source: libcfa/src/heap.cfa@ aa1d13c

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 aa1d13c was b4aa1ab, checked in by Peter A. Buhr <pabuhr@…>, 5 years ago

fix running nested routines on stacks in the heap
  • Property mode set to 100644
File size: 52.5 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 : Sun Dec 13 22:04:10 2020
13// Update Count : 984
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#ifdef __CFA_DEBUG__
467#if __SIZEOF_POINTER__ == 4
468#define MASK 0xdeadbeef
469#else
470#define MASK 0xdeadbeefdeadbeef
471#endif
472#define STRIDE size_t
473
474static void * Memset( void * addr, STRIDE size ) { // debug only
475 if ( size % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, size %zd not multiple of %zd.", size, sizeof(STRIDE) );
476 if ( (STRIDE)addr % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, addr %p not multiple of %zd.", addr, sizeof(STRIDE) );
477
478 STRIDE * end = (STRIDE *)addr + size / sizeof(STRIDE);
479 for ( STRIDE * p = (STRIDE *)addr; p < end; p += 1 ) *p = MASK;
480 return addr;
481} // Memset
482#endif // __CFA_DEBUG__
483
484#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
485
486static inline void * extend( size_t size ) with( heapManager ) {
487 lock( extlock __cfaabi_dbg_ctx2 );
488 ptrdiff_t rem = heapRemaining - size;
489 if ( rem < 0 ) {
490 // If the size requested is bigger than the current remaining storage, increase the size of the heap.
491
492 size_t increase = ceiling2( size > heapExpand ? size : heapExpand, pageSize );
493 if ( sbrk( increase ) == (void *)-1 ) { // failed, no memory ?
494 unlock( extlock );
495 abort( NO_MEMORY_MSG, size ); // give up
496 } // if
497 if ( mprotect( (char *)heapEnd + heapRemaining, increase, PROT_READ | PROT_WRITE | PROT_EXEC ) ) {
498 enum { BufferSize = 128 };
499 char helpText[BufferSize];
500 // Do not call strerror( errno ) as it may call malloc.
501 int len = snprintf( helpText, BufferSize, "internal error, extend(), mprotect failure, heapEnd:%p size:%zd, errno:%d.", heapEnd, increase, errno );
502 __cfaabi_bits_write( STDERR_FILENO, helpText, len );
503 } // if
504 #ifdef __STATISTICS__
505 sbrk_calls += 1;
506 sbrk_storage += increase;
507 #endif // __STATISTICS__
508 #ifdef __CFA_DEBUG__
509 // Set new memory to garbage so subsequent uninitialized usages might fail.
510 //memset( (char *)heapEnd + heapRemaining, '\377', increase );
511 Memset( (char *)heapEnd + heapRemaining, increase );
512 #endif // __CFA_DEBUG__
513 rem = heapRemaining + increase - size;
514 } // if
515
516 HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
517 heapRemaining = rem;
518 heapEnd = (char *)heapEnd + size;
519 unlock( extlock );
520 return block;
521} // extend
522
523
524static inline void * doMalloc( size_t size ) with( heapManager ) {
525 HeapManager.Storage * block; // pointer to new block of storage
526
527 // Look up size in the size list. Make sure the user request includes space for the header that must be allocated
528 // along with the block and is a multiple of the alignment size.
529
530 if ( unlikely( size > ULONG_MAX - sizeof(HeapManager.Storage) ) ) return 0p;
531 size_t tsize = size + sizeof(HeapManager.Storage);
532 if ( likely( tsize < mmapStart ) ) { // small size => sbrk
533 size_t posn;
534 #ifdef FASTLOOKUP
535 if ( tsize < LookupSizes ) posn = lookup[tsize];
536 else
537 #endif // FASTLOOKUP
538 posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
539 HeapManager.FreeHeader * freeElem = &freeLists[posn];
540 verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
541 verify( tsize <= freeElem->blockSize ); // search failure ?
542 tsize = freeElem->blockSize; // total space needed for request
543
544 // Spin until the lock is acquired for this particular size of block.
545
546 #if BUCKETLOCK == SPINLOCK
547 lock( freeElem->lock __cfaabi_dbg_ctx2 );
548 block = freeElem->freeList; // remove node from stack
549 #else
550 block = pop( freeElem->freeList );
551 #endif // BUCKETLOCK
552 if ( unlikely( block == 0p ) ) { // no free block ?
553 #if BUCKETLOCK == SPINLOCK
554 unlock( freeElem->lock );
555 #endif // BUCKETLOCK
556
557 // Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
558 // and then carve it off.
559
560 block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
561 #if BUCKETLOCK == SPINLOCK
562 } else {
563 freeElem->freeList = block->header.kind.real.next;
564 unlock( freeElem->lock );
565 #endif // BUCKETLOCK
566 } // if
567
568 block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
569 } else { // large size => mmap
570 if ( unlikely( size > ULONG_MAX - pageSize ) ) return 0p;
571 tsize = ceiling2( tsize, pageSize ); // must be multiple of page size
572 #ifdef __STATISTICS__
573 __atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
574 __atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
575 #endif // __STATISTICS__
576
577 block = (HeapManager.Storage *)mmap( 0, tsize, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, mmapFd, 0 );
578 if ( block == (HeapManager.Storage *)MAP_FAILED ) { // failed ?
579 if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
580 // Do not call strerror( errno ) as it may call malloc.
581 abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu errno:%d.", &heapManager, tsize, errno );
582 } //if
583 #ifdef __CFA_DEBUG__
584 // Set new memory to garbage so subsequent uninitialized usages might fail.
585 //memset( block, '\377', tsize );
586 Memset( block, tsize );
587 #endif // __CFA_DEBUG__
588 block->header.kind.real.blockSize = tsize; // storage size for munmap
589 } // if
590
591 block->header.kind.real.size = size; // store allocation size
592 void * addr = &(block->data); // adjust off header to user bytes
593 verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
594
595 #ifdef __CFA_DEBUG__
596 __atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
597 if ( traceHeap() ) {
598 enum { BufferSize = 64 };
599 char helpText[BufferSize];
600 int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize );
601 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
602 } // if
603 #endif // __CFA_DEBUG__
604
605 return addr;
606} // doMalloc
607
608
609static inline void doFree( void * addr ) with( heapManager ) {
610 #ifdef __CFA_DEBUG__
611 if ( unlikely( heapManager.heapBegin == 0p ) ) {
612 abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
613 } // if
614 #endif // __CFA_DEBUG__
615
616 HeapManager.Storage.Header * header;
617 HeapManager.FreeHeader * freeElem;
618 size_t size, alignment; // not used (see realloc)
619
620 if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
621 #ifdef __STATISTICS__
622 __atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
623 __atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
624 #endif // __STATISTICS__
625 if ( munmap( header, size ) == -1 ) {
626 #ifdef __CFA_DEBUG__
627 abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
628 "Possible cause is invalid pointer.",
629 addr );
630 #endif // __CFA_DEBUG__
631 } // if
632 } else {
633 #ifdef __CFA_DEBUG__
634 // Set free memory to garbage so subsequent usages might fail.
635 //memset( ((HeapManager.Storage *)header)->data, '\377', freeElem->blockSize - sizeof( HeapManager.Storage ) );
636 Memset( ((HeapManager.Storage *)header)->data, freeElem->blockSize - sizeof( HeapManager.Storage ) );
637 #endif // __CFA_DEBUG__
638
639 #ifdef __STATISTICS__
640 free_storage += size;
641 #endif // __STATISTICS__
642 #if BUCKETLOCK == SPINLOCK
643 lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
644 header->kind.real.next = freeElem->freeList; // push on stack
645 freeElem->freeList = (HeapManager.Storage *)header;
646 unlock( freeElem->lock ); // release spin lock
647 #else
648 push( freeElem->freeList, *(HeapManager.Storage *)header );
649 #endif // BUCKETLOCK
650 } // if
651
652 #ifdef __CFA_DEBUG__
653 __atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
654 if ( traceHeap() ) {
655 char helpText[64];
656 int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
657 __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
658 } // if
659 #endif // __CFA_DEBUG__
660} // doFree
661
662
663size_t prtFree( HeapManager & manager ) with( manager ) {
664 size_t total = 0;
665 #ifdef __STATISTICS__
666 __cfaabi_bits_acquire();
667 __cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
668 #endif // __STATISTICS__
669 for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
670 size_t size = freeLists[i].blockSize;
671 #ifdef __STATISTICS__
672 unsigned int N = 0;
673 #endif // __STATISTICS__
674
675 #if BUCKETLOCK == SPINLOCK
676 for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
677 #else
678 for(;;) {
679// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
680// for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
681// HeapManager.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
682// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
683// p = temp;
684 #endif // BUCKETLOCK
685 total += size;
686 #ifdef __STATISTICS__
687 N += 1;
688 #endif // __STATISTICS__
689 } // for
690
691 #ifdef __STATISTICS__
692 __cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
693 if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
694 #endif // __STATISTICS__
695 } // for
696 #ifdef __STATISTICS__
697 __cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
698 __cfaabi_bits_release();
699 #endif // __STATISTICS__
700 return (char *)heapEnd - (char *)heapBegin - total;
701} // prtFree
702
703
704static void ?{}( HeapManager & manager ) with( manager ) {
705 pageSize = sysconf( _SC_PAGESIZE );
706
707 for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
708 freeLists[i].blockSize = bucketSizes[i];
709 } // for
710
711 #ifdef FASTLOOKUP
712 unsigned int idx = 0;
713 for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
714 if ( i > bucketSizes[idx] ) idx += 1;
715 lookup[i] = idx;
716 } // for
717 #endif // FASTLOOKUP
718
719 if ( ! setMmapStart( default_mmap_start() ) ) {
720 abort( "HeapManager : internal error, mmap start initialization failure." );
721 } // if
722 heapExpand = default_heap_expansion();
723
724 char * end = (char *)sbrk( 0 );
725 heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, pageSize ) - end ); // move start of heap to multiple of alignment
726} // HeapManager
727
728
729static void ^?{}( HeapManager & ) {
730 #ifdef __STATISTICS__
731 if ( traceHeapTerm() ) {
732 printStats();
733 // prtUnfreed() called in heapAppStop()
734 } // if
735 #endif // __STATISTICS__
736} // ~HeapManager
737
738
739static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
740void memory_startup( void ) {
741 #ifdef __CFA_DEBUG__
742 if ( heapBoot ) { // check for recursion during system boot
743 // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
744 abort( "boot() : internal error, recursively invoked during system boot." );
745 } // if
746 heapBoot = true;
747 #endif // __CFA_DEBUG__
748
749 //verify( heapManager.heapBegin != 0 );
750 //heapManager{};
751 if ( heapManager.heapBegin == 0p ) heapManager{}; // sanity check
752} // memory_startup
753
754static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
755void memory_shutdown( void ) {
756 ^heapManager{};
757} // memory_shutdown
758
759
760static inline void * mallocNoStats( size_t size ) { // necessary for malloc statistics
761 verify( heapManager.heapBegin != 0p ); // called before memory_startup ?
762 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
763
764#if __SIZEOF_POINTER__ == 8
765 verify( size < ((typeof(size_t))1 << 48) );
766#endif // __SIZEOF_POINTER__ == 8
767 return doMalloc( size );
768} // mallocNoStats
769
770
771static inline void * callocNoStats( size_t dim, size_t elemSize ) {
772 size_t size = dim * elemSize;
773 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
774 char * addr = (char *)mallocNoStats( size );
775
776 HeapManager.Storage.Header * header;
777 HeapManager.FreeHeader * freeElem;
778 size_t bsize, alignment;
779 #ifndef __CFA_DEBUG__
780 bool mapped =
781 #endif // __CFA_DEBUG__
782 headers( "calloc", addr, header, freeElem, bsize, alignment );
783 #ifndef __CFA_DEBUG__
784
785 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
786 if ( ! mapped )
787 #endif // __CFA_DEBUG__
788 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
789 // `-header`-addr `-size
790 memset( addr, '\0', size ); // set to zeros
791
792 header->kind.real.blockSize |= 2; // mark as zero filled
793 return addr;
794} // callocNoStats
795
796
797static inline void * memalignNoStats( size_t alignment, size_t size ) {
798 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
799
800 #ifdef __CFA_DEBUG__
801 checkAlign( alignment ); // check alignment
802 #endif // __CFA_DEBUG__
803
804 // if alignment <= default alignment, do normal malloc as two headers are unnecessary
805 if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );
806
807 // Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
808 // administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
809 // .-------------v-----------------v----------------v----------,
810 // | Real Header | ... padding ... | Fake Header | data ... |
811 // `-------------^-----------------^-+--------------^----------'
812 // |<--------------------------------' offset/align |<-- alignment boundary
813
814 // subtract libAlign() because it is already the minimum alignment
815 // add sizeof(Storage) for fake header
816 char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(HeapManager.Storage) );
817
818 // address in the block of the "next" alignment address
819 char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(HeapManager.Storage)), alignment );
820
821 // address of header from malloc
822 HeapManager.Storage.Header * realHeader = headerAddr( addr );
823 realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
824 // address of fake header * before* the alignment location
825 HeapManager.Storage.Header * fakeHeader = headerAddr( user );
826 // SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
827 fakeHeader->kind.fake.offset = (char *)fakeHeader - (char *)realHeader;
828 // SKULLDUGGERY: odd alignment imples fake header
829 fakeHeader->kind.fake.alignment = alignment | 1;
830
831 return user;
832} // memalignNoStats
833
834
835static inline void * cmemalignNoStats( size_t alignment, size_t dim, size_t elemSize ) {
836 size_t size = dim * elemSize;
837 if ( unlikely( size ) == 0 ) return 0p; // 0 BYTE ALLOCATION RETURNS NULL POINTER
838 char * addr = (char *)memalignNoStats( alignment, size );
839
840 HeapManager.Storage.Header * header;
841 HeapManager.FreeHeader * freeElem;
842 size_t bsize;
843 #ifndef __CFA_DEBUG__
844 bool mapped =
845 #endif // __CFA_DEBUG__
846 headers( "cmemalign", addr, header, freeElem, bsize, alignment );
847
848 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
849 #ifndef __CFA_DEBUG__
850 if ( ! mapped )
851 #endif // __CFA_DEBUG__
852 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
853 // `-header`-addr `-size
854 memset( addr, '\0', size ); // set to zeros
855
856 header->kind.real.blockSize |= 2; // mark as zero filled
857 return addr;
858} // cmemalignNoStats
859
860
861extern "C" {
862 // Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
863 // then malloc() returns a unique pointer value that can later be successfully passed to free().
864 void * malloc( size_t size ) {
865 #ifdef __STATISTICS__
866 __atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
867 __atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
868 #endif // __STATISTICS__
869
870 return mallocNoStats( size );
871 } // malloc
872
873
874 // Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
875 void * aalloc( size_t dim, size_t elemSize ) {
876 size_t size = dim * elemSize;
877 #ifdef __STATISTICS__
878 __atomic_add_fetch( &aalloc_calls, 1, __ATOMIC_SEQ_CST );
879 __atomic_add_fetch( &aalloc_storage, size, __ATOMIC_SEQ_CST );
880 #endif // __STATISTICS__
881
882 return mallocNoStats( size );
883 } // aalloc
884
885
886 // Same as aalloc() with memory set to zero.
887 void * calloc( size_t dim, size_t elemSize ) {
888 #ifdef __STATISTICS__
889 __atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
890 __atomic_add_fetch( &calloc_storage, dim * elemSize, __ATOMIC_SEQ_CST );
891 #endif // __STATISTICS__
892
893 return callocNoStats( dim, elemSize );
894 } // calloc
895
896
897 // Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
898 // 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
899 // not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
900 // call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
901 void * resize( void * oaddr, size_t size ) {
902 #ifdef __STATISTICS__
903 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
904 #endif // __STATISTICS__
905
906 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
907 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
908 if ( unlikely( oaddr == 0p ) ) {
909 #ifdef __STATISTICS__
910 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
911 #endif // __STATISTICS__
912 return mallocNoStats( size );
913 } // if
914
915 HeapManager.Storage.Header * header;
916 HeapManager.FreeHeader * freeElem;
917 size_t bsize, oalign;
918 headers( "resize", oaddr, header, freeElem, bsize, oalign );
919 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
920
921 // same size, DO NOT preserve STICKY PROPERTIES.
922 if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
923 header->kind.real.blockSize &= -2; // no alignment and turn off 0 fill
924 header->kind.real.size = size; // reset allocation size
925 return oaddr;
926 } // if
927
928 #ifdef __STATISTICS__
929 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
930 #endif // __STATISTICS__
931
932 // change size, DO NOT preserve STICKY PROPERTIES.
933 free( oaddr );
934 return mallocNoStats( size ); // create new area
935 } // resize
936
937
938 // Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
939 // the old and new sizes.
940 void * realloc( void * oaddr, size_t size ) {
941 #ifdef __STATISTICS__
942 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
943 #endif // __STATISTICS__
944
945 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
946 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
947 if ( unlikely( oaddr == 0p ) ) {
948 #ifdef __STATISTICS__
949 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
950 #endif // __STATISTICS__
951 return mallocNoStats( size );
952 } // if
953
954 HeapManager.Storage.Header * header;
955 HeapManager.FreeHeader * freeElem;
956 size_t bsize, oalign;
957 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
958
959 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
960 size_t osize = header->kind.real.size; // old allocation size
961 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
962 if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
963 header->kind.real.size = size; // reset allocation size
964 if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
965 memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
966 } // if
967 return oaddr;
968 } // if
969
970 #ifdef __STATISTICS__
971 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
972 #endif // __STATISTICS__
973
974 // change size and copy old content to new storage
975
976 void * naddr;
977 if ( likely( oalign == libAlign() ) ) { // previous request not aligned ?
978 naddr = mallocNoStats( size ); // create new area
979 } else {
980 naddr = memalignNoStats( oalign, size ); // create new aligned area
981 } // if
982
983 headers( "realloc", naddr, header, freeElem, bsize, oalign );
984 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
985 free( oaddr );
986
987 if ( unlikely( ozfill ) ) { // previous request zero fill ?
988 header->kind.real.blockSize |= 2; // mark new request as zero filled
989 if ( size > osize ) { // previous request larger ?
990 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
991 } // if
992 } // if
993 return naddr;
994 } // realloc
995
996
997 // Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
998 void * memalign( size_t alignment, size_t size ) {
999 #ifdef __STATISTICS__
1000 __atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
1001 __atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
1002 #endif // __STATISTICS__
1003
1004 return memalignNoStats( alignment, size );
1005 } // memalign
1006
1007
1008 // Same as aalloc() with memory alignment.
1009 void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
1010 size_t size = dim * elemSize;
1011 #ifdef __STATISTICS__
1012 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1013 __atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
1014 #endif // __STATISTICS__
1015
1016 return memalignNoStats( alignment, size );
1017 } // amemalign
1018
1019
1020 // Same as calloc() with memory alignment.
1021 void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
1022 #ifdef __STATISTICS__
1023 __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
1024 __atomic_add_fetch( &cmemalign_storage, dim * elemSize, __ATOMIC_SEQ_CST );
1025 #endif // __STATISTICS__
1026
1027 return cmemalignNoStats( alignment, dim, elemSize );
1028 } // cmemalign
1029
1030 // Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1031 // of alignment. This requirement is universally ignored.
1032 void * aligned_alloc( size_t alignment, size_t size ) {
1033 return memalign( alignment, size );
1034 } // aligned_alloc
1035
1036
1037 // Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1038 // memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1039 // is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1040 // free(3).
1041 int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
1042 if ( alignment < libAlign() || ! is_pow2( alignment ) ) return EINVAL; // check alignment
1043 * memptr = memalign( alignment, size );
1044 return 0;
1045 } // posix_memalign
1046
1047 // Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1048 // page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1049 void * valloc( size_t size ) {
1050 return memalign( pageSize, size );
1051 } // valloc
1052
1053
1054 // Same as valloc but rounds size to multiple of page size.
1055 void * pvalloc( size_t size ) {
1056 return memalign( pageSize, ceiling2( size, pageSize ) );
1057 } // pvalloc
1058
1059
1060 // Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1061 // or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1062 // 0p, no operation is performed.
1063 void free( void * addr ) {
1064 #ifdef __STATISTICS__
1065 __atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
1066 #endif // __STATISTICS__
1067
1068 if ( unlikely( addr == 0p ) ) { // special case
1069 // #ifdef __CFA_DEBUG__
1070 // if ( traceHeap() ) {
1071 // #define nullmsg "Free( 0x0 ) size:0\n"
1072 // // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
1073 // __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
1074 // } // if
1075 // #endif // __CFA_DEBUG__
1076 return;
1077 } // exit
1078
1079 doFree( addr );
1080 } // free
1081
1082
1083 // Returns the alignment of an allocation.
1084 size_t malloc_alignment( void * addr ) {
1085 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1086 HeapManager.Storage.Header * header = headerAddr( addr );
1087 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1088 return header->kind.fake.alignment & -2; // remove flag from value
1089 } else {
1090 return libAlign(); // minimum alignment
1091 } // if
1092 } // malloc_alignment
1093
1094
1095 // Set the alignment for an the allocation and return previous alignment or 0 if no alignment.
1096 size_t $malloc_alignment_set( void * addr, size_t alignment ) {
1097 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1098 size_t ret;
1099 HeapManager.Storage.Header * header = headerAddr( addr );
1100 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1101 ret = header->kind.fake.alignment & -2; // remove flag from old value
1102 header->kind.fake.alignment = alignment | 1; // add flag to new value
1103 } else {
1104 ret = 0; // => no alignment to change
1105 } // if
1106 return ret;
1107 } // $malloc_alignment_set
1108
1109
1110 // Returns true if the allocation is zero filled, e.g., allocated by calloc().
1111 bool malloc_zero_fill( void * addr ) {
1112 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1113 HeapManager.Storage.Header * header = headerAddr( addr );
1114 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1115 header = realHeader( header ); // backup from fake to real header
1116 } // if
1117 return (header->kind.real.blockSize & 2) != 0; // zero filled ?
1118 } // malloc_zero_fill
1119
1120 // Set allocation is zero filled and return previous zero filled.
1121 bool $malloc_zero_fill_set( void * addr ) {
1122 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1123 HeapManager.Storage.Header * header = headerAddr( addr );
1124 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1125 header = realHeader( header ); // backup from fake to real header
1126 } // if
1127 bool ret = (header->kind.real.blockSize & 2) != 0; // zero filled ?
1128 header->kind.real.blockSize |= 2; // mark as zero filled
1129 return ret;
1130 } // $malloc_zero_fill_set
1131
1132
1133 // Returns original total allocation size (not bucket size) => array size is dimension * sizeif(T).
1134 size_t malloc_size( void * addr ) {
1135 if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1136 HeapManager.Storage.Header * header = headerAddr( addr );
1137 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1138 header = realHeader( header ); // backup from fake to real header
1139 } // if
1140 return header->kind.real.size;
1141 } // malloc_size
1142
1143 // Set allocation size and return previous size.
1144 size_t $malloc_size_set( void * addr, size_t size ) {
1145 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1146 HeapManager.Storage.Header * header = headerAddr( addr );
1147 if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
1148 header = realHeader( header ); // backup from fake to real header
1149 } // if
1150 size_t ret = header->kind.real.size;
1151 header->kind.real.size = size;
1152 return ret;
1153 } // $malloc_size_set
1154
1155
1156 // Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1157 // malloc or a related function.
1158 size_t malloc_usable_size( void * addr ) {
1159 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1160 HeapManager.Storage.Header * header;
1161 HeapManager.FreeHeader * freeElem;
1162 size_t bsize, alignment;
1163
1164 headers( "malloc_usable_size", addr, header, freeElem, bsize, alignment );
1165 return dataStorage( bsize, addr, header ); // data storage in bucket
1166 } // malloc_usable_size
1167
1168
1169 // Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1170 void malloc_stats( void ) {
1171 #ifdef __STATISTICS__
1172 printStats();
1173 if ( prtFree() ) prtFree( heapManager );
1174 #endif // __STATISTICS__
1175 } // malloc_stats
1176
1177
1178 // Changes the file descripter where malloc_stats() writes statistics.
1179 int malloc_stats_fd( int fd __attribute__(( unused )) ) {
1180 #ifdef __STATISTICS__
1181 int temp = stat_fd;
1182 stat_fd = fd;
1183 return temp;
1184 #else
1185 return -1;
1186 #endif // __STATISTICS__
1187 } // malloc_stats_fd
1188
1189
1190 // Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1191 // specifies the parameter to be modified, and value specifies the new value for that parameter.
1192 int mallopt( int option, int value ) {
1193 choose( option ) {
1194 case M_TOP_PAD:
1195 heapExpand = ceiling2( value, pageSize ); return 1;
1196 case M_MMAP_THRESHOLD:
1197 if ( setMmapStart( value ) ) return 1;
1198 break;
1199 } // switch
1200 return 0; // error, unsupported
1201 } // mallopt
1202
1203
1204 // Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1205 int malloc_trim( size_t ) {
1206 return 0; // => impossible to release memory
1207 } // malloc_trim
1208
1209
1210 // Exports an XML string that describes the current state of the memory-allocation implementation in the caller.
1211 // The string is printed on the file stream stream. The exported string includes information about all arenas (see
1212 // malloc).
1213 int malloc_info( int options, FILE * stream ) {
1214 if ( options != 0 ) { errno = EINVAL; return -1; }
1215 #ifdef __STATISTICS__
1216 return printStatsXML( stream );
1217 #else
1218 return 0; // unsupported
1219 #endif // __STATISTICS__
1220 } // malloc_info
1221
1222
1223 // Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1224 // or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1225 // structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1226 // result. (The caller must free this memory.)
1227 void * malloc_get_state( void ) {
1228 return 0p; // unsupported
1229 } // malloc_get_state
1230
1231
1232 // Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1233 // structure pointed to by state.
1234 int malloc_set_state( void * ) {
1235 return 0; // unsupported
1236 } // malloc_set_state
1237} // extern "C"
1238
1239
1240// Must have CFA linkage to overload with C linkage realloc.
1241void * resize( void * oaddr, size_t nalign, size_t size ) {
1242 #ifdef __STATISTICS__
1243 __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
1244 #endif // __STATISTICS__
1245
1246 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1247 #ifdef __CFA_DEBUG__
1248 else
1249 checkAlign( nalign ); // check alignment
1250 #endif // __CFA_DEBUG__
1251
1252 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1253 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
1254 if ( unlikely( oaddr == 0p ) ) {
1255 #ifdef __STATISTICS__
1256 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1257 #endif // __STATISTICS__
1258 return memalignNoStats( nalign, size );
1259 } // if
1260
1261 // Attempt to reuse existing alignment.
1262 HeapManager.Storage.Header * header = headerAddr( oaddr );
1263 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1264 size_t oalign;
1265 if ( isFakeHeader ) {
1266 oalign = header->kind.fake.alignment & -2; // old alignment
1267 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1268 && ( oalign <= nalign // going down
1269 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1270 ) {
1271 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1272 HeapManager.FreeHeader * freeElem;
1273 size_t bsize, oalign;
1274 headers( "resize", oaddr, header, freeElem, bsize, oalign );
1275 size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1276
1277 if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1278 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1279
1280 header->kind.real.blockSize &= -2; // turn off 0 fill
1281 header->kind.real.size = size; // reset allocation size
1282 return oaddr;
1283 } // if
1284 } // if
1285 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1286 && nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1287 return resize( oaddr, size ); // duplicate special case checks
1288 } // if
1289
1290 #ifdef __STATISTICS__
1291 __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
1292 #endif // __STATISTICS__
1293
1294 // change size, DO NOT preserve STICKY PROPERTIES.
1295 free( oaddr );
1296 return memalignNoStats( nalign, size ); // create new aligned area
1297} // resize
1298
1299
1300void * realloc( void * oaddr, size_t nalign, size_t size ) {
1301 if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
1302 #ifdef __CFA_DEBUG__
1303 else
1304 checkAlign( nalign ); // check alignment
1305 #endif // __CFA_DEBUG__
1306
1307 // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1308 if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
1309 if ( unlikely( oaddr == 0p ) ) {
1310 #ifdef __STATISTICS__
1311 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1312 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1313 #endif // __STATISTICS__
1314 return memalignNoStats( nalign, size );
1315 } // if
1316
1317 // Attempt to reuse existing alignment.
1318 HeapManager.Storage.Header * header = headerAddr( oaddr );
1319 bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1320 size_t oalign;
1321 if ( isFakeHeader ) {
1322 oalign = header->kind.fake.alignment & -2; // old alignment
1323 if ( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1324 && ( oalign <= nalign // going down
1325 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1326 ) {
1327 headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
1328 return realloc( oaddr, size ); // duplicate alignment and special case checks
1329 } // if
1330 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1331 && nalign == libAlign() ) // new alignment also on libAlign => no fake header needed
1332 return realloc( oaddr, size ); // duplicate alignment and special case checks
1333
1334 #ifdef __STATISTICS__
1335 __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
1336 __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
1337 #endif // __STATISTICS__
1338
1339 HeapManager.FreeHeader * freeElem;
1340 size_t bsize;
1341 headers( "realloc", oaddr, header, freeElem, bsize, oalign );
1342
1343 // change size and copy old content to new storage
1344
1345 size_t osize = header->kind.real.size; // old allocation size
1346 bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1347
1348 void * naddr = memalignNoStats( nalign, size ); // create new aligned area
1349
1350 headers( "realloc", naddr, header, freeElem, bsize, oalign );
1351 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1352 free( oaddr );
1353
1354 if ( unlikely( ozfill ) ) { // previous request zero fill ?
1355 header->kind.real.blockSize |= 2; // mark new request as zero filled
1356 if ( size > osize ) { // previous request larger ?
1357 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1358 } // if
1359 } // if
1360 return naddr;
1361} // realloc
1362
1363
1364// Local Variables: //
1365// tab-width: 4 //
1366// compile-command: "cfa -nodebug -O2 heap.cfa" //
1367// End: //
Note: See TracBrowser for help on using the repository browser.