source: libcfa/src/heap.cfa@ 199894e

ADT ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since 199894e was b38b22f, checked in by Peter A. Buhr <pabuhr@…>, 4 years ago

add missing heap statistic counters for free calls and free storage
  • Property mode set to 100644
File size: 54.7 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2017 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// heap.cfa --
8//
9// Author : Peter A. Buhr
10// Created On : Tue Dec 19 21:58:35 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : Sat May 22 08:46:39 2021
13// Update Count : 1036
14//
15
16#include <unistd.h> // sbrk, sysconf
17#include <stdlib.h> // EXIT_FAILURE
18#include <stdbool.h> // true, false
19#include <stdio.h> // snprintf, fileno
20#include <errno.h> // errno
21#include <string.h> // memset, memcpy
22#include <limits.h> // ULONG_MAX
23#include <malloc.h> // memalign, malloc_usable_size
24#include <sys/mman.h> // mmap, munmap
25
26#include "bits/align.hfa" // libAlign
27#include "bits/defs.hfa" // likely, unlikely
28#include "bits/locks.hfa" // __spinlock_t
29#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
30#include "math.hfa" // min
31#include "bitmanip.hfa" // is_pow2, ceiling2
32
33static bool traceHeap = false;
34
35inline bool traceHeap() { return traceHeap; }
36
37bool traceHeapOn() {
38 bool temp = traceHeap;
39 traceHeap = true;
40 return temp;
41} // traceHeapOn
42
43bool traceHeapOff() {
44 bool temp = traceHeap;
45 traceHeap = false;
46 return temp;
47} // traceHeapOff
48
49bool traceHeapTerm() { return false; }
50
51
52static bool prtFree = false;
53
54bool prtFree() {
55 return prtFree;
56} // prtFree
57
58bool prtFreeOn() {
59 bool temp = prtFree;
60 prtFree = true;
61 return temp;
62} // prtFreeOn
63
64bool prtFreeOff() {
65 bool temp = prtFree;
66 prtFree = false;
67 return temp;
68} // prtFreeOff
69
70
71enum {
72 // Define the default extension heap amount in units of bytes. When the uC++ supplied heap reaches the brk address,
73 // the brk address is extended by the extension amount.
74 __CFA_DEFAULT_HEAP_EXPANSION__ = (10 * 1024 * 1024),
75
76 // Define the mmap crossover point during allocation. Allocations less than this amount are allocated from buckets;
77 // values greater than or equal to this value are mmap from the operating system.
78 __CFA_DEFAULT_MMAP_START__ = (512 * 1024 + 1),
79};
80
81size_t default_mmap_start() __attribute__(( weak )) {
82 return __CFA_DEFAULT_MMAP_START__;
83} // default_mmap_start
84
85size_t default_heap_expansion() __attribute__(( weak )) {
86 return __CFA_DEFAULT_HEAP_EXPANSION__;
87} // default_heap_expansion
88
89
90#ifdef __CFA_DEBUG__
91static size_t allocUnfreed; // running total of allocations minus frees
92
93static void prtUnfreed() {
94 if ( allocUnfreed != 0 ) {
95 // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
96 char helpText[512];
97 int len = snprintf( helpText, sizeof(helpText), "CFA warning (UNIX pid:%ld) : program terminating with %zu(0x%zx) bytes of storage allocated but not freed.\n"
98 "Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
99 (long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
100 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
101 } // if
102} // prtUnfreed
103
104extern "C" {
105 void heapAppStart() { // called by __cfaabi_appready_startup
106 allocUnfreed = 0;
107 } // heapAppStart
108
109 void heapAppStop() { // called by __cfaabi_appready_startdown
110 fclose( stdin ); fclose( stdout );
111 prtUnfreed();
112 } // heapAppStop
113} // extern "C"
114#endif // __CFA_DEBUG__
115
116
117// statically allocated variables => zero filled.
118size_t __page_size; // architecture pagesize
119int __map_prot; // common mmap/mprotect protection
120static size_t heapExpand; // sbrk advance
121static size_t mmapStart; // cross over point for mmap
122static unsigned int maxBucketsUsed; // maximum number of buckets in use
123
124
125#define SPINLOCK 0
126#define LOCKFREE 1
127#define BUCKETLOCK SPINLOCK
128#if BUCKETLOCK == SPINLOCK
129#elif BUCKETLOCK == LOCKFREE
130#include <stackLockFree.hfa>
131#else
132 #error undefined lock type for bucket lock
133#endif // LOCKFREE
134
135// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
136// Break recusion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
137enum { NoBucketSizes = 91 }; // number of buckets sizes
138
139struct HeapManager {
140 struct Storage {
141 struct Header { // header
142 union Kind {
143 struct RealHeader {
144 union {
145 struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
146 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
147 uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
148 #endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
149
150 union {
151 // FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
152 // 2nd low-order bit => zero filled
153 void * home; // allocated block points back to home locations (must overlay alignment)
154 size_t blockSize; // size for munmap (must overlay alignment)
155 #if BUCKETLOCK == SPINLOCK
156 Storage * next; // freed block points next freed block of same size
157 #endif // SPINLOCK
158 };
159 size_t size; // allocation size in bytes
160
161 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
162 uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
163 #endif // __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
164 };
165 #if BUCKETLOCK == LOCKFREE
166 Link(Storage) next; // freed block points next freed block of same size (double-wide)
167 #endif // LOCKFREE
168 };
169 } real; // RealHeader
170
171 struct FakeHeader {
172 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
173 uint32_t alignment; // 1st low-order bit => fake header & alignment
174 #endif // __ORDER_LITTLE_ENDIAN__
175
176 uint32_t offset;
177
178 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
179 uint32_t alignment; // low-order bits of home/blockSize used for tricks
180 #endif // __ORDER_BIG_ENDIAN__
181 } fake; // FakeHeader
182 } kind; // Kind
183 } header; // Header
184 char pad[libAlign() - sizeof( Header )];
185 char data[0]; // storage
186 }; // Storage
187
188 static_assert( libAlign() >= sizeof( Storage ), "libAlign() < sizeof( Storage )" );
189
190 struct FreeHeader {
191 #if BUCKETLOCK == SPINLOCK
192 __spinlock_t lock; // must be first field for alignment
193 Storage * freeList;
194 #else
195 StackLF(Storage) freeList;
196 #endif // BUCKETLOCK
197 size_t blockSize; // size of allocations on this list
198 }; // FreeHeader
199
200 // must be first fields for alignment
201 __spinlock_t extlock; // protects allocation-buffer extension
202 FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
203
204 void * heapBegin; // start of heap
205 void * heapEnd; // logical end of heap
206 size_t heapRemaining; // amount of storage not allocated in the current chunk
207}; // HeapManager
208
209#if BUCKETLOCK == LOCKFREE
210static inline {
211 Link(HeapManager.Storage) * ?`next( HeapManager.Storage * this ) { return &this->header.kind.real.next; }
212 void ?{}( HeapManager.FreeHeader & ) {}
213 void ^?{}( HeapManager.FreeHeader & ) {}
214} // distribution
215#endif // LOCKFREE
216
217static inline size_t getKey( const HeapManager.FreeHeader & freeheader ) { return freeheader.blockSize; }
218
219
220#define FASTLOOKUP
221#define __STATISTICS__
222
223// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
224// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
225// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
226static const unsigned int bucketSizes[] @= { // different bucket sizes
227 16 + sizeof(HeapManager.Storage), 32 + sizeof(HeapManager.Storage), 48 + sizeof(HeapManager.Storage), 64 + sizeof(HeapManager.Storage), // 4
228 96 + sizeof(HeapManager.Storage), 112 + sizeof(HeapManager.Storage), 128 + sizeof(HeapManager.Storage), // 3
229 160, 192, 224, 256 + sizeof(HeapManager.Storage), // 4
230 320, 384, 448, 512 + sizeof(HeapManager.Storage), // 4
231 640, 768, 896, 1_024 + sizeof(HeapManager.Storage), // 4
232 1_536, 2_048 + sizeof(HeapManager.Storage), // 2
233 2_560, 3_072, 3_584, 4_096 + sizeof(HeapManager.Storage), // 4
234 6_144, 8_192 + sizeof(HeapManager.Storage), // 2
235 9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(HeapManager.Storage), // 8
236 18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(HeapManager.Storage), // 8
237 36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(HeapManager.Storage), // 8
238 73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(HeapManager.Storage), // 8
239 147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(HeapManager.Storage), // 8
240 294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(HeapManager.Storage), // 8
241 655_360, 786_432, 917_504, 1_048_576 + sizeof(HeapManager.Storage), // 4
242 1_179_648, 1_310_720, 1_441_792, 1_572_864, 1_703_936, 1_835_008, 1_966_080, 2_097_152 + sizeof(HeapManager.Storage), // 8
243 2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(HeapManager.Storage), // 4
244};
245
246static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
247
248#ifdef FASTLOOKUP
249enum { LookupSizes = 65_536 + sizeof(HeapManager.Storage) }; // number of fast lookup sizes
250static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
251#endif // FASTLOOKUP
252
253static const off_t mmapFd = -1; // fake or actual fd for anonymous file
254#ifdef __CFA_DEBUG__
255static bool heapBoot = 0; // detect recursion during boot
256#endif // __CFA_DEBUG__
257
258// The constructor for heapManager is called explicitly in memory_startup.
259static HeapManager heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing
260
261
262#ifdef __STATISTICS__
263// Heap statistics counters.
264static unsigned int malloc_zero_calls, malloc_calls;
265static unsigned long long int malloc_storage;
266static unsigned int aalloc_zero_calls, aalloc_calls;
267static unsigned long long int aalloc_storage;
268static unsigned int calloc_zero_calls, calloc_calls;
269static unsigned long long int calloc_storage;
270static unsigned int memalign_zero_calls, memalign_calls;
271static unsigned long long int memalign_storage;
272static unsigned int amemalign_zero_calls, amemalign_calls;
273static unsigned long long int amemalign_storage;
274static unsigned int cmemalign_zero_calls, cmemalign_calls;
275static unsigned long long int cmemalign_storage;
276static unsigned int resize_zero_calls, resize_calls;
277static unsigned long long int resize_storage;
278static unsigned int realloc_zero_calls, realloc_calls;
279static unsigned long long int realloc_storage;
280static unsigned int free_zero_calls, free_calls;
281static unsigned long long int free_storage;
282static unsigned int mmap_calls;
283static unsigned long long int mmap_storage;
284static unsigned int munmap_calls;
285static unsigned long long int munmap_storage;
286static unsigned int sbrk_calls;
287static unsigned long long int sbrk_storage;
288// Statistics file descriptor (changed by malloc_stats_fd).
289static int stats_fd = STDERR_FILENO; // default stderr
290
291// Use "write" because streams may be shutdown when calls are made.
292static void printStats() {
293 char helpText[1024];
294 __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
295 "\nHeap statistics:\n"
296 " malloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
297 " aalloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
298 " calloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
299 " memalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
300 " amemalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
301 " cmemalign 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
302 " resize 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
303 " realloc 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
304 " free 0-calls %'u; >0-calls %'u; storage %'llu bytes\n"
305 " mmap calls %'u; storage %'llu bytes\n"
306 " munmap calls %'u; storage %'llu bytes\n"
307 " sbrk calls %'u; storage %'llu bytes\n",
308 malloc_zero_calls, malloc_calls, malloc_storage,
309 aalloc_zero_calls, aalloc_calls, aalloc_storage,
310 calloc_zero_calls, calloc_calls, calloc_storage,
311 memalign_zero_calls, memalign_calls, memalign_storage,
312 amemalign_zero_calls, amemalign_calls, amemalign_storage,
313 cmemalign_zero_calls, cmemalign_calls, cmemalign_storage,
314 resize_zero_calls, resize_calls, resize_storage,
315 realloc_zero_calls, realloc_calls, realloc_storage,
316 free_zero_calls, free_calls, free_storage,
317 mmap_calls, mmap_storage,
318 munmap_calls, munmap_storage,
319 sbrk_calls, sbrk_storage
320 );
321} // printStats
322
323static int printStatsXML( FILE * stream ) { // see malloc_info
324 char helpText[1024];
325 int len = snprintf( helpText, sizeof(helpText),
326 "<malloc version=\"1\">\n"
327 "<heap nr=\"0\">\n"
328 "<sizes>\n"
329 "</sizes>\n"
330 "<total type=\"malloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
331 "<total type=\"aalloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
332 "<total type=\"calloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
333 "<total type=\"memalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
334 "<total type=\"amemalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
335 "<total type=\"cmemalign\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
336 "<total type=\"resize\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
337 "<total type=\"realloc\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
338 "<total type=\"free\" 0 count=\"%'u;\" >0 count=\"%'u;\" size=\"%'llu\"/> bytes\n"
339 "<total type=\"mmap\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
340 "<total type=\"munmap\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
341 "<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
342 "</malloc>",
343 malloc_zero_calls, malloc_calls, malloc_storage,
344 aalloc_zero_calls, aalloc_calls, aalloc_storage,
345 calloc_zero_calls, calloc_calls, calloc_storage,
346 memalign_zero_calls, memalign_calls, memalign_storage,
347 amemalign_zero_calls, amemalign_calls, amemalign_storage,
348 cmemalign_zero_calls, cmemalign_calls, cmemalign_storage,
349 resize_zero_calls, resize_calls, resize_storage,
350 realloc_zero_calls, realloc_calls, realloc_storage,
351 free_zero_calls, free_calls, free_storage,
352 mmap_calls, mmap_storage,
353 munmap_calls, munmap_storage,
354 sbrk_calls, sbrk_storage
355 );
356 __cfaabi_bits_write( fileno( stream ), helpText, len ); // ensures all bytes written or exit
357 return len;
358} // printStatsXML
359#endif // __STATISTICS__
360
361
362// thunk problem
363size_t Bsearchl( unsigned int key, const unsigned int * vals, size_t dim ) {
364 size_t l = 0, m, h = dim;
365 while ( l < h ) {
366 m = (l + h) / 2;
367 if ( (unsigned int &)(vals[m]) < key ) { // cast away const
368 l = m + 1;
369 } else {
370 h = m;
371 } // if
372 } // while
373 return l;
374} // Bsearchl
375
376
377static inline bool setMmapStart( size_t value ) { // true => mmapped, false => sbrk
378 if ( value < __page_size || bucketSizes[NoBucketSizes - 1] < value ) return false;
379 mmapStart = value; // set global
380
381 // find the closest bucket size less than or equal to the mmapStart size
382 maxBucketsUsed = Bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
383 assert( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
384 assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
385 return true;
386} // setMmapStart
387
388
389// <-------+----------------------------------------------------> bsize (bucket size)
390// |header |addr
391//==================================================================================
392// align/offset |
393// <-----------------<------------+-----------------------------> bsize (bucket size)
394// |fake-header | addr
395#define headerAddr( addr ) ((HeapManager.Storage.Header *)( (char *)addr - sizeof(HeapManager.Storage) ))
396#define realHeader( header ) ((HeapManager.Storage.Header *)((char *)header - header->kind.fake.offset))
397
398// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
399// |header |addr
400//==================================================================================
401// align/offset |
402// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
403// |fake-header |addr
404#define dataStorage( bsize, addr, header ) (bsize - ( (char *)addr - (char *)header ))
405
406
407static inline void checkAlign( size_t alignment ) {
408 if ( alignment < libAlign() || ! is_pow2( alignment ) ) {
409 abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
410 } // if
411} // checkAlign
412
413
414static inline void checkHeader( bool check, const char name[], void * addr ) {
415 if ( unlikely( check ) ) { // bad address ?
416 abort( "Attempt to %s storage %p with address outside the heap.\n"
417 "Possible cause is duplicate free on same block or overwriting of memory.",
418 name, addr );
419 } // if
420} // checkHeader
421
422
423static inline void fakeHeader( HeapManager.Storage.Header *& header, size_t & alignment ) {
424 if ( unlikely( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
425 alignment = header->kind.fake.alignment & -2; // remove flag from value
426 #ifdef __CFA_DEBUG__
427 checkAlign( alignment ); // check alignment
428 #endif // __CFA_DEBUG__
429 header = realHeader( header ); // backup from fake to real header
430 } else {
431 alignment = libAlign(); // => no fake header
432 } // if
433} // fakeHeader
434
435
436static inline bool headers( const char name[] __attribute__(( unused )), void * addr, HeapManager.Storage.Header *& header, HeapManager.FreeHeader *& freeElem,
437 size_t & size, size_t & alignment ) with( heapManager ) {
438 header = headerAddr( addr );
439
440 if ( unlikely( addr < heapBegin || heapEnd < addr ) ) { // mmapped ?
441 fakeHeader( header, alignment );
442 size = header->kind.real.blockSize & -3; // mmap size
443 return true;
444 } // if
445
446 #ifdef __CFA_DEBUG__
447 checkHeader( header < (HeapManager.Storage.Header *)heapBegin, name, addr ); // bad low address ?
448 #endif // __CFA_DEBUG__
449
450 // header may be safe to dereference
451 fakeHeader( header, alignment );
452 #ifdef __CFA_DEBUG__
453 checkHeader( header < (HeapManager.Storage.Header *)heapBegin || (HeapManager.Storage.Header *)heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
454 #endif // __CFA_DEBUG__
455
456 freeElem = (HeapManager.FreeHeader *)((size_t)header->kind.real.home & -3);
457 #ifdef __CFA_DEBUG__
458 if ( freeElem < &freeLists[0] || &freeLists[NoBucketSizes] <= freeElem ) {
459 abort( "Attempt to %s storage %p with corrupted header.\n"
460 "Possible cause is duplicate free on same block or overwriting of header information.",
461 name, addr );
462 } // if
463 #endif // __CFA_DEBUG__
464 size = freeElem->blockSize;
465 return false;
466} // headers
467
468// #ifdef __CFA_DEBUG__
469// #if __SIZEOF_POINTER__ == 4
470// #define MASK 0xdeadbeef
471// #else
472// #define MASK 0xdeadbeefdeadbeef
473// #endif
474// #define STRIDE size_t
475
476// static void * Memset( void * addr, STRIDE size ) { // debug only
477// if ( size % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, size %zd not multiple of %zd.", size, sizeof(STRIDE) );
478// if ( (STRIDE)addr % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, addr %p not multiple of %zd.", addr, sizeof(STRIDE) );
479
480// STRIDE * end = (STRIDE *)addr + size / sizeof(STRIDE);
481// for ( STRIDE * p = (STRIDE *)addr; p < end; p += 1 ) *p = MASK;
482// return addr;
483// } // Memset
484// #endif // __CFA_DEBUG__
485
486
487#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
488
489static inline void * extend( size_t size ) with( heapManager ) {
490 lock( extlock __cfaabi_dbg_ctx2 );
491 ptrdiff_t rem = heapRemaining - size;
492 if ( rem < 0 ) {
493 // If the size requested is bigger than the current remaining storage, increase the size of the heap.
494
495 size_t increase = ceiling2( size > heapExpand ? size : heapExpand, __page_size );
496 // Do not call abort or strerror( errno ) as they may call malloc.
497 if ( sbrk( increase ) == (void *)-1 ) { // failed, no memory ?
498 unlock( extlock );
499 __cfaabi_bits_print_nolock( STDERR_FILENO, NO_MEMORY_MSG, size );
500 _exit( EXIT_FAILURE ); // give up
501 } // if
502 // Make storage executable for thunks.
503 if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
504 unlock( extlock );
505 __cfaabi_bits_print_nolock( STDERR_FILENO, "extend() : internal error, mprotect failure, heapEnd:%p size:%zd, errno:%d.\n", heapEnd, increase, errno );
506 _exit( EXIT_FAILURE );
507 } // if
508 #ifdef __STATISTICS__
509 sbrk_calls += 1;
510 sbrk_storage += increase;
511 #endif // __STATISTICS__
512 #ifdef __CFA_DEBUG__
513 // Set new memory to garbage so subsequent uninitialized usages might fail.
514 memset( (char *)heapEnd + heapRemaining, '\xde', increase );
515 //Memset( (char *)heapEnd + heapRemaining, increase );
516 #endif // __CFA_DEBUG__
517 rem = heapRemaining + increase - size;
518 } // if
519
520 HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
521 heapRemaining = rem;
522 heapEnd = (char *)heapEnd + size;
523 unlock( extlock );
524 return block;
525} // extend
526
527
528static inline void * doMalloc( size_t size ) with( heapManager ) {
529 HeapManager.Storage * block; // pointer to new block of storage
530
531 // Look up size in the size list. Make sure the user request includes space for the header that must be allocated
532 // along with the block and is a multiple of the alignment size.
533
534 if ( unlikely( size > ULONG_MAX - sizeof(HeapManager.Storage) ) ) return 0p;
535 size_t tsize = size + sizeof(HeapManager.Storage);
536 if ( likely( tsize < mmapStart ) ) { // small size => sbrk
537 size_t posn;
538 #ifdef FASTLOOKUP
539 if ( tsize < LookupSizes ) posn = lookup[tsize];
540 else
541 #endif // FASTLOOKUP
542 posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
543 HeapManager.FreeHeader * freeElem = &freeLists[posn];
544 verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
545 verify( tsize <= freeElem->blockSize ); // search failure ?
546 tsize = freeElem->blockSize; // total space needed for request
547
548 // Spin until the lock is acquired for this particular size of block.
549
550 #if BUCKETLOCK == SPINLOCK
551 lock( freeElem->lock __cfaabi_dbg_ctx2 );
552 block = freeElem->freeList; // remove node from stack
553 #else
554 block = pop( freeElem->freeList );
555 #endif // BUCKETLOCK
556 if ( unlikely( block == 0p ) ) { // no free block ?
557 #if BUCKETLOCK == SPINLOCK
558 unlock( freeElem->lock );
559 #endif // BUCKETLOCK
560
561 // Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
562 // and then carve it off.
563
564 block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
565 #if BUCKETLOCK == SPINLOCK
566 } else {
567 freeElem->freeList = block->header.kind.real.next;
568 unlock( freeElem->lock );
569 #endif // BUCKETLOCK
570 } // if
571
572 block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
573 } else { // large size => mmap
574 if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
575 tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
576 #ifdef __STATISTICS__
577 __atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
578 __atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
579 #endif // __STATISTICS__
580
581 block = (HeapManager.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE | MAP_ANONYMOUS, mmapFd, 0 );
582 if ( block == (HeapManager.Storage *)MAP_FAILED ) { // failed ?
583 if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
584 // Do not call strerror( errno ) as it may call malloc.
585 abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu errno:%d.", &heapManager, tsize, errno );
586 } //if
587 #ifdef __CFA_DEBUG__
588 // Set new memory to garbage so subsequent uninitialized usages might fail.
589 memset( block, '\xde', tsize );
590 //Memset( block, tsize );
591 #endif // __CFA_DEBUG__
592 block->header.kind.real.blockSize = tsize; // storage size for munmap
593 } // if
594
595 block->header.kind.real.size = size; // store allocation size
596 void * addr = &(block->data); // adjust off header to user bytes
597 verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
598
599 #ifdef __CFA_DEBUG__
600 __atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
601 if ( traceHeap() ) {
602 enum { BufferSize = 64 };
603 char helpText[BufferSize];
604 int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize );
605 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
606 } // if
607 #endif // __CFA_DEBUG__
608
609 return addr;
610} // doMalloc
611
612
613static inline void doFree( void * addr ) with( heapManager ) {
614 #ifdef __CFA_DEBUG__
615 if ( unlikely( heapManager.heapBegin == 0p ) ) {
616 abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
617 } // if
618 #endif // __CFA_DEBUG__
619
620 HeapManager.Storage.Header * header;
621 HeapManager.FreeHeader * freeElem;
622 size_t size, alignment; // not used (see realloc)
623
624 if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
625 #ifdef __STATISTICS__
626 __atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
627 __atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
628 #endif // __STATISTICS__
629 if ( munmap( header, size ) == -1 ) {
630 abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
631 "Possible cause is invalid pointer.",
632 addr );
633 } // if
634 } else {
635 #ifdef __CFA_DEBUG__
636 // Set free memory to garbage so subsequent usages might fail.
637 memset( ((HeapManager.Storage *)header)->data, '\xde', freeElem->blockSize - sizeof( HeapManager.Storage ) );
638 //Memset( ((HeapManager.Storage *)header)->data, freeElem->blockSize - sizeof( HeapManager.Storage ) );
639 #endif // __CFA_DEBUG__
640
641 #ifdef __STATISTICS__
642 __atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
643 __atomic_add_fetch( &free_storage, size, __ATOMIC_SEQ_CST );
644 #endif // __STATISTICS__
645
646 #if BUCKETLOCK == SPINLOCK
647 lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
648 header->kind.real.next = freeElem->freeList; // push on stack
649 freeElem->freeList = (HeapManager.Storage *)header;
650 unlock( freeElem->lock ); // release spin lock
651 #else
652 push( freeElem->freeList, *(HeapManager.Storage *)header );
653 #endif // BUCKETLOCK
654 } // if
655
656 #ifdef __CFA_DEBUG__
657 __atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
658 if ( traceHeap() ) {
659 char helpText[64];
660 int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
661 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
662 } // if
663 #endif // __CFA_DEBUG__
664} // doFree
665
666
667size_t prtFree( HeapManager & manager ) with( manager ) {
668 size_t total = 0;
669 #ifdef __STATISTICS__
670 __cfaabi_bits_acquire();
671 __cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
672 #endif // __STATISTICS__
673 for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
674 size_t size = freeLists[i].blockSize;
675 #ifdef __STATISTICS__
676 unsigned int N = 0;
677 #endif // __STATISTICS__
678
679 #if BUCKETLOCK == SPINLOCK
680 for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
681 #else
682 for(;;) {
683// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
684// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
685// HeapManager.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
686// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
687// p = temp;
688 #endif // BUCKETLOCK
689 total += size;
690 #ifdef __STATISTICS__
691 N += 1;
692 #endif // __STATISTICS__
693 } // for
694
695 #ifdef __STATISTICS__
696 __cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
697 if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
698 #endif // __STATISTICS__
699 } // for
700 #ifdef __STATISTICS__
701 __cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
702 __cfaabi_bits_release();
703 #endif // __STATISTICS__
704 return (char *)heapEnd - (char *)heapBegin - total;
705} // prtFree
706
707
708static void ?{}( HeapManager & manager ) with( manager ) {
709 __page_size = sysconf( _SC_PAGESIZE );
710 __map_prot = PROT_READ | PROT_WRITE | PROT_EXEC;
711
712 for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
713 freeLists[i].blockSize = bucketSizes[i];
714 } // for
715
716 #ifdef FASTLOOKUP
717 unsigned int idx = 0;
718 for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
719 if ( i > bucketSizes[idx] ) idx += 1;
720 lookup[i] = idx;
721 } // for
722 #endif // FASTLOOKUP
723
724 if ( ! setMmapStart( default_mmap_start() ) ) {
725 abort( "HeapManager : internal error, mmap start initialization failure." );
726 } // if
727 heapExpand = default_heap_expansion();
728
729 char * end = (char *)sbrk( 0 );
730 heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, __page_size ) - end ); // move start of heap to multiple of alignment
731} // HeapManager
732
733
734static void ^?{}( HeapManager & ) {
735 #ifdef __STATISTICS__
736 if ( traceHeapTerm() ) {
737 printStats();
738 // prtUnfreed() called in heapAppStop()
739 } // if
740 #endif // __STATISTICS__
741} // ~HeapManager
742
743
744static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
745void memory_startup( void ) {
746 #ifdef __CFA_DEBUG__
747 if ( heapBoot ) { // check for recursion during system boot
748 abort( "boot() : internal error, recursively invoked during system boot." );
749 } // if
750 heapBoot = true;
751 #endif // __CFA_DEBUG__
752
753 //verify( heapManager.heapBegin != 0 );
754 //heapManager{};
755 if ( heapManager.heapBegin == 0p ) heapManager{}; // sanity check
756} // memory_startup
757
758static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
759void memory_shutdown( void ) {
760 ^heapManager{};
761} // memory_shutdown
762
763
764static inline void * mallocNoStats( size_t size ) { // necessary for malloc statistics
765 verify( heapManager.heapBegin != 0p ); // called before memory_startup ?
766 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
767
768#if __SIZEOF_POINTER__ == 8
769 verify( size < ((typeof(size_t))1 << 48) );
770#endif // __SIZEOF_POINTER__ == 8
771 return doMalloc( size );
772} // mallocNoStats
773
774
775static inline void * memalignNoStats( size_t alignment, size_t size ) {
776 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
777
778 #ifdef __CFA_DEBUG__
779 checkAlign( alignment ); // check alignment
780 #endif // __CFA_DEBUG__
781
782 // if alignment <= default alignment, do normal malloc as two headers are unnecessary
783 if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );
784
785 // Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
786 // administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
787 // .-------------v-----------------v----------------v----------,
788 // | Real Header | ... padding ... | Fake Header | data ... |
789 // `-------------^-----------------^-+--------------^----------'
790 // |<--------------------------------' offset/align |<-- alignment boundary
791
792 // subtract libAlign() because it is already the minimum alignment
793 // add sizeof(Storage) for fake header
794 char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(HeapManager.Storage) );
795
796 // address in the block of the "next" alignment address
797 char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(HeapManager.Storage)), alignment );
798
799 // address of header from malloc
800 HeapManager.Storage.Header * realHeader = headerAddr( addr );
801 realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
802 // address of fake header * before* the alignment location
803 HeapManager.Storage.Header * fakeHeader = headerAddr( user );
804 // SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
805 fakeHeader->kind.fake.offset = (char *)fakeHeader - (char *)realHeader;
806 // SKULLDUGGERY: odd alignment imples fake header
807 fakeHeader->kind.fake.alignment = alignment | 1;
808
809 return user;
810} // memalignNoStats
811
812
813extern "C" {
814 // Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
815 // then malloc() returns a unique pointer value that can later be successfully passed to free().
816 void * malloc( size_t size ) {
817 #ifdef __STATISTICS__
818 if ( likely( size > 0 ) ) {
819 __atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
820 __atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
821 } else {
822 __atomic_add_fetch( &malloc_zero_calls, 1, __ATOMIC_SEQ_CST );
823 } // if
824 #endif // __STATISTICS__
825
826 return mallocNoStats( size );
827 } // malloc
828
829
830 // Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
831 void * aalloc( size_t dim, size_t elemSize ) {
832 size_t size = dim * elemSize;
833 #ifdef __STATISTICS__
834 if ( likely( size > 0 ) ) {
835 __atomic_add_fetch( &aalloc_calls, 1, __ATOMIC_SEQ_CST );
836 __atomic_add_fetch( &aalloc_storage, size, __ATOMIC_SEQ_CST );
837 } else {
838 __atomic_add_fetch( &aalloc_zero_calls, 1, __ATOMIC_SEQ_CST );
839 } // if
840 #endif // __STATISTICS__
841
842 return mallocNoStats( size );
843 } // aalloc
844
845
846 // Same as aalloc() with memory set to zero.
847 void * calloc( size_t dim, size_t elemSize ) {
848 size_t size = dim * elemSize;
849 if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
850 #ifdef __STATISTICS__
851 __atomic_add_fetch( &calloc_zero_calls, 1, __ATOMIC_SEQ_CST );
852 #endif // __STATISTICS__
853 return 0p;
854 } // if
855 #ifdef __STATISTICS__
856 __atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
857 __atomic_add_fetch( &calloc_storage, dim * elemSize, __ATOMIC_SEQ_CST );
858 #endif // __STATISTICS__
859
860 char * addr = (char *)mallocNoStats( size );
861
862 HeapManager.Storage.Header * header;
863 HeapManager.FreeHeader * freeElem;
864 size_t bsize, alignment;
865
866 #ifndef __CFA_DEBUG__
867 bool mapped =
868 #endif // __CFA_DEBUG__
869 headers( "calloc", addr, header, freeElem, bsize, alignment );
870
871 #ifndef __CFA_DEBUG__
872 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
873 if ( ! mapped )
874 #endif // __CFA_DEBUG__
875 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
876 // `-header`-addr `-size
877 memset( addr, '\0', size ); // set to zeros
878
879 header->kind.real.blockSize |= 2; // mark as zero filled
880 return addr;
881 } // calloc
882
883
884 // Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
885 // 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
886 // not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
887 // call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
888 void * resize( void * oaddr, size_t size ) {
889 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
890 if ( unlikely( size == 0 ) ) { // special cases
891 #ifdef __STATISTICS__
892 __atomic_add_fetch( &resize_zero_calls, 1, __ATOMIC_SEQ_CST );
893 #endif // __STATISTICS__
894 free( oaddr );
895 return 0p;
896 } // if
897 #ifdef __STATISTICS__
898 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
899 #endif // __STATISTICS__
900
901 if ( unlikely( oaddr == 0p ) ) {
902 #ifdef __STATISTICS__
903 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
904 #endif // __STATISTICS__
905 return mallocNoStats( size );
906 } // if
907
908 HeapManager.Storage.Header * header;
909 HeapManager.FreeHeader * freeElem;
910 size_t bsize, oalign;
911 headers( "resize", oaddr, header, freeElem, bsize, oalign );
912
913 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
914 // same size, DO NOT preserve STICKY PROPERTIES.
915 if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
916 header->kind.real.blockSize &= -2; // no alignment and turn off 0 fill
917 header->kind.real.size = size; // reset allocation size
918 return oaddr;
919 } // if
920
921 #ifdef __STATISTICS__
922 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
923 #endif // __STATISTICS__
924
925 // change size, DO NOT preserve STICKY PROPERTIES.
926 free( oaddr );
927 return mallocNoStats( size ); // create new area
928 } // resize
929
930
931 // Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
932 // the old and new sizes.
933 void * realloc( void * oaddr, size_t size ) {
934 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
935 if ( unlikely( size == 0 ) ) { // special cases
936 #ifdef __STATISTICS__
937 __atomic_add_fetch( &realloc_zero_calls, 1, __ATOMIC_SEQ_CST );
938 #endif // __STATISTICS__
939 free( oaddr );
940 return 0p;
941 } // if
942 #ifdef __STATISTICS__
943 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
944 #endif // __STATISTICS__
945
946 if ( unlikely( oaddr == 0p ) ) {
947 #ifdef __STATISTICS__
948 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
949 #endif // __STATISTICS__
950 return mallocNoStats( size );
951 } // if
952
953 HeapManager.Storage.Header * header;
954 HeapManager.FreeHeader * freeElem;
955 size_t bsize, oalign;
956 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
957
958 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
959 size_t osize = header->kind.real.size; // old allocation size
960 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
961 if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
962 header->kind.real.size = size; // reset allocation size
963 if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
964 memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
965 } // if
966 return oaddr;
967 } // if
968
969 #ifdef __STATISTICS__
970 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
971 #endif // __STATISTICS__
972
973 // change size and copy old content to new storage
974
975 void * naddr;
976 if ( likely( oalign == libAlign() ) ) { // previous request not aligned ?
977 naddr = mallocNoStats( size ); // create new area
978 } else {
979 naddr = memalignNoStats( oalign, size ); // create new aligned area
980 } // if
981
982 headers( "realloc", naddr, header, freeElem, bsize, oalign );
983 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
984 free( oaddr );
985
986 if ( unlikely( ozfill ) ) { // previous request zero fill ?
987 header->kind.real.blockSize |= 2; // mark new request as zero filled
988 if ( size > osize ) { // previous request larger ?
989 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
990 } // if
991 } // if
992 return naddr;
993 } // realloc
994
995
996 // Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
997 void * memalign( size_t alignment, size_t size ) {
998 #ifdef __STATISTICS__
999 if ( likely( size > 0 ) ) {
1000 __atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
1001 __atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
1002 } else {
1003 __atomic_add_fetch( &memalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1004 } // if
1005 #endif // __STATISTICS__
1006
1007 return memalignNoStats( alignment, size );
1008 } // memalign
1009
1010
1011 // Same as aalloc() with memory alignment.
1012 void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
1013 size_t size = dim * elemSize;
1014 #ifdef __STATISTICS__
1015 if ( likely( size > 0 ) ) {
1016 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1017 __atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
1018 } else {
1019 __atomic_add_fetch( &cmemalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1020 } // if
1021 #endif // __STATISTICS__
1022
1023 return memalignNoStats( alignment, size );
1024 } // amemalign
1025
1026
1027 // Same as calloc() with memory alignment.
1028 void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
1029 size_t size = dim * elemSize;
1030 if ( unlikely( size ) == 0 ) { // 0 BYTE ALLOCATION RETURNS NULL POINTER
1031 #ifdef __STATISTICS__
1032 __atomic_add_fetch( &cmemalign_zero_calls, 1, __ATOMIC_SEQ_CST );
1033 #endif // __STATISTICS__
1034 return 0p;
1035 } // if
1036 #ifdef __STATISTICS__
1037 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1038 __atomic_add_fetch( &cmemalign_storage, dim * elemSize, __ATOMIC_SEQ_CST );
1039 #endif // __STATISTICS__
1040
1041 char * addr = (char *)memalignNoStats( alignment, size );
1042
1043 HeapManager.Storage.Header * header;
1044 HeapManager.FreeHeader * freeElem;
1045 size_t bsize;
1046
1047 #ifndef __CFA_DEBUG__
1048 bool mapped =
1049 #endif // __CFA_DEBUG__
1050 headers( "cmemalign", addr, header, freeElem, bsize, alignment );
1051
1052 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1053 #ifndef __CFA_DEBUG__
1054 if ( ! mapped )
1055 #endif // __CFA_DEBUG__
1056 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1057 // `-header`-addr `-size
1058 memset( addr, '\0', size ); // set to zeros
1059
1060 header->kind.real.blockSize |= 2; // mark as zero filled
1061 return addr;
1062 } // cmemalign
1063
1064
1065 // Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1066 // of alignment. This requirement is universally ignored.
1067 void * aligned_alloc( size_t alignment, size_t size ) {
1068 return memalign( alignment, size );
1069 } // aligned_alloc
1070
1071
1072 // Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1073 // memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1074 // is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1075 // free(3).
1076 int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
1077 if ( alignment < libAlign() || ! is_pow2( alignment ) ) return EINVAL; // check alignment
1078 * memptr = memalign( alignment, size );
1079 return 0;
1080 } // posix_memalign
1081
1082
1083 // Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1084 // page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1085 void * valloc( size_t size ) {
1086 return memalign( __page_size, size );
1087 } // valloc
1088
1089
1090 // Same as valloc but rounds size to multiple of page size.
1091 void * pvalloc( size_t size ) {
1092 return memalign( __page_size, ceiling2( size, __page_size ) );
1093 } // pvalloc
1094
1095
1096 // Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1097 // or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1098 // 0p, no operation is performed.
1099 void free( void * addr ) {
1100 if ( unlikely( addr == 0p ) ) { // special case
1101 #ifdef __STATISTICS__
1102 __atomic_add_fetch( &free_zero_calls, 1, __ATOMIC_SEQ_CST );
1103 #endif // __STATISTICS__
1104
1105 // #ifdef __CFA_DEBUG__
1106 // if ( traceHeap() ) {
1107 // #define nullmsg "Free( 0x0 ) size:0\n"
1108 // // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
1109 // __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
1110 // } // if
1111 // #endif // __CFA_DEBUG__
1112 return;
1113 } // exit
1114
1115 doFree( addr );
1116 } // free
1117
1118
1119 // Returns the alignment of an allocation.
1120 size_t malloc_alignment( void * addr ) {
1121 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1122 HeapManager.Storage.Header * header = headerAddr( addr );
1123 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1124 return header->kind.fake.alignment & -2; // remove flag from value
1125 } else {
1126 return libAlign(); // minimum alignment
1127 } // if
1128 } // malloc_alignment
1129
1130
1131 // Set the alignment for an the allocation and return previous alignment or 0 if no alignment.
1132 size_t malloc_alignment_set$( void * addr, size_t alignment ) {
1133 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1134 size_t ret;
1135 HeapManager.Storage.Header * header = headerAddr( addr );
1136 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1137 ret = header->kind.fake.alignment & -2; // remove flag from old value
1138 header->kind.fake.alignment = alignment | 1; // add flag to new value
1139 } else {
1140 ret = 0; // => no alignment to change
1141 } // if
1142 return ret;
1143 } // malloc_alignment_set$
1144
1145
1146 // Returns true if the allocation is zero filled, e.g., allocated by calloc().
1147 bool malloc_zero_fill( void * addr ) {
1148 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1149 HeapManager.Storage.Header * header = headerAddr( addr );
1150 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1151 header = realHeader( header ); // backup from fake to real header
1152 } // if
1153 return (header->kind.real.blockSize & 2) != 0; // zero filled ?
1154 } // malloc_zero_fill
1155
1156 // Set allocation is zero filled and return previous zero filled.
1157 bool malloc_zero_fill_set$( void * addr ) {
1158 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1159 HeapManager.Storage.Header * header = headerAddr( addr );
1160 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1161 header = realHeader( header ); // backup from fake to real header
1162 } // if
1163 bool ret = (header->kind.real.blockSize & 2) != 0; // zero filled ?
1164 header->kind.real.blockSize |= 2; // mark as zero filled
1165 return ret;
1166 } // malloc_zero_fill_set$
1167
1168
1169 // Returns original total allocation size (not bucket size) => array size is dimension * sizeif(T).
1170 size_t malloc_size( void * addr ) {
1171 if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1172 HeapManager.Storage.Header * header = headerAddr( addr );
1173 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1174 header = realHeader( header ); // backup from fake to real header
1175 } // if
1176 return header->kind.real.size;
1177 } // malloc_size
1178
1179 // Set allocation size and return previous size.
1180 size_t malloc_size_set$( void * addr, size_t size ) {
1181 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1182 HeapManager.Storage.Header * header = headerAddr( addr );
1183 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1184 header = realHeader( header ); // backup from fake to real header
1185 } // if
1186 size_t ret = header->kind.real.size;
1187 header->kind.real.size = size;
1188 return ret;
1189 } // malloc_size_set$
1190
1191
1192 // Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1193 // malloc or a related function.
1194 size_t malloc_usable_size( void * addr ) {
1195 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1196 HeapManager.Storage.Header * header;
1197 HeapManager.FreeHeader * freeElem;
1198 size_t bsize, alignment;
1199
1200 headers( "malloc_usable_size", addr, header, freeElem, bsize, alignment );
1201 return dataStorage( bsize, addr, header ); // data storage in bucket
1202 } // malloc_usable_size
1203
1204
1205 // Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1206 void malloc_stats( void ) {
1207 #ifdef __STATISTICS__
1208 printStats();
1209 if ( prtFree() ) prtFree( heapManager );
1210 #endif // __STATISTICS__
1211 } // malloc_stats
1212
1213
1214 // Changes the file descripter where malloc_stats() writes statistics.
1215 int malloc_stats_fd( int fd __attribute__(( unused )) ) {
1216 #ifdef __STATISTICS__
1217 int temp = stats_fd;
1218 stats_fd = fd;
1219 return temp;
1220 #else
1221 return -1;
1222 #endif // __STATISTICS__
1223 } // malloc_stats_fd
1224
1225
1226 // Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1227 // specifies the parameter to be modified, and value specifies the new value for that parameter.
1228 int mallopt( int option, int value ) {
1229 choose( option ) {
1230 case M_TOP_PAD:
1231 heapExpand = ceiling2( value, __page_size ); return 1;
1232 case M_MMAP_THRESHOLD:
1233 if ( setMmapStart( value ) ) return 1;
1234 break;
1235 } // switch
1236 return 0; // error, unsupported
1237 } // mallopt
1238
1239
1240 // Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1241 int malloc_trim( size_t ) {
1242 return 0; // => impossible to release memory
1243 } // malloc_trim
1244
1245
1246 // Exports an XML string that describes the current state of the memory-allocation implementation in the caller.
1247 // The string is printed on the file stream stream. The exported string includes information about all arenas (see
1248 // malloc).
1249 int malloc_info( int options, FILE * stream __attribute__(( unused )) ) {
1250 if ( options != 0 ) { errno = EINVAL; return -1; }
1251 #ifdef __STATISTICS__
1252 return printStatsXML( stream );
1253 #else
1254 return 0; // unsupported
1255 #endif // __STATISTICS__
1256 } // malloc_info
1257
1258
1259 // Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1260 // or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1261 // structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1262 // result. (The caller must free this memory.)
1263 void * malloc_get_state( void ) {
1264 return 0p; // unsupported
1265 } // malloc_get_state
1266
1267
1268 // Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1269 // structure pointed to by state.
1270 int malloc_set_state( void * ) {
1271 return 0; // unsupported
1272 } // malloc_set_state
1273} // extern "C"
1274
1275
1276// Must have CFA linkage to overload with C linkage realloc.
1277void * resize( void * oaddr, size_t nalign, size_t size ) {
1278 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1279 if ( unlikely( size == 0 ) ) { // special cases
1280 #ifdef __STATISTICS__
1281 __atomic_add_fetch( &resize_zero_calls, 1, __ATOMIC_SEQ_CST );
1282 #endif // __STATISTICS__
1283 free( oaddr );
1284 return 0p;
1285 } // if
1286
1287 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1288 #ifdef __CFA_DEBUG__
1289 else checkAlign( nalign ); // check alignment
1290 #endif // __CFA_DEBUG__
1291
1292 if ( unlikely( oaddr == 0p ) ) {
1293 #ifdef __STATISTICS__
1294 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
1295 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1296 #endif // __STATISTICS__
1297 return memalignNoStats( nalign, size );
1298 } // if
1299
1300 // Attempt to reuse existing alignment.
1301 HeapManager.Storage.Header * header = headerAddr( oaddr );
1302 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1303 size_t oalign;
1304 if ( isFakeHeader ) {
1305 oalign = header->kind.fake.alignment & -2; // old alignment
1306 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1307 && ( oalign <= nalign // going down
1308 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1309 ) {
1310 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1311 HeapManager.FreeHeader * freeElem;
1312 size_t bsize, oalign;
1313 headers( "resize", oaddr, header, freeElem, bsize, oalign );
1314 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1315
1316 if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1317 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1318
1319 header->kind.real.blockSize &= -2; // turn off 0 fill
1320 header->kind.real.size = size; // reset allocation size
1321 return oaddr;
1322 } // if
1323 } // if
1324 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1325 && nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1326 return resize( oaddr, size ); // duplicate special case checks
1327 } // if
1328
1329 #ifdef __STATISTICS__
1330 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1331 #endif // __STATISTICS__
1332
1333 // change size, DO NOT preserve STICKY PROPERTIES.
1334 free( oaddr );
1335 return memalignNoStats( nalign, size ); // create new aligned area
1336} // resize
1337
1338
1339void * realloc( void * oaddr, size_t nalign, size_t size ) {
1340 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1341 if ( unlikely( size == 0 ) ) { // special cases
1342 #ifdef __STATISTICS__
1343 __atomic_add_fetch( &realloc_zero_calls, 1, __ATOMIC_SEQ_CST );
1344 #endif // __STATISTICS__
1345 free( oaddr );
1346 return 0p;
1347 } // if
1348
1349 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1350 #ifdef __CFA_DEBUG__
1351 else checkAlign( nalign ); // check alignment
1352 #endif // __CFA_DEBUG__
1353
1354 if ( unlikely( oaddr == 0p ) ) {
1355 #ifdef __STATISTICS__
1356 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1357 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1358 #endif // __STATISTICS__
1359 return memalignNoStats( nalign, size );
1360 } // if
1361
1362 // Attempt to reuse existing alignment.
1363 HeapManager.Storage.Header * header = headerAddr( oaddr );
1364 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1365 size_t oalign;
1366 if ( isFakeHeader ) {
1367 oalign = header->kind.fake.alignment & -2; // old alignment
1368 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1369 && ( oalign <= nalign // going down
1370 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1371 ) {
1372 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1373 return realloc( oaddr, size ); // duplicate alignment and special case checks
1374 } // if
1375 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1376 && nalign == libAlign() ) // new alignment also on libAlign => no fake header needed
1377 return realloc( oaddr, size ); // duplicate alignment and special case checks
1378
1379 #ifdef __STATISTICS__
1380 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1381 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1382 #endif // __STATISTICS__
1383
1384 HeapManager.FreeHeader * freeElem;
1385 size_t bsize;
1386 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1387
1388 // change size and copy old content to new storage
1389
1390 size_t osize = header->kind.real.size; // old allocation size
1391 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1392
1393 void * naddr = memalignNoStats( nalign, size ); // create new aligned area
1394
1395 headers( "realloc", naddr, header, freeElem, bsize, oalign );
1396 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1397 free( oaddr );
1398
1399 if ( unlikely( ozfill ) ) { // previous request zero fill ?
1400 header->kind.real.blockSize |= 2; // mark new request as zero filled
1401 if ( size > osize ) { // previous request larger ?
1402 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1403 } // if
1404 } // if
1405 return naddr;
1406} // realloc
1407
1408
1409// Local Variables: //
1410// tab-width: 4 //
1411// compile-command: "cfa -nodebug -O2 heap.cfa" //
1412// End: //
Note: See TracBrowser for help on using the repository browser.