source: libcfa/src/heap.cfa@ e873838

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since e873838 was 113d785, checked in by Peter A. Buhr <pabuhr@…>, 5 years ago

formatting
  • Property mode set to 100644
File size: 51.3 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 : Mon Sep 7 22:17:46 2020
13// Update Count : 957
14//
15
16#include <unistd.h> // sbrk, sysconf
17#include <stdbool.h> // true, false
18#include <stdio.h> // snprintf, fileno
19#include <errno.h> // errno
20#include <string.h> // memset, memcpy
21#include <limits.h> // ULONG_MAX
22#include <malloc.h> // memalign, malloc_usable_size
23#include <sys/mman.h> // mmap, munmap
24
25#include "bits/align.hfa" // libAlign
26#include "bits/defs.hfa" // likely, unlikely
27#include "bits/locks.hfa" // __spinlock_t
28#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
29#include "math.hfa" // ceiling
30#include "bitmanip.hfa" // is_pow2, ceiling2
31
32static bool traceHeap = false;
33
34inline bool traceHeap() { return traceHeap; }
35
36bool traceHeapOn() {
37 bool temp = traceHeap;
38 traceHeap = true;
39 return temp;
40} // traceHeapOn
41
42bool traceHeapOff() {
43 bool temp = traceHeap;
44 traceHeap = false;
45 return temp;
46} // traceHeapOff
47
48bool traceHeapTerm() { return false; }
49
50
51static bool prtFree = false;
52
53inline bool prtFree() {
54 return prtFree;
55} // prtFree
56
57bool prtFreeOn() {
58 bool temp = prtFree;
59 prtFree = true;
60 return temp;
61} // prtFreeOn
62
63bool prtFreeOff() {
64 bool temp = prtFree;
65 prtFree = false;
66 return temp;
67} // prtFreeOff
68
69
70enum {
71 // Define the default extension heap amount in units of bytes. When the uC++ supplied heap reaches the brk address,
72 // the brk address is extended by the extension amount.
73 __CFA_DEFAULT_HEAP_EXPANSION__ = (1 * 1024 * 1024),
74
75 // Define the mmap crossover point during allocation. Allocations less than this amount are allocated from buckets;
76 // values greater than or equal to this value are mmap from the operating system.
77 __CFA_DEFAULT_MMAP_START__ = (512 * 1024 + 1),
78};
79
80size_t default_mmap_start() __attribute__(( weak )) {
81 return __CFA_DEFAULT_MMAP_START__;
82} // default_mmap_start
83
84size_t default_heap_expansion() __attribute__(( weak )) {
85 return __CFA_DEFAULT_HEAP_EXPANSION__;
86} // default_heap_expansion
87
88
89#ifdef __CFA_DEBUG__
90static size_t allocUnfreed; // running total of allocations minus frees
91
92static void prtUnfreed() {
93 if ( allocUnfreed != 0 ) {
94 // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
95 char helpText[512];
96 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"
97 "Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
98 (long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
99 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
100 } // if
101} // prtUnfreed
102
103extern "C" {
104 void heapAppStart() { // called by __cfaabi_appready_startup
105 allocUnfreed = 0;
106 } // heapAppStart
107
108 void heapAppStop() { // called by __cfaabi_appready_startdown
109 fclose( stdin ); fclose( stdout );
110 prtUnfreed();
111 } // heapAppStop
112} // extern "C"
113#endif // __CFA_DEBUG__
114
115
116// statically allocated variables => zero filled.
117static size_t pageSize; // architecture pagesize
118static size_t heapExpand; // sbrk advance
119static size_t mmapStart; // cross over point for mmap
120static unsigned int maxBucketsUsed; // maximum number of buckets in use
121
122
123#define SPINLOCK 0
124#define LOCKFREE 1
125#define BUCKETLOCK SPINLOCK
126#if BUCKETLOCK == SPINLOCK
127#elif BUCKETLOCK == LOCKFREE
128#include <stackLockFree.hfa>
129#else
130 #error undefined lock type for bucket lock
131#endif // LOCKFREE
132
133// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
134// Break recusion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
135enum { NoBucketSizes = 91 }; // number of buckets sizes
136
137struct HeapManager {
138 struct Storage {
139 struct Header { // header
140 union Kind {
141 struct RealHeader {
142 union {
143 struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
144 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
145 uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
146 #endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
147
148 union {
149 // FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
150 // 2nd low-order bit => zero filled
151 void * home; // allocated block points back to home locations (must overlay alignment)
152 size_t blockSize; // size for munmap (must overlay alignment)
153 #if BUCKETLOCK == SPINLOCK
154 Storage * next; // freed block points next freed block of same size
155 #endif // SPINLOCK
156 };
157 size_t size; // allocation size in bytes
158
159 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
160 uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
161 #endif // __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
162 };
163 #if BUCKETLOCK == LOCKFREE
164 Link(Storage) next; // freed block points next freed block of same size (double-wide)
165 #endif // LOCKFREE
166 };
167 } real; // RealHeader
168
169 struct FakeHeader {
170 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
171 uint32_t alignment; // 1st low-order bit => fake header & alignment
172 #endif // __ORDER_LITTLE_ENDIAN__
173
174 uint32_t offset;
175
176 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
177 uint32_t alignment; // low-order bits of home/blockSize used for tricks
178 #endif // __ORDER_BIG_ENDIAN__
179 } fake; // FakeHeader
180 } kind; // Kind
181 } header; // Header
182 char pad[libAlign() - sizeof( Header )];
183 char data[0]; // storage
184 }; // Storage
185
186 static_assert( libAlign() >= sizeof( Storage ), "libAlign() < sizeof( Storage )" );
187
188 struct FreeHeader {
189 #if BUCKETLOCK == SPINLOCK
190 __spinlock_t lock; // must be first field for alignment
191 Storage * freeList;
192 #else
193 StackLF(Storage) freeList;
194 #endif // BUCKETLOCK
195 size_t blockSize; // size of allocations on this list
196 }; // FreeHeader
197
198 // must be first fields for alignment
199 __spinlock_t extlock; // protects allocation-buffer extension
200 FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
201
202 void * heapBegin; // start of heap
203 void * heapEnd; // logical end of heap
204 size_t heapRemaining; // amount of storage not allocated in the current chunk
205}; // HeapManager
206
207#if BUCKETLOCK == LOCKFREE
208static inline {
209 Link(HeapManager.Storage) * ?`next( HeapManager.Storage * this ) { return &this->header.kind.real.next; }
210 void ?{}( HeapManager.FreeHeader & ) {}
211 void ^?{}( HeapManager.FreeHeader & ) {}
212} // distribution
213#endif // LOCKFREE
214
215static inline size_t getKey( const HeapManager.FreeHeader & freeheader ) { return freeheader.blockSize; }
216
217
218#define FASTLOOKUP
219#define __STATISTICS__
220
221// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
222// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
223// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
224static const unsigned int bucketSizes[] @= { // different bucket sizes
225 16 + sizeof(HeapManager.Storage), 32 + sizeof(HeapManager.Storage), 48 + sizeof(HeapManager.Storage), 64 + sizeof(HeapManager.Storage), // 4
226 96 + sizeof(HeapManager.Storage), 112 + sizeof(HeapManager.Storage), 128 + sizeof(HeapManager.Storage), // 3
227 160, 192, 224, 256 + sizeof(HeapManager.Storage), // 4
228 320, 384, 448, 512 + sizeof(HeapManager.Storage), // 4
229 640, 768, 896, 1_024 + sizeof(HeapManager.Storage), // 4
230 1_536, 2_048 + sizeof(HeapManager.Storage), // 2
231 2_560, 3_072, 3_584, 4_096 + sizeof(HeapManager.Storage), // 4
232 6_144, 8_192 + sizeof(HeapManager.Storage), // 2
233 9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(HeapManager.Storage), // 8
234 18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(HeapManager.Storage), // 8
235 36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(HeapManager.Storage), // 8
236 73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(HeapManager.Storage), // 8
237 147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(HeapManager.Storage), // 8
238 294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(HeapManager.Storage), // 8
239 655_360, 786_432, 917_504, 1_048_576 + sizeof(HeapManager.Storage), // 4
240 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
241 2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(HeapManager.Storage), // 4
242};
243
244static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
245
246#ifdef FASTLOOKUP
247enum { LookupSizes = 65_536 + sizeof(HeapManager.Storage) }; // number of fast lookup sizes
248static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
249#endif // FASTLOOKUP
250
251static int mmapFd = -1; // fake or actual fd for anonymous file
252#ifdef __CFA_DEBUG__
253static bool heapBoot = 0; // detect recursion during boot
254#endif // __CFA_DEBUG__
255
256// The constructor for heapManager is called explicitly in memory_startup.
257static HeapManager heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing
258
259
260#ifdef __STATISTICS__
261// Heap statistics counters.
262static unsigned int malloc_calls;
263static unsigned long long int malloc_storage;
264static unsigned int aalloc_calls;
265static unsigned long long int aalloc_storage;
266static unsigned int calloc_calls;
267static unsigned long long int calloc_storage;
268static unsigned int memalign_calls;
269static unsigned long long int memalign_storage;
270static unsigned int amemalign_calls;
271static unsigned long long int amemalign_storage;
272static unsigned int cmemalign_calls;
273static unsigned long long int cmemalign_storage;
274static unsigned int resize_calls;
275static unsigned long long int resize_storage;
276static unsigned int realloc_calls;
277static unsigned long long int realloc_storage;
278static unsigned int free_calls;
279static unsigned long long int free_storage;
280static unsigned int mmap_calls;
281static unsigned long long int mmap_storage;
282static unsigned int munmap_calls;
283static unsigned long long int munmap_storage;
284static unsigned int sbrk_calls;
285static unsigned long long int sbrk_storage;
286// Statistics file descriptor (changed by malloc_stats_fd).
287static int stat_fd = STDERR_FILENO; // default stderr
288
289// Use "write" because streams may be shutdown when calls are made.
290static void printStats() {
291 char helpText[1024];
292 __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
293 "\nHeap statistics:\n"
294 " malloc: calls %u / storage %llu\n"
295 " aalloc: calls %u / storage %llu\n"
296 " calloc: calls %u / storage %llu\n"
297 " memalign: calls %u / storage %llu\n"
298 " amemalign: calls %u / storage %llu\n"
299 " cmemalign: calls %u / storage %llu\n"
300 " resize: calls %u / storage %llu\n"
301 " realloc: calls %u / storage %llu\n"
302 " free: calls %u / storage %llu\n"
303 " mmap: calls %u / storage %llu\n"
304 " munmap: calls %u / storage %llu\n"
305 " sbrk: calls %u / storage %llu\n",
306 malloc_calls, malloc_storage,
307 aalloc_calls, aalloc_storage,
308 calloc_calls, calloc_storage,
309 memalign_calls, memalign_storage,
310 amemalign_calls, amemalign_storage,
311 cmemalign_calls, cmemalign_storage,
312 resize_calls, resize_storage,
313 realloc_calls, realloc_storage,
314 free_calls, free_storage,
315 mmap_calls, mmap_storage,
316 munmap_calls, munmap_storage,
317 sbrk_calls, sbrk_storage
318 );
319} // printStats
320
321static int printStatsXML( FILE * stream ) { // see malloc_info
322 char helpText[1024];
323 int len = snprintf( helpText, sizeof(helpText),
324 "<malloc version=\"1\">\n"
325 "<heap nr=\"0\">\n"
326 "<sizes>\n"
327 "</sizes>\n"
328 "<total type=\"malloc\" count=\"%u\" size=\"%llu\"/>\n"
329 "<total type=\"aalloc\" count=\"%u\" size=\"%llu\"/>\n"
330 "<total type=\"calloc\" count=\"%u\" size=\"%llu\"/>\n"
331 "<total type=\"memalign\" count=\"%u\" size=\"%llu\"/>\n"
332 "<total type=\"amemalign\" count=\"%u\" size=\"%llu\"/>\n"
333 "<total type=\"cmemalign\" count=\"%u\" size=\"%llu\"/>\n"
334 "<total type=\"resize\" count=\"%u\" size=\"%llu\"/>\n"
335 "<total type=\"realloc\" count=\"%u\" size=\"%llu\"/>\n"
336 "<total type=\"free\" count=\"%u\" size=\"%llu\"/>\n"
337 "<total type=\"mmap\" count=\"%u\" size=\"%llu\"/>\n"
338 "<total type=\"munmap\" count=\"%u\" size=\"%llu\"/>\n"
339 "<total type=\"sbrk\" count=\"%u\" size=\"%llu\"/>\n"
340 "</malloc>",
341 malloc_calls, malloc_storage,
342 aalloc_calls, aalloc_storage,
343 calloc_calls, calloc_storage,
344 memalign_calls, memalign_storage,
345 amemalign_calls, amemalign_storage,
346 cmemalign_calls, cmemalign_storage,
347 resize_calls, resize_storage,
348 realloc_calls, realloc_storage,
349 free_calls, free_storage,
350 mmap_calls, mmap_storage,
351 munmap_calls, munmap_storage,
352 sbrk_calls, sbrk_storage
353 );
354 __cfaabi_bits_write( fileno( stream ), helpText, len ); // ensures all bytes written or exit
355 return len;
356} // printStatsXML
357#endif // __STATISTICS__
358
359
360// thunk problem
361size_t Bsearchl( unsigned int key, const unsigned int * vals, size_t dim ) {
362 size_t l = 0, m, h = dim;
363 while ( l < h ) {
364 m = (l + h) / 2;
365 if ( (unsigned int &)(vals[m]) < key ) { // cast away const
366 l = m + 1;
367 } else {
368 h = m;
369 } // if
370 } // while
371 return l;
372} // Bsearchl
373
374
375static inline bool setMmapStart( size_t value ) { // true => mmapped, false => sbrk
376 if ( value < pageSize || bucketSizes[NoBucketSizes - 1] < value ) return false;
377 mmapStart = value; // set global
378
379 // find the closest bucket size less than or equal to the mmapStart size
380 maxBucketsUsed = Bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
381 assert( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
382 assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
383 return true;
384} // setMmapStart
385
386
387// <-------+----------------------------------------------------> bsize (bucket size)
388// |header |addr
389//==================================================================================
390// align/offset |
391// <-----------------<------------+-----------------------------> bsize (bucket size)
392// |fake-header | addr
393#define headerAddr( addr ) ((HeapManager.Storage.Header *)( (char *)addr - sizeof(HeapManager.Storage) ))
394#define realHeader( header ) ((HeapManager.Storage.Header *)((char *)header - header->kind.fake.offset))
395
396// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
397// |header |addr
398//==================================================================================
399// align/offset |
400// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
401// |fake-header |addr
402#define dataStorage( bsize, addr, header ) (bsize - ( (char *)addr - (char *)header ))
403
404
405static inline void checkAlign( size_t alignment ) {
406 if ( alignment < libAlign() || ! is_pow2( alignment ) ) {
407 abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
408 } // if
409} // checkAlign
410
411
412static inline void checkHeader( bool check, const char name[], void * addr ) {
413 if ( unlikely( check ) ) { // bad address ?
414 abort( "Attempt to %s storage %p with address outside the heap.\n"
415 "Possible cause is duplicate free on same block or overwriting of memory.",
416 name, addr );
417 } // if
418} // checkHeader
419
420
421static inline void fakeHeader( HeapManager.Storage.Header *& header, size_t & alignment ) {
422 if ( unlikely( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
423 alignment = header->kind.fake.alignment & -2; // remove flag from value
424 #ifdef __CFA_DEBUG__
425 checkAlign( alignment ); // check alignment
426 #endif // __CFA_DEBUG__
427 header = realHeader( header ); // backup from fake to real header
428 } else {
429 alignment = libAlign(); // => no fake header
430 } // if
431} // fakeHeader
432
433
434static inline bool headers( const char name[] __attribute__(( unused )), void * addr, HeapManager.Storage.Header *& header, HeapManager.FreeHeader *& freeElem,
435 size_t & size, size_t & alignment ) with( heapManager ) {
436 header = headerAddr( addr );
437
438 if ( unlikely( heapEnd < addr ) ) { // mmapped ?
439 fakeHeader( header, alignment );
440 size = header->kind.real.blockSize & -3; // mmap size
441 return true;
442 } // if
443
444 #ifdef __CFA_DEBUG__
445 checkHeader( addr < heapBegin, name, addr ); // bad low address ?
446 #endif // __CFA_DEBUG__
447
448 // header may be safe to dereference
449 fakeHeader( header, alignment );
450 #ifdef __CFA_DEBUG__
451 checkHeader( header < (HeapManager.Storage.Header *)heapBegin || (HeapManager.Storage.Header *)heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
452 #endif // __CFA_DEBUG__
453
454 freeElem = (HeapManager.FreeHeader *)((size_t)header->kind.real.home & -3);
455 #ifdef __CFA_DEBUG__
456 if ( freeElem < &freeLists[0] || &freeLists[NoBucketSizes] <= freeElem ) {
457 abort( "Attempt to %s storage %p with corrupted header.\n"
458 "Possible cause is duplicate free on same block or overwriting of header information.",
459 name, addr );
460 } // if
461 #endif // __CFA_DEBUG__
462 size = freeElem->blockSize;
463 return false;
464} // headers
465
466#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
467
468static inline void * extend( size_t size ) with( heapManager ) {
469 lock( extlock __cfaabi_dbg_ctx2 );
470 ptrdiff_t rem = heapRemaining - size;
471 if ( rem < 0 ) {
472 // If the size requested is bigger than the current remaining storage, increase the size of the heap.
473
474 size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
475 if ( sbrk( increase ) == (void *)-1 ) { // failed, no memory ?
476 unlock( extlock );
477 abort( NO_MEMORY_MSG, size ); // give up
478 } // if
479 #ifdef __STATISTICS__
480 sbrk_calls += 1;
481 sbrk_storage += increase;
482 #endif // __STATISTICS__
483 #ifdef __CFA_DEBUG__
484 // Set new memory to garbage so subsequent uninitialized usages might fail.
485 memset( (char *)heapEnd + heapRemaining, '\377', increase );
486 #endif // __CFA_DEBUG__
487 rem = heapRemaining + increase - size;
488 } // if
489
490 HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
491 heapRemaining = rem;
492 heapEnd = (char *)heapEnd + size;
493 unlock( extlock );
494 return block;
495} // extend
496
497
498static inline void * doMalloc( size_t size ) with( heapManager ) {
499 HeapManager.Storage * block; // pointer to new block of storage
500
501 // Look up size in the size list. Make sure the user request includes space for the header that must be allocated
502 // along with the block and is a multiple of the alignment size.
503
504 if ( unlikely( size > ULONG_MAX - sizeof(HeapManager.Storage) ) ) return 0p;
505 size_t tsize = size + sizeof(HeapManager.Storage);
506 if ( likely( tsize < mmapStart ) ) { // small size => sbrk
507 size_t posn;
508 #ifdef FASTLOOKUP
509 if ( tsize < LookupSizes ) posn = lookup[tsize];
510 else
511 #endif // FASTLOOKUP
512 posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
513 HeapManager.FreeHeader * freeElem = &freeLists[posn];
514 verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
515 verify( tsize <= freeElem->blockSize ); // search failure ?
516 tsize = freeElem->blockSize; // total space needed for request
517
518 // Spin until the lock is acquired for this particular size of block.
519
520 #if BUCKETLOCK == SPINLOCK
521 lock( freeElem->lock __cfaabi_dbg_ctx2 );
522 block = freeElem->freeList; // remove node from stack
523 #else
524 block = pop( freeElem->freeList );
525 #endif // BUCKETLOCK
526 if ( unlikely( block == 0p ) ) { // no free block ?
527 #if BUCKETLOCK == SPINLOCK
528 unlock( freeElem->lock );
529 #endif // BUCKETLOCK
530
531 // Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
532 // and then carve it off.
533
534 block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
535 #if BUCKETLOCK == SPINLOCK
536 } else {
537 freeElem->freeList = block->header.kind.real.next;
538 unlock( freeElem->lock );
539 #endif // BUCKETLOCK
540 } // if
541
542 block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
543 } else { // large size => mmap
544 if ( unlikely( size > ULONG_MAX - pageSize ) ) return 0p;
545 tsize = ceiling2( tsize, pageSize ); // must be multiple of page size
546 #ifdef __STATISTICS__
547 __atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
548 __atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
549 #endif // __STATISTICS__
550
551 block = (HeapManager.Storage *)mmap( 0, tsize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, mmapFd, 0 );
552 if ( block == (HeapManager.Storage *)MAP_FAILED ) { // failed ?
553 if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
554 // Do not call strerror( errno ) as it may call malloc.
555 abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu error:%d.", &heapManager, tsize, errno );
556 } //if
557 #ifdef __CFA_DEBUG__
558 // Set new memory to garbage so subsequent uninitialized usages might fail.
559 memset( block, '\377', tsize );
560 #endif // __CFA_DEBUG__
561 block->header.kind.real.blockSize = tsize; // storage size for munmap
562 } // if
563
564 block->header.kind.real.size = size; // store allocation size
565 void * addr = &(block->data); // adjust off header to user bytes
566 verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
567
568 #ifdef __CFA_DEBUG__
569 __atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
570 if ( traceHeap() ) {
571 enum { BufferSize = 64 };
572 char helpText[BufferSize];
573 int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize );
574 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
575 } // if
576 #endif // __CFA_DEBUG__
577
578 return addr;
579} // doMalloc
580
581
582static inline void doFree( void * addr ) with( heapManager ) {
583 #ifdef __CFA_DEBUG__
584 if ( unlikely( heapManager.heapBegin == 0p ) ) {
585 abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
586 } // if
587 #endif // __CFA_DEBUG__
588
589 HeapManager.Storage.Header * header;
590 HeapManager.FreeHeader * freeElem;
591 size_t size, alignment; // not used (see realloc)
592
593 if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
594 #ifdef __STATISTICS__
595 __atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
596 __atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
597 #endif // __STATISTICS__
598 if ( munmap( header, size ) == -1 ) {
599 #ifdef __CFA_DEBUG__
600 abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
601 "Possible cause is invalid pointer.",
602 addr );
603 #endif // __CFA_DEBUG__
604 } // if
605 } else {
606 #ifdef __CFA_DEBUG__
607 // Set free memory to garbage so subsequent usages might fail.
608 memset( ((HeapManager.Storage *)header)->data, '\377', freeElem->blockSize - sizeof( HeapManager.Storage ) );
609 #endif // __CFA_DEBUG__
610
611 #ifdef __STATISTICS__
612 free_storage += size;
613 #endif // __STATISTICS__
614 #if BUCKETLOCK == SPINLOCK
615 lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
616 header->kind.real.next = freeElem->freeList; // push on stack
617 freeElem->freeList = (HeapManager.Storage *)header;
618 unlock( freeElem->lock ); // release spin lock
619 #else
620 push( freeElem->freeList, *(HeapManager.Storage *)header );
621 #endif // BUCKETLOCK
622 } // if
623
624 #ifdef __CFA_DEBUG__
625 __atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
626 if ( traceHeap() ) {
627 char helpText[64];
628 int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
629 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
630 } // if
631 #endif // __CFA_DEBUG__
632} // doFree
633
634
635size_t prtFree( HeapManager & manager ) with( manager ) {
636 size_t total = 0;
637 #ifdef __STATISTICS__
638 __cfaabi_bits_acquire();
639 __cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
640 #endif // __STATISTICS__
641 for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
642 size_t size = freeLists[i].blockSize;
643 #ifdef __STATISTICS__
644 unsigned int N = 0;
645 #endif // __STATISTICS__
646
647 #if BUCKETLOCK == SPINLOCK
648 for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
649 #else
650 // for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
651// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
652 for ( HeapManager.Storage * p ;; /* p = getNext( p )->top */) {
653 HeapManager.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
654// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
655// p = temp;
656 #endif // BUCKETLOCK
657 total += size;
658 #ifdef __STATISTICS__
659 N += 1;
660 #endif // __STATISTICS__
661 } // for
662
663 #ifdef __STATISTICS__
664 __cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
665 if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
666 #endif // __STATISTICS__
667 } // for
668 #ifdef __STATISTICS__
669 __cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
670 __cfaabi_bits_release();
671 #endif // __STATISTICS__
672 return (char *)heapEnd - (char *)heapBegin - total;
673} // prtFree
674
675
676static void ?{}( HeapManager & manager ) with( manager ) {
677 pageSize = sysconf( _SC_PAGESIZE );
678
679 for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
680 freeLists[i].blockSize = bucketSizes[i];
681 } // for
682
683 #ifdef FASTLOOKUP
684 unsigned int idx = 0;
685 for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
686 if ( i > bucketSizes[idx] ) idx += 1;
687 lookup[i] = idx;
688 } // for
689 #endif // FASTLOOKUP
690
691 if ( ! setMmapStart( default_mmap_start() ) ) {
692 abort( "HeapManager : internal error, mmap start initialization failure." );
693 } // if
694 heapExpand = default_heap_expansion();
695
696 char * end = (char *)sbrk( 0 );
697 heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, libAlign() ) - end ); // move start of heap to multiple of alignment
698} // HeapManager
699
700
701static void ^?{}( HeapManager & ) {
702 #ifdef __STATISTICS__
703 if ( traceHeapTerm() ) {
704 printStats();
705 // prtUnfreed() called in heapAppStop()
706 } // if
707 #endif // __STATISTICS__
708} // ~HeapManager
709
710
711static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
712void memory_startup( void ) {
713 #ifdef __CFA_DEBUG__
714 if ( heapBoot ) { // check for recursion during system boot
715 // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
716 abort( "boot() : internal error, recursively invoked during system boot." );
717 } // if
718 heapBoot = true;
719 #endif // __CFA_DEBUG__
720
721 //verify( heapManager.heapBegin != 0 );
722 //heapManager{};
723 if ( heapManager.heapBegin == 0p ) heapManager{}; // sanity check
724} // memory_startup
725
726static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
727void memory_shutdown( void ) {
728 ^heapManager{};
729} // memory_shutdown
730
731
732static inline void * mallocNoStats( size_t size ) { // necessary for malloc statistics
733 verify( heapManager.heapBegin != 0p ); // called before memory_startup ?
734 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
735
736#if __SIZEOF_POINTER__ == 8
737 verify( size < ((typeof(size_t))1 << 48) );
738#endif // __SIZEOF_POINTER__ == 8
739 return doMalloc( size );
740} // mallocNoStats
741
742
743static inline void * callocNoStats( size_t dim, size_t elemSize ) {
744 size_t size = dim * elemSize;
745 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
746 char * addr = (char *)mallocNoStats( size );
747
748 HeapManager.Storage.Header * header;
749 HeapManager.FreeHeader * freeElem;
750 size_t bsize, alignment;
751 #ifndef __CFA_DEBUG__
752 bool mapped =
753 #endif // __CFA_DEBUG__
754 headers( "calloc", addr, header, freeElem, bsize, alignment );
755 #ifndef __CFA_DEBUG__
756
757 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
758 if ( ! mapped )
759 #endif // __CFA_DEBUG__
760 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
761 // `-header`-addr `-size
762 memset( addr, '\0', size ); // set to zeros
763
764 header->kind.real.blockSize |= 2; // mark as zero filled
765 return addr;
766} // callocNoStats
767
768
769static inline void * memalignNoStats( size_t alignment, size_t size ) {
770 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
771
772 #ifdef __CFA_DEBUG__
773 checkAlign( alignment ); // check alignment
774 #endif // __CFA_DEBUG__
775
776 // if alignment <= default alignment, do normal malloc as two headers are unnecessary
777 if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );
778
779 // Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
780 // administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
781 // .-------------v-----------------v----------------v----------,
782 // | Real Header | ... padding ... | Fake Header | data ... |
783 // `-------------^-----------------^-+--------------^----------'
784 // |<--------------------------------' offset/align |<-- alignment boundary
785
786 // subtract libAlign() because it is already the minimum alignment
787 // add sizeof(Storage) for fake header
788 char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(HeapManager.Storage) );
789
790 // address in the block of the "next" alignment address
791 char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(HeapManager.Storage)), alignment );
792
793 // address of header from malloc
794 HeapManager.Storage.Header * realHeader = headerAddr( addr );
795 realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
796 // address of fake header * before* the alignment location
797 HeapManager.Storage.Header * fakeHeader = headerAddr( user );
798 // SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
799 fakeHeader->kind.fake.offset = (char *)fakeHeader - (char *)realHeader;
800 // SKULLDUGGERY: odd alignment imples fake header
801 fakeHeader->kind.fake.alignment = alignment | 1;
802
803 return user;
804} // memalignNoStats
805
806
807static inline void * cmemalignNoStats( size_t alignment, size_t dim, size_t elemSize ) {
808 size_t size = dim * elemSize;
809 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
810 char * addr = (char *)memalignNoStats( alignment, size );
811
812 HeapManager.Storage.Header * header;
813 HeapManager.FreeHeader * freeElem;
814 size_t bsize;
815 #ifndef __CFA_DEBUG__
816 bool mapped =
817 #endif // __CFA_DEBUG__
818 headers( "cmemalign", addr, header, freeElem, bsize, alignment );
819
820 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
821 #ifndef __CFA_DEBUG__
822 if ( ! mapped )
823 #endif // __CFA_DEBUG__
824 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
825 // `-header`-addr `-size
826 memset( addr, '\0', size ); // set to zeros
827
828 header->kind.real.blockSize |= 2; // mark as zero filled
829 return addr;
830} // cmemalignNoStats
831
832
833extern "C" {
834 // Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
835 // then malloc() returns a unique pointer value that can later be successfully passed to free().
836 void * malloc( size_t size ) {
837 #ifdef __STATISTICS__
838 __atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
839 __atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
840 #endif // __STATISTICS__
841
842 return mallocNoStats( size );
843 } // malloc
844
845
846 // Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
847 void * aalloc( size_t dim, size_t elemSize ) {
848 size_t size = dim * elemSize;
849 #ifdef __STATISTICS__
850 __atomic_add_fetch( &aalloc_calls, 1, __ATOMIC_SEQ_CST );
851 __atomic_add_fetch( &aalloc_storage, size, __ATOMIC_SEQ_CST );
852 #endif // __STATISTICS__
853
854 return mallocNoStats( size );
855 } // aalloc
856
857
858 // Same as aalloc() with memory set to zero.
859 void * calloc( size_t dim, size_t elemSize ) {
860 #ifdef __STATISTICS__
861 __atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
862 __atomic_add_fetch( &calloc_storage, dim * elemSize, __ATOMIC_SEQ_CST );
863 #endif // __STATISTICS__
864
865 return callocNoStats( dim, elemSize );
866 } // calloc
867
868
869 // Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
870 // 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
871 // not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
872 // call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
873 void * resize( void * oaddr, size_t size ) {
874 #ifdef __STATISTICS__
875 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
876 #endif // __STATISTICS__
877
878 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
879 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
880 if ( unlikely( oaddr == 0p ) ) {
881 #ifdef __STATISTICS__
882 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
883 #endif // __STATISTICS__
884 return mallocNoStats( size );
885 } // if
886
887 HeapManager.Storage.Header * header;
888 HeapManager.FreeHeader * freeElem;
889 size_t bsize, oalign;
890 headers( "resize", oaddr, header, freeElem, bsize, oalign );
891 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
892
893 // same size, DO NOT preserve STICKY PROPERTIES.
894 if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
895 header->kind.real.blockSize &= -2; // no alignment and turn off 0 fill
896 header->kind.real.size = size; // reset allocation size
897 return oaddr;
898 } // if
899
900 #ifdef __STATISTICS__
901 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
902 #endif // __STATISTICS__
903
904 // change size, DO NOT preserve STICKY PROPERTIES.
905 free( oaddr );
906 return mallocNoStats( size ); // create new area
907 } // resize
908
909
910 // Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
911 // the old and new sizes.
912 void * realloc( void * oaddr, size_t size ) {
913 #ifdef __STATISTICS__
914 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
915 #endif // __STATISTICS__
916
917 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
918 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
919 if ( unlikely( oaddr == 0p ) ) {
920 #ifdef __STATISTICS__
921 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
922 #endif // __STATISTICS__
923 return mallocNoStats( size );
924 } // if
925
926 HeapManager.Storage.Header * header;
927 HeapManager.FreeHeader * freeElem;
928 size_t bsize, oalign;
929 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
930
931 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
932 size_t osize = header->kind.real.size; // old allocation size
933 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
934 if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
935 header->kind.real.size = size; // reset allocation size
936 if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
937 memset( (char *)oaddr + osize, (int)'\0', size - osize ); // initialize added storage
938 } // if
939 return oaddr;
940 } // if
941
942 #ifdef __STATISTICS__
943 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
944 #endif // __STATISTICS__
945
946 // change size and copy old content to new storage
947
948 void * naddr;
949 if ( likely( oalign == libAlign() ) ) { // previous request not aligned ?
950 naddr = mallocNoStats( size ); // create new area
951 } else {
952 naddr = memalignNoStats( oalign, size ); // create new aligned area
953 } // if
954
955 headers( "realloc", naddr, header, freeElem, bsize, oalign );
956 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
957 free( oaddr );
958
959 if ( unlikely( ozfill ) ) { // previous request zero fill ?
960 header->kind.real.blockSize |= 2; // mark new request as zero filled
961 if ( size > osize ) { // previous request larger ?
962 memset( (char *)naddr + osize, (int)'\0', size - osize ); // initialize added storage
963 } // if
964 } // if
965 return naddr;
966 } // realloc
967
968
969 // Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
970 void * memalign( size_t alignment, size_t size ) {
971 #ifdef __STATISTICS__
972 __atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
973 __atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
974 #endif // __STATISTICS__
975
976 return memalignNoStats( alignment, size );
977 } // memalign
978
979
980 // Same as aalloc() with memory alignment.
981 void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
982 size_t size = dim * elemSize;
983 #ifdef __STATISTICS__
984 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
985 __atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
986 #endif // __STATISTICS__
987
988 return memalignNoStats( alignment, size );
989 } // amemalign
990
991
992 // Same as calloc() with memory alignment.
993 void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
994 #ifdef __STATISTICS__
995 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
996 __atomic_add_fetch( &cmemalign_storage, dim * elemSize, __ATOMIC_SEQ_CST );
997 #endif // __STATISTICS__
998
999 return cmemalignNoStats( alignment, dim, elemSize );
1000 } // cmemalign
1001
1002 // Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1003 // of alignment. This requirement is universally ignored.
1004 void * aligned_alloc( size_t alignment, size_t size ) {
1005 return memalign( alignment, size );
1006 } // aligned_alloc
1007
1008
1009 // Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1010 // memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1011 // is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1012 // free(3).
1013 int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
1014 if ( alignment < libAlign() || ! is_pow2( alignment ) ) return EINVAL; // check alignment
1015 * memptr = memalign( alignment, size );
1016 return 0;
1017 } // posix_memalign
1018
1019 // Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1020 // page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1021 void * valloc( size_t size ) {
1022 return memalign( pageSize, size );
1023 } // valloc
1024
1025
1026 // Same as valloc but rounds size to multiple of page size.
1027 void * pvalloc( size_t size ) {
1028 return memalign( pageSize, ceiling2( size, pageSize ) );
1029 } // pvalloc
1030
1031
1032 // Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1033 // or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1034 // 0p, no operation is performed.
1035 void free( void * addr ) {
1036 #ifdef __STATISTICS__
1037 __atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
1038 #endif // __STATISTICS__
1039
1040 if ( unlikely( addr == 0p ) ) { // special case
1041 // #ifdef __CFA_DEBUG__
1042 // if ( traceHeap() ) {
1043 // #define nullmsg "Free( 0x0 ) size:0\n"
1044 // // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
1045 // __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
1046 // } // if
1047 // #endif // __CFA_DEBUG__
1048 return;
1049 } // exit
1050
1051 doFree( addr );
1052 } // free
1053
1054
1055 // Returns the alignment of an allocation.
1056 size_t malloc_alignment( void * addr ) {
1057 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1058 HeapManager.Storage.Header * header = headerAddr( addr );
1059 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1060 return header->kind.fake.alignment & -2; // remove flag from value
1061 } else {
1062 return libAlign(); // minimum alignment
1063 } // if
1064 } // malloc_alignment
1065
1066
1067 // Set the alignment for an the allocation and return previous alignment or 0 if no alignment.
1068 size_t $malloc_alignment_set( void * addr, size_t alignment ) {
1069 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1070 size_t ret;
1071 HeapManager.Storage.Header * header = headerAddr( addr );
1072 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1073 ret = header->kind.fake.alignment & -2; // remove flag from old value
1074 header->kind.fake.alignment = alignment | 1; // add flag to new value
1075 } else {
1076 ret = 0; // => no alignment to change
1077 } // if
1078 return ret;
1079 } // $malloc_alignment_set
1080
1081
1082 // Returns true if the allocation is zero filled, e.g., allocated by calloc().
1083 bool malloc_zero_fill( void * addr ) {
1084 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1085 HeapManager.Storage.Header * header = headerAddr( addr );
1086 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1087 header = realHeader( header ); // backup from fake to real header
1088 } // if
1089 return (header->kind.real.blockSize & 2) != 0; // zero filled ?
1090 } // malloc_zero_fill
1091
1092 // Set allocation is zero filled and return previous zero filled.
1093 bool $malloc_zero_fill_set( void * addr ) {
1094 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1095 HeapManager.Storage.Header * header = headerAddr( addr );
1096 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1097 header = realHeader( header ); // backup from fake to real header
1098 } // if
1099 bool ret = (header->kind.real.blockSize & 2) != 0; // zero filled ?
1100 header->kind.real.blockSize |= 2; // mark as zero filled
1101 return ret;
1102 } // $malloc_zero_fill_set
1103
1104
1105 // Returns original total allocation size (not bucket size) => array size is dimension * sizeif(T).
1106 size_t malloc_size( void * addr ) {
1107 if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1108 HeapManager.Storage.Header * header = headerAddr( addr );
1109 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1110 header = realHeader( header ); // backup from fake to real header
1111 } // if
1112 return header->kind.real.size;
1113 } // malloc_size
1114
1115 // Set allocation size and return previous size.
1116 size_t $malloc_size_set( void * addr, size_t size ) {
1117 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1118 HeapManager.Storage.Header * header = headerAddr( addr );
1119 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1120 header = realHeader( header ); // backup from fake to real header
1121 } // if
1122 size_t ret = header->kind.real.size;
1123 header->kind.real.size = size;
1124 return ret;
1125 } // $malloc_size_set
1126
1127
1128 // Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1129 // malloc or a related function.
1130 size_t malloc_usable_size( void * addr ) {
1131 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1132 HeapManager.Storage.Header * header;
1133 HeapManager.FreeHeader * freeElem;
1134 size_t bsize, alignment;
1135
1136 headers( "malloc_usable_size", addr, header, freeElem, bsize, alignment );
1137 return dataStorage( bsize, addr, header ); // data storage in bucket
1138 } // malloc_usable_size
1139
1140
1141 // Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1142 void malloc_stats( void ) {
1143 #ifdef __STATISTICS__
1144 printStats();
1145 if ( prtFree() ) prtFree( heapManager );
1146 #endif // __STATISTICS__
1147 } // malloc_stats
1148
1149
1150 // Changes the file descripter where malloc_stats() writes statistics.
1151 int malloc_stats_fd( int fd __attribute__(( unused )) ) {
1152 #ifdef __STATISTICS__
1153 int temp = stat_fd;
1154 stat_fd = fd;
1155 return temp;
1156 #else
1157 return -1;
1158 #endif // __STATISTICS__
1159 } // malloc_stats_fd
1160
1161
1162 // Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1163 // specifies the parameter to be modified, and value specifies the new value for that parameter.
1164 int mallopt( int option, int value ) {
1165 choose( option ) {
1166 case M_TOP_PAD:
1167 heapExpand = ceiling2( value, pageSize ); return 1;
1168 case M_MMAP_THRESHOLD:
1169 if ( setMmapStart( value ) ) return 1;
1170 break;
1171 } // switch
1172 return 0; // error, unsupported
1173 } // mallopt
1174
1175
1176 // Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1177 int malloc_trim( size_t ) {
1178 return 0; // => impossible to release memory
1179 } // malloc_trim
1180
1181
1182 // Exports an XML string that describes the current state of the memory-allocation implementation in the caller.
1183 // The string is printed on the file stream stream. The exported string includes information about all arenas (see
1184 // malloc).
1185 int malloc_info( int options, FILE * stream ) {
1186 if ( options != 0 ) { errno = EINVAL; return -1; }
1187 #ifdef __STATISTICS__
1188 return printStatsXML( stream );
1189 #else
1190 return 0; // unsupported
1191 #endif // __STATISTICS__
1192 } // malloc_info
1193
1194
1195 // Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1196 // or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1197 // structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1198 // result. (The caller must free this memory.)
1199 void * malloc_get_state( void ) {
1200 return 0p; // unsupported
1201 } // malloc_get_state
1202
1203
1204 // Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1205 // structure pointed to by state.
1206 int malloc_set_state( void * ) {
1207 return 0; // unsupported
1208 } // malloc_set_state
1209} // extern "C"
1210
1211
1212// Must have CFA linkage to overload with C linkage realloc.
1213void * resize( void * oaddr, size_t nalign, size_t size ) {
1214 #ifdef __STATISTICS__
1215 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
1216 #endif // __STATISTICS__
1217
1218 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1219 #ifdef __CFA_DEBUG__
1220 else
1221 checkAlign( nalign ); // check alignment
1222 #endif // __CFA_DEBUG__
1223
1224 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1225 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
1226 if ( unlikely( oaddr == 0p ) ) {
1227 #ifdef __STATISTICS__
1228 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1229 #endif // __STATISTICS__
1230 return memalignNoStats( nalign, size );
1231 } // if
1232
1233 // Attempt to reuse existing alignment.
1234 HeapManager.Storage.Header * header = headerAddr( oaddr );
1235 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1236 size_t oalign;
1237 if ( isFakeHeader ) {
1238 oalign = header->kind.fake.alignment & -2; // old alignment
1239 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1240 && ( oalign <= nalign // going down
1241 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1242 ) {
1243 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1244 HeapManager.FreeHeader * freeElem;
1245 size_t bsize, oalign;
1246 headers( "resize", oaddr, header, freeElem, bsize, oalign );
1247 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1248
1249 if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1250 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1251
1252 header->kind.real.blockSize &= -2; // turn off 0 fill
1253 header->kind.real.size = size; // reset allocation size
1254 return oaddr;
1255 } // if
1256 } // if
1257 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1258 && nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1259 return resize( oaddr, size ); // duplicate special case checks
1260 } // if
1261
1262 #ifdef __STATISTICS__
1263 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1264 #endif // __STATISTICS__
1265
1266 // change size, DO NOT preserve STICKY PROPERTIES.
1267 free( oaddr );
1268 return memalignNoStats( nalign, size ); // create new aligned area
1269} // resize
1270
1271
1272void * realloc( void * oaddr, size_t nalign, size_t size ) {
1273 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1274 #ifdef __CFA_DEBUG__
1275 else
1276 checkAlign( nalign ); // check alignment
1277 #endif // __CFA_DEBUG__
1278
1279 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1280 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
1281 if ( unlikely( oaddr == 0p ) ) {
1282 #ifdef __STATISTICS__
1283 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1284 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1285 #endif // __STATISTICS__
1286 return memalignNoStats( nalign, size );
1287 } // if
1288
1289 // Attempt to reuse existing alignment.
1290 HeapManager.Storage.Header * header = headerAddr( oaddr );
1291 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1292 size_t oalign;
1293 if ( isFakeHeader ) {
1294 oalign = header->kind.fake.alignment & -2; // old alignment
1295 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1296 && ( oalign <= nalign // going down
1297 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1298 ) {
1299 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1300 return realloc( oaddr, size ); // duplicate alignment and special case checks
1301 } // if
1302 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1303 && nalign == libAlign() ) // new alignment also on libAlign => no fake header needed
1304 return realloc( oaddr, size ); // duplicate alignment and special case checks
1305
1306 #ifdef __STATISTICS__
1307 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1308 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1309 #endif // __STATISTICS__
1310
1311 HeapManager.FreeHeader * freeElem;
1312 size_t bsize;
1313 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1314
1315 // change size and copy old content to new storage
1316
1317 size_t osize = header->kind.real.size; // old allocation size
1318 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1319
1320 void * naddr = memalignNoStats( nalign, size ); // create new aligned area
1321
1322 headers( "realloc", naddr, header, freeElem, bsize, oalign );
1323 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1324 free( oaddr );
1325
1326 if ( unlikely( ozfill ) ) { // previous request zero fill ?
1327 header->kind.real.blockSize |= 2; // mark new request as zero filled
1328 if ( size > osize ) { // previous request larger ?
1329 memset( (char *)naddr + osize, (int)'\0', size - osize ); // initialize added storage
1330 } // if
1331 } // if
1332 return naddr;
1333} // realloc
1334
1335
1336// Local Variables: //
1337// tab-width: 4 //
1338// compile-command: "cfa -nodebug -O2 heap.cfa" //
1339// End: //
Note: See TracBrowser for help on using the repository browser.