libmtmalloc/common/mtmalloc.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */

/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#include <mtmalloc.h>
#include "mtmalloc_impl.h"
#include <unistd.h>
#include <synch.h>
#include <thread.h>
#include <pthread.h>
#include <stdio.h>
#include <limits.h>
#include <errno.h>
#include <string.h>
#include <strings.h>
#include <sys/param.h>
#include <sys/sysmacros.h>

/*
 * To turn on the asserts just compile -DDEBUG
 */

#ifndef DEBUG
#define NDEBUG
#endif

#include <assert.h>

/*
 * The MT hot malloc implementation contained herein is designed to be
 * plug-compatible with the libc version of malloc. It is not intended
 * to replace that implementation until we decide that it is ok to break
 * customer apps (Solaris 3.0).
 *
 * For requests up to 2^^16, the allocator initializes itself into NCPUS
 * worth of chains of caches. When a memory request is made, the calling thread
 * is vectored into one of NCPUS worth of caches.  The LWP id gives us a cheap,
 * contention-reducing index to use, eventually, this should be replaced with
 * the actual CPU sequence number, when an interface to get it is available.
 *
 * Once the thread is vectored into one of the list of caches the real
 * allocation of the memory begins. The size is determined to figure out which
 * bucket the allocation should be satisfied from. The management of free
 * buckets is done via a bitmask. A free bucket is represented by a 1. The
 * first free bit represents the first free bucket. The position of the bit,
 * represents the position of the bucket in the arena.
 *
 * When the memory from the arena is handed out, the address of the cache
 * control structure is written in the word preceeding the returned memory.
 * This cache control address is used during free() to mark the buffer free
 * in the cache control structure.
 *
 * When all available memory in a cache has been depleted, a new chunk of memory
 * is allocated via sbrk(). The new cache is allocated from this chunk of memory
 * and initialized in the function create_cache(). New caches are installed at
 * the front of a singly linked list of the same size memory pools. This helps
 * to ensure that there will tend to be available memory in the beginning of the
 * list.
 *
 * Long linked lists hurt performance. To decrease this effect, there is a
 * tunable, requestsize, that bumps up the sbrk allocation size and thus
 * increases the number of available blocks within an arena.  We also keep
 * a "hint" for each cache list, which is the last cache in the list allocated
 * from.  This lowers the cost of searching if there are a lot of fully
 * allocated blocks at the front of the list.
 *
 * For requests greater than 2^^16 (oversize allocations), there are two pieces
 * of overhead. There is the OVERHEAD used to hold the cache addr
 * (&oversize_list), plus an oversize_t structure to further describe the block.
 *
 * The oversize list is kept as defragmented as possible by coalescing
 * freed oversized allocations with adjacent neighbors.
 *
 * Addresses handed out are stored in a hash table, and are aligned on
 * MTMALLOC_MIN_ALIGN-byte boundaries at both ends. Request sizes are rounded-up
 * where necessary in order to achieve this. This eases the implementation of
 * MTDEBUGPATTERN and MTINITPATTERN, particularly where coalescing occurs.
 *
 * A memalign allocation takes memalign header overhead.  There's two
 * types of memalign headers distinguished by MTMALLOC_MEMALIGN_MAGIC
 * and MTMALLOC_MEMALIGN_MIN_MAGIC.  When the size of memory taken to
 * get to the aligned address from malloc'ed address is the minimum size
 * OVERHEAD, we create a header taking only one OVERHEAD space with magic
 * number MTMALLOC_MEMALIGN_MIN_MAGIC, and we know by subtracting OVERHEAD
 * from memaligned address, we can get to the malloc'ed address. Otherwise,
 * we create a memalign header taking two OVERHEAD space, one stores
 * MTMALLOC_MEMALIGN_MAGIC magic number, the other one points back to the
 * malloc'ed address.
 */

#if defined(__i386) || defined(__amd64)
#include <arpa/inet.h>  /* for htonl() */
#endif

static void * morecore(size_t);
static void create_cache(cache_t *, size_t bufsize, uint_t hunks);
static void * malloc_internal(size_t, percpu_t *);
static void * oversize(size_t);
static oversize_t *find_oversize(size_t);
static void add_oversize(oversize_t *);
static void copy_pattern(uint32_t, void *, size_t);
static void * verify_pattern(uint32_t, void *, size_t);
static void reinit_cpu_list(void);
static void reinit_cache(cache_t *);
static void free_oversize(oversize_t *);
static oversize_t *oversize_header_alloc(uintptr_t, size_t);

/*
 * oversize hash table stuff
 */
#define NUM_BUCKETS 67  /* must be prime */
#define HASH_OVERSIZE(caddr)    ((uintptr_t)(caddr) % NUM_BUCKETS)
oversize_t *ovsz_hashtab[NUM_BUCKETS];

#define ALIGN(x, a) ((((uintptr_t)(x) + ((uintptr_t)(a) - 1)) \
            & ~((uintptr_t)(a) - 1)))

/* need this to deal with little endianess of x86 */
#if defined(__i386) || defined(__amd64)
#define FLIP_EM(x)  htonl((x))
#else
#define FLIP_EM(x)  (x)
#endif

#define INSERT_ONLY         0
#define COALESCE_LEFT           0x00000001
#define COALESCE_RIGHT          0x00000002
#define COALESCE_WITH_BOTH_SIDES    (COALESCE_LEFT | COALESCE_RIGHT)

#define OVERHEAD    8   /* size needed to write cache addr */
#define HUNKSIZE    8192    /* just a multiplier */

#define MAX_CACHED_SHIFT    16  /* 64K is the max cached size */
#define MAX_CACHED      (1 << MAX_CACHED_SHIFT)
#define MIN_CACHED_SHIFT    4   /* smaller requests rounded up */
#define MTMALLOC_MIN_ALIGN  8   /* min guaranteed alignment */

/* maximum size before overflow */
#define MAX_MTMALLOC    (SIZE_MAX - (SIZE_MAX % MTMALLOC_MIN_ALIGN) \
            - OVSZ_HEADER_SIZE)

#define NUM_CACHES  (MAX_CACHED_SHIFT - MIN_CACHED_SHIFT + 1)
#define CACHELIST_SIZE  ALIGN(NUM_CACHES * sizeof (cache_head_t), \
    CACHE_COHERENCY_UNIT)

#define MINSIZE     9   /* for requestsize, tunable */
#define MAXSIZE     256 /* arbitrary, big enough, for requestsize */

#define FREEPATTERN 0xdeadbeef /* debug fill pattern for free buf */
#define INITPATTERN 0xbaddcafe /* debug fill pattern for new buf */

#define misaligned(p)   ((unsigned)(p) & (sizeof (int) - 1))
#define IS_OVERSIZE(x, y)   (((x) < (y)) && (((x) > MAX_CACHED)? 1 : 0))

static long requestsize = MINSIZE; /* 9 pages per cache; tunable; 9 is min */

static uint_t cpu_mask;
static curcpu_func curcpu;

static int32_t debugopt;
static int32_t reinit;

static percpu_t *cpu_list;
static oversize_t oversize_list;
static mutex_t oversize_lock = DEFAULTMUTEX;

static int ncpus = 0;

#define MTMALLOC_OVERSIZE_MAGIC     ((uintptr_t)&oversize_list)
#define MTMALLOC_MEMALIGN_MAGIC     ((uintptr_t)&oversize_list + 1)
#define MTMALLOC_MEMALIGN_MIN_MAGIC ((uintptr_t)&oversize_list + 2)

/*
 * We require allocations handed out to be aligned on MTMALLOC_MIN_ALIGN-byte
 * boundaries. We round up sizeof (oversize_t) (when necessary) to ensure that
 * this is achieved.
 */
#define OVSZ_SIZE       (ALIGN(sizeof (oversize_t), MTMALLOC_MIN_ALIGN))
#define OVSZ_HEADER_SIZE    (OVSZ_SIZE + OVERHEAD)

/*
 * memalign header takes 2 OVERHEAD space.  One for memalign magic, and the
 * other one points back to the start address of originally allocated space.
 */
#define MEMALIGN_HEADER_SIZE    2 * OVERHEAD
#define MEMALIGN_HEADER_ALLOC(x, shift, malloc_addr)\
    if (shift == OVERHEAD)\
        *((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
            MTMALLOC_MEMALIGN_MIN_MAGIC; \
    else {\
        *((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
            MTMALLOC_MEMALIGN_MAGIC; \
        *((uintptr_t *)((caddr_t)x - 2 * OVERHEAD)) = \
            (uintptr_t)malloc_addr; \
    }

/*
 * Add big to the oversize hash table at the head of the relevant bucket.
 */
static void
insert_hash(oversize_t *big)
{
    caddr_t ret = big->addr;
    int bucket = HASH_OVERSIZE(ret);

    assert(MUTEX_HELD(&oversize_lock));
    big->hash_next = ovsz_hashtab[bucket];
    ovsz_hashtab[bucket] = big;
}

void *
malloc(size_t bytes)
{
    percpu_t *list_rotor;
    uint_t  list_index;

    if (bytes > MAX_CACHED)
        return (oversize(bytes));

    list_index = (curcpu() & cpu_mask);

    list_rotor = &cpu_list[list_index];

    return (malloc_internal(bytes, list_rotor));
}

void *
realloc(void * ptr, size_t bytes)
{
    void *new, *data_ptr;
    cache_t *cacheptr;
    caddr_t mem;
    size_t shift = 0;

    if (ptr == NULL)
        return (malloc(bytes));

    if (bytes == 0) {
        free(ptr);
        return (NULL);
    }

    data_ptr = ptr;
    mem = (caddr_t)ptr - OVERHEAD;

    /*
     * Optimization possibility :
     *  p = malloc(64);
     *  q = realloc(p, 64);
     * q can be same as p.
     * Apply this optimization for the normal
     * sized caches for now.
     */
    if (*(uintptr_t *)mem < MTMALLOC_OVERSIZE_MAGIC ||
        *(uintptr_t *)mem > MTMALLOC_MEMALIGN_MIN_MAGIC) {
        cacheptr = (cache_t *)*(uintptr_t *)mem;
        if (bytes <= (cacheptr->mt_size - OVERHEAD))
            return (ptr);
    }

    new = malloc(bytes);

    if (new == NULL)
        return (NULL);

    /*
     * If new == ptr, ptr has previously been freed. Passing a freed pointer
     * to realloc() is not allowed - unless the caller specifically states
     * otherwise, in which case we must avoid freeing ptr (ie new) before we
     * return new. There is (obviously) no requirement to memcpy() ptr to
     * new before we return.
     */
    if (new == ptr) {
        if (!(debugopt & MTDOUBLEFREE))
            abort();
        return (new);
    }

    if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
        mem -= OVERHEAD;
        ptr = (void *)*(uintptr_t *)mem;
        mem = (caddr_t)ptr - OVERHEAD;
        shift = (size_t)((uintptr_t)data_ptr - (uintptr_t)ptr);
    } else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
        ptr = (void *) mem;
        mem -= OVERHEAD;
        shift = OVERHEAD;
    }

    if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
        oversize_t *old;

        old = (oversize_t *)(mem - OVSZ_SIZE);
        (void) memcpy(new, data_ptr, MIN(bytes, old->size - shift));
        free(ptr);
        return (new);
    }

    cacheptr = (cache_t *)*(uintptr_t *)mem;

    (void) memcpy(new, data_ptr,
        MIN(cacheptr->mt_size - OVERHEAD - shift, bytes));
    free(ptr);

    return (new);
}

void *
calloc(size_t nelem, size_t bytes)
{
    void * ptr;
    size_t size;

    if (nelem == 0 || bytes == 0) {
        size = 0;
    } else {
        size = nelem * bytes;

        /* check for overflow */
        if ((size / nelem) != bytes) {
            errno = ENOMEM;
            return (NULL);
        }
    }

    ptr = malloc(size);
    if (ptr == NULL)
        return (NULL);
    (void) memset(ptr, 0, size);

    return (ptr);
}

void
free(void * ptr)
{
    cache_t *cacheptr;
    caddr_t mem;
    int32_t i;
    caddr_t freeblocks;
    uintptr_t offset;
    uchar_t mask;
    int32_t which_bit, num_bytes;

    if (ptr == NULL)
        return;

    mem = (caddr_t)ptr - OVERHEAD;

    if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
        mem -= OVERHEAD;
        ptr = (void *)*(uintptr_t *)mem;
        mem = (caddr_t)ptr - OVERHEAD;
    } else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
        ptr = (void *) mem;
        mem -= OVERHEAD;
    }

    if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
        oversize_t *big, **opp;
        int bucket;

        big = (oversize_t *)(mem - OVSZ_SIZE);
        (void) mutex_lock(&oversize_lock);

        bucket = HASH_OVERSIZE(big->addr);
        for (opp = &ovsz_hashtab[bucket]; *opp != NULL;
            opp = &(*opp)->hash_next)
            if (*opp == big)
                break;

        if (*opp == NULL) {
            if (!(debugopt & MTDOUBLEFREE))
                abort();
            (void) mutex_unlock(&oversize_lock);
            return;
        }

        *opp = big->hash_next;  /* remove big from the hash table */
        big->hash_next = NULL;

        if (debugopt & MTDEBUGPATTERN)
            copy_pattern(FREEPATTERN, ptr, big->size);
        add_oversize(big);
        (void) mutex_unlock(&oversize_lock);
        return;
    }

    cacheptr = (cache_t *)*(uintptr_t *)mem;
    freeblocks = cacheptr->mt_freelist;

    /*
     * This is the distance measured in bits into the arena.
     * The value of offset is in bytes but there is a 1-1 correlation
     * between distance into the arena and distance into the
     * freelist bitmask.
     */
    offset = mem - cacheptr->mt_arena;

    /*
     * i is total number of bits to offset into freelist bitmask.
     */

    i = offset / cacheptr->mt_size;

    num_bytes = i >> 3;

    /*
     * which_bit is the bit offset into the byte in the freelist.
     * if our freelist bitmask looks like 0xf3 and we are freeing
     * block 5 (ie: the 6th block) our mask will be 0xf7 after
     * the free. Things go left to right that's why the mask is 0x80
     * and not 0x01.
     */
    which_bit = i - (num_bytes << 3);

    mask = 0x80 >> which_bit;

    freeblocks += num_bytes;

    if (debugopt & MTDEBUGPATTERN)
        copy_pattern(FREEPATTERN, ptr, cacheptr->mt_size - OVERHEAD);

    (void) mutex_lock(&cacheptr->mt_cache_lock);

    if (*freeblocks & mask) {
        if (!(debugopt & MTDOUBLEFREE))
            abort();
    } else {
        *freeblocks |= mask;
        cacheptr->mt_nfree++;
    }

    (void) mutex_unlock(&cacheptr->mt_cache_lock);
}

void *
memalign(size_t alignment, size_t size)
{
    size_t alloc_size;
    uintptr_t offset;
    void *alloc_buf;
    void *ret_buf;

    if (size == 0 || alignment == 0 || misaligned(alignment) ||
        (alignment & (alignment - 1)) != 0) {
        errno = EINVAL;
        return (NULL);
    }

    /* <= MTMALLOC_MIN_ALIGN, malloc can provide directly */
    if (alignment <= MTMALLOC_MIN_ALIGN)
        return (malloc(size));

    alloc_size = size + alignment - MTMALLOC_MIN_ALIGN;

    if (alloc_size < size) { /* overflow */
        errno = ENOMEM;
        return (NULL);
    }

    alloc_buf = malloc(alloc_size);

    if (alloc_buf == NULL)
        /* malloc sets errno */
        return (NULL);

    /*
     * If alloc_size > MAX_CACHED, malloc() will have returned a multiple of
     * MTMALLOC_MIN_ALIGN, having rounded-up alloc_size if necessary. Since
     * we will use alloc_size to return the excess fragments to the free
     * list, we also round-up alloc_size if necessary.
     */
    if ((alloc_size > MAX_CACHED) &&
        (alloc_size & (MTMALLOC_MIN_ALIGN - 1)))
        alloc_size = ALIGN(alloc_size, MTMALLOC_MIN_ALIGN);

    if ((offset = (uintptr_t)alloc_buf & (alignment - 1)) == 0) {
        /* aligned correctly */

        size_t frag_size = alloc_size -
            (size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);

        /*
         * If the leftover piece of the memory > MAX_CACHED,
         * split off the piece and return it back to the freelist.
         */
        if (IS_OVERSIZE(frag_size, alloc_size)) {
            oversize_t *orig, *tail;
            uintptr_t taddr;
            size_t data_size;
            taddr = ALIGN((uintptr_t)alloc_buf + size,
                MTMALLOC_MIN_ALIGN);
            data_size = taddr - (uintptr_t)alloc_buf;
            orig = (oversize_t *)((uintptr_t)alloc_buf -
                OVSZ_HEADER_SIZE);
            frag_size = orig->size - data_size -
                OVSZ_HEADER_SIZE;
            orig->size = data_size;
            tail = oversize_header_alloc(taddr, frag_size);
            free_oversize(tail);
        }
        ret_buf = alloc_buf;
    } else {
        uchar_t oversize_bits = 0;
        size_t  head_sz, data_sz, tail_sz;
        uintptr_t ret_addr, taddr, shift, tshift;
        oversize_t *orig, *tail, *big;
        size_t tsize;

        /* needs to be aligned */
        shift = alignment - offset;

        assert(shift >= MTMALLOC_MIN_ALIGN);

        ret_addr = ((uintptr_t)alloc_buf + shift);
        ret_buf = (void *)ret_addr;

        if (alloc_size <= MAX_CACHED) {
            MEMALIGN_HEADER_ALLOC(ret_addr, shift, alloc_buf);
            return (ret_buf);
        }

        /*
         * Only check for the fragments when the memory is allocted
         * from oversize_list.  Split off a fragment and return it
         * to the oversize freelist when it's > MAX_CACHED.
         */

        head_sz = shift - MAX(MEMALIGN_HEADER_SIZE, OVSZ_HEADER_SIZE);

        tail_sz = alloc_size -
            (shift + size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);

        oversize_bits |= IS_OVERSIZE(head_sz, alloc_size) |
            IS_OVERSIZE(size, alloc_size) << DATA_SHIFT |
            IS_OVERSIZE(tail_sz, alloc_size) << TAIL_SHIFT;

        switch (oversize_bits) {
            case NONE_OVERSIZE:
            case DATA_OVERSIZE:
                MEMALIGN_HEADER_ALLOC(ret_addr, shift,
                    alloc_buf);
                break;
            case HEAD_OVERSIZE:
                /*
                 * If we can extend data > MAX_CACHED and have
                 * head still > MAX_CACHED, we split head-end
                 * as the case of head-end and data oversized,
                 * otherwise just create memalign header.
                 */
                tsize = (shift + size) - (MAX_CACHED + 8 +
                    MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);

                if (!IS_OVERSIZE(tsize, alloc_size)) {
                    MEMALIGN_HEADER_ALLOC(ret_addr, shift,
                        alloc_buf);
                    break;
                } else {
                    tsize += OVSZ_HEADER_SIZE;
                    taddr = ALIGN((uintptr_t)alloc_buf +
                        tsize, MTMALLOC_MIN_ALIGN);
                    tshift = ret_addr - taddr;
                    MEMALIGN_HEADER_ALLOC(ret_addr, tshift,
                        taddr);
                    ret_addr = taddr;
                    shift = ret_addr - (uintptr_t)alloc_buf;
                }
                /* FALLTHROUGH */
            case HEAD_AND_DATA_OVERSIZE:
                /*
                 * Split off the head fragment and
                 * return it back to oversize freelist.
                 * Create oversize header for the piece
                 * of (data + tail fragment).
                 */
                orig = (oversize_t *)((uintptr_t)alloc_buf -
                    OVSZ_HEADER_SIZE);
                big = oversize_header_alloc(ret_addr -
                    OVSZ_HEADER_SIZE, (orig->size - shift));
                (void) mutex_lock(&oversize_lock);
                insert_hash(big);
                (void) mutex_unlock(&oversize_lock);
                orig->size = shift - OVSZ_HEADER_SIZE;

                /* free up the head fragment */
                free_oversize(orig);
                break;
            case TAIL_OVERSIZE:
                /*
                 * If we can extend data > MAX_CACHED and have
                 * tail-end still > MAX_CACHED, we split tail
                 * end, otherwise just create memalign header.
                 */
                orig = (oversize_t *)((uintptr_t)alloc_buf -
                    OVSZ_HEADER_SIZE);
                tsize =  orig->size - (MAX_CACHED + 8 +
                    shift + OVSZ_HEADER_SIZE +
                    MTMALLOC_MIN_ALIGN);
                if (!IS_OVERSIZE(tsize, alloc_size)) {
                    MEMALIGN_HEADER_ALLOC(ret_addr, shift,
                        alloc_buf);
                    break;
                } else {
                    size = MAX_CACHED + 8;
                }
                /* FALLTHROUGH */
            case DATA_AND_TAIL_OVERSIZE:
                /*
                 * Split off the tail fragment and
                 * return it back to oversize freelist.
                 * Create memalign header and adjust
                 * the size for the piece of
                 * (head fragment + data).
                 */
                taddr = ALIGN(ret_addr + size,
                    MTMALLOC_MIN_ALIGN);
                data_sz = (size_t)(taddr -
                    (uintptr_t)alloc_buf);
                orig = (oversize_t *)((uintptr_t)alloc_buf -
                    OVSZ_HEADER_SIZE);
                tsize = orig->size - data_sz;
                orig->size = data_sz;
                MEMALIGN_HEADER_ALLOC(ret_buf, shift,
                    alloc_buf);
                tsize -= OVSZ_HEADER_SIZE;
                tail = oversize_header_alloc(taddr,  tsize);
                free_oversize(tail);
                break;
            case HEAD_AND_TAIL_OVERSIZE:
                /*
                 * Split off the head fragment.
                 * We try to free up tail-end when we can
                 * extend data size to (MAX_CACHED + 8)
                 * and remain tail-end oversized.
                 * The bottom line is all split pieces
                 * should be oversize in size.
                 */
                orig = (oversize_t *)((uintptr_t)alloc_buf -
                    OVSZ_HEADER_SIZE);
                tsize =  orig->size - (MAX_CACHED + 8 +
                    OVSZ_HEADER_SIZE + shift +
                    MTMALLOC_MIN_ALIGN);

                if (!IS_OVERSIZE(tsize, alloc_size)) {
                    /*
                     * If the chunk is not big enough
                     * to make both data and tail oversize
                     * we just keep them as one piece.
                     */
                    big = oversize_header_alloc(ret_addr -
                        OVSZ_HEADER_SIZE,
                        orig->size - shift);
                    (void) mutex_lock(&oversize_lock);
                    insert_hash(big);
                    (void) mutex_unlock(&oversize_lock);
                    orig->size = shift - OVSZ_HEADER_SIZE;
                    free_oversize(orig);
                    break;
                } else {
                    /*
                     * extend data size > MAX_CACHED
                     * and handle it as head, data, tail
                     * are all oversized.
                     */
                    size = MAX_CACHED + 8;
                }
                /* FALLTHROUGH */
            case ALL_OVERSIZE:
                /*
                 * split off the head and tail fragments,
                 * return them back to the oversize freelist.
                 * Alloc oversize header for data seg.
                 */
                orig = (oversize_t *)((uintptr_t)alloc_buf -
                    OVSZ_HEADER_SIZE);
                tsize = orig->size;
                orig->size = shift - OVSZ_HEADER_SIZE;
                free_oversize(orig);

                taddr = ALIGN(ret_addr + size,
                    MTMALLOC_MIN_ALIGN);
                data_sz = taddr - ret_addr;
                assert(tsize > (shift + data_sz +
                    OVSZ_HEADER_SIZE));
                tail_sz = tsize -
                    (shift + data_sz + OVSZ_HEADER_SIZE);

                /* create oversize header for data seg */
                big = oversize_header_alloc(ret_addr -
                    OVSZ_HEADER_SIZE, data_sz);
                (void) mutex_lock(&oversize_lock);
                insert_hash(big);
                (void) mutex_unlock(&oversize_lock);

                /* create oversize header for tail fragment */
                tail = oversize_header_alloc(taddr, tail_sz);
                free_oversize(tail);
                break;
            default:
                /* should not reach here */
                assert(0);
        }
    }
    return (ret_buf);
}


void *
valloc(size_t size)
{
    static unsigned pagesize;

    if (size == 0)
        return (NULL);

    if (!pagesize)
        pagesize = sysconf(_SC_PAGESIZE);

    return (memalign(pagesize, size));
}

void
mallocctl(int cmd, long value)
{
    switch (cmd) {

    case MTDEBUGPATTERN:
        /*
         * Reinitialize free blocks in case malloc() is called prior
         * to mallocctl().
         */
        if (value && !(debugopt & cmd)) {
            reinit++;
            debugopt |= cmd;
            reinit_cpu_list();
        }
        /*FALLTHRU*/
    case MTDOUBLEFREE:
    case MTINITBUFFER:
        if (value)
            debugopt |= cmd;
        else
            debugopt &= ~cmd;
        break;
    case MTCHUNKSIZE:
        if (value >= MINSIZE && value <= MAXSIZE)
            requestsize = value;
        break;
    default:
        break;
    }
}

/*
 * Initialization function, called from the init section of the library.
 * No locking is required here because we are single-threaded during
 * library initialization.
 */
static void
setup_caches(void)
{
    uintptr_t oldbrk;
    uintptr_t newbrk;

    size_t cache_space_needed;
    size_t padding;

    curcpu_func new_curcpu;
    uint_t new_cpu_mask;
    percpu_t *new_cpu_list;

    uint_t i, j;
    uintptr_t list_addr;

    /*
     * Get a decent "current cpu identifier", to be used to reduce
     * contention.  Eventually, this should be replaced by an interface
     * to get the actual CPU sequence number in libthread/liblwp.
     */
    new_curcpu = (curcpu_func)thr_self;
    if ((ncpus = 2 * sysconf(_SC_NPROCESSORS_CONF)) <= 0)
        ncpus = 4; /* decent default value */

    /* round ncpus up to a power of 2 */
    while (ncpus & (ncpus - 1))
        ncpus++;

    new_cpu_mask = ncpus - 1;   /* create the cpu mask */

    /*
     * We now do some magic with the brk.  What we want to get in the
     * end is a bunch of well-aligned stuff in a big initial allocation.
     * Along the way, we do sanity checks to make sure no one else has
     * touched the brk (which shouldn't happen, but it's always good to
     * check)
     *
     * First, make sure sbrk is sane, and store the current brk in oldbrk.
     */
    oldbrk = (uintptr_t)sbrk(0);
    if ((void *)oldbrk == (void *)-1)
        abort();    /* sbrk is broken -- we're doomed. */

    /*
     * Now, align the brk to a multiple of CACHE_COHERENCY_UNIT, so that
     * the percpu structures and cache lists will be properly aligned.
     *
     *   2.  All hunks will be page-aligned, assuming HUNKSIZE >= PAGESIZE,
     *  so they can be paged out individually.
     */
    newbrk = ALIGN(oldbrk, CACHE_COHERENCY_UNIT);
    if (newbrk != oldbrk && (uintptr_t)sbrk(newbrk - oldbrk) != oldbrk)
        abort();    /* sbrk is broken -- we're doomed. */

    /*
     * For each cpu, there is one percpu_t and a list of caches
     */
    cache_space_needed = ncpus * (sizeof (percpu_t) + CACHELIST_SIZE);

    new_cpu_list = (percpu_t *)sbrk(cache_space_needed);

    if (new_cpu_list == (percpu_t *)-1 ||
        (uintptr_t)new_cpu_list != newbrk)
        abort();    /* sbrk is broken -- we're doomed. */

    /*
     * Finally, align the brk to HUNKSIZE so that all hunks are
     * page-aligned, to avoid edge-effects.
     */

    newbrk = (uintptr_t)new_cpu_list + cache_space_needed;

    padding = ALIGN(newbrk, HUNKSIZE) - newbrk;

    if (padding > 0 && (uintptr_t)sbrk(padding) != newbrk)
        abort();    /* sbrk is broken -- we're doomed. */

    list_addr = ((uintptr_t)new_cpu_list + (sizeof (percpu_t) * ncpus));

    /* initialize the percpu list */
    for (i = 0; i < ncpus; i++) {
        new_cpu_list[i].mt_caches = (cache_head_t *)list_addr;
        for (j = 0; j < NUM_CACHES; j++) {
            new_cpu_list[i].mt_caches[j].mt_cache = NULL;
            new_cpu_list[i].mt_caches[j].mt_hint = NULL;
        }

        (void) mutex_init(&new_cpu_list[i].mt_parent_lock,
            USYNC_THREAD, NULL);

        /* get the correct cache list alignment */
        list_addr += CACHELIST_SIZE;
    }

    /*
     * Initialize oversize listhead
     */
    oversize_list.next_bysize = &oversize_list;
    oversize_list.prev_bysize = &oversize_list;
    oversize_list.next_byaddr = &oversize_list;
    oversize_list.prev_byaddr = &oversize_list;
    oversize_list.addr = NULL;
    oversize_list.size = 0;     /* sentinal */

    /*
     * Now install the global variables.
     */
    curcpu = new_curcpu;
    cpu_mask = new_cpu_mask;
    cpu_list = new_cpu_list;
}

static void
create_cache(cache_t *cp, size_t size, uint_t chunksize)
{
    long nblocks;

    (void) mutex_init(&cp->mt_cache_lock, USYNC_THREAD, NULL);
    cp->mt_size = size;
    cp->mt_freelist = ((caddr_t)cp + sizeof (cache_t));
    cp->mt_span = chunksize * HUNKSIZE - sizeof (cache_t);
    cp->mt_hunks = chunksize;
    /*
     * rough calculation. We will need to adjust later.
     */
    nblocks = cp->mt_span / cp->mt_size;
    nblocks >>= 3;
    if (nblocks == 0) { /* less than 8 free blocks in this pool */
        int32_t numblocks = 0;
        long i = cp->mt_span;
        size_t sub = cp->mt_size;
        uchar_t mask = 0;

        while (i > sub) {
            numblocks++;
            i -= sub;
        }
        nblocks = numblocks;
        cp->mt_arena = (caddr_t)ALIGN(cp->mt_freelist + 8, 8);
        cp->mt_nfree = numblocks;
        while (numblocks--) {
            mask |= 0x80 >> numblocks;
        }
        *(cp->mt_freelist) = mask;
    } else {
        cp->mt_arena = (caddr_t)ALIGN((caddr_t)cp->mt_freelist +
            nblocks, 32);
        /* recompute nblocks */
        nblocks = (uintptr_t)((caddr_t)cp->mt_freelist +
            cp->mt_span - cp->mt_arena) / cp->mt_size;
        cp->mt_nfree = ((nblocks >> 3) << 3);
        /* Set everything to free */
        (void) memset(cp->mt_freelist, 0xff, nblocks >> 3);
    }

    if (debugopt & MTDEBUGPATTERN)
        copy_pattern(FREEPATTERN, cp->mt_arena, cp->mt_size * nblocks);

    cp->mt_next = NULL;
}

static void
reinit_cpu_list(void)
{
    oversize_t *wp = oversize_list.next_bysize;
    percpu_t *cpuptr;
    cache_t *thiscache;
    cache_head_t *cachehead;

    /* Reinitialize free oversize blocks. */
    (void) mutex_lock(&oversize_lock);
    if (debugopt & MTDEBUGPATTERN)
        for (; wp != &oversize_list; wp = wp->next_bysize)
            copy_pattern(FREEPATTERN, wp->addr, wp->size);
    (void) mutex_unlock(&oversize_lock);

    /* Reinitialize free blocks. */
    for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
        (void) mutex_lock(&cpuptr->mt_parent_lock);
        for (cachehead = &cpuptr->mt_caches[0]; cachehead <
            &cpuptr->mt_caches[NUM_CACHES]; cachehead++) {
            for (thiscache = cachehead->mt_cache; thiscache != NULL;
                thiscache = thiscache->mt_next) {
                (void) mutex_lock(&thiscache->mt_cache_lock);
                if (thiscache->mt_nfree == 0) {
                    (void) mutex_unlock(
                        &thiscache->mt_cache_lock);
                    continue;
                }
                if (thiscache != NULL)
                    reinit_cache(thiscache);
                (void) mutex_unlock(&thiscache->mt_cache_lock);
            }
        }
        (void) mutex_unlock(&cpuptr->mt_parent_lock);
    }
    reinit = 0;
}

static void
reinit_cache(cache_t *thiscache)
{
    uint32_t *freeblocks; /* not a uintptr_t on purpose */
    int32_t i, n;
    caddr_t ret;

    freeblocks = (uint32_t *)thiscache->mt_freelist;
    while (freeblocks < (uint32_t *)thiscache->mt_arena) {
        if (*freeblocks & 0xffffffff) {
            for (i = 0; i < 32; i++) {
                if (FLIP_EM(*freeblocks) & (0x80000000 >> i)) {
                    n = (uintptr_t)(((freeblocks -
                        (uint32_t *)thiscache->mt_freelist)
                        << 5) + i) * thiscache->mt_size;
                    ret = thiscache->mt_arena + n;
                    ret += OVERHEAD;
                    copy_pattern(FREEPATTERN, ret,
                        thiscache->mt_size);
                }
            }
        }
        freeblocks++;
    }
}

static void *
malloc_internal(size_t size, percpu_t *cpuptr)
{
    cache_head_t *cachehead;
    cache_t *thiscache, *hintcache;
    int32_t i, n, logsz, bucket;
    uint32_t index;
    uint32_t *freeblocks; /* not a uintptr_t on purpose */
    caddr_t ret;

    logsz = MIN_CACHED_SHIFT;

    while (size > (1 << logsz))
        logsz++;

    bucket = logsz - MIN_CACHED_SHIFT;

    (void) mutex_lock(&cpuptr->mt_parent_lock);

    /*
     * Find a cache of the appropriate size with free buffers.
     *
     * We don't need to lock each cache as we check their mt_nfree count,
     * since:
     *  1.  We are only looking for caches with mt_nfree > 0.  If a
     *     free happens during our search, it will increment mt_nfree,
     *     which will not effect the test.
     *  2.  Allocations can decrement mt_nfree, but they can't happen
     *     as long as we hold mt_parent_lock.
     */

    cachehead = &cpuptr->mt_caches[bucket];

    /* Search through the list, starting at the mt_hint */
    thiscache = cachehead->mt_hint;

    while (thiscache != NULL && thiscache->mt_nfree == 0)
        thiscache = thiscache->mt_next;

    if (thiscache == NULL) {
        /* wrap around -- search up to the hint */
        thiscache = cachehead->mt_cache;
        hintcache = cachehead->mt_hint;

        while (thiscache != NULL && thiscache != hintcache &&
            thiscache->mt_nfree == 0)
            thiscache = thiscache->mt_next;

        if (thiscache == hintcache)
            thiscache = NULL;
    }


    if (thiscache == NULL) { /* there are no free caches */
        int32_t thisrequest = requestsize;
        int32_t buffer_size = (1 << logsz) + OVERHEAD;

        thiscache = (cache_t *)morecore(thisrequest * HUNKSIZE);

        if (thiscache == (cache_t *)-1) {
            (void) mutex_unlock(&cpuptr->mt_parent_lock);
            errno = EAGAIN;
            return (NULL);
        }
        create_cache(thiscache, buffer_size, thisrequest);

        /* link in the new block at the beginning of the list */
        thiscache->mt_next = cachehead->mt_cache;
        cachehead->mt_cache = thiscache;
    }

    /* update the hint to the cache we found or created */
    cachehead->mt_hint = thiscache;

    /* thiscache now points to a cache with available space */
    (void) mutex_lock(&thiscache->mt_cache_lock);

    freeblocks = (uint32_t *)thiscache->mt_freelist;
    while (freeblocks < (uint32_t *)thiscache->mt_arena) {
        if (*freeblocks & 0xffffffff)
            break;
        freeblocks++;
        if (freeblocks < (uint32_t *)thiscache->mt_arena &&
            *freeblocks & 0xffffffff)
            break;
        freeblocks++;
        if (freeblocks < (uint32_t *)thiscache->mt_arena &&
            *freeblocks & 0xffffffff)
            break;
        freeblocks++;
        if (freeblocks < (uint32_t *)thiscache->mt_arena &&
            *freeblocks & 0xffffffff)
            break;
        freeblocks++;
    }

    /*
     * the offset from mt_freelist to freeblocks is the offset into
     * the arena. Be sure to include the offset into freeblocks
     * of the bitmask. n is the offset.
     */
    for (i = 0; i < 32; ) {
        if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
            break;
        if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
            break;
        if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
            break;
        if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
            break;
    }
    index = 0x80000000 >> --i;


    *freeblocks &= FLIP_EM(~index);

    thiscache->mt_nfree--;

    (void) mutex_unlock(&thiscache->mt_cache_lock);
    (void) mutex_unlock(&cpuptr->mt_parent_lock);

    n = (uintptr_t)(((freeblocks - (uint32_t *)thiscache->mt_freelist) << 5)
        + i) * thiscache->mt_size;
    /*
     * Now you have the offset in n, you've changed the free mask
     * in the freelist. Nothing left to do but find the block
     * in the arena and put the value of thiscache in the word
     * ahead of the handed out address and return the memory
     * back to the user.
     */
    ret = thiscache->mt_arena + n;

    /* Store the cache addr for this buf. Makes free go fast. */
    *(uintptr_t *)ret = (uintptr_t)thiscache;

    /*
     * This assert makes sure we don't hand out memory that is not
     * owned by this cache.
     */
    assert(ret + thiscache->mt_size <= thiscache->mt_freelist +
        thiscache->mt_span);

    ret += OVERHEAD;

    assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */

    if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
        if (verify_pattern(FREEPATTERN, ret, size))
            abort();    /* reference after free */

    if (debugopt & MTINITBUFFER)
        copy_pattern(INITPATTERN, ret, size);
    return ((void *)ret);
}

static void *
morecore(size_t bytes)
{
    void * ret;

    if (bytes > LONG_MAX) {
        intptr_t wad;
        /*
         * The request size is too big. We need to do this in
         * chunks. Sbrk only takes an int for an arg.
         */
        if (bytes == ULONG_MAX)
            return ((void *)-1);

        ret = sbrk(0);
        wad = LONG_MAX;
        while (wad > 0) {
            if (sbrk(wad) == (void *)-1) {
                if (ret != sbrk(0))
                    (void) sbrk(-LONG_MAX);
                return ((void *)-1);
            }
            bytes -= LONG_MAX;
            wad = bytes;
        }
    } else
        ret = sbrk(bytes);

    return (ret);
}


static void *
oversize(size_t size)
{
    caddr_t ret;
    oversize_t *big;

    /* make sure we will not overflow */
    if (size > MAX_MTMALLOC) {
        errno = ENOMEM;
        return (NULL);
    }

    /*
     * Since we ensure every address we hand back is
     * MTMALLOC_MIN_ALIGN-byte aligned, ALIGNing size ensures that the
     * memory handed out is MTMALLOC_MIN_ALIGN-byte aligned at both ends.
     * This eases the implementation of MTDEBUGPATTERN and MTINITPATTERN,
     * particularly where coalescing occurs.
     */
    size = ALIGN(size, MTMALLOC_MIN_ALIGN);

    /*
     * The idea with the global lock is that we are sure to
     * block in the kernel anyway since given an oversize alloc
     * we are sure to have to call morecore();
     */
    (void) mutex_lock(&oversize_lock);

    if ((big = find_oversize(size)) != NULL) {
        if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
            if (verify_pattern(FREEPATTERN, big->addr, size))
                abort();    /* reference after free */
    } else {
        /* Get more 8-byte aligned memory from heap */
        ret = morecore(size + OVSZ_HEADER_SIZE);
        if (ret == (caddr_t)-1) {
            (void) mutex_unlock(&oversize_lock);
            errno = ENOMEM;
            return (NULL);
        }
        big = oversize_header_alloc((uintptr_t)ret, size);
    }
    ret = big->addr;

    insert_hash(big);

    if (debugopt & MTINITBUFFER)
        copy_pattern(INITPATTERN, ret, size);

    (void) mutex_unlock(&oversize_lock);
    assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */
    return ((void *)ret);
}

static void
insert_oversize(oversize_t *op, oversize_t *nx)
{
    oversize_t *sp;

    /* locate correct insertion point in size-ordered list */
    for (sp = oversize_list.next_bysize;
        sp != &oversize_list && (op->size > sp->size);
        sp = sp->next_bysize)
        ;

    /* link into size-ordered list */
    op->next_bysize = sp;
    op->prev_bysize = sp->prev_bysize;
    op->prev_bysize->next_bysize = op;
    op->next_bysize->prev_bysize = op;

    /*
     * link item into address-ordered list
     * (caller provides insertion point as an optimization)
     */
    op->next_byaddr = nx;
    op->prev_byaddr = nx->prev_byaddr;
    op->prev_byaddr->next_byaddr = op;
    op->next_byaddr->prev_byaddr = op;

}

static void
unlink_oversize(oversize_t *lp)
{
    /* unlink from address list */
    lp->prev_byaddr->next_byaddr = lp->next_byaddr;
    lp->next_byaddr->prev_byaddr = lp->prev_byaddr;

    /* unlink from size list */
    lp->prev_bysize->next_bysize = lp->next_bysize;
    lp->next_bysize->prev_bysize = lp->prev_bysize;
}

static void
position_oversize_by_size(oversize_t *op)
{
    oversize_t *sp;

    if (op->size > op->next_bysize->size ||
        op->size < op->prev_bysize->size) {

        /* unlink from size list */
        op->prev_bysize->next_bysize = op->next_bysize;
        op->next_bysize->prev_bysize = op->prev_bysize;

        /* locate correct insertion point in size-ordered list */
        for (sp = oversize_list.next_bysize;
            sp != &oversize_list && (op->size > sp->size);
            sp = sp->next_bysize)
            ;

        /* link into size-ordered list */
        op->next_bysize = sp;
        op->prev_bysize = sp->prev_bysize;
        op->prev_bysize->next_bysize = op;
        op->next_bysize->prev_bysize = op;
    }
}

static void
add_oversize(oversize_t *lp)
{
    int merge_flags = INSERT_ONLY;
    oversize_t *nx;     /* ptr to item right of insertion point */
    oversize_t *pv;     /* ptr to item left of insertion point */
    uint_t size_lp, size_pv, size_nx;
    uintptr_t endp_lp, endp_pv, endp_nx;

    /*
     * Locate insertion point in address-ordered list
     */

    for (nx = oversize_list.next_byaddr;
        nx != &oversize_list && (lp->addr > nx->addr);
        nx = nx->next_byaddr)
        ;

    /*
     * Determine how to add chunk to oversize freelist
     */

    size_lp = OVSZ_HEADER_SIZE + lp->size;
    endp_lp = ALIGN((uintptr_t)lp + size_lp, MTMALLOC_MIN_ALIGN);
    size_lp = endp_lp - (uintptr_t)lp;

    pv = nx->prev_byaddr;

    if (pv->size) {

        size_pv = OVSZ_HEADER_SIZE + pv->size;
        endp_pv = ALIGN((uintptr_t)pv + size_pv,
            MTMALLOC_MIN_ALIGN);
        size_pv = endp_pv - (uintptr_t)pv;

        /* Check for adjacency with left chunk */
        if ((uintptr_t)lp == endp_pv)
            merge_flags |= COALESCE_LEFT;
    }

    if (nx->size) {

        /* Check for adjacency with right chunk */
        if ((uintptr_t)nx == endp_lp) {
            size_nx = OVSZ_HEADER_SIZE + nx->size;
            endp_nx = ALIGN((uintptr_t)nx + size_nx,
                MTMALLOC_MIN_ALIGN);
            size_nx = endp_nx - (uintptr_t)nx;
            merge_flags |= COALESCE_RIGHT;
        }
    }

    /*
     * If MTDEBUGPATTERN==1, lp->addr will have been overwritten with
     * FREEPATTERN for lp->size bytes. If we can merge, the oversize
     * header(s) that will also become part of the memory available for
     * reallocation (ie lp and/or nx) must also be overwritten with
     * FREEPATTERN or we will SIGABRT when this memory is next reallocated.
     */
    switch (merge_flags) {

    case INSERT_ONLY:       /* Coalescing not possible */
        insert_oversize(lp, nx);
        break;
    case COALESCE_LEFT:
        pv->size += size_lp;
        position_oversize_by_size(pv);
        if (debugopt & MTDEBUGPATTERN)
            copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
        break;
    case COALESCE_RIGHT:
        unlink_oversize(nx);
        lp->size += size_nx;
        insert_oversize(lp, pv->next_byaddr);
        if (debugopt & MTDEBUGPATTERN)
            copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
        break;
    case COALESCE_WITH_BOTH_SIDES:  /* Merge (with right) to the left */
        pv->size += size_lp + size_nx;
        unlink_oversize(nx);
        position_oversize_by_size(pv);
        if (debugopt & MTDEBUGPATTERN) {
            copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
            copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
        }
        break;
    }
}

/*
 * Find memory on our list that is at least size big. If we find a block that is
 * big enough, we break it up and return the associated oversize_t struct back
 * to the calling client. Any leftover piece of that block is returned to the
 * freelist.
 */
static oversize_t *
find_oversize(size_t size)
{
    oversize_t *wp = oversize_list.next_bysize;
    while (wp != &oversize_list && size > wp->size)
        wp = wp->next_bysize;

    if (wp == &oversize_list) /* empty list or nothing big enough */
        return (NULL);
    /* breaking up a chunk of memory */
    if ((long)((wp->size - (size + OVSZ_HEADER_SIZE + MTMALLOC_MIN_ALIGN)))
        > MAX_CACHED) {
        caddr_t off;
        oversize_t *np;
        size_t osize;
        off = (caddr_t)ALIGN(wp->addr + size,
            MTMALLOC_MIN_ALIGN);
        osize = wp->size;
        wp->size = (size_t)(off - wp->addr);
        np = oversize_header_alloc((uintptr_t)off,
            osize - (wp->size + OVSZ_HEADER_SIZE));
        if ((long)np->size < 0)
            abort();
        unlink_oversize(wp);
        add_oversize(np);
    } else {
        unlink_oversize(wp);
    }
    return (wp);
}

static void
copy_pattern(uint32_t pattern, void *buf_arg, size_t size)
{
    uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
    uint32_t *buf = buf_arg;

    while (buf < bufend - 3) {
        buf[3] = buf[2] = buf[1] = buf[0] = pattern;
        buf += 4;
    }
    while (buf < bufend)
        *buf++ = pattern;
}

static void *
verify_pattern(uint32_t pattern, void *buf_arg, size_t size)
{
    uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
    uint32_t *buf;

    for (buf = buf_arg; buf < bufend; buf++)
        if (*buf != pattern)
            return (buf);
    return (NULL);
}

static void
free_oversize(oversize_t *ovp)
{
    assert(((uintptr_t)ovp->addr & 7) == 0); /* are we 8 byte aligned */
    assert(ovp->size > MAX_CACHED);

    ovp->next_bysize = ovp->prev_bysize = NULL;
    ovp->next_byaddr = ovp->prev_byaddr = NULL;
    (void) mutex_lock(&oversize_lock);
    add_oversize(ovp);
    (void) mutex_unlock(&oversize_lock);
}

static oversize_t *
oversize_header_alloc(uintptr_t mem, size_t size)
{
    oversize_t *ovsz_hdr;

    assert(size > MAX_CACHED);

    ovsz_hdr = (oversize_t *)mem;
    ovsz_hdr->prev_bysize = NULL;
    ovsz_hdr->next_bysize = NULL;
    ovsz_hdr->prev_byaddr = NULL;
    ovsz_hdr->next_byaddr = NULL;
    ovsz_hdr->hash_next = NULL;
    ovsz_hdr->size = size;
    mem += OVSZ_SIZE;
    *(uintptr_t *)mem = MTMALLOC_OVERSIZE_MAGIC;
    mem += OVERHEAD;
    assert(((uintptr_t)mem & 7) == 0); /* are we 8 byte aligned */
    ovsz_hdr->addr = (caddr_t)mem;
    return (ovsz_hdr);
}

static void
malloc_prepare()
{
    percpu_t *cpuptr;
    cache_head_t *cachehead;
    cache_t *thiscache;

    (void) mutex_lock(&oversize_lock);
    for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
        (void) mutex_lock(&cpuptr->mt_parent_lock);
        for (cachehead = &cpuptr->mt_caches[0];
            cachehead < &cpuptr->mt_caches[NUM_CACHES];
            cachehead++) {
            for (thiscache = cachehead->mt_cache;
                thiscache != NULL;
                thiscache = thiscache->mt_next) {
                (void) mutex_lock(
                    &thiscache->mt_cache_lock);
            }
        }
    }
}

static void
malloc_release()
{
    percpu_t *cpuptr;
    cache_head_t *cachehead;
    cache_t *thiscache;

    for (cpuptr = &cpu_list[ncpus - 1]; cpuptr >= &cpu_list[0]; cpuptr--) {
        for (cachehead = &cpuptr->mt_caches[NUM_CACHES - 1];
            cachehead >= &cpuptr->mt_caches[0];
            cachehead--) {
            for (thiscache = cachehead->mt_cache;
                thiscache != NULL;
                thiscache = thiscache->mt_next) {
                (void) mutex_unlock(
                    &thiscache->mt_cache_lock);
            }
        }
        (void) mutex_unlock(&cpuptr->mt_parent_lock);
    }
    (void) mutex_unlock(&oversize_lock);
}

#pragma init(malloc_init)
static void
malloc_init(void)
{
    /*
     * This works in the init section for this library
     * because setup_caches() doesn't call anything in libc
     * that calls malloc().  If it did, disaster would ensue.
     *
     * For this to work properly, this library must be the first
     * one to have its init section called (after libc) by the
     * dynamic linker.  If some other library's init section
     * ran first and called malloc(), disaster would ensue.
     * Because this is an interposer library for malloc(), the
     * dynamic linker arranges for its init section to run first.
     */
    (void) setup_caches();

    (void) pthread_atfork(malloc_prepare, malloc_release, malloc_release);
}