+               else
+                       add_large_page_to_freelist(head, ph, sp_bits);
        }
 }
 
 unsigned long alloc_size(void *pool, unsigned long poolsize, void *p)
 {
-       unsigned long len, pagenum;
-       struct metaheader *mh;
-
-       assert(poolsize >= MIN_SIZE);
+       struct header *head = pool;
+       unsigned int pgnum, sp_bits;
+       unsigned long offset = (char *)p - (char *)pool;
+       struct page_header *ph;
 
-       mh = first_mheader(pool, poolsize);
-       assert((char *)p >= (char *)(mh + 1));
-       assert((char *)pool + poolsize > (char *)p);
+       if (poolsize < MIN_USEFUL_SIZE)
+               return tiny_alloc_size(pool, poolsize, p);
 
-       pagenum = pool_offset(pool, p) / getpagesize();
+       /* Get page header. */
+       sp_bits = small_page_bits(poolsize);
+       pgnum = offset >> sp_bits;
 
-       if (get_page_state(pool, pagenum) == SPECIAL) {
-               unsigned long off = (unsigned long)p % getpagesize();
-               uint8_t *metadata = get_page_metadata(pool, pagenum);
-               enum sub_metadata_type type = get_bit_pair(metadata, 0);
+       /* Big page? Round down further. */
+       if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE))
+               pgnum &= ~(SMALL_PAGES_PER_LARGE_PAGE - 1);
 
-               assert(off < SUBPAGE_METAOFF);
-
-               switch (type) {
-               case BITMAP:
-                       assert(off % BITMAP_GRANULARITY == 0);
-                       off /= BITMAP_GRANULARITY;
-
-                       /* Offset by one because first bit used for header. */
-                       off++;
-                       len = BITMAP_GRANULARITY;
-                       while (++off < SUBPAGE_METAOFF / BITMAP_GRANULARITY
-                              && get_bit_pair(metadata, off) == TAKEN)
-                               len += BITMAP_GRANULARITY;
-                       break;
-               case UNIFORM:
-                       len = decode_usize(metadata);
-                       break;
-               default:
-                       assert(0);
-               }
-       } else {
-               len = getpagesize();
-               while (get_page_state(pool, ++pagenum) == TAKEN)
-                       len += getpagesize();
-       }
+       /* Step back to page header. */
+       ph = from_pgnum(head, pgnum, sp_bits);
+       if ((void *)ph == p)
+               return huge_size(head, p);
 
-       return len;
+       return bucket_to_size(ph->bucket);
 }
 
-static bool is_metadata_page(void *pool, unsigned long poolsize,
-                            unsigned long page)
+/* Useful for gdb breakpoints. */
+static bool check_fail(void)
 {
-       struct metaheader *mh;
-
-       for (mh = first_mheader(pool,poolsize); mh; mh = next_mheader(pool,mh)){
-               unsigned long start, end;
-
-               start = pool_offset(pool, mh);
-               end = pool_offset(pool, (char *)(mh+1)
-                                 + get_metalen(pool, poolsize, mh));
-               if (page >= start/getpagesize() && page < end/getpagesize())
-                       return true;
-       }
        return false;
 }
 
-static bool check_bitmap_metadata(void *pool, unsigned long *mhoff)
+static unsigned long count_bits(const unsigned long bitmap[],
+                               unsigned long limit)
 {
-       enum alloc_state last_state = FREE;
-       unsigned int i;
-
-       for (i = 0; i < SUBPAGE_METAOFF / BITMAP_GRANULARITY; i++) {
-               enum alloc_state state;
+       unsigned long i, count = 0;
 
-               /* +1 because header is the first byte. */
-               state = get_bit_pair((uint8_t *)pool + *mhoff, i+1);
-               switch (state) {
-               case SPECIAL:
-                       return false;
-               case TAKEN:
-                       if (last_state == FREE)
-                               return false;
-                       break;
-               default:
-                       break;
-               }
-               last_state = state;
+       while (limit >= BITS_PER_LONG) {
+               count += popcount(bitmap[0]);
+               bitmap++;
+               limit -= BITS_PER_LONG;
        }
-       return true;
+
+       for (i = 0; i < limit; i++)
+               if (test_bit(bitmap, i))
+                       count++;
+       return count;
 }
 
-static bool check_uniform_metadata(void *pool, unsigned long *mhoff)
+static bool out_of_bounds(unsigned long pgnum,
+                         unsigned int sp_bits,
+                         unsigned long pagesize,
+                         unsigned long poolsize)
 {
-       uint8_t *meta = (uint8_t *)pool + *mhoff;
-       unsigned int i, usize;
-       struct uniform_cache *uc = pool;
-
-       usize = decode_usize(meta);
-       if (usize == 0 || suitable_for_uc(usize, 1) != usize)
-               return false;
-
-       /* If it's in uniform cache, make sure that agrees on size. */
-       for (i = 0; i < UNIFORM_CACHE_NUM; i++) {
-               uint8_t *ucm;
-
-               if (!uc->size[i])
-                       continue;
+       if (((pgnum << sp_bits) >> sp_bits) != pgnum)
+               return true;
 
-               ucm = get_page_metadata(pool, uc->page[i]);
-               if (ucm != meta)
-                       continue;
+       if ((pgnum << sp_bits) > poolsize)
+               return true;
 
-               if (usize != uc->size[i])
-                       return false;
-       }
-       return true;
+       return ((pgnum << sp_bits) + pagesize > poolsize);
 }
 
-static bool check_subpage(void *pool, unsigned long poolsize,
-                         unsigned long page)
+static bool check_bucket(struct header *head,
+                        unsigned long poolsize,
+                        unsigned long pages[],
+                        struct bucket_state *bs,
+                        unsigned int bindex)
 {
-       unsigned long *mhoff = metadata_off(pool, page);
-
-       if (*mhoff + sizeof(struct metaheader) > poolsize)
-               return false;
-
-       if (*mhoff % ALIGNOF(struct metaheader) != 0)
-               return false;
-
-       /* It must point to a metadata page. */
-       if (!is_metadata_page(pool, poolsize, *mhoff / getpagesize()))
-               return false;
-
-       /* Header at start of subpage allocation */
-       switch (get_bit_pair((uint8_t *)pool + *mhoff, 0)) {
-       case BITMAP:
-               return check_bitmap_metadata(pool, mhoff);
-       case UNIFORM:
-               return check_uniform_metadata(pool, mhoff);
-       default:
-               return false;
+       bool lp_bucket;
+       struct page_header *ph;
+       unsigned long taken, i, prev, pagesize, sp_bits, lp_bits;
+
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
+
+       lp_bucket = large_page_bucket(bindex, sp_bits);
+
+       pagesize = 1UL << (lp_bucket ? lp_bits : sp_bits);
+
+       /* This many elements fit? */
+       taken = page_header_size(bindex / INTER_BUCKET_SPACE,
+                                bs->elements_per_page);
+       taken += bucket_to_size(bindex) * bs->elements_per_page;
+       if (taken > pagesize)
+               return check_fail();
+
+       /* One more wouldn't fit? */
+       taken = page_header_size(bindex / INTER_BUCKET_SPACE,
+                                bs->elements_per_page + 1);
+       taken += bucket_to_size(bindex) * (bs->elements_per_page + 1);
+       if (taken <= pagesize)
+               return check_fail();
+
+       /* Walk used list. */
+       prev = 0;
+       for (i = bs->page_list; i; i = ph->next) {
+               /* Bad pointer? */
+               if (out_of_bounds(i, sp_bits, pagesize, poolsize))
+                       return check_fail();
+               /* Wrong size page? */
+               if (!!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)
+                   != lp_bucket)
+                       return check_fail();
+               /* Large page not on boundary? */
+               if (lp_bucket && (i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
+                       return check_fail();
+               ph = from_pgnum(head, i, sp_bits);
+               /* Linked list corrupt? */
+               if (ph->prev != prev)
+                       return check_fail();
+               /* Already seen this page? */
+               if (test_bit(pages, i))
+                       return check_fail();
+               set_bit(pages, i);
+               /* Empty or full? */
+               if (ph->elements_used == 0)
+                       return check_fail();
+               if (ph->elements_used >= bs->elements_per_page)
+                       return check_fail();
+               /* Used bits don't agree? */
+               if (ph->elements_used != count_bits(ph->used,
+                                                   bs->elements_per_page))
+                       return check_fail();
+               /* Wrong bucket? */
+               if (ph->bucket != bindex)
+                       return check_fail();
+               prev = i;
        }
 
+       /* Walk full list. */
+       prev = 0;
+       for (i = bs->full_list; i; i = ph->next) {
+               /* Bad pointer? */
+               if (out_of_bounds(i, sp_bits, pagesize, poolsize))
+                       return check_fail();
+               /* Wrong size page? */
+               if (!!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)
+                   != lp_bucket)
+               /* Large page not on boundary? */
+               if (lp_bucket && (i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
+                       return check_fail();
+               ph = from_pgnum(head, i, sp_bits);
+               /* Linked list corrupt? */
+               if (ph->prev != prev)
+                       return check_fail();
+               /* Already seen this page? */
+               if (test_bit(pages, i))
+                       return check_fail();
+               set_bit(pages, i);
+               /* Not full? */
+               if (ph->elements_used != bs->elements_per_page)
+                       return check_fail();
+               /* Used bits don't agree? */
+               if (ph->elements_used != count_bits(ph->used,
+                                                   bs->elements_per_page))
+                       return check_fail();
+               /* Wrong bucket? */
+               if (ph->bucket != bindex)
+                       return check_fail();
+               prev = i;
+       }
+       return true;
 }
 
 bool alloc_check(void *pool, unsigned long poolsize)
 {
-       unsigned long i;
-       struct metaheader *mh;
-       enum alloc_state last_state = FREE;
-       bool was_metadata = false;
-
-       if (poolsize < MIN_SIZE)
-               return true;
-
-       if (get_page_state(pool, 0) != TAKEN_START)
-               return false;
-
-       /* First check metadata pages. */
-       /* Metadata pages will be marked TAKEN. */
-       for (mh = first_mheader(pool,poolsize); mh; mh = next_mheader(pool,mh)){
-               unsigned long start, end;
-
-               start = pool_offset(pool, mh);
-               if (start + sizeof(*mh) > poolsize)
-                       return false;
+       struct header *head = pool;
+       unsigned long prev, i, lp_bits, sp_bits, header_size, num_buckets;
+       struct page_header *ph;
+       struct huge_alloc *ha;
+       unsigned long pages[MAX_SMALL_PAGES / BITS_PER_LONG] = { 0 };
+
+       if (poolsize < MIN_USEFUL_SIZE)
+               return tiny_alloc_check(pool, poolsize);
+
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
+
+       num_buckets = max_bucket(lp_bits);
+
+       header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
+
+       /* First, set all bits taken by header. */
+       for (i = 0; i < header_size; i += (1UL << sp_bits))
+               set_bit(pages, i >> sp_bits);
+
+       /* Check small page free list. */
+       prev = 0;
+       for (i = head->small_free_list; i; i = ph->next) {
+               /* Bad pointer? */
+               if (out_of_bounds(i, sp_bits, 1 << sp_bits, poolsize))
+                       return check_fail();
+               /* Large page? */
+               if (test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE))
+                       return check_fail();
+               ph = from_pgnum(head, i, sp_bits);
+               /* Linked list corrupt? */
+               if (ph->prev != prev)
+                       return check_fail();
+               /* Already seen this page? */
+               if (test_bit(pages, i))
+                       return check_fail();
+               set_bit(pages, i);
+               prev = i;
+       }
 
-               end = pool_offset(pool, (char *)(mh+1)
-                                 + get_metalen(pool, poolsize, mh));
-               if (end > poolsize)
-                       return false;
+       /* Check large page free list. */
+       prev = 0;
+       for (i = head->large_free_list; i; i = ph->next) {
+               /* Bad pointer? */
+               if (out_of_bounds(i, sp_bits, 1 << lp_bits, poolsize))
+                       return check_fail();
+               /* Not large page? */
+               if (!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE))
+                       return check_fail();
+               /* Not page boundary? */
+               if ((i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
+                       return check_fail();
+               ph = from_pgnum(head, i, sp_bits);
+               /* Linked list corrupt? */
+               if (ph->prev != prev)
+                       return check_fail();
+               /* Already seen this page? */
+               if (test_bit(pages, i))
+                       return check_fail();
+               set_bit(pages, i);
+               prev = i;
+       }
 
-               /* Non-first pages should start on a page boundary. */
-               if (mh != first_mheader(pool, poolsize)
-                   && start % getpagesize() != 0)
-                       return false;
+       /* Check the buckets. */
+       for (i = 0; i < max_bucket(lp_bits); i++) {
+               struct bucket_state *bs = &head->bs[i];
 
-               /* It should end on a page boundary. */
-               if (end % getpagesize() != 0)
+               if (!check_bucket(head, poolsize, pages, bs, i))
                        return false;
        }
 
-       for (i = 0; i < poolsize / getpagesize(); i++) {
-               enum alloc_state state = get_page_state(pool, i);
-               bool is_metadata = is_metadata_page(pool, poolsize,i);
+       /* Check the huge alloc list. */
+       prev = 0;
+       for (i = head->huge; i; i = ha->next) {
+               unsigned long pgbits, j;
+
+               /* Bad pointer? */
+               if (i >= poolsize || i + sizeof(*ha) > poolsize)
+                       return check_fail();
+               ha = (void *)((char *)head + i);
+
+               /* Check contents of ha. */
+               if (ha->off > poolsize || ha->off + ha->len > poolsize)
+                       return check_fail();
+
+               /* Large or small page? */
+               pgbits = test_bit(head->pagesize, ha->off >> lp_bits)
+                       ? lp_bits : sp_bits;
+
+               /* Not page boundary? */
+               if ((ha->off % (1UL << pgbits)) != 0)
+                       return check_fail();
+
+               /* Not page length? */
+               if ((ha->len % (1UL << pgbits)) != 0)
+                       return check_fail();
+
+               /* Linked list corrupt? */
+               if (ha->prev != prev)
+                       return check_fail();
+
+               for (j = ha->off; j < ha->off + ha->len; j += (1 << sp_bits)) {
+                       /* Already seen this page? */
+                       if (test_bit(pages, j >> sp_bits))
+                               return check_fail();
+                       set_bit(pages, j >> sp_bits);
+               }
 
-               switch (state) {
-               case FREE:
-                       /* metadata pages are never free. */
-                       if (is_metadata)
-                               return false;
-               case TAKEN_START:
-                       break;
-               case TAKEN:
-                       /* This should continue a previous block. */
-                       if (last_state == FREE)
-                               return false;
-                       if (is_metadata != was_metadata)
-                               return false;
-                       break;
-               case SPECIAL:
-                       /* Check metadata pointer etc. */
-                       if (!check_subpage(pool, poolsize, i))
-                               return false;
+               prev = i;
+       }
+               
+       /* Make sure every page accounted for. */
+       for (i = 0; i < poolsize >> sp_bits; i++) {
+               if (!test_bit(pages, i))
+                       return check_fail();
+               if (test_bit(head->pagesize,
+                            i >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
+                       /* Large page, skip rest. */
+                       i += SMALL_PAGES_PER_LARGE_PAGE - 1;
                }
-               last_state = state;
-               was_metadata = is_metadata;
        }
+
        return true;
 }
 
-void alloc_visualize(FILE *out, void *pool, unsigned long poolsize)
+static unsigned long print_overhead(FILE *out, const char *desc,
+                                   unsigned long bytes,
+                                   unsigned long poolsize)
 {
-       struct metaheader *mh;
-       struct uniform_cache *uc = pool;
-       unsigned long pagebitlen, metadata_pages, count[1<<BITS_PER_PAGE], tot;
-       long i;
-
-       if (poolsize < MIN_SIZE) {
-               fprintf(out, "Pool smaller than %u: no content\n", MIN_SIZE);
-               return;
-       }
-
-       tot = 0;
-       for (i = 0; i < UNIFORM_CACHE_NUM; i++)
-               tot += (uc->size[i] != 0);
-       fprintf(out, "Uniform cache (%lu entries):\n", tot);
-       for (i = 0; i < UNIFORM_CACHE_NUM; i++) {
-               unsigned int j, total = 0;
-               uint8_t *meta;
-
-               if (!uc->size[i])
-                       continue;
-
-               /* First two bytes are header. */
-               meta = get_page_metadata(pool, uc->page[i]) + 2;
-
-               for (j = 0; j < SUBPAGE_METAOFF / uc->size[i]; j++)
-                       if (meta[j / 8] & (1 << (j % 8)))
-                               total++;
-
-               printf("  %u: %u/%zu (%zu%% density)\n",
-                      uc->size[j], total, SUBPAGE_METAOFF / uc->size[i],
-                      (total * 100) / (SUBPAGE_METAOFF / uc->size[i]));
-       }
-
-       memset(count, 0, sizeof(count));
-       for (i = 0; i < poolsize / getpagesize(); i++)
-               count[get_page_state(pool, i)]++;
-
-       mh = first_mheader(pool, poolsize);
-       pagebitlen = (uint8_t *)mh - get_page_statebits(pool);
-       fprintf(out, "%lu bytes of page bits: FREE/TAKEN/TAKEN_START/SUBPAGE = %lu/%lu/%lu/%lu\n",
-               pagebitlen, count[0], count[1], count[2], count[3]);
-
-       /* One metadata page for every page of page bits. */
-       metadata_pages = div_up(pagebitlen, getpagesize());
-
-       /* Now do each metadata page. */
-       for (; mh; mh = next_mheader(pool,mh)) {
-               unsigned long free = 0, bitmapblocks = 0, uniformblocks = 0,
-                       len = 0, uniformlen = 0, bitmaplen = 0, metalen;
-               uint8_t *meta = (uint8_t *)(mh + 1);
-
-               metalen = get_metalen(pool, poolsize, mh);
-               metadata_pages += (sizeof(*mh) + metalen) / getpagesize();
-
-               for (i = 0; i < metalen * METADATA_PER_BYTE; i += len) {
-                       switch (get_bit_pair(meta, i)) {
-                       case FREE:
-                               len = 1;
-                               free++;
-                               break;
-                       case BITMAP:
-                               /* Skip over this allocated part. */
-                               len = BITMAP_METALEN * CHAR_BIT;
-                               bitmapblocks++;
-                               bitmaplen += len;
-                               break;
-                       case UNIFORM:
-                               /* Skip over this part. */
-                               len = decode_usize(meta + i/METADATA_PER_BYTE);
-                               len = uniform_metalen(len) * METADATA_PER_BYTE;
-                               uniformblocks++;
-                               uniformlen += len;
-                               break;
-                       default:
-                               assert(0);
-                       }
-               }
+       fprintf(out, "Overhead (%s): %lu bytes (%.3g%%)\n",
+               desc, bytes, 100.0 * bytes / poolsize);
+       return bytes;
+}
 
-               fprintf(out, "Metadata %lu-%lu: %lu free, %lu bitmapblocks, %lu uniformblocks, %lu%% density\n",
-                       pool_offset(pool, mh),
-                       pool_offset(pool, (char *)(mh+1) + metalen),
-                       free, bitmapblocks, uniformblocks,
-                       (bitmaplen + uniformlen) * 100
-                       / (free + bitmaplen + uniformlen));
+static unsigned long count_list(struct header *head,
+                               u16 pgnum,
+                               unsigned int sp_bits,
+                               unsigned long *total_elems)
+{
+       struct page_header *p;
+       unsigned long ret = 0;
+
+       while (pgnum) {
+               p = from_pgnum(head, pgnum, sp_bits);
+               if (total_elems)
+                       (*total_elems) += p->elements_used;
+               ret++;
+               pgnum = p->next;
        }
+       return ret;
+}
 
-       /* Account for total pages allocated. */
-       tot = (count[1] + count[2] - metadata_pages) * getpagesize();
+static unsigned long visualize_bucket(FILE *out, struct header *head,
+                                     unsigned int bucket,
+                                     unsigned long poolsize,
+                                     unsigned int sp_bits)
+{
+       unsigned long num_full, num_partial, num_pages, page_size,
+               elems, hdr_min, hdr_size, elems_per_page, overhead = 0;
+
+       elems_per_page = head->bs[bucket].elements_per_page;
+
+       /* If we used byte-based bitmaps, we could get pg hdr to: */
+       hdr_min = sizeof(struct page_header)
+               - sizeof(((struct page_header *)0)->used)
+               + align_up(elems_per_page, CHAR_BIT) / CHAR_BIT;
+       hdr_size = page_header_size(bucket / INTER_BUCKET_SPACE,
+                                   elems_per_page);
+
+       elems = 0;
+       num_full = count_list(head, head->bs[bucket].full_list, sp_bits,
+                             &elems);
+       num_partial = count_list(head, head->bs[bucket].page_list, sp_bits,
+                                &elems);
+       num_pages = num_full + num_partial;
+       if (!num_pages)
+               return 0;
 
-       fprintf(out, "Total metadata bytes = %lu\n",
-               metadata_pages * getpagesize());
+       fprintf(out, "Bucket %u (%lu bytes):"
+               " %lu full, %lu partial = %lu elements\n",
+               bucket, bucket_to_size(bucket), num_full, num_partial, elems);
+       /* Strict requirement of page header size. */
+       overhead += print_overhead(out, "page headers",
+                                  hdr_min * num_pages, poolsize);
+       /* Gap between minimal page header and actual start. */
+       overhead += print_overhead(out, "page post-header alignments",
+                                  (hdr_size - hdr_min) * num_pages, poolsize);
+       /* Between last element and end of page. */
+       page_size = (1 << sp_bits);
+       if (large_page_bucket(bucket, sp_bits))
+               page_size <<= BITS_FROM_SMALL_TO_LARGE_PAGE;
+
+       overhead += print_overhead(out, "page tails",
+                                  (page_size - (hdr_size
+                                                + (elems_per_page
+                                                   * bucket_to_size(bucket))))
+                                  * num_pages, poolsize);
+       return overhead;
+}
 
-       /* Now do every subpage. */
-       for (i = 0; i < poolsize / getpagesize(); i++) {
-               uint8_t *meta;
-               unsigned int j, allocated;
-               enum sub_metadata_type type;
+void alloc_visualize(FILE *out, void *pool, unsigned long poolsize)
+{
+       struct header *head = pool;
+       unsigned long i, lp_bits, sp_bits, header_size, num_buckets, count,
+               overhead = 0;
 
-               if (get_page_state(pool, i) != SPECIAL)
-                       continue;
+       fprintf(out, "Pool %p size %lu: (%s allocator)\n", pool, poolsize,
+               poolsize < MIN_USEFUL_SIZE ? "tiny" : "standard");
 
-               memset(count, 0, sizeof(count));
-
-               meta = get_page_metadata(pool, i);
-               type = get_bit_pair(meta, 0);
-
-               if (type == BITMAP) {
-                       for (j = 0; j < SUBPAGE_METAOFF/BITMAP_GRANULARITY; j++)
-                               count[get_page_state(meta, j)]++;
-                       allocated = (count[1] + count[2]) * BITMAP_GRANULARITY;
-                       fprintf(out, "Subpage bitmap ");
-               } else {
-                       unsigned int usize = decode_usize(meta);
-
-                       assert(type == UNIFORM);
-                       fprintf(out, "Subpage uniform (%u) ", usize);
-                       meta += 2;
-                       for (j = 0; j < SUBPAGE_METAOFF / usize; j++)
-                               count[!!(meta[j / 8] & (1 << (j % 8)))]++;
-                       allocated = count[1] * usize;
-               }
-               fprintf(out, "%lu: FREE/TAKEN/TAKEN_START = %lu/%lu/%lu %u%% density\n",
-                       i, count[0], count[1], count[2],
-                       allocated * 100 / getpagesize());
-               tot += allocated;
+       if (poolsize < MIN_USEFUL_SIZE) {
+               tiny_alloc_visualize(out, pool, poolsize);
+               return;
        }
-
-       /* This is optimistic, since we overalloc in several cases. */
-       fprintf(out, "Best possible allocation density = %lu%%\n",
-               tot * 100 / poolsize);
+       
+       sp_bits = small_page_bits(poolsize);
+       lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
+
+       num_buckets = max_bucket(lp_bits);
+       header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
+
+       fprintf(out, "Large page size %lu, small page size %lu.\n",
+               1UL << lp_bits, 1UL << sp_bits);
+       overhead += print_overhead(out, "unused pool tail",
+                                  poolsize % (1 << lp_bits), poolsize);
+       fprintf(out, "Main header %lu bytes (%lu small pages).\n",
+               header_size, align_up(header_size, 1 << sp_bits) >> sp_bits);
+       overhead += print_overhead(out, "partial header page",
+                                  align_up(header_size, 1 << sp_bits)
+                                  - header_size, poolsize);
+       /* Total large pages. */
+       i = count_bits(head->pagesize, poolsize >> lp_bits);
+       /* Used pages. */
+       count = i - count_list(head, head->large_free_list, sp_bits, NULL);
+       fprintf(out, "%lu/%lu large pages used (%.3g%%)\n",
+               count, i, count ? 100.0 * count / i : 0.0);
+
+       /* Total small pages. */
+       i = ((poolsize >> lp_bits) - i) << BITS_FROM_SMALL_TO_LARGE_PAGE;
+       /* Used pages */
+       count = i - count_list(head, head->small_free_list, sp_bits, NULL);
+       fprintf(out, "%lu/%lu small pages used (%.3g%%)\n",
+               count, i, count ? 100.0 * count / i : 0.0);
+
+       /* Summary of each bucket. */
+       fprintf(out, "%lu buckets:\n", num_buckets);
+       for (i = 0; i < num_buckets; i++)
+               overhead += visualize_bucket(out, head, i, poolsize, sp_bits);
+
+       print_overhead(out, "total", overhead, poolsize);
 }