Context Navigation

heap.cfa @ 3ee8853

Last change on this file since 3ee8853 was a47fe52, checked in by Peter A. Buhr <pabuhr@…>, 15 months ago
missing add of terminating thread-heap statistics to master heap, check for environment variable CFA_MALLOC_STATS and print heap statistics at program termination
Property mode set to `100644`
File size: 67.5 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 : Wed Aug 2 18:48:30 2023
13	// Update Count : 1614
14	//
15
16	#include <stdio.h>
17	#include <string.h> // memset, memcpy
18	#include <limits.h> // ULONG_MAX
19	#include <errno.h> // errno, ENOMEM, EINVAL
20	#include <unistd.h> // STDERR_FILENO, sbrk, sysconf, write
21	#include <sys/mman.h> // mmap, munmap
22	extern "C" {
23	#include <sys/sysinfo.h> // get_nprocs
24	} // extern "C"
25
26	#include "heap.hfa"
27	#include "bits/align.hfa" // libAlign
28	#include "bits/defs.hfa" // likely, unlikely
29	#include "concurrency/kernel/fwd.hfa" // __POLL_PREEMPTION
30	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
31	#include "math.hfa" // ceiling, min
32	#include "bitmanip.hfa" // is_pow2, ceiling2
33
34	// supported mallopt options
35	#ifndef M_MMAP_THRESHOLD
36	#define M_MMAP_THRESHOLD (-1)
37	#endif // M_MMAP_THRESHOLD
38
39	#ifndef M_TOP_PAD
40	#define M_TOP_PAD (-2)
41	#endif // M_TOP_PAD
42
43	#define FASTLOOKUP // use O(1) table lookup from allocation size to bucket size
44	#define OWNERSHIP // return freed memory to owner thread
45	#define RETURNSPIN // toggle spinlock / lockfree queue
46	#if ! defined( OWNERSHIP ) && defined( RETURNSPIN )
47	#warning "RETURNSPIN is ignored without OWNERSHIP; suggest commenting out RETURNSPIN"
48	#endif // ! OWNERSHIP && RETURNSPIN
49
50	#define CACHE_ALIGN 64
51	#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))
52
53	#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
54
55	//#define __STATISTICS__
56
57	enum {
58	// The default extension heap amount in units of bytes. When the current heap reaches the brk address, the brk
59	// address is extended by the extension amount.
60	__CFA_DEFAULT_HEAP_EXPANSION__ = 10 * 1024 * 1024,
61
62	// The mmap crossover point during allocation. Allocations less than this amount are allocated from buckets; values
63	// greater than or equal to this value are mmap from the operating system.
64	__CFA_DEFAULT_MMAP_START__ = 512 * 1024 + 1,
65
66	// The default unfreed storage amount in units of bytes. When the uC++ program ends it subtracts this amount from
67	// the malloc/free counter to adjust for storage the program does not free.
68	__CFA_DEFAULT_HEAP_UNFREED__ = 0
69	}; // enum
70
71
72	//####################### Heap Trace/Print ####################
73
74
75	static bool traceHeap = false;
76
77	inline bool traceHeap() libcfa_public { return traceHeap; }
78
79	bool traceHeapOn() libcfa_public {
80	bool temp = traceHeap;
81	traceHeap = true;
82	return temp;
83	} // traceHeapOn
84
85	bool traceHeapOff() libcfa_public {
86	bool temp = traceHeap;
87	traceHeap = false;
88	return temp;
89	} // traceHeapOff
90
91	bool traceHeapTerm() libcfa_public { return false; }
92
93
94	static bool prtFree = false;
95
96	bool prtFree() {
97	return prtFree;
98	} // prtFree
99
100	bool prtFreeOn() {
101	bool temp = prtFree;
102	prtFree = true;
103	return temp;
104	} // prtFreeOn
105
106	bool prtFreeOff() {
107	bool temp = prtFree;
108	prtFree = false;
109	return temp;
110	} // prtFreeOff
111
112
113	//######################### Helpers #########################
114
115
116	// generic Bsearchl does not inline, so substitute with hand-coded binary-search.
117	inline __attribute__((always_inline))
118	static size_t Bsearchl( unsigned int key, const unsigned int vals[], size_t dim ) {
119	size_t l = 0, m, h = dim;
120	while ( l < h ) {
121	m = (l + h) / 2;
122	if ( (unsigned int &)(vals[m]) < key ) { // cast away const
123	l = m + 1;
124	} else {
125	h = m;
126	} // if
127	} // while
128	return l;
129	} // Bsearchl
130
131
132	// pause to prevent excess processor bus usage
133	#if defined( __i386 ) \|\| defined( __x86_64 )
134	#define Pause() __asm__ __volatile__ ( "pause" : : : )
135	#elif defined(__ARM_ARCH)
136	#define Pause() __asm__ __volatile__ ( "YIELD" : : : )
137	#else
138	#error unsupported architecture
139	#endif
140
141	typedef volatile uintptr_t SpinLock_t;
142
143	static inline __attribute__((always_inline)) void lock( volatile SpinLock_t & slock ) {
144	enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
145	unsigned int spin = SPIN_START;
146
147	for ( unsigned int i = 1;; i += 1 ) {
148	if ( slock == 0 && __atomic_test_and_set( &slock, __ATOMIC_ACQUIRE ) == 0 ) break; // Fence
149	for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
150	spin += spin; // powers of 2
151	//if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
152	if ( spin > SPIN_END ) spin = SPIN_END; // cap spinning
153	} // for
154	} // spin_lock
155
156	static inline __attribute__((always_inline)) void unlock( volatile SpinLock_t & slock ) {
157	__atomic_clear( &slock, __ATOMIC_RELEASE ); // Fence
158	} // spin_unlock
159
160
161	//####################### Heap Statistics ####################
162
163
164	#ifdef __STATISTICS__
165	enum { CntTriples = 12 }; // number of counter triples
166	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC, FREE };
167
168	struct StatsOverlay { // overlay for iteration
169	unsigned int calls, calls_0;
170	unsigned long long int request, alloc;
171	};
172
173	// Heap statistics counters.
174	union HeapStatistics {
175	struct { // minimum qualification
176	unsigned int malloc_calls, malloc_0_calls;
177	unsigned long long int malloc_storage_request, malloc_storage_alloc;
178	unsigned int aalloc_calls, aalloc_0_calls;
179	unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
180	unsigned int calloc_calls, calloc_0_calls;
181	unsigned long long int calloc_storage_request, calloc_storage_alloc;
182	unsigned int memalign_calls, memalign_0_calls;
183	unsigned long long int memalign_storage_request, memalign_storage_alloc;
184	unsigned int amemalign_calls, amemalign_0_calls;
185	unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
186	unsigned int cmemalign_calls, cmemalign_0_calls;
187	unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
188	unsigned int resize_calls, resize_0_calls;
189	unsigned long long int resize_storage_request, resize_storage_alloc;
190	unsigned int realloc_calls, realloc_0_calls;
191	unsigned long long int realloc_storage_request, realloc_storage_alloc;
192	unsigned int free_calls, free_null_calls;
193	unsigned long long int free_storage_request, free_storage_alloc;
194	unsigned int return_pulls, return_pushes;
195	unsigned long long int return_storage_request, return_storage_alloc;
196	unsigned int mmap_calls, mmap_0_calls; // no zero calls
197	unsigned long long int mmap_storage_request, mmap_storage_alloc;
198	unsigned int munmap_calls, munmap_0_calls; // no zero calls
199	unsigned long long int munmap_storage_request, munmap_storage_alloc;
200	};
201	struct StatsOverlay counters[CntTriples]; // overlay for iteration
202	}; // HeapStatistics
203
204	static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(StatsOverlay),
205	"Heap statistics counter-triplets does not match with array size" );
206
207	static void HeapStatisticsCtor( HeapStatistics & stats ) {
208	memset( &stats, '\0', sizeof(stats) ); // very fast
209	// for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
210	// stats.counters[i].calls = stats.counters[i].calls_0 = stats.counters[i].request = stats.counters[i].alloc = 0;
211	// } // for
212	} // HeapStatisticsCtor
213
214	static HeapStatistics & ?+=?( HeapStatistics & lhs, const HeapStatistics & rhs ) {
215	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
216	lhs.counters[i].calls += rhs.counters[i].calls;
217	lhs.counters[i].calls_0 += rhs.counters[i].calls_0;
218	lhs.counters[i].request += rhs.counters[i].request;
219	lhs.counters[i].alloc += rhs.counters[i].alloc;
220	} // for
221	return lhs;
222	} // ?+=?
223	#endif // __STATISTICS__
224
225
226	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
227	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
228	enum { NoBucketSizes = 91 }; // number of buckets sizes
229
230	struct Heap {
231	struct Storage {
232	struct Header { // header
233	union Kind {
234	struct RealHeader {
235	union {
236	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
237	union {
238	// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
239	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
240	void * home; // allocated block points back to home locations (must overlay alignment)
241	size_t blockSize; // size for munmap (must overlay alignment)
242	Storage * next; // freed block points to next freed block of same size
243	};
244	size_t size; // allocation size in bytes
245	};
246	};
247	} real; // RealHeader
248
249	struct FakeHeader {
250	uintptr_t alignment; // 1st low-order bit => fake header & alignment
251	uintptr_t offset;
252	} fake; // FakeHeader
253	} kind; // Kind
254	} header; // Header
255
256	char pad[libAlign() - sizeof( Header )];
257	char data[0]; // storage
258	}; // Storage
259
260	static_assert( libAlign() >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
261
262	struct CALIGN FreeHeader {
263	size_t blockSize CALIGN; // size of allocations on this list
264	#ifdef OWNERSHIP
265	#ifdef RETURNSPIN
266	SpinLock_t returnLock;
267	#endif // RETURNSPIN
268	Storage * returnList; // other thread return list
269	#endif // OWNERSHIP
270
271	Storage * freeList; // thread free list
272	Heap * homeManager; // heap owner (free storage to bucket, from bucket to heap)
273	}; // FreeHeader
274
275	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
276	void * heapBuffer; // start of free storage in buffer
277	size_t heapReserve; // amount of remaining free storage in buffer
278
279	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
280	Heap * nextHeapManager; // intrusive link of existing heaps; traversed to collect statistics or check unfreed storage
281	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
282	Heap * nextFreeHeapManager; // intrusive link of free heaps from terminated threads; reused by new threads
283
284	#ifdef __CFA_DEBUG__
285	ptrdiff_t allocUnfreed; // running total of allocations minus frees; can be negative
286	#endif // __CFA_DEBUG__
287
288	#ifdef __STATISTICS__
289	HeapStatistics stats; // local statistic table for this heap
290	#endif // __STATISTICS__
291	}; // Heap
292
293
294	struct HeapMaster {
295	SpinLock_t extLock; // protects allocation-buffer extension
296	SpinLock_t mgrLock; // protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd
297
298	void * heapBegin; // start of heap
299	void * heapEnd; // logical end of heap
300	size_t heapRemaining; // amount of storage not allocated in the current chunk
301	size_t pageSize; // architecture pagesize
302	size_t heapExpand; // sbrk advance
303	size_t mmapStart; // cross over point for mmap
304	unsigned int maxBucketsUsed; // maximum number of buckets in use
305
306	Heap * heapManagersList; // heap-list head
307	Heap * freeHeapManagersList; // free-list head
308
309	// Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
310	Heap * heapManagersStorage; // next heap to use in heap superblock
311	Heap * heapManagersStorageEnd; // logical heap outside of superblock's end
312
313	#ifdef __STATISTICS__
314	HeapStatistics stats; // global stats for thread-local heaps to add there counters when exiting
315	unsigned long int threads_started, threads_exited; // counts threads that have started and exited
316	unsigned long int reused_heap, new_heap; // counts reusability of heaps
317	unsigned int sbrk_calls;
318	unsigned long long int sbrk_storage;
319	int stats_fd;
320	#endif // __STATISTICS__
321	}; // HeapMaster
322
323
324	#ifdef FASTLOOKUP
325	enum { LookupSizes = 65_536 + sizeof(Heap.Storage) }; // number of fast lookup sizes
326	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
327	#endif // FASTLOOKUP
328
329	static volatile bool heapMasterBootFlag = false; // trigger for first heap
330	static HeapMaster heapMaster @= {}; // program global
331
332	static void heapMasterCtor();
333	static void heapMasterDtor();
334	static Heap * getHeap();
335
336
337	// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
338	// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
339	// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
340	static const unsigned int bucketSizes[] @= { // different bucket sizes
341	16 + sizeof(Heap.Storage), 32 + sizeof(Heap.Storage), 48 + sizeof(Heap.Storage), 64 + sizeof(Heap.Storage), // 4
342	96 + sizeof(Heap.Storage), 112 + sizeof(Heap.Storage), 128 + sizeof(Heap.Storage), // 3
343	160, 192, 224, 256 + sizeof(Heap.Storage), // 4
344	320, 384, 448, 512 + sizeof(Heap.Storage), // 4
345	640, 768, 896, 1_024 + sizeof(Heap.Storage), // 4
346	1_536, 2_048 + sizeof(Heap.Storage), // 2
347	2_560, 3_072, 3_584, 4_096 + sizeof(Heap.Storage), // 4
348	6_144, 8_192 + sizeof(Heap.Storage), // 2
349	9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(Heap.Storage), // 8
350	18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(Heap.Storage), // 8
351	36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(Heap.Storage), // 8
352	73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(Heap.Storage), // 8
353	147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(Heap.Storage), // 8
354	294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(Heap.Storage), // 8
355	655_360, 786_432, 917_504, 1_048_576 + sizeof(Heap.Storage), // 4
356	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(Heap.Storage), // 8
357	2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(Heap.Storage), // 4
358	};
359
360	static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
361
362
363	// extern visibility, used by runtime kernel
364	libcfa_public size_t __page_size; // architecture pagesize
365	libcfa_public int __map_prot; // common mmap/mprotect protection
366
367
368	// Thread-local storage is allocated lazily when the storage is accessed.
369	static __thread size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
370	static __thread Heap * heapManager CALIGN TLSMODEL;
371	static __thread bool heapManagerBootFlag CALIGN TLSMODEL = false;
372	static __thread size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing
373
374
375	// declare helper functions for HeapMaster
376	void noMemory(); // forward, called by "builtin_new" when malloc returns 0
377
378
379	void heapMasterCtor() with( heapMaster ) {
380	// Singleton pattern to initialize heap master
381
382	verify( bucketSizes[0] == (16 + sizeof(Heap.Storage)) );
383
384	__page_size = sysconf( _SC_PAGESIZE );
385	__map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC;
386
387	extLock = 0;
388	mgrLock = 0;
389
390	char * end = (char *)sbrk( 0 );
391	heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, libAlign() ) - end ); // move start of heap to multiple of alignment
392	heapRemaining = 0;
393	heapExpand = malloc_expansion();
394	mmapStart = malloc_mmap_start();
395
396	// find the closest bucket size less than or equal to the mmapStart size
397	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
398
399	verify( (mmapStart >= pageSize) && (bucketSizes[NoBucketSizes - 1] >= mmapStart) );
400	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
401	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
402
403	heapManagersList = 0p;
404	freeHeapManagersList = 0p;
405
406	heapManagersStorage = 0p;
407	heapManagersStorageEnd = 0p;
408
409	#ifdef __STATISTICS__
410	HeapStatisticsCtor( stats ); // clear statistic counters
411	threads_started = threads_exited = 0;
412	reused_heap = new_heap = 0;
413	sbrk_calls = sbrk_storage = 0;
414	stats_fd = STDERR_FILENO;
415	#endif // __STATISTICS__
416
417	#ifdef FASTLOOKUP
418	for ( unsigned int i = 0, idx = 0; i < LookupSizes; i += 1 ) {
419	if ( i > bucketSizes[idx] ) idx += 1;
420	lookup[i] = idx;
421	verify( i <= bucketSizes[idx] );
422	verify( (i <= 32 && idx == 0) \|\| (i > bucketSizes[idx - 1]) );
423	} // for
424	#endif // FASTLOOKUP
425
426	heapMasterBootFlag = true;
427	} // heapMasterCtor
428
429
430	#define NO_MEMORY_MSG "** Error ** insufficient heap memory available to allocate %zd new bytes."
431
432	Heap * getHeap() with( heapMaster ) {
433	Heap * heap;
434	if ( freeHeapManagersList ) { // free heap for reused ?
435	heap = freeHeapManagersList;
436	freeHeapManagersList = heap->nextFreeHeapManager;
437
438	#ifdef __STATISTICS__
439	reused_heap += 1;
440	#endif // __STATISTICS__
441	} else { // free heap not found, create new
442	// Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps *
443	// 12K ~= 120K byte superblock. Where 128-heap superblock handles a medium sized multi-processor server.
444	size_t remaining = heapManagersStorageEnd - heapManagersStorage; // remaining free heaps in superblock
445	if ( ! heapManagersStorage \|\| remaining == 0 ) {
446	// Each block of heaps is a multiple of the number of cores on the computer.
447	int HeapDim = get_nprocs(); // get_nprocs_conf does not work
448	size_t size = HeapDim * sizeof( Heap );
449
450	heapManagersStorage = (Heap *)mmap( 0, size, PROT_READ \| PROT_WRITE, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
451	if ( unlikely( heapManagersStorage == (Heap *)MAP_FAILED ) ) { // failed ?
452	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
453	// Do not call strerror( errno ) as it may call malloc.
454	abort( "** Error ** attempt to allocate block of heaps of size %zu bytes and mmap failed with errno %d.", size, errno );
455	} // if
456	heapManagersStorageEnd = &heapManagersStorage[HeapDim]; // outside array
457	} // if
458
459	heap = heapManagersStorage;
460	heapManagersStorage = heapManagersStorage + 1; // bump next heap
461
462	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
463	heap->nextHeapManager = heapManagersList;
464	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
465	heapManagersList = heap;
466
467	#ifdef __STATISTICS__
468	new_heap += 1;
469	#endif // __STATISTICS__
470
471	with( *heap ) {
472	for ( unsigned int j = 0; j < NoBucketSizes; j += 1 ) { // initialize free lists
473	#ifdef OWNERSHIP
474	#ifdef RETURNSPIN
475	freeLists[j].returnLock = 0;
476	freeLists[j].returnList = 0p;
477	#endif // RETURNSPIN
478	#endif // OWNERSHIP
479
480	freeLists[j].freeList = 0p;
481	freeLists[j].homeManager = heap;
482	freeLists[j].blockSize = bucketSizes[j];
483	} // for
484
485	heapBuffer = 0p;
486	heapReserve = 0;
487	nextFreeHeapManager = 0p;
488	#ifdef __CFA_DEBUG__
489	allocUnfreed = 0;
490	#endif // __CFA_DEBUG__
491	heapManagerBootFlag = true;
492	} // with
493	} // if
494
495	return heap;
496	} // getHeap
497
498
499	void heapManagerCtor() libcfa_public {
500	if ( unlikely( ! heapMasterBootFlag ) ) heapMasterCtor();
501
502	lock( heapMaster.mgrLock ); // protect heapMaster counters
503
504	assert( ! heapManagerBootFlag );
505
506	// get storage for heap manager
507
508	heapManager = getHeap();
509
510	#ifdef __STATISTICS__
511	HeapStatisticsCtor( heapManager->stats ); // heap local
512	heapMaster.threads_started += 1;
513	#endif // __STATISTICS__
514
515	unlock( heapMaster.mgrLock );
516	} // heapManagerCtor
517
518
519	void heapManagerDtor() libcfa_public {
520	if ( unlikely( ! heapManagerBootFlag ) ) return; // thread never used ?
521
522	lock( heapMaster.mgrLock );
523
524	// place heap on list of free heaps for reusability
525	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
526	heapMaster.freeHeapManagersList = heapManager;
527
528	#ifdef __STATISTICS__
529	heapMaster.stats += heapManager->stats; // retain this heap's statistics
530	heapMaster.threads_exited += 1;
531	#endif // __STATISTICS__
532
533	// Do not set heapManager to NULL because it is used after Cforall is shutdown but before the program shuts down.
534
535	heapManagerBootFlag = false;
536	unlock( heapMaster.mgrLock );
537	} // heapManagerDtor
538
539
540	//####################### Memory Allocation Routines Helpers ####################
541
542
543	extern int cfa_main_returned; // from interpose.cfa
544	extern "C" {
545	void memory_startup( void ) { // singleton => called once at start of program
546	if ( ! heapMasterBootFlag ) heapManagerCtor(); // sanity check
547	} // memory_startup
548
549	void memory_shutdown( void ) {
550	heapManagerDtor();
551	} // memory_shutdown
552
553	void heapAppStart() { // called by __cfaabi_appready_startup
554	verify( heapManager );
555	#ifdef __CFA_DEBUG__
556	heapManager->allocUnfreed = 0; // clear prior allocation counts
557	#endif // __CFA_DEBUG__
558
559	#ifdef __STATISTICS__
560	HeapStatisticsCtor( heapManager->stats ); // clear prior statistic counters
561	#endif // __STATISTICS__
562	} // heapAppStart
563
564	void heapAppStop() { // called by __cfaabi_appready_startdown
565	fclose( stdin ); fclose( stdout ); // free buffer storage
566	if ( ! cfa_main_returned ) return; // do not check unfreed storage if exit called
567
568	#ifdef __STATISTICS__
569	if ( getenv( "CFA_MALLOC_STATS" ) ) { // check for external printing
570	malloc_stats();
571	} // if
572	#endif // __STATISTICS__
573
574	#ifdef __CFA_DEBUG__
575	// allocUnfreed is set to 0 when a heap is created and it accumulates any unfreed storage during its multiple thread
576	// usages. At the end, add up each heap allocUnfreed value across all heaps to get the total unfreed storage.
577	ptrdiff_t allocUnfreed = 0;
578	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
579	allocUnfreed += heap->allocUnfreed;
580	} // for
581
582	allocUnfreed -= malloc_unfreed(); // subtract any user specified unfreed storage
583	if ( allocUnfreed > 0 ) {
584	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
585	char helpText[512];
586	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
587	"CFA warning (UNIX pid:%ld) : program terminating with %td(%#tx) bytes of storage allocated but not freed.\n"
588	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
589	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
590	} // if
591	#endif // __CFA_DEBUG__
592	} // heapAppStop
593	} // extern "C"
594
595
596	#ifdef __STATISTICS__
597	static HeapStatistics stats; // zero filled
598
599	#define prtFmt \
600	"\nHeap statistics: (storage request / allocation)\n" \
601	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
602	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
603	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
604	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
605	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
606	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
607	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
608	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
609	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
610	" return pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n" \
611	" sbrk calls %'u; storage %'llu bytes\n" \
612	" mmap calls %'u; storage %'llu / %'llu bytes\n" \
613	" munmap calls %'u; storage %'llu / %'llu bytes\n" \
614	" threads started %'lu; exited %'lu\n" \
615	" heaps new %'lu; reused %'lu\n"
616
617	// Use "write" because streams may be shutdown when calls are made.
618	static int printStats( HeapStatistics & stats ) with( heapMaster, stats ) { // see malloc_stats
619	char helpText[sizeof(prtFmt) + 1024]; // space for message and values
620	return __cfaabi_bits_print_buffer( stats_fd, helpText, sizeof(helpText), prtFmt,
621	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
622	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
623	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
624	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
625	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
626	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
627	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
628	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
629	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
630	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
631	sbrk_calls, sbrk_storage,
632	mmap_calls, mmap_storage_request, mmap_storage_alloc,
633	munmap_calls, munmap_storage_request, munmap_storage_alloc,
634	threads_started, threads_exited,
635	new_heap, reused_heap
636	);
637	} // printStats
638
639	#define prtFmtXML \
640	"<malloc version=\"1\">\n" \
641	"<heap nr=\"0\">\n" \
642	"<sizes>\n" \
643	"</sizes>\n" \
644	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
645	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
646	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
647	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
648	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
649	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
650	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
651	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
652	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
653	"<total type=\"return\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
654	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
655	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
656	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
657	"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n" \
658	"<total type=\"heaps\" new=\"%'lu;\" reused=\"%'lu\"/>\n" \
659	"</malloc>"
660
661	static int printStatsXML( HeapStatistics & stats, FILE * stream ) with( heapMaster, stats ) { // see malloc_info
662	char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
663	return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
664	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
665	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
666	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
667	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
668	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
669	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
670	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
671	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
672	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
673	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
674	sbrk_calls, sbrk_storage,
675	mmap_calls, mmap_storage_request, mmap_storage_alloc,
676	munmap_calls, munmap_storage_request, munmap_storage_alloc,
677	threads_started, threads_exited,
678	new_heap, reused_heap
679	);
680	} // printStatsXML
681
682	static HeapStatistics & collectStats( HeapStatistics & stats ) with( heapMaster ) {
683	lock( mgrLock );
684
685	stats += heapMaster.stats;
686	for ( Heap * heap = heapManagersList; heap; heap = heap->nextHeapManager ) {
687	stats += heap->stats;
688	} // for
689
690	unlock( mgrLock );
691	return stats;
692	} // collectStats
693	#endif // __STATISTICS__
694
695
696	static bool setMmapStart( size_t value ) with( heapMaster ) { // true => mmapped, false => sbrk
697	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
698	mmapStart = value; // set global
699
700	// find the closest bucket size less than or equal to the mmapStart size
701	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
702
703	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
704	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
705	return true;
706	} // setMmapStart
707
708
709	// <-------+----------------------------------------------------> bsize (bucket size)
710	// \|header \|addr
711	//==================================================================================
712	// align/offset \|
713	// <-----------------<------------+-----------------------------> bsize (bucket size)
714	// \|fake-header \| addr
715	#define HeaderAddr( addr ) ((Heap.Storage.Header )( (char )addr - sizeof(Heap.Storage) ))
716	#define RealHeader( header ) ((Heap.Storage.Header )((char )header - header->kind.fake.offset))
717
718	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
719	// \|header \|addr
720	//==================================================================================
721	// align/offset \|
722	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
723	// \|fake-header \|addr
724	#define DataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
725
726
727	inline __attribute__((always_inline))
728	static void checkAlign( size_t alignment ) {
729	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) {
730	abort( "** Error ** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
731	} // if
732	} // checkAlign
733
734
735	inline __attribute__((always_inline))
736	static void checkHeader( bool check, const char name[], void * addr ) {
737	if ( unlikely( check ) ) { // bad address ?
738	abort( "** Error ** attempt to %s storage %p with address outside the heap.\n"
739	"Possible cause is duplicate free on same block or overwriting of memory.",
740	name, addr );
741	} // if
742	} // checkHeader
743
744
745	// Manipulate sticky bits stored in unused 3 low-order bits of an address.
746	// bit0 => alignment => fake header
747	// bit1 => zero filled (calloc)
748	// bit2 => mapped allocation versus sbrk
749	#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
750	#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
751	#define MarkAlignmentBit( align ) ((align) \| 1)
752	#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
753	#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
754	#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
755	#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
756	#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize \|= 2)
757	#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
758	#define MarkMmappedBit( size ) ((size) \| 4)
759
760
761	inline __attribute__((always_inline))
762	static void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
763	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
764	alignment = ClearAlignmentBit( header ); // clear flag from value
765	#ifdef __CFA_DEBUG__
766	checkAlign( alignment ); // check alignment
767	#endif // __CFA_DEBUG__
768	header = RealHeader( header ); // backup from fake to real header
769	} else {
770	alignment = libAlign(); // => no fake header
771	} // if
772	} // fakeHeader
773
774
775	inline __attribute__((always_inline))
776	static bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
777	Heap.FreeHeader & freeHead, size_t & size, size_t & alignment ) with( heapMaster, heapManager ) {
778	header = HeaderAddr( addr );
779
780	#ifdef __CFA_DEBUG__
781	checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
782	#endif // __CFA_DEBUG__
783
784	if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
785	freeHead = (Heap.FreeHeader *)(header->kind.real.home);
786	alignment = libAlign();
787	} else {
788	fakeHeader( header, alignment );
789	if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
790	verify( addr < heapBegin \|\| heapEnd < addr );
791	size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
792	return true;
793	} // if
794
795	freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
796	} // if
797	size = freeHead->blockSize;
798
799	#ifdef __CFA_DEBUG__
800	checkHeader( header < (Heap.Storage.Header )heapBegin \|\| (Heap.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
801
802	Heap * homeManager;
803	if ( unlikely( freeHead == 0p \|\| // freed and only free-list node => null link
804	// freed and link points at another free block not to a bucket in the bucket array.
805	(homeManager = freeHead->homeManager, freeHead < &homeManager->freeLists[0] \|\|
806	&homeManager->freeLists[NoBucketSizes] <= freeHead ) ) ) {
807	abort( "** Error ** attempt to %s storage %p with corrupted header.\n"
808	"Possible cause is duplicate free on same block or overwriting of header information.",
809	name, addr );
810	} // if
811	#endif // __CFA_DEBUG__
812
813	return false;
814	} // headers
815
816
817	static void * master_extend( size_t size ) with( heapMaster ) {
818	lock( extLock );
819
820	ptrdiff_t rem = heapRemaining - size;
821	if ( unlikely( rem < 0 ) ) { // negative ?
822	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
823
824	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
825	if ( unlikely( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
826	unlock( extLock );
827	abort( NO_MEMORY_MSG, size ); // give up
828	} // if
829
830	// Make storage executable for thunks.
831	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
832	unlock( extLock );
833	abort( "** Error ** attempt to make heap storage executable for thunks and mprotect failed with errno %d.", errno );
834	} // if
835
836	rem = heapRemaining + increase - size;
837
838	#ifdef __STATISTICS__
839	sbrk_calls += 1;
840	sbrk_storage += increase;
841	#endif // __STATISTICS__
842	} // if
843
844	Heap.Storage * block = (Heap.Storage *)heapEnd;
845	heapRemaining = rem;
846	heapEnd = (char *)heapEnd + size;
847
848	unlock( extLock );
849	return block;
850	} // master_extend
851
852
853	__attribute__(( noinline ))
854	static void * manager_extend( size_t size ) with( *heapManager ) {
855	ptrdiff_t rem = heapReserve - size;
856
857	if ( unlikely( rem < 0 ) ) { // negative ?
858	// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
859	// smaller freeLists, and increase the reserve.
860
861	rem = heapReserve; // positive
862
863	if ( (unsigned int)rem >= bucketSizes[0] ) { // minimal size ? otherwise ignore
864	size_t bucket;
865	#ifdef FASTLOOKUP
866	if ( likely( rem < LookupSizes ) ) bucket = lookup[rem];
867	#endif // FASTLOOKUP
868	bucket = Bsearchl( rem, bucketSizes, heapMaster.maxBucketsUsed );
869	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
870	Heap.FreeHeader * freeHead = &(freeLists[bucket]);
871
872	// The remaining storage may not be bucket size, whereas all other allocations are. Round down to previous
873	// bucket size in this case.
874	if ( unlikely( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
875	Heap.Storage * block = (Heap.Storage *)heapBuffer;
876
877	block->header.kind.real.next = freeHead->freeList; // push on stack
878	freeHead->freeList = block;
879	} // if
880
881	size_t increase = ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), libAlign() );
882	heapBuffer = master_extend( increase );
883	rem = increase - size;
884	} // if
885
886	Heap.Storage * block = (Heap.Storage *)heapBuffer;
887	heapReserve = rem;
888	heapBuffer = (char *)heapBuffer + size;
889
890	return block;
891	} // manager_extend
892
893
894	#define BOOT_HEAP_MANAGER \
895	if ( unlikely( ! heapMasterBootFlag ) ) { \
896	heapManagerCtor(); /* trigger for first heap */ \
897	} /* if */
898
899	#ifdef __STATISTICS__
900	#define STAT_NAME __counter
901	#define STAT_PARM , unsigned int STAT_NAME
902	#define STAT_ARG( name ) , name
903	#define STAT_0_CNT( counter ) stats.counters[counter].calls_0 += 1
904	#else
905	#define STAT_NAME
906	#define STAT_PARM
907	#define STAT_ARG( name )
908	#define STAT_0_CNT( counter )
909	#endif // __STATISTICS__
910
911	// Uncomment to get allocation addresses for a 0-sized allocation rather than a null pointer.
912	//#define __NONNULL_0_ALLOC__
913	#if ! defined( __NONNULL_0_ALLOC__ )
914	#define __NULL_0_ALLOC__ unlikely( size == 0 ) \|\| /* 0 BYTE ALLOCATION RETURNS NULL POINTER */
915	#else
916	#define __NULL_0_ALLOC__
917	#endif // __NONNULL_0_ALLOC__
918
919	#define PROLOG( counter, ... ) \
920	BOOT_HEAP_MANAGER; \
921	if ( \
922	__NULL_0_ALLOC__ \
923	unlikely( size > ULONG_MAX - sizeof(Heap.Storage) ) ) { /* error check */ \
924	STAT_0_CNT( counter ); \
925	__VA_ARGS__; \
926	return 0p; \
927	} /* if */
928
929
930	#define SCRUB_SIZE 1024lu
931	// Do not use '\xfe' for scrubbing because dereferencing an address composed of it causes a SIGSEGV without a valid IP
932	// pointer in the interrupt frame.
933	#define SCRUB '\xff'
934
935	static void * doMalloc( size_t size STAT_PARM ) libcfa_nopreempt with( *heapManager ) {
936	PROLOG( STAT_NAME );
937
938	verify( heapManager );
939	Heap.Storage * block; // pointer to new block of storage
940
941	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
942	// along with the block and is a multiple of the alignment size.
943	size_t tsize = size + sizeof(Heap.Storage);
944
945	#ifdef __STATISTICS__
946	stats.counters[STAT_NAME].calls += 1;
947	stats.counters[STAT_NAME].request += size;
948	#endif // __STATISTICS__
949
950	#ifdef __CFA_DEBUG__
951	allocUnfreed += size;
952	#endif // __CFA_DEBUG__
953
954	if ( likely( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
955	size_t bucket;
956	#ifdef FASTLOOKUP
957	if ( likely( tsize < LookupSizes ) ) bucket = lookup[tsize];
958	else
959	#endif // FASTLOOKUP
960	bucket = Bsearchl( tsize, bucketSizes, heapMaster.maxBucketsUsed );
961	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
962	Heap.FreeHeader * freeHead = &freeLists[bucket];
963
964	verify( freeHead <= &freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
965	verify( tsize <= freeHead->blockSize ); // search failure ?
966
967	tsize = freeHead->blockSize; // total space needed for request
968	#ifdef __STATISTICS__
969	stats.counters[STAT_NAME].alloc += tsize;
970	#endif // __STATISTICS__
971
972	block = freeHead->freeList; // remove node from stack
973	if ( unlikely( block == 0p ) ) { // no free block ?
974	// Freelist for this size is empty, so check return list (OWNERSHIP), or carve it out of the heap if there
975	// is enough left, or get some more heap storage and carve it off.
976	#ifdef OWNERSHIP
977	if ( unlikely( freeHead->returnList ) ) { // race, get next time if lose race
978	#ifdef RETURNSPIN
979	lock( freeHead->returnLock );
980	block = freeHead->returnList;
981	freeHead->returnList = 0p;
982	unlock( freeHead->returnLock );
983	#else
984	block = __atomic_exchange_n( &freeHead->returnList, 0p, __ATOMIC_SEQ_CST );
985	#endif // RETURNSPIN
986
987	verify( block );
988	#ifdef __STATISTICS__
989	stats.return_pulls += 1;
990	#endif // __STATISTICS__
991
992	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
993
994	freeHead->freeList = block->header.kind.real.next; // merge returnList into freeHead
995	} else {
996	#endif // OWNERSHIP
997	// Do not leave kernel thread as manager_extend accesses heapManager.
998	disable_interrupts();
999	block = (Heap.Storage *)manager_extend( tsize ); // mutual exclusion on call
1000	enable_interrupts( false );
1001
1002	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1003
1004	#ifdef __CFA_DEBUG__
1005	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes.
1006	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1007	#endif // __CFA_DEBUG__
1008	#ifdef OWNERSHIP
1009	} // if
1010	#endif // OWNERSHIP
1011	} else {
1012	// Memory is scrubbed in doFree.
1013	freeHead->freeList = block->header.kind.real.next;
1014	} // if
1015
1016	block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
1017	} else { // large size => mmap
1018	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
1019	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
1020
1021	#ifdef __STATISTICS__
1022	stats.counters[STAT_NAME].alloc += tsize;
1023	stats.mmap_calls += 1;
1024	stats.mmap_storage_request += size;
1025	stats.mmap_storage_alloc += tsize;
1026	#endif // __STATISTICS__
1027
1028	disable_interrupts();
1029	block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
1030	enable_interrupts( false );
1031
1032	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1033
1034	if ( unlikely( block == (Heap.Storage *)MAP_FAILED ) ) { // failed ?
1035	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
1036	// Do not call strerror( errno ) as it may call malloc.
1037	abort( "** Error ** attempt to allocate large object (> %zu) of size %zu bytes and mmap failed with errno %d.",
1038	size, heapMaster.mmapStart, errno );
1039	} // if
1040	block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
1041
1042	#ifdef __CFA_DEBUG__
1043	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes. The
1044	// rest of the storage set to 0 by mmap.
1045	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1046	#endif // __CFA_DEBUG__
1047	} // if
1048
1049	block->header.kind.real.size = size; // store allocation size
1050	void * addr = &(block->data); // adjust off header to user bytes
1051	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
1052
1053	#ifdef __CFA_DEBUG__
1054	if ( traceHeap() ) {
1055	char helpText[64];
1056	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1057	"%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
1058	} // if
1059	#endif // __CFA_DEBUG__
1060
1061	// poll_interrupts(); // call rollforward
1062
1063	return addr;
1064	} // doMalloc
1065
1066
1067	static void doFree( void * addr ) libcfa_nopreempt with( *heapManager ) {
1068	verify( addr );
1069
1070	// detect free after thread-local storage destruction and use global stats in that case
1071
1072	Heap.Storage.Header * header;
1073	Heap.FreeHeader * freeHead;
1074	size_t size, alignment;
1075
1076	bool mapped = headers( "free", addr, header, freeHead, size, alignment );
1077	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
1078	size_t rsize = header->kind.real.size; // optimization
1079	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
1080
1081	#ifdef __STATISTICS__
1082	stats.free_storage_request += rsize;
1083	stats.free_storage_alloc += size;
1084	#endif // __STATISTICS__
1085
1086	#ifdef __CFA_DEBUG__
1087	allocUnfreed -= rsize;
1088	#endif // __CFA_DEBUG__
1089
1090	if ( unlikely( mapped ) ) { // mmapped ?
1091	#ifdef __STATISTICS__
1092	stats.munmap_calls += 1;
1093	stats.munmap_storage_request += rsize;
1094	stats.munmap_storage_alloc += size;
1095	#endif // __STATISTICS__
1096
1097	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1098
1099	// Does not matter where this storage is freed.
1100	if ( unlikely( munmap( header, size ) == -1 ) ) {
1101	// Do not call strerror( errno ) as it may call malloc.
1102	abort( "** Error ** attempt to deallocate large object %p and munmap failed with errno %d.\n"
1103	"Possible cause is invalid delete pointer: either not allocated or with corrupt header.",
1104	addr, errno );
1105	} // if
1106	} else {
1107	#ifdef __CFA_DEBUG__
1108	// memset is NOT always inlined!
1109	disable_interrupts();
1110	// Scrub old memory so subsequent usages might fail. Only scrub the first/last SCRUB_SIZE bytes.
1111	char * data = ((Heap.Storage *)header)->data; // data address
1112	size_t dsize = size - sizeof(Heap.Storage); // data size
1113	if ( dsize <= SCRUB_SIZE * 2 ) {
1114	memset( data, SCRUB, dsize ); // scrub all
1115	} else {
1116	memset( data, SCRUB, SCRUB_SIZE ); // scrub front
1117	memset( data + dsize - SCRUB_SIZE, SCRUB, SCRUB_SIZE ); // scrub back
1118	} // if
1119	enable_interrupts( false );
1120	#endif // __CFA_DEBUG__
1121
1122	#ifdef OWNERSHIP
1123	if ( likely( heapManager == freeHead->homeManager ) ) { // belongs to this thread
1124	header->kind.real.next = freeHead->freeList; // push on stack
1125	freeHead->freeList = (Heap.Storage *)header;
1126	} else { // return to thread owner
1127	verify( heapManager );
1128
1129	#ifdef RETURNSPIN
1130	lock( freeHead->returnLock );
1131	header->kind.real.next = freeHead->returnList; // push to bucket return list
1132	freeHead->returnList = (Heap.Storage *)header;
1133	unlock( freeHead->returnLock );
1134	#else // lock free
1135	header->kind.real.next = freeHead->returnList; // link new node to top node
1136	// CAS resets header->kind.real.next = freeHead->returnList on failure
1137	while ( ! __atomic_compare_exchange_n( &freeHead->returnList, &header->kind.real.next, (Heap.Storage *)header,
1138	false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
1139
1140	#ifdef __STATISTICS__
1141	stats.return_pushes += 1;
1142	stats.return_storage_request += rsize;
1143	stats.return_storage_alloc += size;
1144	#endif // __STATISTICS__
1145	#endif // RETURNSPIN
1146	} // if
1147
1148	#else // no OWNERSHIP
1149
1150	// kind.real.home is address in owner thread's freeLists, so compute the equivalent position in this thread's freeList.
1151	freeHead = &freeLists[ClearStickyBits( (Heap.FreeHeader *)(header->kind.real.home) ) - &freeHead->homeManager->freeLists[0]];
1152	header->kind.real.next = freeHead->freeList; // push on stack
1153	freeHead->freeList = (Heap.Storage *)header;
1154	#endif // ! OWNERSHIP
1155
1156	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1157	} // if
1158
1159	#ifdef __CFA_DEBUG__
1160	if ( traceHeap() ) {
1161	char helpText[64];
1162	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1163	"Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
1164	} // if
1165	#endif // __CFA_DEBUG__
1166
1167	// poll_interrupts(); // call rollforward
1168	} // doFree
1169
1170
1171	size_t prtFree( Heap & manager ) with( manager ) {
1172	size_t total = 0;
1173	#ifdef __STATISTICS__
1174	__cfaabi_bits_acquire();
1175	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
1176	#endif // __STATISTICS__
1177	for ( unsigned int i = 0; i < heapMaster.maxBucketsUsed; i += 1 ) {
1178	size_t size = freeLists[i].blockSize;
1179	#ifdef __STATISTICS__
1180	unsigned int N = 0;
1181	#endif // __STATISTICS__
1182
1183	for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
1184	total += size;
1185	#ifdef __STATISTICS__
1186	N += 1;
1187	#endif // __STATISTICS__
1188	} // for
1189
1190	#ifdef __STATISTICS__
1191	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
1192	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
1193	#endif // __STATISTICS__
1194	} // for
1195	#ifdef __STATISTICS__
1196	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
1197	__cfaabi_bits_release();
1198	#endif // __STATISTICS__
1199	return (char )heapMaster.heapEnd - (char )heapMaster.heapBegin - total;
1200	} // prtFree
1201
1202
1203	#ifdef __STATISTICS__
1204	static void incCalls( size_t statName ) libcfa_nopreempt {
1205	heapManager->stats.counters[statName].calls += 1;
1206	} // incCalls
1207
1208	static void incZeroCalls( size_t statName ) libcfa_nopreempt {
1209	heapManager->stats.counters[statName].calls_0 += 1;
1210	} // incZeroCalls
1211	#endif // __STATISTICS__
1212
1213	#ifdef __CFA_DEBUG__
1214	static void incUnfreed( intptr_t offset ) libcfa_nopreempt {
1215	heapManager->allocUnfreed += offset;
1216	} // incUnfreed
1217	#endif // __CFA_DEBUG__
1218
1219
1220	static void * memalignNoStats( size_t alignment, size_t size STAT_PARM ) {
1221	checkAlign( alignment ); // check alignment
1222
1223	// if alignment <= default alignment or size == 0, do normal malloc as two headers are unnecessary
1224	if ( unlikely( alignment <= libAlign() \|\| size == 0 ) ) return doMalloc( size STAT_ARG( STAT_NAME ) );
1225
1226	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1227	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1228	// .-------------v-----------------v----------------v----------,
1229	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
1230	// `-------------^-----------------^-+--------------^----------'
1231	// \|<--------------------------------' offset/align \|<-- alignment boundary
1232
1233	// subtract libAlign() because it is already the minimum alignment
1234	// add sizeof(Storage) for fake header
1235	size_t offset = alignment - libAlign() + sizeof(Heap.Storage);
1236	char * addr = (char *)doMalloc( size + offset STAT_ARG( STAT_NAME ) );
1237
1238	// address in the block of the "next" alignment address
1239	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
1240
1241	// address of header from malloc
1242	Heap.Storage.Header * realHeader = HeaderAddr( addr );
1243	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1244	#ifdef __CFA_DEBUG__
1245	incUnfreed( -offset ); // adjustment off the offset from call to doMalloc
1246	#endif // __CFA_DEBUG__
1247
1248	// address of fake header before the alignment location
1249	Heap.Storage.Header * fakeHeader = HeaderAddr( user );
1250
1251	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1252	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
1253	// SKULLDUGGERY: odd alignment implies fake header
1254	fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
1255
1256	return user;
1257	} // memalignNoStats
1258
1259
1260	//####################### Memory Allocation Routines ####################
1261
1262
1263	extern "C" {
1264	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1265	// then malloc() returns a unique pointer value that can later be successfully passed to free().
1266	void * malloc( size_t size ) libcfa_public {
1267	return doMalloc( size STAT_ARG( MALLOC ) );
1268	} // malloc
1269
1270
1271	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1272	void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
1273	return doMalloc( dim * elemSize STAT_ARG( AALLOC ) );
1274	} // aalloc
1275
1276
1277	// Same as aalloc() with memory set to zero.
1278	void * calloc( size_t dim, size_t elemSize ) libcfa_public {
1279	size_t size = dim * elemSize;
1280	char * addr = (char *)doMalloc( size STAT_ARG( CALLOC ) );
1281
1282	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1283
1284	Heap.Storage.Header * header;
1285	Heap.FreeHeader * freeHead;
1286	size_t bsize, alignment;
1287
1288	#ifndef __CFA_DEBUG__
1289	bool mapped =
1290	#endif // __CFA_DEBUG__
1291	headers( "calloc", addr, header, freeHead, bsize, alignment );
1292
1293	#ifndef __CFA_DEBUG__
1294	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1295	if ( likely( ! mapped ) )
1296	#endif // __CFA_DEBUG__
1297	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1298	// `-header`-addr `-size
1299	memset( addr, '\0', size ); // set to zeros
1300
1301	MarkZeroFilledBit( header ); // mark as zero fill
1302	return addr;
1303	} // calloc
1304
1305
1306	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1307	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1308	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
1309	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1310	void * resize( void * oaddr, size_t size ) libcfa_public {
1311	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1312	return doMalloc( size STAT_ARG( RESIZE ) );
1313	} // if
1314
1315	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1316
1317	Heap.Storage.Header * header;
1318	Heap.FreeHeader * freeHead;
1319	size_t bsize, oalign;
1320	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1321
1322	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1323	// same size, DO NOT preserve STICKY PROPERTIES.
1324	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1325	ClearZeroFillBit( header ); // no alignment and turn off 0 fill
1326	#ifdef __CFA_DEBUG__
1327	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1328	#endif // __CFA_DEBUG__
1329	header->kind.real.size = size; // reset allocation size
1330	#ifdef __STATISTICS__
1331	incCalls( RESIZE );
1332	#endif // __STATISTICS__
1333	return oaddr;
1334	} // if
1335
1336	// change size, DO NOT preserve STICKY PROPERTIES.
1337	doFree( oaddr ); // free previous storage
1338
1339	return doMalloc( size STAT_ARG( RESIZE ) ); // create new area
1340	} // resize
1341
1342
1343	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1344	// the old and new sizes.
1345	void * realloc( void * oaddr, size_t size ) libcfa_public {
1346	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1347	return doMalloc( size STAT_ARG( REALLOC ) );
1348	} // if
1349
1350	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1351
1352	Heap.Storage.Header * header;
1353	Heap.FreeHeader * freeHead;
1354	size_t bsize, oalign;
1355	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1356
1357	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1358	size_t osize = header->kind.real.size; // old allocation size
1359	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1360	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1361	#ifdef __CFA_DEBUG__
1362	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1363	#endif // __CFA_DEBUG__
1364	header->kind.real.size = size; // reset allocation size
1365	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1366	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1367	} // if
1368	#ifdef __STATISTICS__
1369	incCalls( REALLOC );
1370	#endif // __STATISTICS__
1371	return oaddr;
1372	} // if
1373
1374	// change size and copy old content to new storage
1375
1376	void * naddr;
1377	if ( likely( oalign <= libAlign() ) ) { // previous request not aligned ?
1378	naddr = doMalloc( size STAT_ARG( REALLOC ) ); // create new area
1379	} else {
1380	naddr = memalignNoStats( oalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1381	} // if
1382
1383	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1384	// To preserve prior fill, the entire bucket must be copied versus the size.
1385	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1386	doFree( oaddr ); // free previous storage
1387
1388	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1389	MarkZeroFilledBit( header ); // mark new request as zero filled
1390	if ( size > osize ) { // previous request larger ?
1391	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1392	} // if
1393	} // if
1394	return naddr;
1395	} // realloc
1396
1397
1398	// Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1399	void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1400	return realloc( oaddr, dim * elemSize );
1401	} // reallocarray
1402
1403
1404	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1405	void * memalign( size_t alignment, size_t size ) libcfa_public {
1406	return memalignNoStats( alignment, size STAT_ARG( MEMALIGN ) );
1407	} // memalign
1408
1409
1410	// Same as aalloc() with memory alignment.
1411	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1412	return memalignNoStats( alignment, dim * elemSize STAT_ARG( AMEMALIGN ) );
1413	} // amemalign
1414
1415
1416	// Same as calloc() with memory alignment.
1417	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1418	size_t size = dim * elemSize;
1419	char * addr = (char *)memalignNoStats( alignment, size STAT_ARG( CMEMALIGN ) );
1420
1421	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1422
1423	Heap.Storage.Header * header;
1424	Heap.FreeHeader * freeHead;
1425	size_t bsize;
1426
1427	#ifndef __CFA_DEBUG__
1428	bool mapped =
1429	#endif // __CFA_DEBUG__
1430	headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1431
1432	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1433	#ifndef __CFA_DEBUG__
1434	if ( ! mapped )
1435	#endif // __CFA_DEBUG__
1436	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1437	// `-header`-addr `-size
1438	memset( addr, '\0', size ); // set to zeros
1439
1440	MarkZeroFilledBit( header ); // mark as zero filled
1441	return addr;
1442	} // cmemalign
1443
1444
1445	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1446	// of alignment. This requirement is universally ignored.
1447	void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1448	return memalign( alignment, size );
1449	} // aligned_alloc
1450
1451
1452	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1453	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1454	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1455	// free(3).
1456	int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1457	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1458	*memptr = memalign( alignment, size );
1459	return 0;
1460	} // posix_memalign
1461
1462
1463	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1464	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1465	void * valloc( size_t size ) libcfa_public {
1466	return memalign( __page_size, size );
1467	} // valloc
1468
1469
1470	// Same as valloc but rounds size to multiple of page size.
1471	void * pvalloc( size_t size ) libcfa_public {
1472	return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1473	} // pvalloc
1474
1475
1476	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1477	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1478	// 0p, no operation is performed.
1479	void free( void * addr ) libcfa_public {
1480	if ( unlikely( addr == 0p ) ) { // special case
1481	#ifdef __STATISTICS__
1482	if ( heapManager )
1483	incZeroCalls( FREE );
1484	#endif // __STATISTICS__
1485	return;
1486	} // if
1487
1488	#ifdef __STATISTICS__
1489	incCalls( FREE );
1490	#endif // __STATISTICS__
1491
1492	doFree( addr ); // handles heapManager == nullptr
1493	} // free
1494
1495
1496	// Returns the alignment of an allocation.
1497	size_t malloc_alignment( void * addr ) libcfa_public {
1498	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1499	Heap.Storage.Header * header = HeaderAddr( addr );
1500	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1501	return ClearAlignmentBit( header ); // clear flag from value
1502	} else {
1503	return libAlign(); // minimum alignment
1504	} // if
1505	} // malloc_alignment
1506
1507
1508	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1509	bool malloc_zero_fill( void * addr ) libcfa_public {
1510	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1511	Heap.Storage.Header * header = HeaderAddr( addr );
1512	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1513	header = RealHeader( header ); // backup from fake to real header
1514	} // if
1515	return ZeroFillBit( header ); // zero filled ?
1516	} // malloc_zero_fill
1517
1518
1519	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1520	size_t malloc_size( void * addr ) libcfa_public {
1521	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1522	Heap.Storage.Header * header = HeaderAddr( addr );
1523	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1524	header = RealHeader( header ); // backup from fake to real header
1525	} // if
1526	return header->kind.real.size;
1527	} // malloc_size
1528
1529
1530	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1531	// malloc or a related function.
1532	size_t malloc_usable_size( void * addr ) libcfa_public {
1533	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1534	Heap.Storage.Header * header;
1535	Heap.FreeHeader * freeHead;
1536	size_t bsize, alignment;
1537
1538	headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1539	return DataStorage( bsize, addr, header ); // data storage in bucket
1540	} // malloc_usable_size
1541
1542
1543	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1544	void malloc_stats( void ) libcfa_public {
1545	#ifdef __STATISTICS__
1546	HeapStatistics stats;
1547	HeapStatisticsCtor( stats );
1548	if ( printStats( collectStats( stats ) ) == -1 ) {
1549	#else
1550	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1551	if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1552	#endif // __STATISTICS__
1553	abort( "** Error ** write failed in malloc_stats" );
1554	} // if
1555	} // malloc_stats
1556
1557
1558	// Changes the file descriptor where malloc_stats() writes statistics.
1559	int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1560	#ifdef __STATISTICS__
1561	int temp = heapMaster.stats_fd;
1562	heapMaster.stats_fd = fd;
1563	return temp;
1564	#else
1565	return -1; // unsupported
1566	#endif // __STATISTICS__
1567	} // malloc_stats_fd
1568
1569
1570	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1571	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1572	// malloc).
1573	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1574	if ( options != 0 ) { errno = EINVAL; return -1; }
1575	#ifdef __STATISTICS__
1576	HeapStatistics stats;
1577	HeapStatisticsCtor( stats );
1578	return printStatsXML( collectStats( stats ), stream ); // returns bytes written or -1
1579	#else
1580	return 0; // unsupported
1581	#endif // __STATISTICS__
1582	} // malloc_info
1583
1584
1585	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1586	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1587	int mallopt( int option, int value ) libcfa_public {
1588	if ( value < 0 ) return 0;
1589	choose( option ) {
1590	case M_TOP_PAD:
1591	heapMaster.heapExpand = ceiling2( value, __page_size );
1592	return 1;
1593	case M_MMAP_THRESHOLD:
1594	if ( setMmapStart( value ) ) return 1;
1595	} // choose
1596	return 0; // error, unsupported
1597	} // mallopt
1598
1599
1600	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1601	int malloc_trim( size_t ) libcfa_public {
1602	return 0; // => impossible to release memory
1603	} // malloc_trim
1604
1605
1606	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1607	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1608	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1609	// result. (The caller must free this memory.)
1610	void * malloc_get_state( void ) libcfa_public {
1611	return 0p; // unsupported
1612	} // malloc_get_state
1613
1614
1615	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1616	// structure pointed to by state.
1617	int malloc_set_state( void * ) libcfa_public {
1618	return 0; // unsupported
1619	} // malloc_set_state
1620
1621
1622	// Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1623	__attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1624
1625	// Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1626	__attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1627
1628	// Amount subtracted to adjust for unfreed program storage (debug only).
1629	__attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1630	} // extern "C"
1631
1632
1633	// Must have CFA linkage to overload with C linkage realloc.
1634	void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1635	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1636	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) );
1637	} // if
1638
1639	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1640
1641	// Attempt to reuse existing alignment.
1642	Heap.Storage.Header * header = HeaderAddr( oaddr );
1643	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1644	size_t oalign;
1645
1646	if ( unlikely( isFakeHeader ) ) {
1647	checkAlign( nalign ); // check alignment
1648	oalign = ClearAlignmentBit( header ); // old alignment
1649	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1650	&& ( oalign <= nalign // going down
1651	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1652	) ) {
1653	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1654	Heap.FreeHeader * freeHead;
1655	size_t bsize, oalign;
1656	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1657	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1658
1659	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1660	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1661	ClearZeroFillBit( header ); // turn off 0 fill
1662	#ifdef __CFA_DEBUG__
1663	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1664	#endif // __CFA_DEBUG__
1665	header->kind.real.size = size; // reset allocation size
1666	#ifdef __STATISTICS__
1667	incCalls( RESIZE );
1668	#endif // __STATISTICS__
1669	return oaddr;
1670	} // if
1671	} // if
1672	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1673	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1674	return resize( oaddr, size ); // duplicate special case checks
1675	} // if
1676
1677	// change size, DO NOT preserve STICKY PROPERTIES.
1678	doFree( oaddr ); // free previous storage
1679	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) ); // create new aligned area
1680	} // resize
1681
1682
1683	void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1684	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1685	return memalignNoStats( nalign, size STAT_ARG( REALLOC ) );
1686	} // if
1687
1688	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1689
1690	// Attempt to reuse existing alignment.
1691	Heap.Storage.Header * header = HeaderAddr( oaddr );
1692	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1693	size_t oalign;
1694	if ( unlikely( isFakeHeader ) ) {
1695	checkAlign( nalign ); // check alignment
1696	oalign = ClearAlignmentBit( header ); // old alignment
1697	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1698	&& ( oalign <= nalign // going down
1699	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1700	) ) {
1701	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1702	return realloc( oaddr, size ); // duplicate special case checks
1703	} // if
1704	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1705	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1706	return realloc( oaddr, size ); // duplicate special case checks
1707	} // if
1708
1709	Heap.FreeHeader * freeHead;
1710	size_t bsize;
1711	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1712
1713	// change size and copy old content to new storage
1714
1715	size_t osize = header->kind.real.size; // old allocation size
1716	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1717
1718	void * naddr = memalignNoStats( nalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1719
1720	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1721	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1722	doFree( oaddr ); // free previous storage
1723
1724	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1725	MarkZeroFilledBit( header ); // mark new request as zero filled
1726	if ( size > osize ) { // previous request larger ?
1727	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1728	} // if
1729	} // if
1730	return naddr;
1731	} // realloc
1732
1733
1734	void * reallocarray( void * oaddr, size_t nalign, size_t dim, size_t elemSize ) __THROW {
1735	return realloc( oaddr, nalign, dim * elemSize );
1736	} // reallocarray
1737
1738
1739	// Local Variables: //
1740	// tab-width: 4 //
1741	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1742	// End: //

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format