Context Navigation

heap.cfa @ c7f12a4

ADTast-experimental

Last change on this file since c7f12a4 was 7a2057a, checked in by Peter A. Buhr <pabuhr@…>, 2 years ago
remove unused BUCKETLOCK, restructure OWNERSHIP and RETURNSPIN
Property mode set to `100644`
File size: 67.0 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 Oct 30 15:33:13 2022
13	// Update Count : 1580
14	//
15
16	#include <stdio.h>
17	#include <string.h> // memset, memcpy
18	#include <limits.h> // ULONG_MAX
19	#include <stdlib.h> // EXIT_FAILURE
20	#include <errno.h> // errno, ENOMEM, EINVAL
21	#include <unistd.h> // STDERR_FILENO, sbrk, sysconf
22	#include <malloc.h> // memalign, malloc_usable_size
23	#include <sys/mman.h> // mmap, munmap
24	extern "C" {
25	#include <sys/sysinfo.h> // get_nprocs
26	} // extern "C"
27
28	#include "bits/align.hfa" // libAlign
29	#include "bits/defs.hfa" // likely, unlikely
30	#include "concurrency/kernel/fwd.hfa" // __POLL_PREEMPTION
31	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
32	#include "math.hfa" // ceiling, min
33	#include "bitmanip.hfa" // is_pow2, ceiling2
34
35	// supported mallopt options
36	#ifndef M_MMAP_THRESHOLD
37	#define M_MMAP_THRESHOLD (-1)
38	#endif // M_MMAP_THRESHOLD
39
40	#ifndef M_TOP_PAD
41	#define M_TOP_PAD (-2)
42	#endif // M_TOP_PAD
43
44	#define FASTLOOKUP // use O(1) table lookup from allocation size to bucket size
45	#define OWNERSHIP // return freed memory to owner thread
46	#define RETURNSPIN // toggle spinlock / lockfree queue
47	#if ! defined( OWNERSHIP ) && defined( RETURNSPIN )
48	#warning "RETURNSPIN is ignored without OWNERSHIP; suggest commenting out RETURNSPIN"
49	#endif // ! OWNERSHIP && RETURNSPIN
50
51	#define CACHE_ALIGN 64
52	#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))
53
54	#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
55
56	//#define __STATISTICS__
57
58	enum {
59	// The default extension heap amount in units of bytes. When the current heap reaches the brk address, the brk
60	// address is extended by the extension amount.
61	__CFA_DEFAULT_HEAP_EXPANSION__ = 10 * 1024 * 1024,
62
63	// The mmap crossover point during allocation. Allocations less than this amount are allocated from buckets; values
64	// greater than or equal to this value are mmap from the operating system.
65	__CFA_DEFAULT_MMAP_START__ = 512 * 1024 + 1,
66
67	// The default unfreed storage amount in units of bytes. When the uC++ program ends it subtracts this amount from
68	// the malloc/free counter to adjust for storage the program does not free.
69	__CFA_DEFAULT_HEAP_UNFREED__ = 0
70	}; // enum
71
72
73	//####################### Heap Trace/Print ####################
74
75
76	static bool traceHeap = false;
77
78	inline bool traceHeap() libcfa_public { return traceHeap; }
79
80	bool traceHeapOn() libcfa_public {
81	bool temp = traceHeap;
82	traceHeap = true;
83	return temp;
84	} // traceHeapOn
85
86	bool traceHeapOff() libcfa_public {
87	bool temp = traceHeap;
88	traceHeap = false;
89	return temp;
90	} // traceHeapOff
91
92	bool traceHeapTerm() libcfa_public { return false; }
93
94
95	static bool prtFree = false;
96
97	bool prtFree() {
98	return prtFree;
99	} // prtFree
100
101	bool prtFreeOn() {
102	bool temp = prtFree;
103	prtFree = true;
104	return temp;
105	} // prtFreeOn
106
107	bool prtFreeOff() {
108	bool temp = prtFree;
109	prtFree = false;
110	return temp;
111	} // prtFreeOff
112
113
114	//######################### Helpers #########################
115
116
117	// generic Bsearchl does not inline, so substitute with hand-coded binary-search.
118	inline __attribute__((always_inline))
119	static size_t Bsearchl( unsigned int key, const unsigned int vals[], size_t dim ) {
120	size_t l = 0, m, h = dim;
121	while ( l < h ) {
122	m = (l + h) / 2;
123	if ( (unsigned int &)(vals[m]) < key ) { // cast away const
124	l = m + 1;
125	} else {
126	h = m;
127	} // if
128	} // while
129	return l;
130	} // Bsearchl
131
132
133	// pause to prevent excess processor bus usage
134	#if defined( __i386 ) \|\| defined( __x86_64 )
135	#define Pause() __asm__ __volatile__ ( "pause" : : : )
136	#elif defined(__ARM_ARCH)
137	#define Pause() __asm__ __volatile__ ( "YIELD" : : : )
138	#else
139	#error unsupported architecture
140	#endif
141
142	typedef volatile uintptr_t SpinLock_t CALIGN; // aligned addressable word-size
143
144	static inline __attribute__((always_inline)) void lock( volatile SpinLock_t & slock ) {
145	enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
146	unsigned int spin = SPIN_START;
147
148	for ( unsigned int i = 1;; i += 1 ) {
149	if ( slock == 0 && __atomic_test_and_set( &slock, __ATOMIC_SEQ_CST ) == 0 ) break; // Fence
150	for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
151	spin += spin; // powers of 2
152	//if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
153	if ( spin > SPIN_END ) spin = SPIN_END; // cap spinning
154	} // for
155	} // spin_lock
156
157	static inline __attribute__((always_inline)) void unlock( volatile SpinLock_t & slock ) {
158	__atomic_clear( &slock, __ATOMIC_SEQ_CST ); // Fence
159	} // spin_unlock
160
161
162	//####################### Heap Statistics ####################
163
164
165	#ifdef __STATISTICS__
166	enum { CntTriples = 12 }; // number of counter triples
167	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC, FREE };
168
169	struct StatsOverlay { // overlay for iteration
170	unsigned int calls, calls_0;
171	unsigned long long int request, alloc;
172	};
173
174	// Heap statistics counters.
175	union HeapStatistics {
176	struct { // minimum qualification
177	unsigned int malloc_calls, malloc_0_calls;
178	unsigned long long int malloc_storage_request, malloc_storage_alloc;
179	unsigned int aalloc_calls, aalloc_0_calls;
180	unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
181	unsigned int calloc_calls, calloc_0_calls;
182	unsigned long long int calloc_storage_request, calloc_storage_alloc;
183	unsigned int memalign_calls, memalign_0_calls;
184	unsigned long long int memalign_storage_request, memalign_storage_alloc;
185	unsigned int amemalign_calls, amemalign_0_calls;
186	unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
187	unsigned int cmemalign_calls, cmemalign_0_calls;
188	unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
189	unsigned int resize_calls, resize_0_calls;
190	unsigned long long int resize_storage_request, resize_storage_alloc;
191	unsigned int realloc_calls, realloc_0_calls;
192	unsigned long long int realloc_storage_request, realloc_storage_alloc;
193	unsigned int free_calls, free_null_calls;
194	unsigned long long int free_storage_request, free_storage_alloc;
195	unsigned int return_pulls, return_pushes;
196	unsigned long long int return_storage_request, return_storage_alloc;
197	unsigned int mmap_calls, mmap_0_calls; // no zero calls
198	unsigned long long int mmap_storage_request, mmap_storage_alloc;
199	unsigned int munmap_calls, munmap_0_calls; // no zero calls
200	unsigned long long int munmap_storage_request, munmap_storage_alloc;
201	};
202	struct StatsOverlay counters[CntTriples]; // overlay for iteration
203	}; // HeapStatistics
204
205	static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(StatsOverlay),
206	"Heap statistics counter-triplets does not match with array size" );
207
208	static void HeapStatisticsCtor( HeapStatistics & stats ) {
209	memset( &stats, '\0', sizeof(stats) ); // very fast
210	// for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
211	// stats.counters[i].calls = stats.counters[i].calls_0 = stats.counters[i].request = stats.counters[i].alloc = 0;
212	// } // for
213	} // HeapStatisticsCtor
214
215	static HeapStatistics & ?+=?( HeapStatistics & lhs, const HeapStatistics & rhs ) {
216	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
217	lhs.counters[i].calls += rhs.counters[i].calls;
218	lhs.counters[i].calls_0 += rhs.counters[i].calls_0;
219	lhs.counters[i].request += rhs.counters[i].request;
220	lhs.counters[i].alloc += rhs.counters[i].alloc;
221	} // for
222	return lhs;
223	} // ?+=?
224	#endif // __STATISTICS__
225
226
227	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
228	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
229	enum { NoBucketSizes = 91 }; // number of buckets sizes
230
231	struct Heap {
232	struct Storage {
233	struct Header { // header
234	union Kind {
235	struct RealHeader {
236	union {
237	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
238	union {
239	// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
240	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
241	void * home; // allocated block points back to home locations (must overlay alignment)
242	size_t blockSize; // size for munmap (must overlay alignment)
243	Storage * next; // freed block points to next freed block of same size
244	};
245	size_t size; // allocation size in bytes
246	};
247	};
248	} real; // RealHeader
249
250	struct FakeHeader {
251	uintptr_t alignment; // 1st low-order bit => fake header & alignment
252	uintptr_t offset;
253	} fake; // FakeHeader
254	} kind; // Kind
255	} header; // Header
256
257	char pad[libAlign() - sizeof( Header )];
258	char data[0]; // storage
259	}; // Storage
260
261	static_assert( libAlign() >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
262
263	struct __attribute__(( aligned (8) )) FreeHeader {
264	size_t blockSize __attribute__(( aligned(8) )); // size of allocations on this list
265	#ifdef OWNERSHIP
266	#ifdef RETURNSPIN
267	SpinLock_t returnLock;
268	#endif // RETURNSPIN
269	Storage * returnList; // other thread return list
270	#endif // OWNERSHIP
271
272	Storage * freeList; // thread free list
273	Heap * homeManager; // heap owner (free storage to bucket, from bucket to heap)
274	}; // FreeHeader
275
276	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
277	void * heapBuffer; // start of free storage in buffer
278	size_t heapReserve; // amount of remaining free storage in buffer
279
280	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
281	Heap * nextHeapManager; // intrusive link of existing heaps; traversed to collect statistics or check unfreed storage
282	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
283	Heap * nextFreeHeapManager; // intrusive link of free heaps from terminated threads; reused by new threads
284
285	#ifdef __CFA_DEBUG__
286	int64_t allocUnfreed; // running total of allocations minus frees; can be negative
287	#endif // __CFA_DEBUG__
288
289	#ifdef __STATISTICS__
290	HeapStatistics stats; // local statistic table for this heap
291	#endif // __STATISTICS__
292	}; // Heap
293
294
295	struct HeapMaster {
296	SpinLock_t extLock; // protects allocation-buffer extension
297	SpinLock_t mgrLock; // protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd
298
299	void * heapBegin; // start of heap
300	void * heapEnd; // logical end of heap
301	size_t heapRemaining; // amount of storage not allocated in the current chunk
302	size_t pageSize; // architecture pagesize
303	size_t heapExpand; // sbrk advance
304	size_t mmapStart; // cross over point for mmap
305	unsigned int maxBucketsUsed; // maximum number of buckets in use
306
307	Heap * heapManagersList; // heap-list head
308	Heap * freeHeapManagersList; // free-list head
309
310	// Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
311	Heap * heapManagersStorage; // next heap to use in heap superblock
312	Heap * heapManagersStorageEnd; // logical heap outside of superblock's end
313
314	#ifdef __STATISTICS__
315	HeapStatistics stats; // global stats for thread-local heaps to add there counters when exiting
316	unsigned long int threads_started, threads_exited; // counts threads that have started and exited
317	unsigned long int reused_heap, new_heap; // counts reusability of heaps
318	unsigned int sbrk_calls;
319	unsigned long long int sbrk_storage;
320	int stats_fd;
321	#endif // __STATISTICS__
322	}; // HeapMaster
323
324
325	#ifdef FASTLOOKUP
326	enum { LookupSizes = 65_536 + sizeof(Heap.Storage) }; // number of fast lookup sizes
327	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
328	#endif // FASTLOOKUP
329
330	static volatile bool heapMasterBootFlag = false; // trigger for first heap
331	static HeapMaster heapMaster @= {}; // program global
332
333	static void heapMasterCtor();
334	static void heapMasterDtor();
335	static Heap * getHeap();
336
337
338	// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
339	// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
340	// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
341	static const unsigned int bucketSizes[] @= { // different bucket sizes
342	16 + sizeof(Heap.Storage), 32 + sizeof(Heap.Storage), 48 + sizeof(Heap.Storage), 64 + sizeof(Heap.Storage), // 4
343	96 + sizeof(Heap.Storage), 112 + sizeof(Heap.Storage), 128 + sizeof(Heap.Storage), // 3
344	160, 192, 224, 256 + sizeof(Heap.Storage), // 4
345	320, 384, 448, 512 + sizeof(Heap.Storage), // 4
346	640, 768, 896, 1_024 + sizeof(Heap.Storage), // 4
347	1_536, 2_048 + sizeof(Heap.Storage), // 2
348	2_560, 3_072, 3_584, 4_096 + sizeof(Heap.Storage), // 4
349	6_144, 8_192 + sizeof(Heap.Storage), // 2
350	9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(Heap.Storage), // 8
351	18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(Heap.Storage), // 8
352	36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(Heap.Storage), // 8
353	73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(Heap.Storage), // 8
354	147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(Heap.Storage), // 8
355	294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(Heap.Storage), // 8
356	655_360, 786_432, 917_504, 1_048_576 + sizeof(Heap.Storage), // 4
357	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
358	2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(Heap.Storage), // 4
359	};
360
361	static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
362
363
364	// extern visibility, used by runtime kernel
365	libcfa_public size_t __page_size; // architecture pagesize
366	libcfa_public int __map_prot; // common mmap/mprotect protection
367
368
369	// Thread-local storage is allocated lazily when the storage is accessed.
370	static __thread size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
371	static __thread Heap * volatile heapManager CALIGN TLSMODEL;
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	} // with
492	} // if
493
494	return heap;
495	} // getHeap
496
497
498	void heapManagerCtor() libcfa_public {
499	if ( unlikely( ! heapMasterBootFlag ) ) heapMasterCtor();
500
501	lock( heapMaster.mgrLock ); // protect heapMaster counters
502
503	// get storage for heap manager
504
505	heapManager = getHeap();
506
507	#ifdef __STATISTICS__
508	HeapStatisticsCtor( heapManager->stats ); // heap local
509	heapMaster.threads_started += 1;
510	#endif // __STATISTICS__
511
512	unlock( heapMaster.mgrLock );
513	} // heapManagerCtor
514
515
516	void heapManagerDtor() libcfa_public {
517	lock( heapMaster.mgrLock );
518
519	// place heap on list of free heaps for reusability
520	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
521	heapMaster.freeHeapManagersList = heapManager;
522
523	#ifdef __STATISTICS__
524	heapMaster.threads_exited += 1;
525	#endif // __STATISTICS__
526
527	// Do not set heapManager to NULL because it is used after Cforall is shutdown but before the program shuts down.
528
529	unlock( heapMaster.mgrLock );
530	} // heapManagerDtor
531
532
533	//####################### Memory Allocation Routines Helpers ####################
534
535
536	extern int cfa_main_returned; // from interpose.cfa
537	extern "C" {
538	void memory_startup( void ) {
539	if ( ! heapMasterBootFlag ) heapManagerCtor(); // sanity check
540	} // memory_startup
541
542	void memory_shutdown( void ) {
543	heapManagerDtor();
544	} // memory_shutdown
545
546	void heapAppStart() { // called by __cfaabi_appready_startup
547	verify( heapManager );
548	#ifdef __CFA_DEBUG__
549	heapManager->allocUnfreed = 0; // clear prior allocation counts
550	#endif // __CFA_DEBUG__
551
552	#ifdef __STATISTICS__
553	HeapStatisticsCtor( heapManager->stats ); // clear prior statistic counters
554	#endif // __STATISTICS__
555	} // heapAppStart
556
557	void heapAppStop() { // called by __cfaabi_appready_startdown
558	fclose( stdin ); fclose( stdout ); // free buffer storage
559	if ( ! cfa_main_returned ) return; // do not check unfreed storage if exit called
560
561	#ifdef __CFA_DEBUG__
562	// allocUnfreed is set to 0 when a heap is created and it accumulates any unfreed storage during its multiple thread
563	// usages. At the end, add up each heap allocUnfreed value across all heaps to get the total unfreed storage.
564	int64_t allocUnfreed = 0;
565	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
566	allocUnfreed += heap->allocUnfreed;
567	} // for
568
569	allocUnfreed -= malloc_unfreed(); // subtract any user specified unfreed storage
570	if ( allocUnfreed > 0 ) {
571	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
572	char helpText[512];
573	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
574	"CFA warning (UNIX pid:%ld) : program terminating with %ju(0x%jx) bytes of storage allocated but not freed.\n"
575	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
576	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
577	} // if
578	#endif // __CFA_DEBUG__
579	} // heapAppStop
580	} // extern "C"
581
582
583	#ifdef __STATISTICS__
584	static HeapStatistics stats; // zero filled
585
586	#define prtFmt \
587	"\nHeap statistics: (storage request / allocation)\n" \
588	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
589	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
590	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
591	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
592	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
593	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
594	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
595	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
596	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
597	" return pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n" \
598	" sbrk calls %'u; storage %'llu bytes\n" \
599	" mmap calls %'u; storage %'llu / %'llu bytes\n" \
600	" munmap calls %'u; storage %'llu / %'llu bytes\n" \
601	" threads started %'lu; exited %'lu\n" \
602	" heaps new %'lu; reused %'lu\n"
603
604	// Use "write" because streams may be shutdown when calls are made.
605	static int printStats( HeapStatistics & stats ) with( heapMaster, stats ) { // see malloc_stats
606	char helpText[sizeof(prtFmt) + 1024]; // space for message and values
607	return __cfaabi_bits_print_buffer( stats_fd, helpText, sizeof(helpText), prtFmt,
608	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
609	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
610	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
611	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
612	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
613	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
614	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
615	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
616	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
617	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
618	sbrk_calls, sbrk_storage,
619	mmap_calls, mmap_storage_request, mmap_storage_alloc,
620	munmap_calls, munmap_storage_request, munmap_storage_alloc,
621	threads_started, threads_exited,
622	new_heap, reused_heap
623	);
624	} // printStats
625
626	#define prtFmtXML \
627	"<malloc version=\"1\">\n" \
628	"<heap nr=\"0\">\n" \
629	"<sizes>\n" \
630	"</sizes>\n" \
631	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
632	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
633	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
634	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
635	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
636	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
637	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
638	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
639	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
640	"<total type=\"return\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
641	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
642	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
643	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
644	"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n" \
645	"<total type=\"heaps\" new=\"%'lu;\" reused=\"%'lu\"/>\n" \
646	"</malloc>"
647
648	static int printStatsXML( HeapStatistics & stats, FILE * stream ) with( heapMaster, stats ) { // see malloc_info
649	char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
650	return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
651	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
652	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
653	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
654	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
655	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
656	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
657	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
658	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
659	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
660	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
661	sbrk_calls, sbrk_storage,
662	mmap_calls, mmap_storage_request, mmap_storage_alloc,
663	munmap_calls, munmap_storage_request, munmap_storage_alloc,
664	threads_started, threads_exited,
665	new_heap, reused_heap
666	);
667	} // printStatsXML
668
669	static HeapStatistics & collectStats( HeapStatistics & stats ) with( heapMaster ) {
670	lock( mgrLock );
671
672	stats += heapMaster.stats;
673	for ( Heap * heap = heapManagersList; heap; heap = heap->nextHeapManager ) {
674	stats += heap->stats;
675	} // for
676
677	unlock( mgrLock );
678	return stats;
679	} // collectStats
680	#endif // __STATISTICS__
681
682
683	static bool setMmapStart( size_t value ) with( heapMaster ) { // true => mmapped, false => sbrk
684	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
685	mmapStart = value; // set global
686
687	// find the closest bucket size less than or equal to the mmapStart size
688	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
689
690	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
691	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
692	return true;
693	} // setMmapStart
694
695
696	// <-------+----------------------------------------------------> bsize (bucket size)
697	// \|header \|addr
698	//==================================================================================
699	// align/offset \|
700	// <-----------------<------------+-----------------------------> bsize (bucket size)
701	// \|fake-header \| addr
702	#define HeaderAddr( addr ) ((Heap.Storage.Header )( (char )addr - sizeof(Heap.Storage) ))
703	#define RealHeader( header ) ((Heap.Storage.Header )((char )header - header->kind.fake.offset))
704
705	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
706	// \|header \|addr
707	//==================================================================================
708	// align/offset \|
709	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
710	// \|fake-header \|addr
711	#define DataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
712
713
714	inline __attribute__((always_inline))
715	static void checkAlign( size_t alignment ) {
716	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) {
717	abort( "** Error ** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
718	} // if
719	} // checkAlign
720
721
722	inline __attribute__((always_inline))
723	static void checkHeader( bool check, const char name[], void * addr ) {
724	if ( unlikely( check ) ) { // bad address ?
725	abort( "** Error ** attempt to %s storage %p with address outside the heap.\n"
726	"Possible cause is duplicate free on same block or overwriting of memory.",
727	name, addr );
728	} // if
729	} // checkHeader
730
731
732	// Manipulate sticky bits stored in unused 3 low-order bits of an address.
733	// bit0 => alignment => fake header
734	// bit1 => zero filled (calloc)
735	// bit2 => mapped allocation versus sbrk
736	#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
737	#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
738	#define MarkAlignmentBit( align ) ((align) \| 1)
739	#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
740	#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
741	#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
742	#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
743	#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize \|= 2)
744	#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
745	#define MarkMmappedBit( size ) ((size) \| 4)
746
747
748	inline __attribute__((always_inline))
749	static void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
750	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
751	alignment = ClearAlignmentBit( header ); // clear flag from value
752	#ifdef __CFA_DEBUG__
753	checkAlign( alignment ); // check alignment
754	#endif // __CFA_DEBUG__
755	header = RealHeader( header ); // backup from fake to real header
756	} else {
757	alignment = libAlign(); // => no fake header
758	} // if
759	} // fakeHeader
760
761
762	inline __attribute__((always_inline))
763	static bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
764	Heap.FreeHeader & freeHead, size_t & size, size_t & alignment ) with( heapMaster, heapManager ) {
765	header = HeaderAddr( addr );
766
767	#ifdef __CFA_DEBUG__
768	checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
769	#endif // __CFA_DEBUG__
770
771	if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
772	freeHead = (Heap.FreeHeader *)(header->kind.real.home);
773	alignment = libAlign();
774	} else {
775	fakeHeader( header, alignment );
776	if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
777	verify( addr < heapBegin \|\| heapEnd < addr );
778	size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
779	return true;
780	} // if
781
782	freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
783	} // if
784	size = freeHead->blockSize;
785
786	#ifdef __CFA_DEBUG__
787	checkHeader( header < (Heap.Storage.Header )heapBegin \|\| (Heap.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
788
789	Heap * homeManager;
790	if ( unlikely( freeHead == 0p \|\| // freed and only free-list node => null link
791	// freed and link points at another free block not to a bucket in the bucket array.
792	(homeManager = freeHead->homeManager, freeHead < &homeManager->freeLists[0] \|\|
793	&homeManager->freeLists[NoBucketSizes] <= freeHead ) ) ) {
794	abort( "** Error ** attempt to %s storage %p with corrupted header.\n"
795	"Possible cause is duplicate free on same block or overwriting of header information.",
796	name, addr );
797	} // if
798	#endif // __CFA_DEBUG__
799
800	return false;
801	} // headers
802
803
804	static void * master_extend( size_t size ) with( heapMaster ) {
805	lock( extLock );
806
807	ptrdiff_t rem = heapRemaining - size;
808	if ( unlikely( rem < 0 ) ) {
809	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
810
811	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
812	// Do not call abort or strerror( errno ) as they may call malloc.
813	if ( unlikely( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
814	unlock( extLock );
815	abort( NO_MEMORY_MSG, size ); // no memory
816	} // if
817
818	// Make storage executable for thunks.
819	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
820	unlock( extLock );
821	abort( "** Error ** attempt to make heap storage executable for thunks and mprotect failed with errno %d.", errno );
822	} // if
823
824	rem = heapRemaining + increase - size;
825
826	#ifdef __STATISTICS__
827	sbrk_calls += 1;
828	sbrk_storage += increase;
829	#endif // __STATISTICS__
830	} // if
831
832	Heap.Storage * block = (Heap.Storage *)heapEnd;
833	heapRemaining = rem;
834	heapEnd = (char *)heapEnd + size;
835
836	unlock( extLock );
837	return block;
838	} // master_extend
839
840
841	__attribute__(( noinline ))
842	static void * manager_extend( size_t size ) with( *heapManager ) {
843	ptrdiff_t rem = heapReserve - size;
844
845	if ( unlikely( rem < 0 ) ) { // negative
846	// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
847	// smaller freeLists, and increase the reserve.
848
849	rem = heapReserve; // positive
850
851	if ( rem >= bucketSizes[0] ) { // minimal size ? otherwise ignore
852	size_t bucket;
853	#ifdef FASTLOOKUP
854	if ( likely( rem < LookupSizes ) ) bucket = lookup[rem];
855	#endif // FASTLOOKUP
856	bucket = Bsearchl( rem, bucketSizes, heapMaster.maxBucketsUsed );
857	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
858	Heap.FreeHeader * freeHead = &(freeLists[bucket]);
859
860	// The remaining storage many not be bucket size, whereas all other allocations are. Round down to previous
861	// bucket size in this case.
862	if ( unlikely( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
863	Heap.Storage * block = (Heap.Storage *)heapBuffer;
864
865	block->header.kind.real.next = freeHead->freeList; // push on stack
866	freeHead->freeList = block;
867	} // if
868
869	size_t increase = ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), libAlign() );
870	heapBuffer = master_extend( increase );
871	rem = increase - size;
872	} // if
873
874	Heap.Storage * block = (Heap.Storage *)heapBuffer;
875	heapReserve = rem;
876	heapBuffer = (char *)heapBuffer + size;
877
878	return block;
879	} // manager_extend
880
881
882	#define BOOT_HEAP_MANAGER \
883	if ( unlikely( ! heapMasterBootFlag ) ) { \
884	heapManagerCtor(); /* trigger for first heap */ \
885	} /* if */
886
887	#ifdef __STATISTICS__
888	#define STAT_NAME __counter
889	#define STAT_PARM , unsigned int STAT_NAME
890	#define STAT_ARG( name ) , name
891	#define STAT_0_CNT( counter ) stats.counters[counter].calls_0 += 1
892	#else
893	#define STAT_NAME
894	#define STAT_PARM
895	#define STAT_ARG( name )
896	#define STAT_0_CNT( counter )
897	#endif // __STATISTICS__
898
899	#define PROLOG( counter, ... ) \
900	BOOT_HEAP_MANAGER; \
901	if ( unlikely( size == 0 ) \|\| /* 0 BYTE ALLOCATION RETURNS NULL POINTER */ \
902	unlikely( size > ULONG_MAX - sizeof(Heap.Storage) ) ) { /* error check */ \
903	STAT_0_CNT( counter ); \
904	__VA_ARGS__; \
905	return 0p; \
906	} /* if */
907
908
909	#define SCRUB_SIZE 1024lu
910	// Do not use '\xfe' for scrubbing because dereferencing an address composed of it causes a SIGSEGV without a valid IP
911	// pointer in the interrupt frame.
912	#define SCRUB '\xff'
913
914	static void * doMalloc( size_t size STAT_PARM ) libcfa_nopreempt with( *heapManager ) {
915	PROLOG( STAT_NAME );
916
917	verify( heapManager );
918	Heap.Storage * block; // pointer to new block of storage
919
920	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
921	// along with the block and is a multiple of the alignment size.
922	size_t tsize = size + sizeof(Heap.Storage);
923
924	#ifdef __STATISTICS__
925	stats.counters[STAT_NAME].calls += 1;
926	stats.counters[STAT_NAME].request += size;
927	#endif // __STATISTICS__
928
929	#ifdef __CFA_DEBUG__
930	allocUnfreed += size;
931	#endif // __CFA_DEBUG__
932
933	if ( likely( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
934	size_t bucket;
935	#ifdef FASTLOOKUP
936	if ( likely( tsize < LookupSizes ) ) bucket = lookup[tsize];
937	else
938	#endif // FASTLOOKUP
939	bucket = Bsearchl( tsize, bucketSizes, heapMaster.maxBucketsUsed );
940	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
941	Heap.FreeHeader * freeHead = &freeLists[bucket];
942
943	verify( freeHead <= &freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
944	verify( tsize <= freeHead->blockSize ); // search failure ?
945
946	tsize = freeHead->blockSize; // total space needed for request
947	#ifdef __STATISTICS__
948	stats.counters[STAT_NAME].alloc += tsize;
949	#endif // __STATISTICS__
950
951	block = freeHead->freeList; // remove node from stack
952	if ( unlikely( block == 0p ) ) { // no free block ?
953	// Freelist for this size is empty, so check return list (OWNERSHIP), carve it out of the heap, if there
954	// is enough left, or get some more heap storage and carve it off.
955	#ifdef OWNERSHIP
956	if ( unlikely( freeHead->returnList ) ) { // race, get next time if lose race
957	#ifdef RETURNSPIN
958	lock( freeHead->returnLock );
959	block = freeHead->returnList;
960	freeHead->returnList = 0p;
961	unlock( freeHead->returnLock );
962	#else
963	block = __atomic_exchange_n( &freeHead->returnList, 0p, __ATOMIC_SEQ_CST );
964	#endif // RETURNSPIN
965
966	verify( block );
967	#ifdef __STATISTICS__
968	stats.return_pulls += 1;
969	#endif // __STATISTICS__
970
971	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
972
973	freeHead->freeList = block->header.kind.real.next; // merge returnList into freeHead
974	} else {
975	#endif // OWNERSHIP
976	// Do not leave kernel thread as manager_extend accesses heapManager.
977	disable_interrupts();
978	block = (Heap.Storage *)manager_extend( tsize ); // mutual exclusion on call
979	enable_interrupts( false );
980
981	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
982
983	#ifdef __CFA_DEBUG__
984	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes.
985	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
986	#endif // __CFA_DEBUG__
987	#ifdef OWNERSHIP
988	} // if
989	#endif // OWNERSHIP
990	} else {
991	// Memory is scrubbed in doFree.
992	freeHead->freeList = block->header.kind.real.next;
993	} // if
994
995	block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
996	} else { // large size => mmap
997	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
998	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
999
1000	#ifdef __STATISTICS__
1001	stats.counters[STAT_NAME].alloc += tsize;
1002	stats.mmap_calls += 1;
1003	stats.mmap_storage_request += size;
1004	stats.mmap_storage_alloc += tsize;
1005	#endif // __STATISTICS__
1006
1007	disable_interrupts();
1008	block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
1009	enable_interrupts( false );
1010
1011	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1012
1013	if ( unlikely( block == (Heap.Storage *)MAP_FAILED ) ) { // failed ?
1014	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
1015	// Do not call strerror( errno ) as it may call malloc.
1016	abort( "** Error ** attempt to allocate large object (> %zu) of size %zu bytes and mmap failed with errno %d.",
1017	size, heapMaster.mmapStart, errno );
1018	} // if
1019	block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
1020
1021	#ifdef __CFA_DEBUG__
1022	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes. The
1023	// rest of the storage set to 0 by mmap.
1024	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1025	#endif // __CFA_DEBUG__
1026	} // if
1027
1028	block->header.kind.real.size = size; // store allocation size
1029	void * addr = &(block->data); // adjust off header to user bytes
1030	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
1031
1032	#ifdef __CFA_DEBUG__
1033	if ( traceHeap() ) {
1034	char helpText[64];
1035	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1036	"%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
1037	} // if
1038	#endif // __CFA_DEBUG__
1039
1040	// poll_interrupts(); // call rollforward
1041
1042	return addr;
1043	} // doMalloc
1044
1045
1046	static void doFree( void * addr ) libcfa_nopreempt with( *heapManager ) {
1047	verify( addr );
1048
1049	// detect free after thread-local storage destruction and use global stats in that case
1050
1051	Heap.Storage.Header * header;
1052	Heap.FreeHeader * freeHead;
1053	size_t size, alignment;
1054
1055	bool mapped = headers( "free", addr, header, freeHead, size, alignment );
1056	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
1057	size_t rsize = header->kind.real.size; // optimization
1058	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
1059
1060	#ifdef __STATISTICS__
1061	stats.free_storage_request += rsize;
1062	stats.free_storage_alloc += size;
1063	#endif // __STATISTICS__
1064
1065	#ifdef __CFA_DEBUG__
1066	allocUnfreed -= rsize;
1067	#endif // __CFA_DEBUG__
1068
1069	if ( unlikely( mapped ) ) { // mmapped ?
1070	#ifdef __STATISTICS__
1071	stats.munmap_calls += 1;
1072	stats.munmap_storage_request += rsize;
1073	stats.munmap_storage_alloc += size;
1074	#endif // __STATISTICS__
1075
1076	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1077
1078	// Does not matter where this storage is freed.
1079	if ( unlikely( munmap( header, size ) == -1 ) ) {
1080	// Do not call strerror( errno ) as it may call malloc.
1081	abort( "** Error ** attempt to deallocate large object %p and munmap failed with errno %d.\n"
1082	"Possible cause is invalid delete pointer: either not allocated or with corrupt header.",
1083	addr, errno );
1084	} // if
1085	} else {
1086	#ifdef __CFA_DEBUG__
1087	// memset is NOT always inlined!
1088	disable_interrupts();
1089	// Scrub old memory so subsequent usages might fail. Only scrub the first/last SCRUB_SIZE bytes.
1090	char * data = ((Heap.Storage *)header)->data; // data address
1091	size_t dsize = size - sizeof(Heap.Storage); // data size
1092	if ( dsize <= SCRUB_SIZE * 2 ) {
1093	memset( data, SCRUB, dsize ); // scrub all
1094	} else {
1095	memset( data, SCRUB, SCRUB_SIZE ); // scrub front
1096	memset( data + dsize - SCRUB_SIZE, SCRUB, SCRUB_SIZE ); // scrub back
1097	} // if
1098	enable_interrupts( false );
1099	#endif // __CFA_DEBUG__
1100
1101	#ifdef OWNERSHIP
1102	if ( likely( heapManager == freeHead->homeManager ) ) { // belongs to this thread
1103	header->kind.real.next = freeHead->freeList; // push on stack
1104	freeHead->freeList = (Heap.Storage *)header;
1105	} else { // return to thread owner
1106	verify( heapManager );
1107
1108	#ifdef RETURNSPIN
1109	lock( freeHead->returnLock );
1110	header->kind.real.next = freeHead->returnList; // push to bucket return list
1111	freeHead->returnList = (Heap.Storage *)header;
1112	unlock( freeHead->returnLock );
1113	#else // lock free
1114	header->kind.real.next = freeHead->returnList; // link new node to top node
1115	// CAS resets header->kind.real.next = freeHead->returnList on failure
1116	while ( ! __atomic_compare_exchange_n( &freeHead->returnList, &header->kind.real.next, (Heap.Storage *)header,
1117	false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
1118	#endif // RETURNSPIN
1119	} // if
1120
1121	#else // no OWNERSHIP
1122
1123	// kind.real.home is address in owner thread's freeLists, so compute the equivalent position in this thread's freeList.
1124	freeHead = &freeLists[ClearStickyBits( (Heap.FreeHeader *)(header->kind.real.home) ) - &freeHead->homeManager->freeLists[0]];
1125	header->kind.real.next = freeHead->freeList; // push on stack
1126	freeHead->freeList = (Heap.Storage *)header;
1127	#endif // ! OWNERSHIP
1128
1129	#ifdef __U_STATISTICS__
1130	stats.return_pushes += 1;
1131	stats.return_storage_request += rsize;
1132	stats.return_storage_alloc += size;
1133	#endif // __U_STATISTICS__
1134
1135	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1136	} // if
1137
1138	#ifdef __CFA_DEBUG__
1139	if ( traceHeap() ) {
1140	char helpText[64];
1141	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1142	"Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
1143	} // if
1144	#endif // __CFA_DEBUG__
1145
1146	// poll_interrupts(); // call rollforward
1147	} // doFree
1148
1149
1150	size_t prtFree( Heap & manager ) with( manager ) {
1151	size_t total = 0;
1152	#ifdef __STATISTICS__
1153	__cfaabi_bits_acquire();
1154	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
1155	#endif // __STATISTICS__
1156	for ( unsigned int i = 0; i < heapMaster.maxBucketsUsed; i += 1 ) {
1157	size_t size = freeLists[i].blockSize;
1158	#ifdef __STATISTICS__
1159	unsigned int N = 0;
1160	#endif // __STATISTICS__
1161
1162	for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
1163	total += size;
1164	#ifdef __STATISTICS__
1165	N += 1;
1166	#endif // __STATISTICS__
1167	} // for
1168
1169	#ifdef __STATISTICS__
1170	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
1171	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
1172	#endif // __STATISTICS__
1173	} // for
1174	#ifdef __STATISTICS__
1175	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
1176	__cfaabi_bits_release();
1177	#endif // __STATISTICS__
1178	return (char )heapMaster.heapEnd - (char )heapMaster.heapBegin - total;
1179	} // prtFree
1180
1181
1182	#ifdef __STATISTICS__
1183	static void incCalls( intptr_t statName ) libcfa_nopreempt {
1184	heapManager->stats.counters[statName].calls += 1;
1185	} // incCalls
1186
1187	static void incZeroCalls( intptr_t statName ) libcfa_nopreempt {
1188	heapManager->stats.counters[statName].calls_0 += 1;
1189	} // incZeroCalls
1190	#endif // __STATISTICS__
1191
1192	#ifdef __CFA_DEBUG__
1193	static void incUnfreed( intptr_t offset ) libcfa_nopreempt {
1194	heapManager->allocUnfreed += offset;
1195	} // incUnfreed
1196	#endif // __CFA_DEBUG__
1197
1198
1199	static void * memalignNoStats( size_t alignment, size_t size STAT_PARM ) {
1200	checkAlign( alignment ); // check alignment
1201
1202	// if alignment <= default alignment or size == 0, do normal malloc as two headers are unnecessary
1203	if ( unlikely( alignment <= libAlign() \|\| size == 0 ) ) return doMalloc( size STAT_ARG( STAT_NAME ) );
1204
1205	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1206	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1207	// .-------------v-----------------v----------------v----------,
1208	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
1209	// `-------------^-----------------^-+--------------^----------'
1210	// \|<--------------------------------' offset/align \|<-- alignment boundary
1211
1212	// subtract libAlign() because it is already the minimum alignment
1213	// add sizeof(Storage) for fake header
1214	size_t offset = alignment - libAlign() + sizeof(Heap.Storage);
1215	char * addr = (char *)doMalloc( size + offset STAT_ARG( STAT_NAME ) );
1216
1217	// address in the block of the "next" alignment address
1218	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
1219
1220	// address of header from malloc
1221	Heap.Storage.Header * realHeader = HeaderAddr( addr );
1222	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1223	#ifdef __CFA_DEBUG__
1224	incUnfreed( -offset ); // adjustment off the offset from call to doMalloc
1225	#endif // __CFA_DEBUG__
1226
1227	// address of fake header before the alignment location
1228	Heap.Storage.Header * fakeHeader = HeaderAddr( user );
1229
1230	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1231	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
1232	// SKULLDUGGERY: odd alignment implies fake header
1233	fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
1234
1235	return user;
1236	} // memalignNoStats
1237
1238
1239	//####################### Memory Allocation Routines ####################
1240
1241
1242	extern "C" {
1243	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1244	// then malloc() returns a unique pointer value that can later be successfully passed to free().
1245	void * malloc( size_t size ) libcfa_public {
1246	return doMalloc( size STAT_ARG( MALLOC ) );
1247	} // malloc
1248
1249
1250	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1251	void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
1252	return doMalloc( dim * elemSize STAT_ARG( AALLOC ) );
1253	} // aalloc
1254
1255
1256	// Same as aalloc() with memory set to zero.
1257	void * calloc( size_t dim, size_t elemSize ) libcfa_public {
1258	size_t size = dim * elemSize;
1259	char * addr = (char *)doMalloc( size STAT_ARG( CALLOC ) );
1260
1261	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1262
1263	Heap.Storage.Header * header;
1264	Heap.FreeHeader * freeHead;
1265	size_t bsize, alignment;
1266
1267	#ifndef __CFA_DEBUG__
1268	bool mapped =
1269	#endif // __CFA_DEBUG__
1270	headers( "calloc", addr, header, freeHead, bsize, alignment );
1271
1272	#ifndef __CFA_DEBUG__
1273	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1274	if ( likely( ! mapped ) )
1275	#endif // __CFA_DEBUG__
1276	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1277	// `-header`-addr `-size
1278	memset( addr, '\0', size ); // set to zeros
1279
1280	MarkZeroFilledBit( header ); // mark as zero fill
1281	return addr;
1282	} // calloc
1283
1284
1285	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1286	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1287	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
1288	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1289	void * resize( void * oaddr, size_t size ) libcfa_public {
1290	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1291	return doMalloc( size STAT_ARG( RESIZE ) );
1292	} // if
1293
1294	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1295
1296	Heap.Storage.Header * header;
1297	Heap.FreeHeader * freeHead;
1298	size_t bsize, oalign;
1299	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1300
1301	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1302	// same size, DO NOT preserve STICKY PROPERTIES.
1303	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1304	ClearZeroFillBit( header ); // no alignment and turn off 0 fill
1305	#ifdef __CFA_DEBUG__
1306	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1307	#endif // __CFA_DEBUG__
1308	header->kind.real.size = size; // reset allocation size
1309	#ifdef __STATISTICS__
1310	incCalls( RESIZE );
1311	#endif // __STATISTICS__
1312	return oaddr;
1313	} // if
1314
1315	// change size, DO NOT preserve STICKY PROPERTIES.
1316	doFree( oaddr ); // free previous storage
1317
1318	return doMalloc( size STAT_ARG( RESIZE ) ); // create new area
1319	} // resize
1320
1321
1322	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1323	// the old and new sizes.
1324	void * realloc( void * oaddr, size_t size ) libcfa_public {
1325	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1326	return doMalloc( size STAT_ARG( REALLOC ) );
1327	} // if
1328
1329	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1330
1331	Heap.Storage.Header * header;
1332	Heap.FreeHeader * freeHead;
1333	size_t bsize, oalign;
1334	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1335
1336	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1337	size_t osize = header->kind.real.size; // old allocation size
1338	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1339	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1340	#ifdef __CFA_DEBUG__
1341	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1342	#endif // __CFA_DEBUG__
1343	header->kind.real.size = size; // reset allocation size
1344	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1345	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1346	} // if
1347	#ifdef __STATISTICS__
1348	incCalls( REALLOC );
1349	#endif // __STATISTICS__
1350	return oaddr;
1351	} // if
1352
1353	// change size and copy old content to new storage
1354
1355	void * naddr;
1356	if ( likely( oalign <= libAlign() ) ) { // previous request not aligned ?
1357	naddr = doMalloc( size STAT_ARG( REALLOC ) ); // create new area
1358	} else {
1359	naddr = memalignNoStats( oalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1360	} // if
1361
1362	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1363	// To preserve prior fill, the entire bucket must be copied versus the size.
1364	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1365	doFree( oaddr ); // free previous storage
1366
1367	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1368	MarkZeroFilledBit( header ); // mark new request as zero filled
1369	if ( size > osize ) { // previous request larger ?
1370	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1371	} // if
1372	} // if
1373	return naddr;
1374	} // realloc
1375
1376
1377	// Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1378	void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1379	return realloc( oaddr, dim * elemSize );
1380	} // reallocarray
1381
1382
1383	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1384	void * memalign( size_t alignment, size_t size ) libcfa_public {
1385	return memalignNoStats( alignment, size STAT_ARG( MEMALIGN ) );
1386	} // memalign
1387
1388
1389	// Same as aalloc() with memory alignment.
1390	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1391	return memalignNoStats( alignment, dim * elemSize STAT_ARG( AMEMALIGN ) );
1392	} // amemalign
1393
1394
1395	// Same as calloc() with memory alignment.
1396	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1397	size_t size = dim * elemSize;
1398	char * addr = (char *)memalignNoStats( alignment, size STAT_ARG( CMEMALIGN ) );
1399
1400	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1401
1402	Heap.Storage.Header * header;
1403	Heap.FreeHeader * freeHead;
1404	size_t bsize;
1405
1406	#ifndef __CFA_DEBUG__
1407	bool mapped =
1408	#endif // __CFA_DEBUG__
1409	headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1410
1411	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1412	#ifndef __CFA_DEBUG__
1413	if ( ! mapped )
1414	#endif // __CFA_DEBUG__
1415	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1416	// `-header`-addr `-size
1417	memset( addr, '\0', size ); // set to zeros
1418
1419	MarkZeroFilledBit( header ); // mark as zero filled
1420	return addr;
1421	} // cmemalign
1422
1423
1424	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1425	// of alignment. This requirement is universally ignored.
1426	void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1427	return memalign( alignment, size );
1428	} // aligned_alloc
1429
1430
1431	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1432	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1433	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1434	// free(3).
1435	int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1436	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1437	*memptr = memalign( alignment, size );
1438	return 0;
1439	} // posix_memalign
1440
1441
1442	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1443	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1444	void * valloc( size_t size ) libcfa_public {
1445	return memalign( __page_size, size );
1446	} // valloc
1447
1448
1449	// Same as valloc but rounds size to multiple of page size.
1450	void * pvalloc( size_t size ) libcfa_public {
1451	return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1452	} // pvalloc
1453
1454
1455	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1456	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1457	// 0p, no operation is performed.
1458	void free( void * addr ) libcfa_public {
1459	// verify( heapManager );
1460
1461	if ( unlikely( addr == 0p ) ) { // special case
1462	#ifdef __STATISTICS__
1463	if ( heapManager )
1464	incZeroCalls( FREE );
1465	#endif // __STATISTICS__
1466	return;
1467	} // if
1468
1469	#ifdef __STATISTICS__
1470	incCalls( FREE );
1471	#endif // __STATISTICS__
1472
1473	doFree( addr ); // handles heapManager == nullptr
1474	} // free
1475
1476
1477	// Returns the alignment of an allocation.
1478	size_t malloc_alignment( void * addr ) libcfa_public {
1479	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1480	Heap.Storage.Header * header = HeaderAddr( addr );
1481	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1482	return ClearAlignmentBit( header ); // clear flag from value
1483	} else {
1484	return libAlign(); // minimum alignment
1485	} // if
1486	} // malloc_alignment
1487
1488
1489	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1490	bool malloc_zero_fill( void * addr ) libcfa_public {
1491	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1492	Heap.Storage.Header * header = HeaderAddr( addr );
1493	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1494	header = RealHeader( header ); // backup from fake to real header
1495	} // if
1496	return ZeroFillBit( header ); // zero filled ?
1497	} // malloc_zero_fill
1498
1499
1500	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1501	size_t malloc_size( void * addr ) libcfa_public {
1502	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1503	Heap.Storage.Header * header = HeaderAddr( addr );
1504	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1505	header = RealHeader( header ); // backup from fake to real header
1506	} // if
1507	return header->kind.real.size;
1508	} // malloc_size
1509
1510
1511	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1512	// malloc or a related function.
1513	size_t malloc_usable_size( void * addr ) libcfa_public {
1514	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1515	Heap.Storage.Header * header;
1516	Heap.FreeHeader * freeHead;
1517	size_t bsize, alignment;
1518
1519	headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1520	return DataStorage( bsize, addr, header ); // data storage in bucket
1521	} // malloc_usable_size
1522
1523
1524	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1525	void malloc_stats( void ) libcfa_public {
1526	#ifdef __STATISTICS__
1527	HeapStatistics stats;
1528	HeapStatisticsCtor( stats );
1529	if ( printStats( collectStats( stats ) ) == -1 ) {
1530	#else
1531	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1532	if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1533	#endif // __STATISTICS__
1534	abort( "** Error ** write failed in malloc_stats" );
1535	} // if
1536	} // malloc_stats
1537
1538
1539	// Changes the file descriptor where malloc_stats() writes statistics.
1540	int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1541	#ifdef __STATISTICS__
1542	int temp = heapMaster.stats_fd;
1543	heapMaster.stats_fd = fd;
1544	return temp;
1545	#else
1546	return -1; // unsupported
1547	#endif // __STATISTICS__
1548	} // malloc_stats_fd
1549
1550
1551	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1552	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1553	// malloc).
1554	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1555	if ( options != 0 ) { errno = EINVAL; return -1; }
1556	#ifdef __STATISTICS__
1557	HeapStatistics stats;
1558	HeapStatisticsCtor( stats );
1559	return printStatsXML( collectStats( stats ), stream ); // returns bytes written or -1
1560	#else
1561	return 0; // unsupported
1562	#endif // __STATISTICS__
1563	} // malloc_info
1564
1565
1566	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1567	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1568	int mallopt( int option, int value ) libcfa_public {
1569	if ( value < 0 ) return 0;
1570	choose( option ) {
1571	case M_TOP_PAD:
1572	heapMaster.heapExpand = ceiling2( value, __page_size );
1573	return 1;
1574	case M_MMAP_THRESHOLD:
1575	if ( setMmapStart( value ) ) return 1;
1576	} // choose
1577	return 0; // error, unsupported
1578	} // mallopt
1579
1580
1581	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1582	int malloc_trim( size_t ) libcfa_public {
1583	return 0; // => impossible to release memory
1584	} // malloc_trim
1585
1586
1587	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1588	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1589	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1590	// result. (The caller must free this memory.)
1591	void * malloc_get_state( void ) libcfa_public {
1592	return 0p; // unsupported
1593	} // malloc_get_state
1594
1595
1596	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1597	// structure pointed to by state.
1598	int malloc_set_state( void * ) libcfa_public {
1599	return 0; // unsupported
1600	} // malloc_set_state
1601
1602
1603	// Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1604	__attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1605
1606	// Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1607	__attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1608
1609	// Amount subtracted to adjust for unfreed program storage (debug only).
1610	__attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1611	} // extern "C"
1612
1613
1614	// Must have CFA linkage to overload with C linkage realloc.
1615	void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1616	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1617	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) );
1618	} // if
1619
1620	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1621
1622	// Attempt to reuse existing alignment.
1623	Heap.Storage.Header * header = HeaderAddr( oaddr );
1624	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1625	size_t oalign;
1626
1627	if ( unlikely( isFakeHeader ) ) {
1628	checkAlign( nalign ); // check alignment
1629	oalign = ClearAlignmentBit( header ); // old alignment
1630	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1631	&& ( oalign <= nalign // going down
1632	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1633	) ) {
1634	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1635	Heap.FreeHeader * freeHead;
1636	size_t bsize, oalign;
1637	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1638	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1639
1640	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1641	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1642	ClearZeroFillBit( header ); // turn off 0 fill
1643	#ifdef __CFA_DEBUG__
1644	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1645	#endif // __CFA_DEBUG__
1646	header->kind.real.size = size; // reset allocation size
1647	#ifdef __STATISTICS__
1648	incCalls( RESIZE );
1649	#endif // __STATISTICS__
1650	return oaddr;
1651	} // if
1652	} // if
1653	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1654	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1655	return resize( oaddr, size ); // duplicate special case checks
1656	} // if
1657
1658	// change size, DO NOT preserve STICKY PROPERTIES.
1659	doFree( oaddr ); // free previous storage
1660	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) ); // create new aligned area
1661	} // resize
1662
1663
1664	void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1665	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1666	return memalignNoStats( nalign, size STAT_ARG( REALLOC ) );
1667	} // if
1668
1669	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1670
1671	// Attempt to reuse existing alignment.
1672	Heap.Storage.Header * header = HeaderAddr( oaddr );
1673	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1674	size_t oalign;
1675	if ( unlikely( isFakeHeader ) ) {
1676	checkAlign( nalign ); // check alignment
1677	oalign = ClearAlignmentBit( header ); // old alignment
1678	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1679	&& ( oalign <= nalign // going down
1680	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1681	) ) {
1682	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1683	return realloc( oaddr, size ); // duplicate special case checks
1684	} // if
1685	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1686	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1687	return realloc( oaddr, size ); // duplicate special case checks
1688	} // if
1689
1690	Heap.FreeHeader * freeHead;
1691	size_t bsize;
1692	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1693
1694	// change size and copy old content to new storage
1695
1696	size_t osize = header->kind.real.size; // old allocation size
1697	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1698
1699	void * naddr = memalignNoStats( nalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1700
1701	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1702	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1703	doFree( oaddr ); // free previous storage
1704
1705	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1706	MarkZeroFilledBit( header ); // mark new request as zero filled
1707	if ( size > osize ) { // previous request larger ?
1708	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1709	} // if
1710	} // if
1711	return naddr;
1712	} // realloc
1713
1714
1715	void * reallocarray( void * oaddr, size_t nalign, size_t dim, size_t elemSize ) __THROW {
1716	return realloc( oaddr, nalign, dim * elemSize );
1717	} // reallocarray
1718
1719
1720	// Local Variables: //
1721	// tab-width: 4 //
1722	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1723	// End: //

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

Download in other formats:

Original Format