Context Navigation

heap.cfa @ da6db1a2

Last change on this file since da6db1a2 was a7662b8, checked in by Peter A. Buhr <pabuhr@…>, 23 months ago
add cast from int to unsigned
Property mode set to `100644`
File size: 66.7 KB

Line
1	//
2	// Cforall Version 1.0.0 Copyright (C) 2017 University of Waterloo
3	//
4	// The contents of this file are covered under the licence agreement in the
5	// file "LICENCE" distributed with Cforall.
6	//
7	// heap.cfa --
8	//
9	// Author : Peter A. Buhr
10	// Created On : Tue Dec 19 21:58:35 2017
11	// Last Modified By : Peter A. Buhr
12	// Last Modified On : Fri Dec 30 08:37:37 2022
13	// Update Count : 1605
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 size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing
372
373
374	// declare helper functions for HeapMaster
375	void noMemory(); // forward, called by "builtin_new" when malloc returns 0
376
377
378	void heapMasterCtor() with( heapMaster ) {
379	// Singleton pattern to initialize heap master
380
381	verify( bucketSizes[0] == (16 + sizeof(Heap.Storage)) );
382
383	__page_size = sysconf( _SC_PAGESIZE );
384	__map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC;
385
386	extLock = 0;
387	mgrLock = 0;
388
389	char * end = (char *)sbrk( 0 );
390	heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, libAlign() ) - end ); // move start of heap to multiple of alignment
391	heapRemaining = 0;
392	heapExpand = malloc_expansion();
393	mmapStart = malloc_mmap_start();
394
395	// find the closest bucket size less than or equal to the mmapStart size
396	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
397
398	verify( (mmapStart >= pageSize) && (bucketSizes[NoBucketSizes - 1] >= mmapStart) );
399	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
400	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
401
402	heapManagersList = 0p;
403	freeHeapManagersList = 0p;
404
405	heapManagersStorage = 0p;
406	heapManagersStorageEnd = 0p;
407
408	#ifdef __STATISTICS__
409	HeapStatisticsCtor( stats ); // clear statistic counters
410	threads_started = threads_exited = 0;
411	reused_heap = new_heap = 0;
412	sbrk_calls = sbrk_storage = 0;
413	stats_fd = STDERR_FILENO;
414	#endif // __STATISTICS__
415
416	#ifdef FASTLOOKUP
417	for ( unsigned int i = 0, idx = 0; i < LookupSizes; i += 1 ) {
418	if ( i > bucketSizes[idx] ) idx += 1;
419	lookup[i] = idx;
420	verify( i <= bucketSizes[idx] );
421	verify( (i <= 32 && idx == 0) \|\| (i > bucketSizes[idx - 1]) );
422	} // for
423	#endif // FASTLOOKUP
424
425	heapMasterBootFlag = true;
426	} // heapMasterCtor
427
428
429	#define NO_MEMORY_MSG "** Error ** insufficient heap memory available to allocate %zd new bytes."
430
431	Heap * getHeap() with( heapMaster ) {
432	Heap * heap;
433	if ( freeHeapManagersList ) { // free heap for reused ?
434	heap = freeHeapManagersList;
435	freeHeapManagersList = heap->nextFreeHeapManager;
436
437	#ifdef __STATISTICS__
438	reused_heap += 1;
439	#endif // __STATISTICS__
440	} else { // free heap not found, create new
441	// Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps *
442	// 12K ~= 120K byte superblock. Where 128-heap superblock handles a medium sized multi-processor server.
443	size_t remaining = heapManagersStorageEnd - heapManagersStorage; // remaining free heaps in superblock
444	if ( ! heapManagersStorage \|\| remaining == 0 ) {
445	// Each block of heaps is a multiple of the number of cores on the computer.
446	int HeapDim = get_nprocs(); // get_nprocs_conf does not work
447	size_t size = HeapDim * sizeof( Heap );
448
449	heapManagersStorage = (Heap *)mmap( 0, size, PROT_READ \| PROT_WRITE, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
450	if ( unlikely( heapManagersStorage == (Heap *)MAP_FAILED ) ) { // failed ?
451	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
452	// Do not call strerror( errno ) as it may call malloc.
453	abort( "** Error ** attempt to allocate block of heaps of size %zu bytes and mmap failed with errno %d.", size, errno );
454	} // if
455	heapManagersStorageEnd = &heapManagersStorage[HeapDim]; // outside array
456	} // if
457
458	heap = heapManagersStorage;
459	heapManagersStorage = heapManagersStorage + 1; // bump next heap
460
461	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
462	heap->nextHeapManager = heapManagersList;
463	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
464	heapManagersList = heap;
465
466	#ifdef __STATISTICS__
467	new_heap += 1;
468	#endif // __STATISTICS__
469
470	with( *heap ) {
471	for ( unsigned int j = 0; j < NoBucketSizes; j += 1 ) { // initialize free lists
472	#ifdef OWNERSHIP
473	#ifdef RETURNSPIN
474	freeLists[j].returnLock = 0;
475	freeLists[j].returnList = 0p;
476	#endif // RETURNSPIN
477	#endif // OWNERSHIP
478
479	freeLists[j].freeList = 0p;
480	freeLists[j].homeManager = heap;
481	freeLists[j].blockSize = bucketSizes[j];
482	} // for
483
484	heapBuffer = 0p;
485	heapReserve = 0;
486	nextFreeHeapManager = 0p;
487	#ifdef __CFA_DEBUG__
488	allocUnfreed = 0;
489	#endif // __CFA_DEBUG__
490	} // with
491	} // if
492
493	return heap;
494	} // getHeap
495
496
497	void heapManagerCtor() libcfa_public {
498	if ( unlikely( ! heapMasterBootFlag ) ) heapMasterCtor();
499
500	lock( heapMaster.mgrLock ); // protect heapMaster counters
501
502	// get storage for heap manager
503
504	heapManager = getHeap();
505
506	#ifdef __STATISTICS__
507	HeapStatisticsCtor( heapManager->stats ); // heap local
508	heapMaster.threads_started += 1;
509	#endif // __STATISTICS__
510
511	unlock( heapMaster.mgrLock );
512	} // heapManagerCtor
513
514
515	void heapManagerDtor() libcfa_public {
516	lock( heapMaster.mgrLock );
517
518	// place heap on list of free heaps for reusability
519	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
520	heapMaster.freeHeapManagersList = heapManager;
521
522	#ifdef __STATISTICS__
523	heapMaster.threads_exited += 1;
524	#endif // __STATISTICS__
525
526	// Do not set heapManager to NULL because it is used after Cforall is shutdown but before the program shuts down.
527
528	unlock( heapMaster.mgrLock );
529	} // heapManagerDtor
530
531
532	//####################### Memory Allocation Routines Helpers ####################
533
534
535	extern int cfa_main_returned; // from interpose.cfa
536	extern "C" {
537	void memory_startup( void ) {
538	if ( ! heapMasterBootFlag ) heapManagerCtor(); // sanity check
539	} // memory_startup
540
541	void memory_shutdown( void ) {
542	heapManagerDtor();
543	} // memory_shutdown
544
545	void heapAppStart() { // called by __cfaabi_appready_startup
546	verify( heapManager );
547	#ifdef __CFA_DEBUG__
548	heapManager->allocUnfreed = 0; // clear prior allocation counts
549	#endif // __CFA_DEBUG__
550
551	#ifdef __STATISTICS__
552	HeapStatisticsCtor( heapManager->stats ); // clear prior statistic counters
553	#endif // __STATISTICS__
554	} // heapAppStart
555
556	void heapAppStop() { // called by __cfaabi_appready_startdown
557	fclose( stdin ); fclose( stdout ); // free buffer storage
558	if ( ! cfa_main_returned ) return; // do not check unfreed storage if exit called
559
560	#ifdef __CFA_DEBUG__
561	// allocUnfreed is set to 0 when a heap is created and it accumulates any unfreed storage during its multiple thread
562	// usages. At the end, add up each heap allocUnfreed value across all heaps to get the total unfreed storage.
563	ptrdiff_t allocUnfreed = 0;
564	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
565	allocUnfreed += heap->allocUnfreed;
566	} // for
567
568	allocUnfreed -= malloc_unfreed(); // subtract any user specified unfreed storage
569	if ( allocUnfreed > 0 ) {
570	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
571	char helpText[512];
572	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
573	"CFA warning (UNIX pid:%ld) : program terminating with %td(%#tx) bytes of storage allocated but not freed.\n"
574	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
575	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
576	} // if
577	#endif // __CFA_DEBUG__
578	} // heapAppStop
579	} // extern "C"
580
581
582	#ifdef __STATISTICS__
583	static HeapStatistics stats; // zero filled
584
585	#define prtFmt \
586	"\nHeap statistics: (storage request / allocation)\n" \
587	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
588	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
589	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
590	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
591	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
592	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
593	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
594	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
595	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
596	" return pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n" \
597	" sbrk calls %'u; storage %'llu bytes\n" \
598	" mmap calls %'u; storage %'llu / %'llu bytes\n" \
599	" munmap calls %'u; storage %'llu / %'llu bytes\n" \
600	" threads started %'lu; exited %'lu\n" \
601	" heaps new %'lu; reused %'lu\n"
602
603	// Use "write" because streams may be shutdown when calls are made.
604	static int printStats( HeapStatistics & stats ) with( heapMaster, stats ) { // see malloc_stats
605	char helpText[sizeof(prtFmt) + 1024]; // space for message and values
606	return __cfaabi_bits_print_buffer( stats_fd, helpText, sizeof(helpText), prtFmt,
607	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
608	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
609	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
610	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
611	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
612	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
613	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
614	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
615	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
616	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
617	sbrk_calls, sbrk_storage,
618	mmap_calls, mmap_storage_request, mmap_storage_alloc,
619	munmap_calls, munmap_storage_request, munmap_storage_alloc,
620	threads_started, threads_exited,
621	new_heap, reused_heap
622	);
623	} // printStats
624
625	#define prtFmtXML \
626	"<malloc version=\"1\">\n" \
627	"<heap nr=\"0\">\n" \
628	"<sizes>\n" \
629	"</sizes>\n" \
630	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
631	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
632	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
633	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
634	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
635	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
636	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
637	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
638	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
639	"<total type=\"return\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
640	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
641	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
642	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
643	"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n" \
644	"<total type=\"heaps\" new=\"%'lu;\" reused=\"%'lu\"/>\n" \
645	"</malloc>"
646
647	static int printStatsXML( HeapStatistics & stats, FILE * stream ) with( heapMaster, stats ) { // see malloc_info
648	char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
649	return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
650	malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
651	aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
652	calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
653	memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
654	amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
655	cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
656	resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
657	realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
658	free_calls, free_null_calls, free_storage_request, free_storage_alloc,
659	return_pulls, return_pushes, return_storage_request, return_storage_alloc,
660	sbrk_calls, sbrk_storage,
661	mmap_calls, mmap_storage_request, mmap_storage_alloc,
662	munmap_calls, munmap_storage_request, munmap_storage_alloc,
663	threads_started, threads_exited,
664	new_heap, reused_heap
665	);
666	} // printStatsXML
667
668	static HeapStatistics & collectStats( HeapStatistics & stats ) with( heapMaster ) {
669	lock( mgrLock );
670
671	stats += heapMaster.stats;
672	for ( Heap * heap = heapManagersList; heap; heap = heap->nextHeapManager ) {
673	stats += heap->stats;
674	} // for
675
676	unlock( mgrLock );
677	return stats;
678	} // collectStats
679	#endif // __STATISTICS__
680
681
682	static bool setMmapStart( size_t value ) with( heapMaster ) { // true => mmapped, false => sbrk
683	if ( value < __page_size \|\| bucketSizes[NoBucketSizes - 1] < value ) return false;
684	mmapStart = value; // set global
685
686	// find the closest bucket size less than or equal to the mmapStart size
687	maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
688
689	verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
690	verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
691	return true;
692	} // setMmapStart
693
694
695	// <-------+----------------------------------------------------> bsize (bucket size)
696	// \|header \|addr
697	//==================================================================================
698	// align/offset \|
699	// <-----------------<------------+-----------------------------> bsize (bucket size)
700	// \|fake-header \| addr
701	#define HeaderAddr( addr ) ((Heap.Storage.Header )( (char )addr - sizeof(Heap.Storage) ))
702	#define RealHeader( header ) ((Heap.Storage.Header )((char )header - header->kind.fake.offset))
703
704	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
705	// \|header \|addr
706	//==================================================================================
707	// align/offset \|
708	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
709	// \|fake-header \|addr
710	#define DataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
711
712
713	inline __attribute__((always_inline))
714	static void checkAlign( size_t alignment ) {
715	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) {
716	abort( "** Error ** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
717	} // if
718	} // checkAlign
719
720
721	inline __attribute__((always_inline))
722	static void checkHeader( bool check, const char name[], void * addr ) {
723	if ( unlikely( check ) ) { // bad address ?
724	abort( "** Error ** attempt to %s storage %p with address outside the heap.\n"
725	"Possible cause is duplicate free on same block or overwriting of memory.",
726	name, addr );
727	} // if
728	} // checkHeader
729
730
731	// Manipulate sticky bits stored in unused 3 low-order bits of an address.
732	// bit0 => alignment => fake header
733	// bit1 => zero filled (calloc)
734	// bit2 => mapped allocation versus sbrk
735	#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
736	#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
737	#define MarkAlignmentBit( align ) ((align) \| 1)
738	#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
739	#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
740	#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
741	#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
742	#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize \|= 2)
743	#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
744	#define MarkMmappedBit( size ) ((size) \| 4)
745
746
747	inline __attribute__((always_inline))
748	static void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
749	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
750	alignment = ClearAlignmentBit( header ); // clear flag from value
751	#ifdef __CFA_DEBUG__
752	checkAlign( alignment ); // check alignment
753	#endif // __CFA_DEBUG__
754	header = RealHeader( header ); // backup from fake to real header
755	} else {
756	alignment = libAlign(); // => no fake header
757	} // if
758	} // fakeHeader
759
760
761	inline __attribute__((always_inline))
762	static bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
763	Heap.FreeHeader & freeHead, size_t & size, size_t & alignment ) with( heapMaster, heapManager ) {
764	header = HeaderAddr( addr );
765
766	#ifdef __CFA_DEBUG__
767	checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
768	#endif // __CFA_DEBUG__
769
770	if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
771	freeHead = (Heap.FreeHeader *)(header->kind.real.home);
772	alignment = libAlign();
773	} else {
774	fakeHeader( header, alignment );
775	if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
776	verify( addr < heapBegin \|\| heapEnd < addr );
777	size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
778	return true;
779	} // if
780
781	freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
782	} // if
783	size = freeHead->blockSize;
784
785	#ifdef __CFA_DEBUG__
786	checkHeader( header < (Heap.Storage.Header )heapBegin \|\| (Heap.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
787
788	Heap * homeManager;
789	if ( unlikely( freeHead == 0p \|\| // freed and only free-list node => null link
790	// freed and link points at another free block not to a bucket in the bucket array.
791	(homeManager = freeHead->homeManager, freeHead < &homeManager->freeLists[0] \|\|
792	&homeManager->freeLists[NoBucketSizes] <= freeHead ) ) ) {
793	abort( "** Error ** attempt to %s storage %p with corrupted header.\n"
794	"Possible cause is duplicate free on same block or overwriting of header information.",
795	name, addr );
796	} // if
797	#endif // __CFA_DEBUG__
798
799	return false;
800	} // headers
801
802
803	static void * master_extend( size_t size ) with( heapMaster ) {
804	lock( extLock );
805
806	ptrdiff_t rem = heapRemaining - size;
807	if ( unlikely( rem < 0 ) ) { // negative ?
808	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
809
810	size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
811	if ( unlikely( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
812	unlock( extLock );
813	abort( NO_MEMORY_MSG, size ); // give up
814	} // if
815
816	// Make storage executable for thunks.
817	if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
818	unlock( extLock );
819	abort( "** Error ** attempt to make heap storage executable for thunks and mprotect failed with errno %d.", errno );
820	} // if
821
822	rem = heapRemaining + increase - size;
823
824	#ifdef __STATISTICS__
825	sbrk_calls += 1;
826	sbrk_storage += increase;
827	#endif // __STATISTICS__
828	} // if
829
830	Heap.Storage * block = (Heap.Storage *)heapEnd;
831	heapRemaining = rem;
832	heapEnd = (char *)heapEnd + size;
833
834	unlock( extLock );
835	return block;
836	} // master_extend
837
838
839	__attribute__(( noinline ))
840	static void * manager_extend( size_t size ) with( *heapManager ) {
841	ptrdiff_t rem = heapReserve - size;
842
843	if ( unlikely( rem < 0 ) ) { // negative ?
844	// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
845	// smaller freeLists, and increase the reserve.
846
847	rem = heapReserve; // positive
848
849	if ( (unsigned int)rem >= bucketSizes[0] ) { // minimal size ? otherwise ignore
850	size_t bucket;
851	#ifdef FASTLOOKUP
852	if ( likely( rem < LookupSizes ) ) bucket = lookup[rem];
853	#endif // FASTLOOKUP
854	bucket = Bsearchl( rem, bucketSizes, heapMaster.maxBucketsUsed );
855	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
856	Heap.FreeHeader * freeHead = &(freeLists[bucket]);
857
858	// The remaining storage may not be bucket size, whereas all other allocations are. Round down to previous
859	// bucket size in this case.
860	if ( unlikely( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
861	Heap.Storage * block = (Heap.Storage *)heapBuffer;
862
863	block->header.kind.real.next = freeHead->freeList; // push on stack
864	freeHead->freeList = block;
865	} // if
866
867	size_t increase = ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), libAlign() );
868	heapBuffer = master_extend( increase );
869	rem = increase - size;
870	} // if
871
872	Heap.Storage * block = (Heap.Storage *)heapBuffer;
873	heapReserve = rem;
874	heapBuffer = (char *)heapBuffer + size;
875
876	return block;
877	} // manager_extend
878
879
880	#define BOOT_HEAP_MANAGER \
881	if ( unlikely( ! heapMasterBootFlag ) ) { \
882	heapManagerCtor(); /* trigger for first heap */ \
883	} /* if */
884
885	#ifdef __STATISTICS__
886	#define STAT_NAME __counter
887	#define STAT_PARM , unsigned int STAT_NAME
888	#define STAT_ARG( name ) , name
889	#define STAT_0_CNT( counter ) stats.counters[counter].calls_0 += 1
890	#else
891	#define STAT_NAME
892	#define STAT_PARM
893	#define STAT_ARG( name )
894	#define STAT_0_CNT( counter )
895	#endif // __STATISTICS__
896
897	#define PROLOG( counter, ... ) \
898	BOOT_HEAP_MANAGER; \
899	if ( unlikely( size == 0 ) \|\| /* 0 BYTE ALLOCATION RETURNS NULL POINTER */ \
900	unlikely( size > ULONG_MAX - sizeof(Heap.Storage) ) ) { /* error check */ \
901	STAT_0_CNT( counter ); \
902	__VA_ARGS__; \
903	return 0p; \
904	} /* if */
905
906
907	#define SCRUB_SIZE 1024lu
908	// Do not use '\xfe' for scrubbing because dereferencing an address composed of it causes a SIGSEGV without a valid IP
909	// pointer in the interrupt frame.
910	#define SCRUB '\xff'
911
912	static void * doMalloc( size_t size STAT_PARM ) libcfa_nopreempt with( *heapManager ) {
913	PROLOG( STAT_NAME );
914
915	verify( heapManager );
916	Heap.Storage * block; // pointer to new block of storage
917
918	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
919	// along with the block and is a multiple of the alignment size.
920	size_t tsize = size + sizeof(Heap.Storage);
921
922	#ifdef __STATISTICS__
923	stats.counters[STAT_NAME].calls += 1;
924	stats.counters[STAT_NAME].request += size;
925	#endif // __STATISTICS__
926
927	#ifdef __CFA_DEBUG__
928	allocUnfreed += size;
929	#endif // __CFA_DEBUG__
930
931	if ( likely( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
932	size_t bucket;
933	#ifdef FASTLOOKUP
934	if ( likely( tsize < LookupSizes ) ) bucket = lookup[tsize];
935	else
936	#endif // FASTLOOKUP
937	bucket = Bsearchl( tsize, bucketSizes, heapMaster.maxBucketsUsed );
938	verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
939	Heap.FreeHeader * freeHead = &freeLists[bucket];
940
941	verify( freeHead <= &freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
942	verify( tsize <= freeHead->blockSize ); // search failure ?
943
944	tsize = freeHead->blockSize; // total space needed for request
945	#ifdef __STATISTICS__
946	stats.counters[STAT_NAME].alloc += tsize;
947	#endif // __STATISTICS__
948
949	block = freeHead->freeList; // remove node from stack
950	if ( unlikely( block == 0p ) ) { // no free block ?
951	// Freelist for this size is empty, so check return list (OWNERSHIP), or carve it out of the heap if there
952	// is enough left, or get some more heap storage and carve it off.
953	#ifdef OWNERSHIP
954	if ( unlikely( freeHead->returnList ) ) { // race, get next time if lose race
955	#ifdef RETURNSPIN
956	lock( freeHead->returnLock );
957	block = freeHead->returnList;
958	freeHead->returnList = 0p;
959	unlock( freeHead->returnLock );
960	#else
961	block = __atomic_exchange_n( &freeHead->returnList, 0p, __ATOMIC_SEQ_CST );
962	#endif // RETURNSPIN
963
964	verify( block );
965	#ifdef __STATISTICS__
966	stats.return_pulls += 1;
967	#endif // __STATISTICS__
968
969	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
970
971	freeHead->freeList = block->header.kind.real.next; // merge returnList into freeHead
972	} else {
973	#endif // OWNERSHIP
974	// Do not leave kernel thread as manager_extend accesses heapManager.
975	disable_interrupts();
976	block = (Heap.Storage *)manager_extend( tsize ); // mutual exclusion on call
977	enable_interrupts( false );
978
979	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
980
981	#ifdef __CFA_DEBUG__
982	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes.
983	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
984	#endif // __CFA_DEBUG__
985	#ifdef OWNERSHIP
986	} // if
987	#endif // OWNERSHIP
988	} else {
989	// Memory is scrubbed in doFree.
990	freeHead->freeList = block->header.kind.real.next;
991	} // if
992
993	block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
994	} else { // large size => mmap
995	if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
996	tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
997
998	#ifdef __STATISTICS__
999	stats.counters[STAT_NAME].alloc += tsize;
1000	stats.mmap_calls += 1;
1001	stats.mmap_storage_request += size;
1002	stats.mmap_storage_alloc += tsize;
1003	#endif // __STATISTICS__
1004
1005	disable_interrupts();
1006	block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, -1, 0 );
1007	enable_interrupts( false );
1008
1009	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1010
1011	if ( unlikely( block == (Heap.Storage *)MAP_FAILED ) ) { // failed ?
1012	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
1013	// Do not call strerror( errno ) as it may call malloc.
1014	abort( "** Error ** attempt to allocate large object (> %zu) of size %zu bytes and mmap failed with errno %d.",
1015	size, heapMaster.mmapStart, errno );
1016	} // if
1017	block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
1018
1019	#ifdef __CFA_DEBUG__
1020	// Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first SCRUB_SIZE bytes. The
1021	// rest of the storage set to 0 by mmap.
1022	memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1023	#endif // __CFA_DEBUG__
1024	} // if
1025
1026	block->header.kind.real.size = size; // store allocation size
1027	void * addr = &(block->data); // adjust off header to user bytes
1028	verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
1029
1030	#ifdef __CFA_DEBUG__
1031	if ( traceHeap() ) {
1032	char helpText[64];
1033	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1034	"%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
1035	} // if
1036	#endif // __CFA_DEBUG__
1037
1038	// poll_interrupts(); // call rollforward
1039
1040	return addr;
1041	} // doMalloc
1042
1043
1044	static void doFree( void * addr ) libcfa_nopreempt with( *heapManager ) {
1045	verify( addr );
1046
1047	// detect free after thread-local storage destruction and use global stats in that case
1048
1049	Heap.Storage.Header * header;
1050	Heap.FreeHeader * freeHead;
1051	size_t size, alignment;
1052
1053	bool mapped = headers( "free", addr, header, freeHead, size, alignment );
1054	#if defined( __STATISTICS__ ) \|\| defined( __CFA_DEBUG__ )
1055	size_t rsize = header->kind.real.size; // optimization
1056	#endif // __STATISTICS__ \|\| __CFA_DEBUG__
1057
1058	#ifdef __STATISTICS__
1059	stats.free_storage_request += rsize;
1060	stats.free_storage_alloc += size;
1061	#endif // __STATISTICS__
1062
1063	#ifdef __CFA_DEBUG__
1064	allocUnfreed -= rsize;
1065	#endif // __CFA_DEBUG__
1066
1067	if ( unlikely( mapped ) ) { // mmapped ?
1068	#ifdef __STATISTICS__
1069	stats.munmap_calls += 1;
1070	stats.munmap_storage_request += rsize;
1071	stats.munmap_storage_alloc += size;
1072	#endif // __STATISTICS__
1073
1074	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1075
1076	// Does not matter where this storage is freed.
1077	if ( unlikely( munmap( header, size ) == -1 ) ) {
1078	// Do not call strerror( errno ) as it may call malloc.
1079	abort( "** Error ** attempt to deallocate large object %p and munmap failed with errno %d.\n"
1080	"Possible cause is invalid delete pointer: either not allocated or with corrupt header.",
1081	addr, errno );
1082	} // if
1083	} else {
1084	#ifdef __CFA_DEBUG__
1085	// memset is NOT always inlined!
1086	disable_interrupts();
1087	// Scrub old memory so subsequent usages might fail. Only scrub the first/last SCRUB_SIZE bytes.
1088	char * data = ((Heap.Storage *)header)->data; // data address
1089	size_t dsize = size - sizeof(Heap.Storage); // data size
1090	if ( dsize <= SCRUB_SIZE * 2 ) {
1091	memset( data, SCRUB, dsize ); // scrub all
1092	} else {
1093	memset( data, SCRUB, SCRUB_SIZE ); // scrub front
1094	memset( data + dsize - SCRUB_SIZE, SCRUB, SCRUB_SIZE ); // scrub back
1095	} // if
1096	enable_interrupts( false );
1097	#endif // __CFA_DEBUG__
1098
1099	#ifdef OWNERSHIP
1100	if ( likely( heapManager == freeHead->homeManager ) ) { // belongs to this thread
1101	header->kind.real.next = freeHead->freeList; // push on stack
1102	freeHead->freeList = (Heap.Storage *)header;
1103	} else { // return to thread owner
1104	verify( heapManager );
1105
1106	#ifdef RETURNSPIN
1107	lock( freeHead->returnLock );
1108	header->kind.real.next = freeHead->returnList; // push to bucket return list
1109	freeHead->returnList = (Heap.Storage *)header;
1110	unlock( freeHead->returnLock );
1111	#else // lock free
1112	header->kind.real.next = freeHead->returnList; // link new node to top node
1113	// CAS resets header->kind.real.next = freeHead->returnList on failure
1114	while ( ! __atomic_compare_exchange_n( &freeHead->returnList, &header->kind.real.next, (Heap.Storage *)header,
1115	false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
1116
1117	#ifdef __STATISTICS__
1118	stats.return_pushes += 1;
1119	stats.return_storage_request += rsize;
1120	stats.return_storage_alloc += size;
1121	#endif // __STATISTICS__
1122	#endif // RETURNSPIN
1123	} // if
1124
1125	#else // no OWNERSHIP
1126
1127	// kind.real.home is address in owner thread's freeLists, so compute the equivalent position in this thread's freeList.
1128	freeHead = &freeLists[ClearStickyBits( (Heap.FreeHeader *)(header->kind.real.home) ) - &freeHead->homeManager->freeLists[0]];
1129	header->kind.real.next = freeHead->freeList; // push on stack
1130	freeHead->freeList = (Heap.Storage *)header;
1131	#endif // ! OWNERSHIP
1132
1133	// OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1134	} // if
1135
1136	#ifdef __CFA_DEBUG__
1137	if ( traceHeap() ) {
1138	char helpText[64];
1139	__cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1140	"Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
1141	} // if
1142	#endif // __CFA_DEBUG__
1143
1144	// poll_interrupts(); // call rollforward
1145	} // doFree
1146
1147
1148	size_t prtFree( Heap & manager ) with( manager ) {
1149	size_t total = 0;
1150	#ifdef __STATISTICS__
1151	__cfaabi_bits_acquire();
1152	__cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
1153	#endif // __STATISTICS__
1154	for ( unsigned int i = 0; i < heapMaster.maxBucketsUsed; i += 1 ) {
1155	size_t size = freeLists[i].blockSize;
1156	#ifdef __STATISTICS__
1157	unsigned int N = 0;
1158	#endif // __STATISTICS__
1159
1160	for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
1161	total += size;
1162	#ifdef __STATISTICS__
1163	N += 1;
1164	#endif // __STATISTICS__
1165	} // for
1166
1167	#ifdef __STATISTICS__
1168	__cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
1169	if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
1170	#endif // __STATISTICS__
1171	} // for
1172	#ifdef __STATISTICS__
1173	__cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
1174	__cfaabi_bits_release();
1175	#endif // __STATISTICS__
1176	return (char )heapMaster.heapEnd - (char )heapMaster.heapBegin - total;
1177	} // prtFree
1178
1179
1180	#ifdef __STATISTICS__
1181	static void incCalls( size_t statName ) libcfa_nopreempt {
1182	heapManager->stats.counters[statName].calls += 1;
1183	} // incCalls
1184
1185	static void incZeroCalls( size_t statName ) libcfa_nopreempt {
1186	heapManager->stats.counters[statName].calls_0 += 1;
1187	} // incZeroCalls
1188	#endif // __STATISTICS__
1189
1190	#ifdef __CFA_DEBUG__
1191	static void incUnfreed( intptr_t offset ) libcfa_nopreempt {
1192	heapManager->allocUnfreed += offset;
1193	} // incUnfreed
1194	#endif // __CFA_DEBUG__
1195
1196
1197	static void * memalignNoStats( size_t alignment, size_t size STAT_PARM ) {
1198	checkAlign( alignment ); // check alignment
1199
1200	// if alignment <= default alignment or size == 0, do normal malloc as two headers are unnecessary
1201	if ( unlikely( alignment <= libAlign() \|\| size == 0 ) ) return doMalloc( size STAT_ARG( STAT_NAME ) );
1202
1203	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1204	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1205	// .-------------v-----------------v----------------v----------,
1206	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
1207	// `-------------^-----------------^-+--------------^----------'
1208	// \|<--------------------------------' offset/align \|<-- alignment boundary
1209
1210	// subtract libAlign() because it is already the minimum alignment
1211	// add sizeof(Storage) for fake header
1212	size_t offset = alignment - libAlign() + sizeof(Heap.Storage);
1213	char * addr = (char *)doMalloc( size + offset STAT_ARG( STAT_NAME ) );
1214
1215	// address in the block of the "next" alignment address
1216	char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
1217
1218	// address of header from malloc
1219	Heap.Storage.Header * realHeader = HeaderAddr( addr );
1220	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1221	#ifdef __CFA_DEBUG__
1222	incUnfreed( -offset ); // adjustment off the offset from call to doMalloc
1223	#endif // __CFA_DEBUG__
1224
1225	// address of fake header before the alignment location
1226	Heap.Storage.Header * fakeHeader = HeaderAddr( user );
1227
1228	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1229	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
1230	// SKULLDUGGERY: odd alignment implies fake header
1231	fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
1232
1233	return user;
1234	} // memalignNoStats
1235
1236
1237	//####################### Memory Allocation Routines ####################
1238
1239
1240	extern "C" {
1241	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1242	// then malloc() returns a unique pointer value that can later be successfully passed to free().
1243	void * malloc( size_t size ) libcfa_public {
1244	return doMalloc( size STAT_ARG( MALLOC ) );
1245	} // malloc
1246
1247
1248	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1249	void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
1250	return doMalloc( dim * elemSize STAT_ARG( AALLOC ) );
1251	} // aalloc
1252
1253
1254	// Same as aalloc() with memory set to zero.
1255	void * calloc( size_t dim, size_t elemSize ) libcfa_public {
1256	size_t size = dim * elemSize;
1257	char * addr = (char *)doMalloc( size STAT_ARG( CALLOC ) );
1258
1259	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1260
1261	Heap.Storage.Header * header;
1262	Heap.FreeHeader * freeHead;
1263	size_t bsize, alignment;
1264
1265	#ifndef __CFA_DEBUG__
1266	bool mapped =
1267	#endif // __CFA_DEBUG__
1268	headers( "calloc", addr, header, freeHead, bsize, alignment );
1269
1270	#ifndef __CFA_DEBUG__
1271	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1272	if ( likely( ! mapped ) )
1273	#endif // __CFA_DEBUG__
1274	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1275	// `-header`-addr `-size
1276	memset( addr, '\0', size ); // set to zeros
1277
1278	MarkZeroFilledBit( header ); // mark as zero fill
1279	return addr;
1280	} // calloc
1281
1282
1283	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1284	// 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1285	// not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
1286	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1287	void * resize( void * oaddr, size_t size ) libcfa_public {
1288	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1289	return doMalloc( size STAT_ARG( RESIZE ) );
1290	} // if
1291
1292	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1293
1294	Heap.Storage.Header * header;
1295	Heap.FreeHeader * freeHead;
1296	size_t bsize, oalign;
1297	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1298
1299	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1300	// same size, DO NOT preserve STICKY PROPERTIES.
1301	if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1302	ClearZeroFillBit( header ); // no alignment and turn off 0 fill
1303	#ifdef __CFA_DEBUG__
1304	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1305	#endif // __CFA_DEBUG__
1306	header->kind.real.size = size; // reset allocation size
1307	#ifdef __STATISTICS__
1308	incCalls( RESIZE );
1309	#endif // __STATISTICS__
1310	return oaddr;
1311	} // if
1312
1313	// change size, DO NOT preserve STICKY PROPERTIES.
1314	doFree( oaddr ); // free previous storage
1315
1316	return doMalloc( size STAT_ARG( RESIZE ) ); // create new area
1317	} // resize
1318
1319
1320	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1321	// the old and new sizes.
1322	void * realloc( void * oaddr, size_t size ) libcfa_public {
1323	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1324	return doMalloc( size STAT_ARG( REALLOC ) );
1325	} // if
1326
1327	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1328
1329	Heap.Storage.Header * header;
1330	Heap.FreeHeader * freeHead;
1331	size_t bsize, oalign;
1332	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1333
1334	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1335	size_t osize = header->kind.real.size; // old allocation size
1336	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1337	if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1338	#ifdef __CFA_DEBUG__
1339	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1340	#endif // __CFA_DEBUG__
1341	header->kind.real.size = size; // reset allocation size
1342	if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1343	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1344	} // if
1345	#ifdef __STATISTICS__
1346	incCalls( REALLOC );
1347	#endif // __STATISTICS__
1348	return oaddr;
1349	} // if
1350
1351	// change size and copy old content to new storage
1352
1353	void * naddr;
1354	if ( likely( oalign <= libAlign() ) ) { // previous request not aligned ?
1355	naddr = doMalloc( size STAT_ARG( REALLOC ) ); // create new area
1356	} else {
1357	naddr = memalignNoStats( oalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1358	} // if
1359
1360	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1361	// To preserve prior fill, the entire bucket must be copied versus the size.
1362	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1363	doFree( oaddr ); // free previous storage
1364
1365	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1366	MarkZeroFilledBit( header ); // mark new request as zero filled
1367	if ( size > osize ) { // previous request larger ?
1368	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1369	} // if
1370	} // if
1371	return naddr;
1372	} // realloc
1373
1374
1375	// Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1376	void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1377	return realloc( oaddr, dim * elemSize );
1378	} // reallocarray
1379
1380
1381	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1382	void * memalign( size_t alignment, size_t size ) libcfa_public {
1383	return memalignNoStats( alignment, size STAT_ARG( MEMALIGN ) );
1384	} // memalign
1385
1386
1387	// Same as aalloc() with memory alignment.
1388	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1389	return memalignNoStats( alignment, dim * elemSize STAT_ARG( AMEMALIGN ) );
1390	} // amemalign
1391
1392
1393	// Same as calloc() with memory alignment.
1394	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1395	size_t size = dim * elemSize;
1396	char * addr = (char *)memalignNoStats( alignment, size STAT_ARG( CMEMALIGN ) );
1397
1398	if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1399
1400	Heap.Storage.Header * header;
1401	Heap.FreeHeader * freeHead;
1402	size_t bsize;
1403
1404	#ifndef __CFA_DEBUG__
1405	bool mapped =
1406	#endif // __CFA_DEBUG__
1407	headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1408
1409	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1410	#ifndef __CFA_DEBUG__
1411	if ( ! mapped )
1412	#endif // __CFA_DEBUG__
1413	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1414	// `-header`-addr `-size
1415	memset( addr, '\0', size ); // set to zeros
1416
1417	MarkZeroFilledBit( header ); // mark as zero filled
1418	return addr;
1419	} // cmemalign
1420
1421
1422	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1423	// of alignment. This requirement is universally ignored.
1424	void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1425	return memalign( alignment, size );
1426	} // aligned_alloc
1427
1428
1429	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1430	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1431	// is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1432	// free(3).
1433	int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1434	if ( unlikely( alignment < libAlign() \|\| ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1435	*memptr = memalign( alignment, size );
1436	return 0;
1437	} // posix_memalign
1438
1439
1440	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1441	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1442	void * valloc( size_t size ) libcfa_public {
1443	return memalign( __page_size, size );
1444	} // valloc
1445
1446
1447	// Same as valloc but rounds size to multiple of page size.
1448	void * pvalloc( size_t size ) libcfa_public {
1449	return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1450	} // pvalloc
1451
1452
1453	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1454	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1455	// 0p, no operation is performed.
1456	void free( void * addr ) libcfa_public {
1457	if ( unlikely( addr == 0p ) ) { // special case
1458	#ifdef __STATISTICS__
1459	if ( heapManager )
1460	incZeroCalls( FREE );
1461	#endif // __STATISTICS__
1462	return;
1463	} // if
1464
1465	#ifdef __STATISTICS__
1466	incCalls( FREE );
1467	#endif // __STATISTICS__
1468
1469	doFree( addr ); // handles heapManager == nullptr
1470	} // free
1471
1472
1473	// Returns the alignment of an allocation.
1474	size_t malloc_alignment( void * addr ) libcfa_public {
1475	if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1476	Heap.Storage.Header * header = HeaderAddr( addr );
1477	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1478	return ClearAlignmentBit( header ); // clear flag from value
1479	} else {
1480	return libAlign(); // minimum alignment
1481	} // if
1482	} // malloc_alignment
1483
1484
1485	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1486	bool malloc_zero_fill( void * addr ) libcfa_public {
1487	if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1488	Heap.Storage.Header * header = HeaderAddr( addr );
1489	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1490	header = RealHeader( header ); // backup from fake to real header
1491	} // if
1492	return ZeroFillBit( header ); // zero filled ?
1493	} // malloc_zero_fill
1494
1495
1496	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1497	size_t malloc_size( void * addr ) libcfa_public {
1498	if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1499	Heap.Storage.Header * header = HeaderAddr( addr );
1500	if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1501	header = RealHeader( header ); // backup from fake to real header
1502	} // if
1503	return header->kind.real.size;
1504	} // malloc_size
1505
1506
1507	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1508	// malloc or a related function.
1509	size_t malloc_usable_size( void * addr ) libcfa_public {
1510	if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1511	Heap.Storage.Header * header;
1512	Heap.FreeHeader * freeHead;
1513	size_t bsize, alignment;
1514
1515	headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1516	return DataStorage( bsize, addr, header ); // data storage in bucket
1517	} // malloc_usable_size
1518
1519
1520	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1521	void malloc_stats( void ) libcfa_public {
1522	#ifdef __STATISTICS__
1523	HeapStatistics stats;
1524	HeapStatisticsCtor( stats );
1525	if ( printStats( collectStats( stats ) ) == -1 ) {
1526	#else
1527	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1528	if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1529	#endif // __STATISTICS__
1530	abort( "** Error ** write failed in malloc_stats" );
1531	} // if
1532	} // malloc_stats
1533
1534
1535	// Changes the file descriptor where malloc_stats() writes statistics.
1536	int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1537	#ifdef __STATISTICS__
1538	int temp = heapMaster.stats_fd;
1539	heapMaster.stats_fd = fd;
1540	return temp;
1541	#else
1542	return -1; // unsupported
1543	#endif // __STATISTICS__
1544	} // malloc_stats_fd
1545
1546
1547	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1548	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1549	// malloc).
1550	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1551	if ( options != 0 ) { errno = EINVAL; return -1; }
1552	#ifdef __STATISTICS__
1553	HeapStatistics stats;
1554	HeapStatisticsCtor( stats );
1555	return printStatsXML( collectStats( stats ), stream ); // returns bytes written or -1
1556	#else
1557	return 0; // unsupported
1558	#endif // __STATISTICS__
1559	} // malloc_info
1560
1561
1562	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1563	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1564	int mallopt( int option, int value ) libcfa_public {
1565	if ( value < 0 ) return 0;
1566	choose( option ) {
1567	case M_TOP_PAD:
1568	heapMaster.heapExpand = ceiling2( value, __page_size );
1569	return 1;
1570	case M_MMAP_THRESHOLD:
1571	if ( setMmapStart( value ) ) return 1;
1572	} // choose
1573	return 0; // error, unsupported
1574	} // mallopt
1575
1576
1577	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1578	int malloc_trim( size_t ) libcfa_public {
1579	return 0; // => impossible to release memory
1580	} // malloc_trim
1581
1582
1583	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1584	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1585	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1586	// result. (The caller must free this memory.)
1587	void * malloc_get_state( void ) libcfa_public {
1588	return 0p; // unsupported
1589	} // malloc_get_state
1590
1591
1592	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1593	// structure pointed to by state.
1594	int malloc_set_state( void * ) libcfa_public {
1595	return 0; // unsupported
1596	} // malloc_set_state
1597
1598
1599	// Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1600	__attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1601
1602	// Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1603	__attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1604
1605	// Amount subtracted to adjust for unfreed program storage (debug only).
1606	__attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1607	} // extern "C"
1608
1609
1610	// Must have CFA linkage to overload with C linkage realloc.
1611	void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1612	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1613	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) );
1614	} // if
1615
1616	PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1617
1618	// Attempt to reuse existing alignment.
1619	Heap.Storage.Header * header = HeaderAddr( oaddr );
1620	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1621	size_t oalign;
1622
1623	if ( unlikely( isFakeHeader ) ) {
1624	checkAlign( nalign ); // check alignment
1625	oalign = ClearAlignmentBit( header ); // old alignment
1626	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1627	&& ( oalign <= nalign // going down
1628	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1629	) ) {
1630	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1631	Heap.FreeHeader * freeHead;
1632	size_t bsize, oalign;
1633	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1634	size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1635
1636	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1637	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1638	ClearZeroFillBit( header ); // turn off 0 fill
1639	#ifdef __CFA_DEBUG__
1640	incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1641	#endif // __CFA_DEBUG__
1642	header->kind.real.size = size; // reset allocation size
1643	#ifdef __STATISTICS__
1644	incCalls( RESIZE );
1645	#endif // __STATISTICS__
1646	return oaddr;
1647	} // if
1648	} // if
1649	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1650	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1651	return resize( oaddr, size ); // duplicate special case checks
1652	} // if
1653
1654	// change size, DO NOT preserve STICKY PROPERTIES.
1655	doFree( oaddr ); // free previous storage
1656	return memalignNoStats( nalign, size STAT_ARG( RESIZE ) ); // create new aligned area
1657	} // resize
1658
1659
1660	void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1661	if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1662	return memalignNoStats( nalign, size STAT_ARG( REALLOC ) );
1663	} // if
1664
1665	PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1666
1667	// Attempt to reuse existing alignment.
1668	Heap.Storage.Header * header = HeaderAddr( oaddr );
1669	bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1670	size_t oalign;
1671	if ( unlikely( isFakeHeader ) ) {
1672	checkAlign( nalign ); // check alignment
1673	oalign = ClearAlignmentBit( header ); // old alignment
1674	if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1675	&& ( oalign <= nalign // going down
1676	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1677	) ) {
1678	HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1679	return realloc( oaddr, size ); // duplicate special case checks
1680	} // if
1681	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1682	&& nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1683	return realloc( oaddr, size ); // duplicate special case checks
1684	} // if
1685
1686	Heap.FreeHeader * freeHead;
1687	size_t bsize;
1688	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1689
1690	// change size and copy old content to new storage
1691
1692	size_t osize = header->kind.real.size; // old allocation size
1693	bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1694
1695	void * naddr = memalignNoStats( nalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1696
1697	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1698	memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1699	doFree( oaddr ); // free previous storage
1700
1701	if ( unlikely( ozfill ) ) { // previous request zero fill ?
1702	MarkZeroFilledBit( header ); // mark new request as zero filled
1703	if ( size > osize ) { // previous request larger ?
1704	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1705	} // if
1706	} // if
1707	return naddr;
1708	} // realloc
1709
1710
1711	void * reallocarray( void * oaddr, size_t nalign, size_t dim, size_t elemSize ) __THROW {
1712	return realloc( oaddr, nalign, dim * elemSize );
1713	} // reallocarray
1714
1715
1716	// Local Variables: //
1717	// tab-width: 4 //
1718	// compile-command: "cfa -nodebug -O2 heap.cfa" //
1719	// End: //

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

Download in other formats:

Original Format