/*
 * jmemmgr.c
 *
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
 * Copyright (C) 2016, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
 * This file contains the JPEG system-independent memory management
 * routines.  This code is usable across a wide variety of machines; most
 * of the system dependencies have been isolated in a separate file.
 * The major functions provided here are:
 *   * pool-based allocation and freeing of memory;
 *   * policy decisions about how to divide available memory among the
 *     virtual arrays;
 *   * control logic for swapping virtual arrays between main memory and
 *     backing storage.
 * The separate system-dependent file provides the actual backing-storage
 * access code, and it contains the policy decision about how much total
 * main memory to use.
 * This file is system-dependent in the sense that some of its functions
 * are unnecessary in some systems.  For example, if there is enough virtual
 * memory so that backing storage will never be used, much of the virtual
 * array control logic could be removed.  (Of course, if you have that much
 * memory then you shouldn't care about a little bit of unused code...)
 */

#define JPEG_INTERNALS
#define AM_MEMORY_MANAGER       /* we define jvirt_Xarray_control structs */
#include "jinclude.h"
#include "jpeglib.h"
#include "jmemsys.h"            /* import the system-dependent declarations */
#if !defined(_MSC_VER) || _MSC_VER > 1600
#include <stdint.h>
#endif
#include <limits.h>

#ifndef NO_GETENV
#ifndef HAVE_STDLIB_H           /* <stdlib.h> should declare getenv() */
extern char *getenv(const char *name);
#endif
#endif


LOCAL(size_t)
round_up_pow2(size_t a, size_t b)
/* a rounded up to the next multiple of b, i.e. ceil(a/b)*b */
/* Assumes a >= 0, b > 0, and b is a power of 2 */
{
  return ((a + b - 1) & (~(b - 1)));
}


/*
 * Some important notes:
 *   The allocation routines provided here must never return NULL.
 *   They should exit to error_exit if unsuccessful.
 *
 *   It's not a good idea to try to merge the sarray and barray routines,
 *   even though they are textually almost the same, because samples are
 *   usually stored as bytes while coefficients are shorts or ints.  Thus,
 *   in machines where byte pointers have a different representation from
 *   word pointers, the resulting machine code could not be the same.
 */


/*
 * Many machines require storage alignment: longs must start on 4-byte
 * boundaries, doubles on 8-byte boundaries, etc.  On such machines, malloc()
 * always returns pointers that are multiples of the worst-case alignment
 * requirement, and we had better do so too.
 * There isn't any really portable way to determine the worst-case alignment
 * requirement.  This module assumes that the alignment requirement is
 * multiples of ALIGN_SIZE.
 * By default, we define ALIGN_SIZE as sizeof(double).  This is necessary on
 * some workstations (where doubles really do need 8-byte alignment) and will
 * work fine on nearly everything.  If your machine has lesser alignment needs,
 * you can save a few bytes by making ALIGN_SIZE smaller.
 * The only place I know of where this will NOT work is certain Macintosh
 * 680x0 compilers that define double as a 10-byte IEEE extended float.
 * Doing 10-byte alignment is counterproductive because longwords won't be
 * aligned well.  Put "#define ALIGN_SIZE 4" in jconfig.h if you have
 * such a compiler.
 */

#ifndef ALIGN_SIZE              /* so can override from jconfig.h */
#ifndef WITH_SIMD
#define ALIGN_SIZE  sizeof(double)
#else
#define ALIGN_SIZE  32 /* Most of the SIMD instructions we support require
                          16-byte (128-bit) alignment, but AVX2 requires
                          32-byte alignment. */
#endif
#endif

#ifdef WITH_SIMD
#if (ALIGN_SIZE % 16)
  #error "ALIGN_SIZE is not a multiple of 16 bytes - required for SIMD instructions."
#endif
#if defined(__AVX2__) && (ALIGN_SIZE % 32)
  #error "AVX2 requires 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes."
#elif defined(__ARM_NEON) && (ALIGN_SIZE % 32)
  /* 32-byte alignment allows us to extract more performance from */
  /* fancy-upsampling algorithms when using NEON. */
  #error "NEON optimizations rely on 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes."
#endif
#endif

/*
 * We allocate objects from "pools", where each pool is gotten with a single
 * request to jpeg_get_small() or jpeg_get_large().  There is no per-object
 * overhead within a pool, except for alignment padding.  Each pool has a
 * header with a link to the next pool of the same class.
 * Small and large pool headers are identical.
 */

typedef struct small_pool_struct *small_pool_ptr;

typedef struct small_pool_struct {
  small_pool_ptr next;          /* next in list of pools */
  size_t bytes_used;            /* how many bytes already used within pool */
  size_t bytes_left;            /* bytes still available in this pool */
} small_pool_hdr;

typedef struct large_pool_struct *large_pool_ptr;

typedef struct large_pool_struct {
  large_pool_ptr next;          /* next in list of pools */
  size_t bytes_used;            /* how many bytes already used within pool */
  size_t bytes_left;            /* bytes still available in this pool */
} large_pool_hdr;

/*
 * Here is the full definition of a memory manager object.
 */

typedef struct {
  struct jpeg_memory_mgr pub;   /* public fields */

  /* Each pool identifier (lifetime class) names a linked list of pools. */
  small_pool_ptr small_list[JPOOL_NUMPOOLS];
  large_pool_ptr large_list[JPOOL_NUMPOOLS];

  /* Since we only have one lifetime class of virtual arrays, only one
   * linked list is necessary (for each datatype).  Note that the virtual
   * array control blocks being linked together are actually stored somewhere
   * in the small-pool list.
   */
  jvirt_sarray_ptr virt_sarray_list;
  jvirt_barray_ptr virt_barray_list;

  /* This counts total space obtained from jpeg_get_small/large */
  size_t total_space_allocated;

  /* alloc_sarray and alloc_barray set this value for use by virtual
   * array routines.
   */
  JDIMENSION last_rowsperchunk; /* from most recent alloc_sarray/barray */
} my_memory_mgr;

typedef my_memory_mgr *my_mem_ptr;


/*
 * The control blocks for virtual arrays.
 * Note that these blocks are allocated in the "small" pool area.
 * System-dependent info for the associated backing store (if any) is hidden
 * inside the backing_store_info struct.
 */

struct jvirt_sarray_control {
  JSAMPARRAY mem_buffer;        /* => the in-memory buffer */
  JDIMENSION rows_in_array;     /* total virtual array height */
  JDIMENSION samplesperrow;     /* width of array (and of memory buffer) */
  JDIMENSION maxaccess;         /* max rows accessed by access_virt_sarray */
  JDIMENSION rows_in_mem;       /* height of memory buffer */
  JDIMENSION rowsperchunk;      /* allocation chunk size in mem_buffer */
  JDIMENSION cur_start_row;     /* first logical row # in the buffer */
  JDIMENSION first_undef_row;   /* row # of first uninitialized row */
  boolean pre_zero;             /* pre-zero mode requested? */
  boolean dirty;                /* do current buffer contents need written? */
  boolean b_s_open;             /* is backing-store data valid? */
  jvirt_sarray_ptr next;        /* link to next virtual sarray control block */
  backing_store_info b_s_info;  /* System-dependent control info */
};

struct jvirt_barray_control {
  JBLOCKARRAY mem_buffer;       /* => the in-memory buffer */
  JDIMENSION rows_in_array;     /* total virtual array height */
  JDIMENSION blocksperrow;      /* width of array (and of memory buffer) */
  JDIMENSION maxaccess;         /* max rows accessed by access_virt_barray */
  JDIMENSION rows_in_mem;       /* height of memory buffer */
  JDIMENSION rowsperchunk;      /* allocation chunk size in mem_buffer */
  JDIMENSION cur_start_row;     /* first logical row # in the buffer */
  JDIMENSION first_undef_row;   /* row # of first uninitialized row */
  boolean pre_zero;             /* pre-zero mode requested? */
  boolean dirty;                /* do current buffer contents need written? */
  boolean b_s_open;             /* is backing-store data valid? */
  jvirt_barray_ptr next;        /* link to next virtual barray control block */
  backing_store_info b_s_info;  /* System-dependent control info */
};


#ifdef MEM_STATS                /* optional extra stuff for statistics */

LOCAL(void)
print_mem_stats(j_common_ptr cinfo, int pool_id)
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  small_pool_ptr shdr_ptr;
  large_pool_ptr lhdr_ptr;

  /* Since this is only a debugging stub, we can cheat a little by using
   * fprintf directly rather than going through the trace message code.
   * This is helpful because message parm array can't handle longs.
   */
  fprintf(stderr, "Freeing pool %d, total space = %ld\n",
          pool_id, mem->total_space_allocated);

  for (lhdr_ptr = mem->large_list[pool_id]; lhdr_ptr != NULL;
       lhdr_ptr = lhdr_ptr->next) {
    fprintf(stderr, "  Large chunk used %ld\n", (long)lhdr_ptr->bytes_used);
  }

  for (shdr_ptr = mem->small_list[pool_id]; shdr_ptr != NULL;
       shdr_ptr = shdr_ptr->next) {
    fprintf(stderr, "  Small chunk used %ld free %ld\n",
            (long)shdr_ptr->bytes_used, (long)shdr_ptr->bytes_left);
  }
}

#endif /* MEM_STATS */


LOCAL(void)
out_of_memory(j_common_ptr cinfo, int which)
/* Report an out-of-memory error and stop execution */
/* If we compiled MEM_STATS support, report alloc requests before dying */
{
#ifdef MEM_STATS
  cinfo->err->trace_level = 2;  /* force self_destruct to report stats */
#endif
  ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, which);
}


/*
 * Allocation of "small" objects.
 *
 * For these, we use pooled storage.  When a new pool must be created,
 * we try to get enough space for the current request plus a "slop" factor,
 * where the slop will be the amount of leftover space in the new pool.
 * The speed vs. space tradeoff is largely determined by the slop values.
 * A different slop value is provided for each pool class (lifetime),
 * and we also distinguish the first pool of a class from later ones.
 * NOTE: the values given work fairly well on both 16- and 32-bit-int
 * machines, but may be too small if longs are 64 bits or more.
 *
 * Since we do not know what alignment malloc() gives us, we have to
 * allocate ALIGN_SIZE-1 extra space per pool to have room for alignment
 * adjustment.
 */

static const size_t first_pool_slop[JPOOL_NUMPOOLS] = {
  1600,                         /* first PERMANENT pool */
  16000                         /* first IMAGE pool */
};

static const size_t extra_pool_slop[JPOOL_NUMPOOLS] = {
  0,                            /* additional PERMANENT pools */
  5000                          /* additional IMAGE pools */
};

#define MIN_SLOP  50            /* greater than 0 to avoid futile looping */


METHODDEF(void *)
alloc_small(j_common_ptr cinfo, int pool_id, size_t sizeofobject)
/* Allocate a "small" object */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  small_pool_ptr hdr_ptr, prev_hdr_ptr;
  char *data_ptr;
  size_t min_request, slop;

  /*
   * Round up the requested size to a multiple of ALIGN_SIZE in order
   * to assure alignment for the next object allocated in the same pool
   * and so that algorithms can straddle outside the proper area up
   * to the next alignment.
   */
  if (sizeofobject > MAX_ALLOC_CHUNK) {
    /* This prevents overflow/wrap-around in round_up_pow2() if sizeofobject
       is close to SIZE_MAX. */
    out_of_memory(cinfo, 7);
  }
  sizeofobject = round_up_pow2(sizeofobject, ALIGN_SIZE);

  /* Check for unsatisfiable request (do now to ensure no overflow below) */
  if ((sizeof(small_pool_hdr) + sizeofobject + ALIGN_SIZE - 1) >
      MAX_ALLOC_CHUNK)
    out_of_memory(cinfo, 1);    /* request exceeds malloc's ability */

  /* See if space is available in any existing pool */
  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */
  prev_hdr_ptr = NULL;
  hdr_ptr = mem->small_list[pool_id];
  while (hdr_ptr != NULL) {
    if (hdr_ptr->bytes_left >= sizeofobject)
      break;                    /* found pool with enough space */
    prev_hdr_ptr = hdr_ptr;
    hdr_ptr = hdr_ptr->next;
  }

  /* Time to make a new pool? */
  if (hdr_ptr == NULL) {
    /* min_request is what we need now, slop is what will be leftover */
    min_request = sizeof(small_pool_hdr) + sizeofobject + ALIGN_SIZE - 1;
    if (prev_hdr_ptr == NULL)   /* first pool in class? */
      slop = first_pool_slop[pool_id];
    else
      slop = extra_pool_slop[pool_id];
    /* Don't ask for more than MAX_ALLOC_CHUNK */
    if (slop > (size_t)(MAX_ALLOC_CHUNK - min_request))
      slop = (size_t)(MAX_ALLOC_CHUNK - min_request);
    /* Try to get space, if fail reduce slop and try again */
    for (;;) {
      hdr_ptr = (small_pool_ptr)jpeg_get_small(cinfo, min_request + slop);
      if (hdr_ptr != NULL)
        break;
      slop /= 2;
      if (slop < MIN_SLOP)      /* give up when it gets real small */
        out_of_memory(cinfo, 2); /* jpeg_get_small failed */
    }
    mem->total_space_allocated += min_request + slop;
    /* Success, initialize the new pool header and add to end of list */
    hdr_ptr->next = NULL;
    hdr_ptr->bytes_used = 0;
    hdr_ptr->bytes_left = sizeofobject + slop;
    if (prev_hdr_ptr == NULL)   /* first pool in class? */
      mem->small_list[pool_id] = hdr_ptr;
    else
      prev_hdr_ptr->next = hdr_ptr;
  }

  /* OK, allocate the object from the current pool */
  data_ptr = (char *)hdr_ptr; /* point to first data byte in pool... */
  data_ptr += sizeof(small_pool_hdr); /* ...by skipping the header... */
  if ((size_t)data_ptr % ALIGN_SIZE) /* ...and adjust for alignment */
    data_ptr += ALIGN_SIZE - (size_t)data_ptr % ALIGN_SIZE;
  data_ptr += hdr_ptr->bytes_used; /* point to place for object */
  hdr_ptr->bytes_used += sizeofobject;
  hdr_ptr->bytes_left -= sizeofobject;

  return (void *)data_ptr;
}


/*
 * Allocation of "large" objects.
 *
 * The external semantics of these are the same as "small" objects.  However,
 * the pool management heuristics are quite different.  We assume that each
 * request is large enough that it may as well be passed directly to
 * jpeg_get_large; the pool management just links everything together
 * so that we can free it all on demand.
 * Note: the major use of "large" objects is in JSAMPARRAY and JBLOCKARRAY
 * structures.  The routines that create these structures (see below)
 * deliberately bunch rows together to ensure a large request size.
 */

METHODDEF(void *)
alloc_large(j_common_ptr cinfo, int pool_id, size_t sizeofobject)
/* Allocate a "large" object */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  large_pool_ptr hdr_ptr;
  char *data_ptr;

  /*
   * Round up the requested size to a multiple of ALIGN_SIZE so that
   * algorithms can straddle outside the proper area up to the next
   * alignment.
   */
  if (sizeofobject > MAX_ALLOC_CHUNK) {
    /* This prevents overflow/wrap-around in round_up_pow2() if sizeofobject
       is close to SIZE_MAX. */
    out_of_memory(cinfo, 8);
  }
  sizeofobject = round_up_pow2(sizeofobject, ALIGN_SIZE);

  /* Check for unsatisfiable request (do now to ensure no overflow below) */
  if ((sizeof(large_pool_hdr) + sizeofobject + ALIGN_SIZE - 1) >
      MAX_ALLOC_CHUNK)
    out_of_memory(cinfo, 3);    /* request exceeds malloc's ability */

  /* Always make a new pool */
  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */

  hdr_ptr = (large_pool_ptr)jpeg_get_large(cinfo, sizeofobject +
                                           sizeof(large_pool_hdr) +
                                           ALIGN_SIZE - 1);
  if (hdr_ptr == NULL)
    out_of_memory(cinfo, 4);    /* jpeg_get_large failed */
  mem->total_space_allocated += sizeofobject + sizeof(large_pool_hdr) +
                                ALIGN_SIZE - 1;

  /* Success, initialize the new pool header and add to list */
  hdr_ptr->next = mem->large_list[pool_id];
  /* We maintain space counts in each pool header for statistical purposes,
   * even though they are not needed for allocation.
   */
  hdr_ptr->bytes_used = sizeofobject;
  hdr_ptr->bytes_left = 0;
  mem->large_list[pool_id] = hdr_ptr;

  data_ptr = (char *)hdr_ptr; /* point to first data byte in pool... */
  data_ptr += sizeof(small_pool_hdr); /* ...by skipping the header... */
  if ((size_t)data_ptr % ALIGN_SIZE) /* ...and adjust for alignment */
    data_ptr += ALIGN_SIZE - (size_t)data_ptr % ALIGN_SIZE;

  return (void *)data_ptr;
}


/*
 * Creation of 2-D sample arrays.
 *
 * To minimize allocation overhead and to allow I/O of large contiguous
 * blocks, we allocate the sample rows in groups of as many rows as possible
 * without exceeding MAX_ALLOC_CHUNK total bytes per allocation request.
 * NB: the virtual array control routines, later in this file, know about
 * this chunking of rows.  The rowsperchunk value is left in the mem manager
 * object so that it can be saved away if this sarray is the workspace for
 * a virtual array.
 *
 * Since we are often upsampling with a factor 2, we align the size (not
 * the start) to 2 * ALIGN_SIZE so that the upsampling routines don't have
 * to be as careful about size.
 */

METHODDEF(JSAMPARRAY)
alloc_sarray(j_common_ptr cinfo, int pool_id, JDIMENSION samplesperrow,
             JDIMENSION numrows)
/* Allocate a 2-D sample array */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  JSAMPARRAY result;
  JSAMPROW workspace;
  JDIMENSION rowsperchunk, currow, i;
  long ltemp;

  /* Make sure each row is properly aligned */
  if ((ALIGN_SIZE % sizeof(JSAMPLE)) != 0)
    out_of_memory(cinfo, 5);    /* safety check */

  if (samplesperrow > MAX_ALLOC_CHUNK) {
    /* This prevents overflow/wrap-around in round_up_pow2() if sizeofobject
       is close to SIZE_MAX. */
    out_of_memory(cinfo, 9);
  }
  samplesperrow = (JDIMENSION)round_up_pow2(samplesperrow, (2 * ALIGN_SIZE) /
                                                           sizeof(JSAMPLE));

  /* Calculate max # of rows allowed in one allocation chunk */
  ltemp = (MAX_ALLOC_CHUNK - sizeof(large_pool_hdr)) /
          ((long)samplesperrow * sizeof(JSAMPLE));
  if (ltemp <= 0)
    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
  if (ltemp < (long)numrows)
    rowsperchunk = (JDIMENSION)ltemp;
  else
    rowsperchunk = numrows;
  mem->last_rowsperchunk = rowsperchunk;

  /* Get space for row pointers (small object) */
  result = (JSAMPARRAY)alloc_small(cinfo, pool_id,
                                   (size_t)(numrows * sizeof(JSAMPROW)));

  /* Get the rows themselves (large objects) */
  currow = 0;
  while (currow < numrows) {
    rowsperchunk = MIN(rowsperchunk, numrows - currow);
    workspace = (JSAMPROW)alloc_large(cinfo, pool_id,
      (size_t)((size_t)rowsperchunk * (size_t)samplesperrow *
               sizeof(JSAMPLE)));
    for (i = rowsperchunk; i > 0; i--) {
      result[currow++] = workspace;
      workspace += samplesperrow;
    }
  }

  return result;
}


/*
 * Creation of 2-D coefficient-block arrays.
 * This is essentially the same as the code for sample arrays, above.
 */

METHODDEF(JBLOCKARRAY)
alloc_barray(j_common_ptr cinfo, int pool_id, JDIMENSION blocksperrow,
             JDIMENSION numrows)
/* Allocate a 2-D coefficient-block array */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  JBLOCKARRAY result;
  JBLOCKROW workspace;
  JDIMENSION rowsperchunk, currow, i;
  long ltemp;

  /* Make sure each row is properly aligned */
  if ((sizeof(JBLOCK) % ALIGN_SIZE) != 0)
    out_of_memory(cinfo, 6);    /* safety check */

  /* Calculate max # of rows allowed in one allocation chunk */
  ltemp = (MAX_ALLOC_CHUNK - sizeof(large_pool_hdr)) /
          ((long)blocksperrow * sizeof(JBLOCK));
  if (ltemp <= 0)
    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
  if (ltemp < (long)numrows)
    rowsperchunk = (JDIMENSION)ltemp;
  else
    rowsperchunk = numrows;
  mem->last_rowsperchunk = rowsperchunk;

  /* Get space for row pointers (small object) */
  result = (JBLOCKARRAY)alloc_small(cinfo, pool_id,
                                    (size_t)(numrows * sizeof(JBLOCKROW)));

  /* Get the rows themselves (large objects) */
  currow = 0;
  while (currow < numrows) {
    rowsperchunk = MIN(rowsperchunk, numrows - currow);
    workspace = (JBLOCKROW)alloc_large(cinfo, pool_id,
        (size_t)((size_t)rowsperchunk * (size_t)blocksperrow *
                  sizeof(JBLOCK)));
    for (i = rowsperchunk; i > 0; i--) {
      result[currow++] = workspace;
      workspace += blocksperrow;
    }
  }

  return result;
}


/*
 * About virtual array management:
 *
 * The above "normal" array routines are only used to allocate strip buffers
 * (as wide as the image, but just a few rows high).  Full-image-sized buffers
 * are handled as "virtual" arrays.  The array is still accessed a strip at a
 * time, but the memory manager must save the whole array for repeated
 * accesses.  The intended implementation is that there is a strip buffer in
 * memory (as high as is possible given the desired memory limit), plus a
 * backing file that holds the rest of the array.
 *
 * The request_virt_array routines are told the total size of the image and
 * the maximum number of rows that will be accessed at once.  The in-memory
 * buffer must be at least as large as the maxaccess value.
 *
 * The request routines create control blocks but not the in-memory buffers.
 * That is postponed until realize_virt_arrays is called.  At that time the
 * total amount of space needed is known (approximately, anyway), so free
 * memory can be divided up fairly.
 *
 * The access_virt_array routines are responsible for making a specific strip
 * area accessible (after reading or writing the backing file, if necessary).
 * Note that the access routines are told whether the caller intends to modify
 * the accessed strip; during a read-only pass this saves having to rewrite
 * data to disk.  The access routines are also responsible for pre-zeroing
 * any newly accessed rows, if pre-zeroing was requested.
 *
 * In current usage, the access requests are usually for nonoverlapping
 * strips; that is, successive access start_row numbers differ by exactly
 * num_rows = maxaccess.  This means we can get good performance with simple
 * buffer dump/reload logic, by making the in-memory buffer be a multiple
 * of the access height; then there will never be accesses across bufferload
 * boundaries.  The code will still work with overlapping access requests,
 * but it doesn't handle bufferload overlaps very efficiently.
 */


METHODDEF(jvirt_sarray_ptr)
request_virt_sarray(j_common_ptr cinfo, int pool_id, boolean pre_zero,
                    JDIMENSION samplesperrow, JDIMENSION numrows,
                    JDIMENSION maxaccess)
/* Request a virtual 2-D sample array */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  jvirt_sarray_ptr result;

  /* Only IMAGE-lifetime virtual arrays are currently supported */
  if (pool_id != JPOOL_IMAGE)
    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */

  /* get control block */
  result = (jvirt_sarray_ptr)alloc_small(cinfo, pool_id,
                                         sizeof(struct jvirt_sarray_control));

  result->mem_buffer = NULL;    /* marks array not yet realized */
  result->rows_in_array = numrows;
  result->samplesperrow = samplesperrow;
  result->maxaccess = maxaccess;
  result->pre_zero = pre_zero;
  result->b_s_open = FALSE;     /* no associated backing-store object */
  result->next = mem->virt_sarray_list; /* add to list of virtual arrays */
  mem->virt_sarray_list = result;

  return result;
}


METHODDEF(jvirt_barray_ptr)
request_virt_barray(j_common_ptr cinfo, int pool_id, boolean pre_zero,
                    JDIMENSION blocksperrow, JDIMENSION numrows,
                    JDIMENSION maxaccess)
/* Request a virtual 2-D coefficient-block array */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  jvirt_barray_ptr result;

  /* Only IMAGE-lifetime virtual arrays are currently supported */
  if (pool_id != JPOOL_IMAGE)
    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */

  /* get control block */
  result = (jvirt_barray_ptr)alloc_small(cinfo, pool_id,
                                         sizeof(struct jvirt_barray_control));

  result->mem_buffer = NULL;    /* marks array not yet realized */
  result->rows_in_array = numrows;
  result->blocksperrow = blocksperrow;
  result->maxaccess = maxaccess;
  result->pre_zero = pre_zero;
  result->b_s_open = FALSE;     /* no associated backing-store object */
  result->next = mem->virt_barray_list; /* add to list of virtual arrays */
  mem->virt_barray_list = result;

  return result;
}


METHODDEF(void)
realize_virt_arrays(j_common_ptr cinfo)
/* Allocate the in-memory buffers for any unrealized virtual arrays */
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  size_t space_per_minheight, maximum_space, avail_mem;
  size_t minheights, max_minheights;
  jvirt_sarray_ptr sptr;
  jvirt_barray_ptr bptr;

  /* Compute the minimum space needed (maxaccess rows in each buffer)
   * and the maximum space needed (full image height in each buffer).
   * These may be of use to the system-dependent jpeg_mem_available routine.
   */
  space_per_minheight = 0;
  maximum_space = 0;
  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
    if (sptr->mem_buffer == NULL) { /* if not realized yet */
      size_t new_space = (long)sptr->rows_in_array *
                         (long)sptr->samplesperrow * sizeof(JSAMPLE);

      space_per_minheight += (long)sptr->maxaccess *
                             (long)sptr->samplesperrow * sizeof(JSAMPLE);
      if (SIZE_MAX - maximum_space < new_space)
        out_of_memory(cinfo, 10);
      maximum_space += new_space;
    }
  }
  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
    if (bptr->mem_buffer == NULL) { /* if not realized yet */
      size_t new_space = (long)bptr->rows_in_array *
                         (long)bptr->blocksperrow * sizeof(JBLOCK);

      space_per_minheight += (long)bptr->maxaccess *
                             (long)bptr->blocksperrow * sizeof(JBLOCK);
      if (SIZE_MAX - maximum_space < new_space)
        out_of_memory(cinfo, 11);
      maximum_space += new_space;
    }
  }

  if (space_per_minheight <= 0)
    return;                     /* no unrealized arrays, no work */

  /* Determine amount of memory to actually use; this is system-dependent. */
  avail_mem = jpeg_mem_available(cinfo, space_per_minheight, maximum_space,
                                 mem->total_space_allocated);

  /* If the maximum space needed is available, make all the buffers full
   * height; otherwise parcel it out with the same number of minheights
   * in each buffer.
   */
  if (avail_mem >= maximum_space)
    max_minheights = 1000000000L;
  else {
    max_minheights = avail_mem / space_per_minheight;
    /* If there doesn't seem to be enough space, try to get the minimum
     * anyway.  This allows a "stub" implementation of jpeg_mem_available().
     */
    if (max_minheights <= 0)
      max_minheights = 1;
  }

  /* Allocate the in-memory buffers and initialize backing store as needed. */

  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
    if (sptr->mem_buffer == NULL) { /* if not realized yet */
      minheights = ((long)sptr->rows_in_array - 1L) / sptr->maxaccess + 1L;
      if (minheights <= max_minheights) {
        /* This buffer fits in memory */
        sptr->rows_in_mem = sptr->rows_in_array;
      } else {
        /* It doesn't fit in memory, create backing store. */
        sptr->rows_in_mem = (JDIMENSION)(max_minheights * sptr->maxaccess);
        jpeg_open_backing_store(cinfo, &sptr->b_s_info,
                                (long)sptr->rows_in_array *
                                (long)sptr->samplesperrow *
                                (long)sizeof(JSAMPLE));
        sptr->b_s_open = TRUE;
      }
      sptr->mem_buffer = alloc_sarray(cinfo, JPOOL_IMAGE,
                                      sptr->samplesperrow, sptr->rows_in_mem);
      sptr->rowsperchunk = mem->last_rowsperchunk;
      sptr->cur_start_row = 0;
      sptr->first_undef_row = 0;
      sptr->dirty = FALSE;
    }
  }

  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
    if (bptr->mem_buffer == NULL) { /* if not realized yet */
      minheights = ((long)bptr->rows_in_array - 1L) / bptr->maxaccess + 1L;
      if (minheights <= max_minheights) {
        /* This buffer fits in memory */
        bptr->rows_in_mem = bptr->rows_in_array;
      } else {
        /* It doesn't fit in memory, create backing store. */
        bptr->rows_in_mem = (JDIMENSION)(max_minheights * bptr->maxaccess);
        jpeg_open_backing_store(cinfo, &bptr->b_s_info,
                                (long)bptr->rows_in_array *
                                (long)bptr->blocksperrow *
                                (long)sizeof(JBLOCK));
        bptr->b_s_open = TRUE;
      }
      bptr->mem_buffer = alloc_barray(cinfo, JPOOL_IMAGE,
                                      bptr->blocksperrow, bptr->rows_in_mem);
      bptr->rowsperchunk = mem->last_rowsperchunk;
      bptr->cur_start_row = 0;
      bptr->first_undef_row = 0;
      bptr->dirty = FALSE;
    }
  }
}


LOCAL(void)
do_sarray_io(j_common_ptr cinfo, jvirt_sarray_ptr ptr, boolean writing)
/* Do backing store read or write of a virtual sample array */
{
  long bytesperrow, file_offset, byte_count, rows, thisrow, i;

  bytesperrow = (long)ptr->samplesperrow * sizeof(JSAMPLE);
  file_offset = ptr->cur_start_row * bytesperrow;
  /* Loop to read or write each allocation chunk in mem_buffer */
  for (i = 0; i < (long)ptr->rows_in_mem; i += ptr->rowsperchunk) {
    /* One chunk, but check for short chunk at end of buffer */
    rows = MIN((long)ptr->rowsperchunk, (long)ptr->rows_in_mem - i);
    /* Transfer no more than is currently defined */
    thisrow = (long)ptr->cur_start_row + i;
    rows = MIN(rows, (long)ptr->first_undef_row - thisrow);
    /* Transfer no more than fits in file */
    rows = MIN(rows, (long)ptr->rows_in_array - thisrow);
    if (rows <= 0)              /* this chunk might be past end of file! */
      break;
    byte_count = rows * bytesperrow;
    if (writing)
      (*ptr->b_s_info.write_backing_store) (cinfo, &ptr->b_s_info,
                                            (void *)ptr->mem_buffer[i],
                                            file_offset, byte_count);
    else
      (*ptr->b_s_info.read_backing_store) (cinfo, &ptr->b_s_info,
                                           (void *)ptr->mem_buffer[i],
                                           file_offset, byte_count);
    file_offset += byte_count;
  }
}


LOCAL(void)
do_barray_io(j_common_ptr cinfo, jvirt_barray_ptr ptr, boolean writing)
/* Do backing store read or write of a virtual coefficient-block array */
{
  long bytesperrow, file_offset, byte_count, rows, thisrow, i;

  bytesperrow = (long)ptr->blocksperrow * sizeof(JBLOCK);
  file_offset = ptr->cur_start_row * bytesperrow;
  /* Loop to read or write each allocation chunk in mem_buffer */
  for (i = 0; i < (long)ptr->rows_in_mem; i += ptr->rowsperchunk) {
    /* One chunk, but check for short chunk at end of buffer */
    rows = MIN((long)ptr->rowsperchunk, (long)ptr->rows_in_mem - i);
    /* Transfer no more than is currently defined */
    thisrow = (long)ptr->cur_start_row + i;
    rows = MIN(rows, (long)ptr->first_undef_row - thisrow);
    /* Transfer no more than fits in file */
    rows = MIN(rows, (long)ptr->rows_in_array - thisrow);
    if (rows <= 0)              /* this chunk might be past end of file! */
      break;
    byte_count = rows * bytesperrow;
    if (writing)
      (*ptr->b_s_info.write_backing_store) (cinfo, &ptr->b_s_info,
                                            (void *)ptr->mem_buffer[i],
                                            file_offset, byte_count);
    else
      (*ptr->b_s_info.read_backing_store) (cinfo, &ptr->b_s_info,
                                           (void *)ptr->mem_buffer[i],
                                           file_offset, byte_count);
    file_offset += byte_count;
  }
}


METHODDEF(JSAMPARRAY)
access_virt_sarray(j_common_ptr cinfo, jvirt_sarray_ptr ptr,
                   JDIMENSION start_row, JDIMENSION num_rows, boolean writable)
/* Access the part of a virtual sample array starting at start_row */
/* and extending for num_rows rows.  writable is true if  */
/* caller intends to modify the accessed area. */
{
  JDIMENSION end_row = start_row + num_rows;
  JDIMENSION undef_row;

  /* debugging check */
  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
      ptr->mem_buffer == NULL)
    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);

  /* Make the desired part of the virtual array accessible */
  if (start_row < ptr->cur_start_row ||
      end_row > ptr->cur_start_row + ptr->rows_in_mem) {
    if (!ptr->b_s_open)
      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
    /* Flush old buffer contents if necessary */
    if (ptr->dirty) {
      do_sarray_io(cinfo, ptr, TRUE);
      ptr->dirty = FALSE;
    }
    /* Decide what part of virtual array to access.
     * Algorithm: if target address > current window, assume forward scan,
     * load starting at target address.  If target address < current window,
     * assume backward scan, load so that target area is top of window.
     * Note that when switching from forward write to forward read, will have
     * start_row = 0, so the limiting case applies and we load from 0 anyway.
     */
    if (start_row > ptr->cur_start_row) {
      ptr->cur_start_row = start_row;
    } else {
      /* use long arithmetic here to avoid overflow & unsigned problems */
      long ltemp;

      ltemp = (long)end_row - (long)ptr->rows_in_mem;
      if (ltemp < 0)
        ltemp = 0;              /* don't fall off front end of file */
      ptr->cur_start_row = (JDIMENSION)ltemp;
    }
    /* Read in the selected part of the array.
     * During the initial write pass, we will do no actual read
     * because the selected part is all undefined.
     */
    do_sarray_io(cinfo, ptr, FALSE);
  }
  /* Ensure the accessed part of the array is defined; prezero if needed.
   * To improve locality of access, we only prezero the part of the array
   * that the caller is about to access, not the entire in-memory array.
   */
  if (ptr->first_undef_row < end_row) {
    if (ptr->first_undef_row < start_row) {
      if (writable)             /* writer skipped over a section of array */
        ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
      undef_row = start_row;    /* but reader is allowed to read ahead */
    } else {
      undef_row = ptr->first_undef_row;
    }
    if (writable)
      ptr->first_undef_row = end_row;
    if (ptr->pre_zero) {
      size_t bytesperrow = (size_t)ptr->samplesperrow * sizeof(JSAMPLE);
      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
      end_row -= ptr->cur_start_row;
      while (undef_row < end_row) {
        jzero_far((void *)ptr->mem_buffer[undef_row], bytesperrow);
        undef_row++;
      }
    } else {
      if (!writable)            /* reader looking at undefined data */
        ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
    }
  }
  /* Flag the buffer dirty if caller will write in it */
  if (writable)
    ptr->dirty = TRUE;
  /* Return address of proper part of the buffer */
  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
}


METHODDEF(JBLOCKARRAY)
access_virt_barray(j_common_ptr cinfo, jvirt_barray_ptr ptr,
                   JDIMENSION start_row, JDIMENSION num_rows, boolean writable)
/* Access the part of a virtual block array starting at start_row */
/* and extending for num_rows rows.  writable is true if  */
/* caller intends to modify the accessed area. */
{
  JDIMENSION end_row = start_row + num_rows;
  JDIMENSION undef_row;

  /* debugging check */
  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
      ptr->mem_buffer == NULL)
    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);

  /* Make the desired part of the virtual array accessible */
  if (start_row < ptr->cur_start_row ||
      end_row > ptr->cur_start_row + ptr->rows_in_mem) {
    if (!ptr->b_s_open)
      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
    /* Flush old buffer contents if necessary */
    if (ptr->dirty) {
      do_barray_io(cinfo, ptr, TRUE);
      ptr->dirty = FALSE;
    }
    /* Decide what part of virtual array to access.
     * Algorithm: if target address > current window, assume forward scan,
     * load starting at target address.  If target address < current window,
     * assume backward scan, load so that target area is top of window.
     * Note that when switching from forward write to forward read, will have
     * start_row = 0, so the limiting case applies and we load from 0 anyway.
     */
    if (start_row > ptr->cur_start_row) {
      ptr->cur_start_row = start_row;
    } else {
      /* use long arithmetic here to avoid overflow & unsigned problems */
      long ltemp;

      ltemp = (long)end_row - (long)ptr->rows_in_mem;
      if (ltemp < 0)
        ltemp = 0;              /* don't fall off front end of file */
      ptr->cur_start_row = (JDIMENSION)ltemp;
    }
    /* Read in the selected part of the array.
     * During the initial write pass, we will do no actual read
     * because the selected part is all undefined.
     */
    do_barray_io(cinfo, ptr, FALSE);
  }
  /* Ensure the accessed part of the array is defined; prezero if needed.
   * To improve locality of access, we only prezero the part of the array
   * that the caller is about to access, not the entire in-memory array.
   */
  if (ptr->first_undef_row < end_row) {
    if (ptr->first_undef_row < start_row) {
      if (writable)             /* writer skipped over a section of array */
        ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
      undef_row = start_row;    /* but reader is allowed to read ahead */
    } else {
      undef_row = ptr->first_undef_row;
    }
    if (writable)
      ptr->first_undef_row = end_row;
    if (ptr->pre_zero) {
      size_t bytesperrow = (size_t)ptr->blocksperrow * sizeof(JBLOCK);
      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
      end_row -= ptr->cur_start_row;
      while (undef_row < end_row) {
        jzero_far((void *)ptr->mem_buffer[undef_row], bytesperrow);
        undef_row++;
      }
    } else {
      if (!writable)            /* reader looking at undefined data */
        ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
    }
  }
  /* Flag the buffer dirty if caller will write in it */
  if (writable)
    ptr->dirty = TRUE;
  /* Return address of proper part of the buffer */
  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
}


/*
 * Release all objects belonging to a specified pool.
 */

METHODDEF(void)
free_pool(j_common_ptr cinfo, int pool_id)
{
  my_mem_ptr mem = (my_mem_ptr)cinfo->mem;
  small_pool_ptr shdr_ptr;
  large_pool_ptr lhdr_ptr;
  size_t space_freed;

  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */

#ifdef MEM_STATS
  if (cinfo->err->trace_level > 1)
    print_mem_stats(cinfo, pool_id); /* print pool's memory usage statistics */
#endif

  /* If freeing IMAGE pool, close any virtual arrays first */
  if (pool_id == JPOOL_IMAGE) {
    jvirt_sarray_ptr sptr;
    jvirt_barray_ptr bptr;

    for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
      if (sptr->b_s_open) {     /* there may be no backing store */
        sptr->b_s_open = FALSE; /* prevent recursive close if error */
        (*sptr->b_s_info.close_backing_store) (cinfo, &sptr->b_s_info);
      }
    }
    mem->virt_sarray_list = NULL;
    for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
      if (bptr->b_s_open) {     /* there may be no backing store */
        bptr->b_s_open = FALSE; /* prevent recursive close if error */
        (*bptr->b_s_info.close_backing_store) (cinfo, &bptr->b_s_info);
      }
    }
    mem->virt_barray_list = NULL;
  }

  /* Release large objects */
  lhdr_ptr = mem->large_list[pool_id];
  mem->large_list[pool_id] = NULL;

  while (lhdr_ptr != NULL) {
    large_pool_ptr next_lhdr_ptr = lhdr_ptr->next;
    space_freed = lhdr_ptr->bytes_used +
                  lhdr_ptr->bytes_left +
                  sizeof(large_pool_hdr);
    jpeg_free_large(cinfo, (void *)lhdr_ptr, space_freed);
    mem->total_space_allocated -= space_freed;
    lhdr_ptr = next_lhdr_ptr;
  }

  /* Release small objects */
  shdr_ptr = mem->small_list[pool_id];
  mem->small_list[pool_id] = NULL;

  while (shdr_ptr != NULL) {
    small_pool_ptr next_shdr_ptr = shdr_ptr->next;
    space_freed = shdr_ptr->bytes_used + shdr_ptr->bytes_left +
                  sizeof(small_pool_hdr);
    jpeg_free_small(cinfo, (void *)shdr_ptr, space_freed);
    mem->total_space_allocated -= space_freed;
    shdr_ptr = next_shdr_ptr;
  }
}


/*
 * Close up shop entirely.
 * Note that this cannot be called unless cinfo->mem is non-NULL.
 */

METHODDEF(void)
self_destruct(j_common_ptr cinfo)
{
  int pool;

  /* Close all backing store, release all memory.
   * Releasing pools in reverse order might help avoid fragmentation
   * with some (brain-damaged) malloc libraries.
   */
  for (pool = JPOOL_NUMPOOLS - 1; pool >= JPOOL_PERMANENT; pool--) {
    free_pool(cinfo, pool);
  }

  /* Release the memory manager control block too. */
  jpeg_free_small(cinfo, (void *)cinfo->mem, sizeof(my_memory_mgr));
  cinfo->mem = NULL;            /* ensures I will be called only once */

  jpeg_mem_term(cinfo);         /* system-dependent cleanup */
}


/*
 * Memory manager initialization.
 * When this is called, only the error manager pointer is valid in cinfo!
 */

GLOBAL(void)
jinit_memory_mgr(j_common_ptr cinfo)
{
  my_mem_ptr mem;
  long max_to_use;
  int pool;
  size_t test_mac;

  cinfo->mem = NULL;            /* for safety if init fails */

  /* Check for configuration errors.
   * sizeof(ALIGN_TYPE) should be a power of 2; otherwise, it probably
   * doesn't reflect any real hardware alignment requirement.
   * The test is a little tricky: for X>0, X and X-1 have no one-bits
   * in common if and only if X is a power of 2, ie has only one one-bit.
   * Some compilers may give an "unreachable code" warning here; ignore it.
   */
  if ((ALIGN_SIZE & (ALIGN_SIZE - 1)) != 0)
    ERREXIT(cinfo, JERR_BAD_ALIGN_TYPE);
  /* MAX_ALLOC_CHUNK must be representable as type size_t, and must be
   * a multiple of ALIGN_SIZE.
   * Again, an "unreachable code" warning may be ignored here.
   * But a "constant too large" warning means you need to fix MAX_ALLOC_CHUNK.
   */
  test_mac = (size_t)MAX_ALLOC_CHUNK;
  if ((long)test_mac != MAX_ALLOC_CHUNK ||
      (MAX_ALLOC_CHUNK % ALIGN_SIZE) != 0)
    ERREXIT(cinfo, JERR_BAD_ALLOC_CHUNK);

  max_to_use = jpeg_mem_init(cinfo); /* system-dependent initialization */

  /* Attempt to allocate memory manager's control block */
  mem = (my_mem_ptr)jpeg_get_small(cinfo, sizeof(my_memory_mgr));

  if (mem == NULL) {
    jpeg_mem_term(cinfo);       /* system-dependent cleanup */
    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 0);
  }

  /* OK, fill in the method pointers */
  mem->pub.alloc_small = alloc_small;
  mem->pub.alloc_large = alloc_large;
  mem->pub.alloc_sarray = alloc_sarray;
  mem->pub.alloc_barray = alloc_barray;
  mem->pub.request_virt_sarray = request_virt_sarray;
  mem->pub.request_virt_barray = request_virt_barray;
  mem->pub.realize_virt_arrays = realize_virt_arrays;
  mem->pub.access_virt_sarray = access_virt_sarray;
  mem->pub.access_virt_barray = access_virt_barray;
  mem->pub.free_pool = free_pool;
  mem->pub.self_destruct = self_destruct;

  /* Make MAX_ALLOC_CHUNK accessible to other modules */
  mem->pub.max_alloc_chunk = MAX_ALLOC_CHUNK;

  /* Initialize working state */
  mem->pub.max_memory_to_use = max_to_use;

  for (pool = JPOOL_NUMPOOLS - 1; pool >= JPOOL_PERMANENT; pool--) {
    mem->small_list[pool] = NULL;
    mem->large_list[pool] = NULL;
  }
  mem->virt_sarray_list = NULL;
  mem->virt_barray_list = NULL;

  mem->total_space_allocated = sizeof(my_memory_mgr);

  /* Declare ourselves open for business */
  cinfo->mem = &mem->pub;

  /* Check for an environment variable JPEGMEM; if found, override the
   * default max_memory setting from jpeg_mem_init.  Note that the
   * surrounding application may again override this value.
   * If your system doesn't support getenv(), define NO_GETENV to disable
   * this feature.
   */
#ifndef NO_GETENV
  {
    char *memenv;

    if ((memenv = getenv("JPEGMEM")) != NULL) {
      char ch = 'x';

      if (sscanf(memenv, "%ld%c", &max_to_use, &ch) > 0) {
        if (ch == 'm' || ch == 'M')
          max_to_use *= 1000L;
        mem->pub.max_memory_to_use = max_to_use * 1000L;
      }
    }
  }
#endif

}