/* * jmemmgr.c * * Copyright (C) 1991-1997, Thomas G. Lane. * This file is part of the Independent JPEG Group's software. * For conditions of distribution and use, see the accompanying README 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" /* * 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 sizeof(ALIGN_TYPE). * By default, we define ALIGN_TYPE as 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_TYPE 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_TYPE long" in jconfig.h if you have * such a compiler. */ #ifndef ALIGN_TYPE /* so can override from jconfig.h */ #define ALIGN_TYPE double #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 except that the latter's * link pointer must be FAR on 80x86 machines. * Notice that the "real" header fields are union'ed with a dummy ALIGN_TYPE * field. This forces the compiler to make SIZEOF(small_pool_hdr) a multiple * of the alignment requirement of ALIGN_TYPE. */ typedef union small_pool_struct * small_pool_ptr; typedef union small_pool_struct { 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 */ } hdr; ALIGN_TYPE dummy; /* included in union to ensure alignment */ } small_pool_hdr; typedef union large_pool_struct * large_pool_ptr; typedef union large_pool_struct { 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 */ } hdr; ALIGN_TYPE dummy; /* included in union to ensure alignment */ } 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_barray_ptr virt_barray_list; /* 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_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? */ jvirt_barray_ptr next; /* link to next virtual barray control block */ }; LOCAL(void) out_of_memory (j_common_ptr cinfo, int which) /* Report an out-of-memory error and stop execution */ { 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. */ 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 odd_bytes, min_request, slop; /* Round up the requested size to a multiple of SIZEOF(ALIGN_TYPE) */ odd_bytes = sizeofobject % SIZEOF(ALIGN_TYPE); if (odd_bytes > 0) sizeofobject += SIZEOF(ALIGN_TYPE) - odd_bytes; /* 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->hdr.bytes_left >= sizeofobject) break; /* found pool with enough space */ prev_hdr_ptr = hdr_ptr; hdr_ptr = hdr_ptr->hdr.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 = sizeofobject + SIZEOF(small_pool_hdr); if (prev_hdr_ptr == NULL) /* first pool in class? */ slop = first_pool_slop[pool_id]; else slop = extra_pool_slop[pool_id]; /* Try to get space, if fail reduce slop and try again */ for (;;) { hdr_ptr = (small_pool_ptr) malloc(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 */ } /* Success, initialize the new pool header and add to end of list */ hdr_ptr->hdr.next = NULL; hdr_ptr->hdr.bytes_used = 0; hdr_ptr->hdr.bytes_left = sizeofobject + slop; if (prev_hdr_ptr == NULL) /* first pool in class? */ mem->small_list[pool_id] = hdr_ptr; else prev_hdr_ptr->hdr.next = hdr_ptr; } /* OK, allocate the object from the current pool */ data_ptr = (char *) (hdr_ptr + 1); /* point to first data byte in pool */ data_ptr += hdr_ptr->hdr.bytes_used; /* point to place for object */ hdr_ptr->hdr.bytes_used += sizeofobject; hdr_ptr->hdr.bytes_left -= sizeofobject; return (void *) data_ptr; } /* * Allocation of "large" objects. * * The external semantics of these are the same as "small" objects, * except that FAR pointers are used on 80x86. 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; size_t odd_bytes; /* Round up the requested size to a multiple of SIZEOF(ALIGN_TYPE) */ odd_bytes = sizeofobject % SIZEOF(ALIGN_TYPE); if (odd_bytes > 0) sizeofobject += SIZEOF(ALIGN_TYPE) - odd_bytes; /* 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) malloc(sizeofobject + SIZEOF(large_pool_hdr)); if (hdr_ptr == NULL) out_of_memory(cinfo, 4); /* jpeg_get_large failed */ /* Success, initialize the new pool header and add to list */ hdr_ptr->hdr.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->hdr.bytes_used = sizeofobject; hdr_ptr->hdr.bytes_left = 0; mem->large_list[pool_id] = hdr_ptr; return (void *) (hdr_ptr + 1); /* point to first data byte in pool */ } /* * Creation of 2-D sample arrays. * The pointers are in near heap, the samples themselves in FAR heap. * * 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. */ 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 i; /* Calculate max # of rows allowed in one allocation chunk */ mem->last_rowsperchunk = numrows; /* 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) */ workspace = (JSAMPROW) alloc_large(cinfo, pool_id, (size_t) ((size_t) numrows * (size_t) samplesperrow * SIZEOF(JSAMPLE))); for (i = 0; i < numrows; i++) { result[i] = 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 i; /* Calculate max # of rows allowed in one allocation chunk */ mem->last_rowsperchunk = numrows; /* 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) */ workspace = (JBLOCKROW) alloc_large(cinfo, pool_id, (size_t) ((size_t) numrows * (size_t) blocksperrow * SIZEOF(JBLOCK))); for (i = 0; i < numrows; i++) { result[i] = 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_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->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; long space_per_minheight; long minheights; 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; for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) { if (bptr->mem_buffer == NULL) { /* if not realized yet */ space_per_minheight += (long) bptr->maxaccess * (long) bptr->blocksperrow * SIZEOF(JBLOCK); } } if (space_per_minheight <= 0) return; /* no unrealized arrays, no work */ /* Allocate the in-memory buffers and initialize backing store as needed. */ 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; bptr->rows_in_mem = bptr->rows_in_array; 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; } } } 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) ERREXIT(cinfo, JERR_VIRTUAL_BUG); /* 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) { MEMZERO((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; if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS) ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id); /* safety check */ /* 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->hdr.next; free(lhdr_ptr); 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->hdr.next; free(shdr_ptr); 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. */ free(cinfo->mem); cinfo->mem = NULL; /* ensures I will be called only once */ } /* * 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; int pool; 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 ((SIZEOF(ALIGN_TYPE) & (SIZEOF(ALIGN_TYPE)-1)) != 0) ERREXIT(cinfo, JERR_BAD_ALIGN_TYPE); /* Attempt to allocate memory manager's control block */ mem = (my_mem_ptr) malloc(SIZEOF(my_memory_mgr)); if (mem == NULL) { 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_barray = request_virt_barray; mem->pub.realize_virt_arrays = realize_virt_arrays; mem->pub.access_virt_barray = access_virt_barray; mem->pub.free_pool = free_pool; mem->pub.self_destruct = self_destruct; /* Initialize working state */ for (pool = JPOOL_NUMPOOLS-1; pool >= JPOOL_PERMANENT; pool--) { mem->small_list[pool] = NULL; mem->large_list[pool] = NULL; } mem->virt_barray_list = NULL; /* Declare ourselves open for business */ cinfo->mem = & mem->pub; }