Context Navigation

source: libcfa/src/heap.cc @ 8ee163e2

ADTast-experimentalenumpthread-emulationqualifiedEnum

Last change on this file since 8ee163e2 was 5614a191, checked in by m3zulfiq <m3zulfiq@…>, 2 years ago
removed old memory allocator and replaced it with the concurrent allocator
Property mode set to `100644`
File size: 67.4 KB

Line
1	#include "heap.h"
2
3	#include <algorithm> // lower_bound, min
4	#include <cstring> // strlen, memset, memcpy
5	#include <climits> // ULONG_MAX
6	#include <cstdarg> // va_start, va_end
7	#include <cerrno> // errno, ENOMEM, EINVAL
8	#include <cassert>
9	#include <unistd.h> // STDERR_FILENO, sbrk, sysconf, write
10	#include <sys/mman.h> // mmap, munmap
11	#include <cstdint> // uintptr_t, uint64_t, uint32_t
12
13	#define TLS
14	#define AWAYSPIN // toggle spinlock / lockfree stack
15	#define FASTLOOKUP // use O(1) table lookup from allocation size to bucket size
16	#define CFA_THREADSAFE_HEAP
17
18	#ifdef __CFA_DEBUG__
19	#define __DEBUG__
20	#endif // __CFA_DEBUG__
21
22	#define LIKELY(x) __builtin_expect(!!(x), 1)
23	#define UNLIKELY(x) __builtin_expect(!!(x), 0)
24
25	#define str(s) #s
26	#define xstr(s) str(s)
27	#define WARNING( s ) xstr( GCC diagnostic ignored str( -W ## s ) )
28	#define NOWARNING( statement, warning ) \
29	_Pragma( "GCC diagnostic push" ) \
30	_Pragma( WARNING( warning ) ) \
31	statement ; \
32	_Pragma ( "GCC diagnostic pop" )
33
34	enum { __ALIGN__ = 16, // minimum allocation alignment, bytes
35	__DEFAULT_HEAP_EXPANSION__ = 2 * 1024 * 1024, // sbrk extension amount when full
36	__DEFAULT_MMAP_START__ = 512 * 1024 + 1, // crossover allocation size from sbrk to mmap
37	__DEFAULT_HEAP_UNFREED__ = 0, // amount subtracted to adjust for unfreed program storage
38	}; // enum
39
40	#ifdef TLS
41	#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
42	#else
43	#define TLSMODEL
44	#endif // TLS
45
46	#ifdef CFA_THREADSAFE_HEAP
47	extern "C" {
48	void enable_interrupts();
49	void disable_interrupts();
50	}
51	#define ENABLE_INTERRUPTS (void)enable_interrupts()
52	#define DISABLE_INTERRUPTS (void)disable_interrupts()
53	#else
54	#define ENABLE_INTERRUPTS
55	#define DISABLE_INTERRUPTS
56	#endif // CFA_THREADSAFE_HEAP
57
58	//######################### Helpers #########################
59
60
61	// Called by macro assert in assert.h. Replace to prevent recursive call to malloc.
62	/*
63	void __assert_fail( const char assertion[], const char file[], unsigned int line, const char function[] ) {
64	extern const char * __progname; // global name of running executable (argv[0])
65	char helpText[1024];
66	int len = snprintf( helpText, sizeof(helpText), "Internal assertion error \"%s\" from program \"%s\" in \"%s\" at line %d in file \"%s.\n",
67	assertion, __progname, function, line, file );
68	NOWARNING( write( STDERR_FILENO, helpText, len ), unused-result );
69	abort();
70	// CONTROL NEVER REACHES HERE!
71	} // __assert_fail
72	FIXME */
73
74	void abort( const char fmt[], ... ) __attribute__(( format(printf, 1, 2), __nothrow__, __leaf__, __noreturn__ ));
75	void abort( const char fmt[], ... ) { // overload real abort
76	va_list args;
77	va_start( args, fmt );
78	vfprintf( stderr, fmt, args );
79	if ( fmt[strlen( fmt ) - 1] != '\n' ) { // add optional newline if missing at the end of the format text
80	vfprintf( stderr, "\n", args ); // g++-10 does not allow nullptr for va_list
81	} // if
82	va_end( args );
83	abort(); // call the real abort
84	// CONTROL NEVER REACHES HERE!
85	} // abort
86
87	static inline bool Pow2( unsigned long int value ) {
88	// clears all bits below value, rounding value down to the next lower multiple of value
89	return (value & (value - 1)) == 0;
90	} // Pow2
91
92	static inline unsigned long int Floor( unsigned long int value, unsigned long int align ) {
93	assert( Pow2( align ) );
94	// clears all bits above or equal to align, getting (value % align), the phase of value with regards to align
95	return value & -align;
96	} // Floor
97
98	static inline unsigned long int Ceiling( unsigned long int value, unsigned long int align ) {
99	assert( Pow2( align ) );
100	// "negate, round down, negate" is the same as round up
101	return -Floor( -value, align );
102	} // Ceiling
103
104	template< typename T > static inline T AtomicFetchAdd( volatile T & counter, int increment ) {
105	return __atomic_fetch_add( &counter, increment, __ATOMIC_SEQ_CST );
106	} // AtomicFetchAdd
107
108
109	//######################### Spin Lock #########################
110
111
112	#define CACHE_ALIGN 128 // Intel recommendation
113	#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))
114
115	// pause to prevent excess processor bus usage
116	#if defined( __i386 ) \|\| defined( __x86_64 )
117	#define Pause() __asm__ __volatile__ ( "pause" : : : )
118	#elif defined(__ARM_ARCH)
119	#define Pause() __asm__ __volatile__ ( "YIELD" : : : )
120	#else
121	#error unsupported architecture
122	#endif
123
124	typedef volatile uintptr_t SpinLock_t CALIGN; // aligned addressable word-size
125
126	void spin_acquire( volatile SpinLock_t * lock ) {
127	enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
128	unsigned int spin = SPIN_START;
129
130	for ( unsigned int i = 1;; i += 1 ) {
131	if ( *lock == 0 && __atomic_test_and_set( lock, __ATOMIC_SEQ_CST ) == 0 ) break; // Fence
132	for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
133	spin += spin; // powers of 2
134	//if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
135	if ( spin > SPIN_END ) spin = SPIN_END; // cap spinning
136	} // for
137	} // spin_lock
138
139	void spin_release( volatile SpinLock_t * lock ) {
140	__atomic_clear( lock, __ATOMIC_SEQ_CST ); // Fence
141	} // spin_unlock
142
143
144	//####################### Heap Statistics ####################
145
146
147	#ifdef __STATISTICS__
148	enum { CntTriples = 12 }; // number of counter triples
149	struct HeapStatistics {
150	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC };
151	union {
152	struct {
153	unsigned int malloc_calls, malloc_0_calls;
154	unsigned long long int malloc_storage_request, malloc_storage_alloc;
155	unsigned int aalloc_calls, aalloc_0_calls;
156	unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
157	unsigned int calloc_calls, calloc_0_calls;
158	unsigned long long int calloc_storage_request, calloc_storage_alloc;
159	unsigned int memalign_calls, memalign_0_calls;
160	unsigned long long int memalign_storage_request, memalign_storage_alloc;
161	unsigned int amemalign_calls, amemalign_0_calls;
162	unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
163	unsigned int cmemalign_calls, cmemalign_0_calls;
164	unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
165	unsigned int resize_calls, resize_0_calls;
166	unsigned long long int resize_storage_request, resize_storage_alloc;
167	unsigned int realloc_calls, realloc_0_calls;
168	unsigned long long int realloc_storage_request, realloc_storage_alloc;
169	unsigned int free_calls, free_null_calls;
170	unsigned long long int free_storage_request, free_storage_alloc;
171	unsigned int away_pulls, away_pushes;
172	unsigned long long int away_storage_request, away_storage_alloc;
173	unsigned int mmap_calls, mmap_0_calls; // no zero calls
174	unsigned long long int mmap_storage_request, mmap_storage_alloc;
175	unsigned int munmap_calls, munmap_0_calls; // no zero calls
176	unsigned long long int munmap_storage_request, munmap_storage_alloc;
177	};
178	struct { // overlay for iteration
179	unsigned int cnt1, cnt2;
180	unsigned long long int cnt3, cnt4;
181	} counters[CntTriples];
182	};
183
184	HeapStatistics() {
185	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
186	counters[i].cnt1 = counters[i].cnt2 = counters[i].cnt3 = counters[i].cnt4 = 0;
187	} // for
188	} // HeapStatistics::HeapStatistics
189
190	friend HeapStatistics & operator+=( HeapStatistics & lhs, const HeapStatistics & rhs ) {
191	for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
192	lhs.counters[i].cnt1 += rhs.counters[i].cnt1;
193	lhs.counters[i].cnt2 += rhs.counters[i].cnt2;
194	lhs.counters[i].cnt3 += rhs.counters[i].cnt3;
195	lhs.counters[i].cnt4 += rhs.counters[i].cnt4;
196	} // for
197	return lhs;
198	} // HeapStatistics::operator+=
199	}; // HeapStatistics
200
201	static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(HeapStatistics::counters[0] ),
202	"Heap statistics counter-triplets does not match with array size" );
203	#endif // __STATISTICS__
204
205
206	//####################### Heap Structure ####################
207
208
209	struct Heap {
210	struct FreeHeader; // forward declaration
211
212	struct Storage {
213	struct Header { // header
214	union Kind {
215	struct RealHeader {
216	union {
217	struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
218	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
219	uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
220	#endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
221
222	union {
223	// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
224	FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
225	size_t blockSize; // size for munmap (must overlay alignment)
226	Storage * next; // freed block points to next freed block of same size
227	};
228	size_t size; // allocation size in bytes
229
230	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
231	uint64_t padding; // unused, force home/blocksize to overlay alignment in fake header
232	#endif // __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
233	};
234	};
235	} real; // RealHeader
236	struct FakeHeader {
237	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
238	uint32_t alignment; // 1st low-order bit => fake header & alignment
239	#endif // __ORDER_LITTLE_ENDIAN__
240
241	uint32_t offset;
242
243	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
244	uint32_t alignment; // 1st low-order bit => fake header & alignment
245	#endif // __ORDER_BIG_ENDIAN__
246	} fake; // FakeHeader
247	} kind; // Kind
248	} header; // Header
249
250	char pad[__ALIGN__ - sizeof( Header )];
251	char data[0]; // storage
252	}; // Storage
253
254	static_assert( __ALIGN__ >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
255
256	struct FreeHeader {
257	#ifdef AWAYSPIN
258	SpinLock_t awayLock; // LOCK(S) MUST BE FIRST FIELD(S) FOR ALIGNMENT
259	#endif // AWAYSPIN
260
261	Storage * freeList; // thread free list
262	Storage * awayList; // other thread return list
263
264	Heap * homeManager; // heap owner (free storage to bucket, from bucket to heap)
265	size_t blockSize; // size of allocations on this list
266
267	bool operator<( const size_t bsize ) const { return blockSize < bsize; }
268	}; // FreeHeader
269
270	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
271	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
272	enum {
273	#ifdef FASTLOOKUP
274	LookupSizes = 65'536 + sizeof(Storage), // number of fast lookup sizes '
275	#endif // FASTLOOKUP
276	NoBucketSizes = 91, // number of bucket sizes
277	}; // enum
278
279	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
280	void * heapBuffer;
281	size_t heapReserve;
282
283	Heap * nextHeapManager; // intrusive link of existing heaps; traversed to collect statistics
284	Heap * nextFreeHeapManager; // intrusive link of free heaps from terminated threads; reused by new threads
285
286	#ifdef __DEBUG__
287	long long int allocUnfreed; // running total of allocations minus frees; can be negative
288	#endif // __DEBUG__
289
290	#ifdef __STATISTICS__
291	HeapStatistics stats; // local statistic table for this heap
292	#endif // __STATISTICS__
293
294	static void heapManagerCtor(
295	#ifdef __DEBUG__
296	size_t size
297	#endif // __DEBUG__
298	);
299	static void heapManagerDtor();
300	}; // Heap
301
302
303	struct ThreadManager {
304	int dummy; // used to trigger allocation of storage
305	~ThreadManager() { Heap::heapManagerDtor(); } // called automagically when thread terminates
306	}; // ThreadManager
307
308
309	struct HeapMaster {
310	SpinLock_t masterExtLock; // protects allocation-buffer extension
311	SpinLock_t masterMgrLock; // protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd
312
313	#ifdef FASTLOOKUP
314	unsigned char lookup[Heap::LookupSizes]; // O(1) lookup for small sizes
315	#endif // FASTLOOKUP
316
317	static const unsigned int bucketSizes[]; // initialized statically, outside constructor
318	void * heapBegin; // start of heap
319	void * heapEnd; // logical end of heap
320	size_t heapRemaining; // amount of storage not allocated in the current chunk
321	size_t pageSize; // architecture pagesize
322	size_t heapExpand; // sbrk advance
323	size_t mmapStart; // cross over point for mmap
324	unsigned int maxBucketsUsed; // maximum number of buckets in use
325	static const off_t mmapFd; // fake or actual fd for anonymous file. initialized statically, outside constructor
326
327	Heap * heapManagersList; // heap-list head
328	Heap * freeHeapManagersList; // free-list head
329
330	// Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
331	Heap * heapManagersStorage; // next heap to use in heap superblock
332	Heap * heapManagersStorageEnd; // logical heap outside of superblock's end
333
334	#ifdef __STATISTICS__
335	unsigned long int threads_started, threads_exited; // counts threads that have started and exited
336	unsigned long int reused_heap, new_heap; // counts reusability of heaps
337	unsigned int sbrk_calls;
338	unsigned long long int sbrk_storage;
339	int stats_fd;
340	HeapStatistics stats; // global stats for thread-local heaps to add there counters when exiting
341	#endif // __STATISTICS__
342
343	// Prevents two threads from constructing heapMaster.
344	static volatile bool heapMasterBootFlag; // trigger for first heap
345
346	#ifdef __DEBUG__
347	long long int allocUnfreed;
348	#endif // __DEBUG__
349
350	static void heapMasterCtor();
351	static void heapMasterDtor();
352	}; // HeapMaster
353
354	extern "C" {
355	int __map_prot = PROT_READ \| PROT_WRITE \| PROT_EXEC; // common mmap/mprotect protection
356	size_t __cfa_page_size; // architecture pagesize
357	}
358
359	volatile bool HeapMaster::heapMasterBootFlag = false;
360	static HeapMaster heapMaster; // program global
361
362	// Thread-local storage is allocated lazily when the storage is accessed.
363	static thread_local size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
364	static thread_local ThreadManager threadManager CALIGN TLSMODEL;
365	// Do not put heapManager in ThreadManager because thread-local destructor results in extra access code.
366	static thread_local Heap * heapManager CALIGN TLSMODEL;
367	static thread_local bool heapManagerBootFlag CALIGN TLSMODEL = false;
368	static thread_local size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing
369
370
371	#ifdef __DEBUG__
372	extern "C" {
373	void heapAppStart( void ) {
374	assert( heapManager );
375	heapManager->allocUnfreed = 0;
376	} // heapAppStart
377
378	void heapAppStop( void ) {
379	long long int allocUnfreed = heapMaster.allocUnfreed;
380	for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
381	allocUnfreed += heap->allocUnfreed;
382	} // for
383
384	allocUnfreed -= malloc_unfreed();
385	if ( allocUnfreed > 0 ) {
386	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
387	char helpText[512];
388	int len = snprintf( helpText, sizeof(helpText), "Runtime warning (UNIX pid:%ld) : program terminating with %llu(0x%llx) bytes of storage allocated but not freed.\n"
389	"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
390	(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
391	NOWARNING( write( STDERR_FILENO, helpText, len ), unused-result );
392	} // if
393	} // heapAppStop
394	} // extern "C"
395	#endif // __DEBUG__
396
397
398	// declare helper functions for HeapMaster
399	void noMemory(); // forward, called by "builtin_new" when malloc returns 0
400
401	void HeapMaster::heapMasterCtor() {
402	// Singleton pattern to initialize heap master
403	__cfa_page_size = sysconf( _SC_PAGESIZE );
404
405	assert( heapMaster.mmapFd == -1 );
406	assert( heapMaster.bucketSizes[0] == (16 + sizeof(Heap::Storage)) );
407
408	heapMaster.masterExtLock = 0;
409	heapMaster.masterMgrLock = 0;
410
411	char * end = (char *)sbrk( 0 );
412	heapMaster.heapBegin = heapMaster.heapEnd = sbrk( (char *)Ceiling( (long unsigned int)end, __ALIGN__ ) - end ); // move start of heap to multiple of alignment
413	heapMaster.heapRemaining = 0;
414	heapMaster.heapExpand = malloc_expansion();
415	heapMaster.mmapStart = malloc_mmap_start();
416
417	// find the closest bucket size less than or equal to the mmapStart size
418	heapMaster.maxBucketsUsed = std::lower_bound( heapMaster.bucketSizes, heapMaster.bucketSizes + (Heap::NoBucketSizes - 1), heapMaster.mmapStart ) - heapMaster.bucketSizes; // binary search
419
420	assert( (heapMaster.mmapStart >= __cfa_page_size) && (heapMaster.bucketSizes[Heap::NoBucketSizes - 1] >= heapMaster.mmapStart) );
421	assert( heapMaster.maxBucketsUsed < Heap::NoBucketSizes ); // subscript failure ?
422	assert( heapMaster.mmapStart <= heapMaster.bucketSizes[heapMaster.maxBucketsUsed] ); // search failure ?
423
424	heapMaster.heapManagersList = nullptr;
425	heapMaster.freeHeapManagersList = nullptr;
426
427	heapMaster.heapManagersStorage = nullptr;
428	heapMaster.heapManagersStorageEnd = nullptr;
429
430	#ifdef __STATISTICS__
431	heapMaster.threads_started = heapMaster.threads_exited = 0;
432	heapMaster.reused_heap = heapMaster.new_heap = 0;
433	heapMaster.sbrk_calls = heapMaster.sbrk_storage = 0;
434	heapMaster.stats_fd = STDERR_FILENO;
435	#endif // __STATISTICS__
436
437	#ifdef __DEBUG__
438	heapMaster.allocUnfreed = 0;
439	#endif // __DEBUG__
440
441	#ifdef FASTLOOKUP
442	for ( unsigned int i = 0, idx = 0; i < Heap::LookupSizes; i += 1 ) {
443	if ( i > heapMaster.bucketSizes[idx] ) idx += 1;
444	heapMaster.lookup[i] = idx;
445	assert( i <= heapMaster.bucketSizes[idx] );
446	assert( (i <= 32 && idx == 0) \|\| (i > heapMaster.bucketSizes[idx - 1]) );
447	} // for
448	#endif // FASTLOOKUP
449
450	std::set_new_handler( noMemory ); // do not throw exception as the default
451
452	HeapMaster::heapMasterBootFlag = true;
453	} // HeapMaster::heapMasterCtor
454
455
456	#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."
457
458	void Heap::heapManagerCtor(
459	#ifdef __DEBUG__
460	size_t size
461	#endif // __DEBUG__
462	) {
463	if ( UNLIKELY( ! HeapMaster::heapMasterBootFlag ) ) HeapMaster::heapMasterCtor();
464
465	// Trigger thread_local storage implicit allocation (causes recursive call)
466	volatile int dummy __attribute__(( unused )) = threadManager.dummy;
467
468	spin_acquire( &heapMaster.masterMgrLock ); // protect heapMaster counters
469	// The atomic test-and-set instruction is a fence so heapManagerBootFlag is read after the magic recursive call to
470	// initialize thread-local storage. Hence, heapManagerBootFlag is NOT declared as volatile.
471	if ( heapManagerBootFlag ) { // singleton
472	spin_release( &heapMaster.masterMgrLock );
473	return; // always return on recursive initiation
474	} // if
475
476	assert( ! heapManagerBootFlag );
477
478	// get storage for heap manager
479
480	if ( heapMaster.freeHeapManagersList ) { // free heap for reused ?
481	heapManager = heapMaster.freeHeapManagersList;
482	heapMaster.freeHeapManagersList = heapManager->nextFreeHeapManager;
483
484	#ifdef __STATISTICS__
485	heapMaster.reused_heap += 1;
486	#endif // __STATISTICS__
487	} else { // free heap not found, create new
488	// Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps * 12K ~= 120K byte superblock.
489	// Where 128-heap superblock handles a medium sized multi-processor server.
490	enum { HeapDim = 128 }; // number of heaps in superblock
491	size_t remaining = heapMaster.heapManagersStorageEnd - heapMaster.heapManagersStorage; // remaining free heaps in superblock
492	if ( ! heapMaster.heapManagersStorage \|\| remaining != 0 ) {
493	size_t size = HeapDim * sizeof( Heap );
494	heapMaster.heapManagersStorage = (Heap *)mmap( 0, size, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, heapMaster.mmapFd, 0 );
495	if ( UNLIKELY( heapMaster.heapManagersStorage == MAP_FAILED ) ) { // failed ?
496	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
497	// Do not call strerror( errno ) as it may call malloc.
498	abort( "heapManagerCtor() : internal error, mmap failure, size:%zu error %d.",
499	size, errno );
500	} // if
501	heapMaster.heapManagersStorageEnd = &heapMaster.heapManagersStorage[HeapDim]; // outside array
502	} // if
503
504	heapManager = heapMaster.heapManagersStorage;
505	heapMaster.heapManagersStorage = heapMaster.heapManagersStorage + 1; // bump next heap
506
507	heapManager->nextHeapManager = heapMaster.heapManagersList;
508	heapMaster.heapManagersList = heapManager;
509
510	#ifdef __STATISTICS__
511	heapMaster.new_heap += 1;
512	#endif // __STATISTICS__
513	} // if
514
515	#ifdef __STATISTICS__
516	heapMaster.threads_started += 1;
517	#endif // __STATISTICS__
518
519	#ifdef __DEBUG__
520	heapManager->allocUnfreed -= size;
521	#endif // __DEBUG__
522
523	spin_release( &heapMaster.masterMgrLock );
524
525	for ( unsigned int j = 0; j < Heap::NoBucketSizes; j += 1 ) { // initialize free lists
526	heapManager->freeLists[j] = (Heap::FreeHeader){
527	#ifdef AWAYSPIN
528	.awayLock = 0,
529	#endif // AWAYSPIN
530	.freeList = nullptr,
531	.awayList = nullptr,
532	.homeManager = heapManager,
533	.blockSize = heapMaster.bucketSizes[j],
534	};
535	} // for
536
537	heapManager->heapBuffer = nullptr;
538	heapManager->heapReserve = 0;
539	heapManager->nextFreeHeapManager = nullptr;
540	heapManagerBootFlag = true;
541	} // Heap::heapManagerCtor
542
543
544	void Heap::heapManagerDtor() {
545	if ( UNLIKELY( ! heapManagerBootFlag ) ) return;
546
547	DISABLE_INTERRUPTS;
548
549	spin_acquire( &heapMaster.masterMgrLock );
550
551	// place heap on list of free heaps for reusability
552	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
553	heapMaster.freeHeapManagersList = heapManager;
554
555	// SKULLDUGGERY: The thread heap ends BEFORE the last free(s) occurs from the thread-local storage allocations for
556	// the thread. This final allocation must be handled in doFree for this thread and its terminated heap. However,
557	// this heap has just been put on the heap freelist, and hence there is a race returning the thread-local storage
558	// and a new thread using this heap. The current thread detects it is executing its last free in doFree via
559	// heapManager being null. The trick is for this thread to placed the last free onto the current heap's away-list as
560	// the free-storage header points are this heap. Now, even if other threads are pushing to the away list, it is safe
561	// because of the locking.
562	heapManager = nullptr;
563
564	#ifdef __STATISTICS__
565	heapMaster.threads_exited += 1;
566	#endif // __STATISTICS__
567
568	spin_release( &heapMaster.masterMgrLock );
569
570	ENABLE_INTERRUPTS;
571	} // Heap::heapManagerDtor
572
573
574	const off_t HeapMaster::mmapFd = -1;
575	// Bucket size must be multiple of 16.
576	// Powers of 2 are common allocation sizes, so make powers of 2 generate the minimum required size.
577	const unsigned int HeapMaster::bucketSizes[] = { // different bucket sizes
578	16 + sizeof(Heap::Storage), 32 + sizeof(Heap::Storage), 48 + sizeof(Heap::Storage), 64 + sizeof(Heap::Storage), // 4
579	96 + sizeof(Heap::Storage), 112 + sizeof(Heap::Storage), 128 + sizeof(Heap::Storage), // 3
580	160, 192, 224, 256 + sizeof(Heap::Storage), // 4
581	320, 384, 448, 512 + sizeof(Heap::Storage), // 4
582	640, 768, 896, 1'024 + sizeof(Heap::Storage), // 4
583	1'536, 2'048 + sizeof(Heap::Storage), // 2
584	2'560, 3'072, 3'584, 4'096 + sizeof(Heap::Storage), // 4
585	6'144, 8'192 + sizeof(Heap::Storage), // 2
586	9'216, 10'240, 11'264, 12'288, 13'312, 14'336, 15'360, 16'384 + sizeof(Heap::Storage), // 8
587	18'432, 20'480, 22'528, 24'576, 26'624, 28'672, 30'720, 32'768 + sizeof(Heap::Storage), // 8
588	36'864, 40'960, 45'056, 49'152, 53'248, 57'344, 61'440, 65'536 + sizeof(Heap::Storage), // 8
589	73'728, 81'920, 90'112, 98'304, 106'496, 114'688, 122'880, 131'072 + sizeof(Heap::Storage), // 8
590	147'456, 163'840, 180'224, 196'608, 212'992, 229'376, 245'760, 262'144 + sizeof(Heap::Storage), // 8
591	294'912, 327'680, 360'448, 393'216, 425'984, 458'752, 491'520, 524'288 + sizeof(Heap::Storage), // 8
592	655'360, 786'432, 917'504, 1'048'576 + sizeof(Heap::Storage), // 4
593	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
594	2'621'440, 3'145'728, 3'670'016, 4'194'304 + sizeof(Heap::Storage), // 4
595	};
596
597	static_assert( Heap::NoBucketSizes == sizeof(HeapMaster::bucketSizes) / sizeof(HeapMaster::bucketSizes[0]), "size of bucket array wrong" );
598
599
600	//####################### Memory Allocation Routines' Helpers ####################
601
602
603	#ifdef __STATISTICS__
604	static inline HeapStatistics & collectStats( HeapStatistics & stats ) {
605	spin_acquire( &heapMaster.masterMgrLock );
606
607	stats += heapMaster.stats;
608	for ( Heap * node = heapMaster.heapManagersList; node; node = node->nextHeapManager ) {
609	stats += node->stats;
610	} // for
611
612	spin_release(&heapMaster.masterMgrLock);
613	return stats;
614	} // collectStats
615
616	// Use "write" because streams may be shutdown when calls are made.
617	static void printStats( HeapStatistics & stats ) {
618	char helpText[1024];
619	int len = snprintf( helpText, sizeof(helpText),
620	"\nHeap statistics: (storage request / allocation)\n"
621	" malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
622	" aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
623	" calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
624	" memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
625	" amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
626	" cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
627	" resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
628	" realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
629	" free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n"
630	" away pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n"
631	" sbrk calls %'u; storage %'llu bytes\n"
632	" mmap calls %'u; storage %'llu / %'llu bytes\n"
633	" munmap calls %'u; storage %'llu / %'llu bytes\n"
634	" threads started %'lu; exited %'lu\n"
635	" heaps new %'lu; reused %'lu\n",
636	stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
637	stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
638	stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
639	stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
640	stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
641	stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
642	stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
643	stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
644	stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
645	stats.away_pulls, stats.away_pushes, stats.away_storage_request, stats.away_storage_alloc,
646	heapMaster.sbrk_calls, heapMaster.sbrk_storage,
647	stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
648	stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc,
649	heapMaster.threads_started, heapMaster.threads_exited,
650	heapMaster.new_heap, heapMaster.reused_heap
651	);
652	NOWARNING( write( heapMaster.stats_fd, helpText, len ), unused-result );
653	} // printStats
654
655
656	static int printStatsXML( HeapStatistics & stats, FILE * stream ) {
657	char helpText[1024];
658	int len = snprintf( helpText, sizeof(helpText),
659	"<malloc version=\"1\">\n"
660	"<heap nr=\"0\">\n"
661	"<sizes>\n"
662	"</sizes>\n"
663	"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
664	"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
665	"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
666	"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
667	"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
668	"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
669	"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
670	"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
671	"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
672	"<total type=\"away\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
673	"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
674	"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n"
675	"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
676	"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n"
677	"</malloc>",
678	stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
679	stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
680	stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
681	stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
682	stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
683	stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
684	stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
685	stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
686	stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
687	stats.away_pulls, stats.away_pushes, stats.away_storage_request, stats.away_storage_alloc,
688	heapMaster.sbrk_calls, heapMaster.sbrk_storage,
689	stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
690	stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc,
691	heapMaster.threads_started, heapMaster.threads_exited
692	);
693	NOWARNING( write( fileno(stream), helpText, len ), unused-result );
694	return len;
695	} // printStatsXML
696	#endif // __STATISTICS__
697
698
699	inline void noMemory() {
700	abort( "Heap memory exhausted at %zu bytes.\n"
701	"Possible cause is very large memory allocation and/or large amount of unfreed storage allocated by the program or system/library routines.",
702	((char )(sbrk( 0 )) - (char )(heapMaster.heapBegin)) );
703	} // noMemory
704
705
706	static bool setMmapStart( size_t value ) {
707	if ( value < __cfa_page_size \|\| heapMaster.bucketSizes[Heap::NoBucketSizes - 1] < value ) return false;
708	heapMaster.mmapStart = value; // set global
709
710	// find the closest bucket size less than or equal to the mmapStart size
711	heapMaster.maxBucketsUsed = std::lower_bound( heapMaster.bucketSizes, heapMaster.bucketSizes + (Heap::NoBucketSizes - 1), heapMaster.mmapStart ) - heapMaster.bucketSizes; // binary search
712	assert( heapMaster.maxBucketsUsed < Heap::NoBucketSizes ); // subscript failure ?
713	assert( heapMaster.mmapStart <= heapMaster.bucketSizes[heapMaster.maxBucketsUsed] ); // search failure ?
714	return true;
715	} // setMmapStart
716
717	// <-------+----------------------------------------------------> bsize (bucket size)
718	// \|header \|addr
719	//==================================================================================
720	// align/offset \|
721	// <-----------------<------------+-----------------------------> bsize (bucket size)
722	// \|fake-header \| addr
723	#define headerAddr( addr ) ((Heap::Storage::Header )( (char )addr - sizeof(Heap::Storage) ))
724	#define realHeader( header ) ((Heap::Storage::Header )((char )header - header->kind.fake.offset))
725
726	// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
727	// \|header \|addr
728	//==================================================================================
729	// align/offset \|
730	// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
731	// \|fake-header \|addr
732	#define dataStorage( bsize, addr, header ) (bsize - ( (char )addr - (char )header ))
733
734
735	static inline void checkAlign( size_t alignment ) {
736	if ( UNLIKELY( alignment < __ALIGN__ \|\| ! Pow2( alignment ) ) ) {
737	abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, __ALIGN__ );
738	} // if
739	} // checkAlign
740
741
742	static inline void checkHeader( bool check, const char name[], void * addr ) {
743	if ( UNLIKELY( check ) ) { // bad address ?
744	abort( "Attempt to %s storage %p with address outside the heap.\n"
745	"Possible cause is duplicate free on same block or overwriting of memory.",
746	name, addr );
747	} // if
748	} // checkHeader
749
750
751	static inline void fakeHeader( Heap::Storage::Header *& header, size_t & alignment ) {
752	if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
753	alignment = header->kind.fake.alignment & -2; // remove flag from value
754	#ifdef __DEBUG__
755	checkAlign( alignment ); // check alignment
756	#endif // __DEBUG__
757	header = realHeader( header ); // backup from fake to real header
758	} else {
759	alignment = __ALIGN__; // => no fake header
760	} // if
761	} // fakeHeader
762
763
764	static inline bool headers( const char name[] __attribute__(( unused )), void * addr, Heap::Storage::Header & header, Heap::FreeHeader & freeHead, size_t & size, size_t & alignment ) {
765	header = headerAddr( addr );
766
767	if ( UNLIKELY( addr < heapMaster.heapBegin \|\| heapMaster.heapEnd < addr ) ) { // mmapped ?
768	fakeHeader( header, alignment );
769	size = header->kind.real.blockSize & -3; // mmap size
770	return true;
771	} // if
772
773	#ifdef __DEBUG__
774	checkHeader( header < heapMaster.heapBegin, name, addr ); // bad low address ?
775	#endif // __DEBUG__
776
777	// header may be safe to dereference
778	fakeHeader( header, alignment );
779	#ifdef __DEBUG__
780	checkHeader( header < heapMaster.heapBegin \|\| heapMaster.heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
781	#endif // __DEBUG__
782
783	freeHead = (Heap::FreeHeader *)((size_t)header->kind.real.home & -3);
784	#ifdef __DEBUG__
785	Heap * homeManager = freeHead->homeManager;
786	if ( UNLIKELY( freeHead < &homeManager->freeLists[0] \|\| &homeManager->freeLists[Heap::NoBucketSizes - 1] < freeHead ) ) {
787	abort( "Attempt to %s storage %p with corrupted header.\n"
788	"Possible cause is duplicate free on same block or overwriting of header information.",
789	name, addr );
790	} // if
791	#endif // __DEBUG__
792	size = freeHead->blockSize;
793	return false;
794	} // headers
795
796
797	static inline void * master_extend( size_t size ) {
798	spin_acquire( &heapMaster.masterExtLock );
799
800	ptrdiff_t rem = heapMaster.heapRemaining - size;
801	if ( UNLIKELY( rem < 0 ) ) {
802	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
803
804	size_t increase = Ceiling( size > heapMaster.heapExpand ? size : heapMaster.heapExpand, __ALIGN__ );
805	if ( UNLIKELY( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
806	spin_release( &heapMaster.masterExtLock );
807	abort( NO_MEMORY_MSG, size ); // give up
808	} // if
809	#ifdef __STATISTICS__
810	heapMaster.sbrk_calls += 1;
811	heapMaster.sbrk_storage += increase;
812	#endif // __STATISTICS__
813	rem = heapMaster.heapRemaining + increase - size;
814	} // if
815
816	Heap::Storage * block = (Heap::Storage *)heapMaster.heapEnd;
817	heapMaster.heapRemaining = rem;
818	heapMaster.heapEnd = (char *)heapMaster.heapEnd + size;
819
820	spin_release( &heapMaster.masterExtLock );
821	return block;
822	} // master_extend
823
824
825	static inline void * manager_extend( size_t size ) {
826	ptrdiff_t rem = heapManager->heapReserve - size;
827
828	if ( UNLIKELY( rem < 0 ) ) { // negative
829	// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
830	// smaller freeLists, and increase the reserve.
831
832	rem = heapManager->heapReserve; // positive
833
834	if ( rem >= heapMaster.bucketSizes[0] ) {
835	Heap::FreeHeader * freeHead =
836	#ifdef FASTLOOKUP
837	rem < Heap::LookupSizes ? &(heapManager->freeLists[heapMaster.lookup[rem]]) :
838	#endif // FASTLOOKUP
839	std::lower_bound( heapManager->freeLists, heapManager->freeLists + heapMaster.maxBucketsUsed, rem ); // binary search
840
841	if ( UNLIKELY( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
842	Heap::Storage * block = (Heap::Storage *)heapManager->heapBuffer;
843
844	block->header.kind.real.next = freeHead->freeList; // push on stack
845	freeHead->freeList = block;
846	} // if
847
848	size_t increase = Ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), __ALIGN__ );
849	heapManager->heapBuffer = master_extend(increase);
850	rem = increase - size;
851	} // if
852
853	Heap::Storage * block = (Heap::Storage *)heapManager->heapBuffer;
854	heapManager->heapReserve = rem;
855	heapManager->heapBuffer = (char *)heapManager->heapBuffer + size;
856
857	return block;
858	} // manager_extend
859
860
861	static inline void * doMalloc( size_t size
862	#ifdef __STATISTICS__
863	, unsigned int counter
864	#endif // __STATISTICS__
865	) {
866	Heap::Storage * block;
867
868	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
869	// along with the block and is a multiple of the alignment size.
870	size_t tsize = size + sizeof(Heap::Storage);
871
872	#ifdef __STATISTICS__
873	heapManager->stats.counters[counter].cnt1 += 1;
874	heapManager->stats.counters[counter].cnt3 += size;
875	#endif // __STATISTICS__
876
877	if ( LIKELY( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
878	Heap::FreeHeader * freeHead =
879	#ifdef FASTLOOKUP
880	LIKELY( tsize < Heap::LookupSizes ) ? &(heapManager->freeLists[heapMaster.lookup[tsize]]) :
881	#endif // FASTLOOKUP
882	std::lower_bound( heapManager->freeLists, heapManager->freeLists + heapMaster.maxBucketsUsed, tsize ); // binary search
883
884	assert( freeHead <= &heapManager->freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
885	assert( tsize <= freeHead->blockSize ); // search failure ?
886	tsize = freeHead->blockSize; // total space needed for request
887	#ifdef __STATISTICS__
888	heapManager->stats.counters[counter].cnt4 += tsize;
889	#endif // __STATISTICS__
890
891	block = freeHead->freeList; // remove node from stack
892	if ( UNLIKELY( block == nullptr ) ) { // no free block ?
893	// Freelist for that size is empty, so carve it out of the heap, if there is enough left, or get some more
894	// and then carve it off.
895
896	#ifdef AWAYSPIN
897	spin_acquire( &freeHead->awayLock );
898	block = freeHead->awayList;
899	freeHead->awayList = nullptr;
900	spin_release( &freeHead->awayLock );
901	#else
902	block = __atomic_exchange_n( &freeHead->awayList, nullptr, __ATOMIC_SEQ_CST );
903	#endif // AWAYSPIN
904	if ( LIKELY( block == nullptr ) ) { // away list also empty?
905	block = (Heap::Storage *)manager_extend( tsize ); // mutual exclusion on call
906	} else { // merge awayList into freeHead
907	#ifdef __STATISTICS__
908	heapManager->stats.away_pulls += 1;
909	#endif // __STATISTICS__
910	freeHead->freeList = block->header.kind.real.next;
911	} // if
912	} else {
913	freeHead->freeList = block->header.kind.real.next;
914	} // if
915
916	block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
917	} else { // large size => mmap
918	if ( UNLIKELY( size > ULONG_MAX - __cfa_page_size ) ) return nullptr; // error check
919	tsize = Ceiling( tsize, __cfa_page_size ); // must be multiple of page size
920	#ifdef __STATISTICS__
921	heapManager->stats.counters[counter].cnt4 += tsize;
922	heapManager->stats.mmap_calls += 1;
923	heapManager->stats.mmap_storage_request += size;
924	heapManager->stats.mmap_storage_alloc += tsize;
925	#endif // __STATISTICS__
926
927	block = (Heap::Storage *)::mmap( 0, tsize, __map_prot, MAP_PRIVATE \| MAP_ANONYMOUS, heapMaster.mmapFd, 0 );
928	if ( UNLIKELY( block == MAP_FAILED ) ) { // failed ?
929	if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
930	// Do not call strerror( errno ) as it may call malloc.
931	abort( "(Heap &)0x%p.doMalloc() : internal error, mmap failure, size:%zu %lu %lu error %d.",
932	&heapManager, tsize, size, heapMaster.mmapStart, errno );
933	} // if
934	block->header.kind.real.blockSize = tsize; // storage size for munmap
935	} // if
936
937	block->header.kind.real.size = size; // store allocation size
938	void * addr = &(block->data); // adjust off header to user bytes
939	assert( ((uintptr_t)addr & (__ALIGN__ - 1)) == 0 ); // minimum alignment ?
940
941	#ifdef __DEBUG__
942	heapManager->allocUnfreed += size;
943	#endif // __DEBUG__
944
945	return addr;
946	} // doMalloc
947
948
949	static inline void doFree( void * addr ) {
950	DISABLE_INTERRUPTS;
951
952	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
953	#ifdef __DEBUG__
954	0
955	#endif // __DEBUG__
956	);
957
958	Heap::Storage::Header * header;
959	Heap::FreeHeader * freeHead;
960	size_t size, alignment; // not used (see realloc)
961
962	bool mapped = headers( "free", addr, header, freeHead, size, alignment );
963
964	if ( UNLIKELY( mapped ) ) { // mmapped ?
965	#ifdef __STATISTICS__
966	heapManager->stats.munmap_calls += 1;
967	heapManager->stats.munmap_storage_request += header->kind.real.size;
968	heapManager->stats.munmap_storage_alloc += size;
969	#endif // __STATISTICS__
970	if ( UNLIKELY( munmap( header, size ) == -1 ) ) {
971	abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
972	"Possible cause is invalid pointer.",
973	addr );
974	} // if
975	} else {
976	if ( LIKELY( heapManager == freeHead->homeManager ) ) { // belongs to this thread
977	header->kind.real.next = freeHead->freeList; // push on stack
978	freeHead->freeList = (Heap::Storage *)header;
979	} else { // return to thread owner
980	#ifdef AWAYSPIN
981	spin_acquire( &freeHead->awayLock );
982	header->kind.real.next = freeHead->awayList; // push to bucket away list
983	freeHead->awayList = (Heap::Storage *)header;
984	spin_release( &freeHead->awayLock );
985	#else // lock free
986	header->kind.real.next = freeHead->awayList; // link new node to top node
987	// CAS resets header->kind.real.next = freeHead->awayList on failure
988	while ( ! __atomic_compare_exchange_n( &freeHead->awayList, &header->kind.real.next, header,
989	false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
990	#endif // AWAYSPIN
991
992	// detect free after thread-local storage destruction and use global stats in that case
993	if ( UNLIKELY( heapManager == nullptr ) ) {
994	#ifdef __STATISTICS__
995	AtomicFetchAdd( heapMaster.stats.free_storage_request, header->kind.real.size );
996	AtomicFetchAdd( heapMaster.stats.free_storage_alloc, size );
997	#endif // __STATISTICS__
998	// away push counters are not incremented because this is a self-away push, and there is no
999	// corresponding pull counter that needs to match.
1000	ENABLE_INTERRUPTS;
1001	return;
1002	} // if
1003
1004	#ifdef __STATISTICS__
1005	heapManager->stats.away_pushes += 1;
1006	heapManager->stats.away_storage_request += header->kind.real.size;
1007	heapManager->stats.away_storage_alloc += size;
1008	#endif // __STATISTICS__
1009	} // if
1010	} // if
1011
1012	#ifdef __STATISTICS__
1013	heapManager->stats.free_storage_request += header->kind.real.size;
1014	heapManager->stats.free_storage_alloc += size;
1015	#endif // __STATISTICS__
1016
1017	#ifdef __DEBUG__
1018	heapManager->allocUnfreed -= header->kind.real.size;
1019	#endif // __DEBUG__
1020
1021	ENABLE_INTERRUPTS;
1022	} // doFree
1023
1024
1025	static inline void * mallocNoStats( size_t size
1026	#ifdef __STATISTICS__
1027	, unsigned int counter
1028	#endif // __STATISTICS__
1029	) {
1030	DISABLE_INTERRUPTS;
1031
1032	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
1033	#ifdef __DEBUG__
1034	size
1035	#endif // __DEBUG__
1036	);
1037
1038	if ( UNLIKELY( size ) == 0 \|\| // 0 BYTE ALLOCATION RETURNS NULL POINTER
1039	UNLIKELY( size > ULONG_MAX - sizeof(Heap::Storage) ) ) { // error check
1040	#ifdef __STATISTICS__
1041	heapManager->stats.counters[counter].cnt2 += 1;
1042	#endif // __STATISTICS__
1043	ENABLE_INTERRUPTS;
1044	return nullptr;
1045	} // if
1046
1047	void * addr = doMalloc( size
1048	#ifdef __STATISTICS__
1049	, counter
1050	#endif // __STATISTICS__
1051	);
1052
1053	ENABLE_INTERRUPTS;
1054	return addr;
1055	} // mallocNoStats
1056
1057
1058	static inline void * memalignNoStats( size_t alignment, size_t size
1059	#ifdef __STATISTICS__
1060	, unsigned int counter
1061	#endif // __STATISTICS__
1062	) {
1063	DISABLE_INTERRUPTS;
1064
1065	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
1066	#ifdef __DEBUG__
1067	size
1068	#endif // __DEBUG__
1069	);
1070
1071	if ( UNLIKELY( size ) == 0 \|\| // 0 BYTE ALLOCATION RETURNS NULL POINTER
1072	UNLIKELY( size > ULONG_MAX - sizeof(Heap::Storage) ) ) { // error check
1073	#ifdef __STATISTICS__
1074	heapManager->stats.counters[counter].cnt2 += 1;
1075	#endif // __STATISTICS__
1076
1077	ENABLE_INTERRUPTS;
1078	return nullptr;
1079	} // if
1080
1081	#ifdef __DEBUG__
1082	checkAlign( alignment ); // check alignment
1083	#endif // __DEBUG__
1084
1085	// if alignment <= default alignment, do normal malloc as two headers are unnecessary
1086	if ( UNLIKELY( alignment <= __ALIGN__ ) ) {
1087	void * addr = doMalloc( size
1088	#ifdef __STATISTICS__
1089	, counter
1090	#endif // __STATISTICS__
1091	);
1092
1093	ENABLE_INTERRUPTS;
1094	return addr;
1095	}
1096
1097
1098	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1099	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1100	// .-------------v-----------------v----------------v----------,
1101	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
1102	// `-------------^-----------------^-+--------------^----------'
1103	// \|<--------------------------------' offset/align \|<-- alignment boundary
1104
1105	// subtract __ALIGN__ because it is already the minimum alignment
1106	// add sizeof(Heap::Storage) for fake header
1107	char * addr = (char *)doMalloc( size + alignment - __ALIGN__ + sizeof(Heap::Storage)
1108	#ifdef __STATISTICS__
1109	, counter
1110	#endif // __STATISTICS__
1111	);
1112
1113	// address in the block of the "next" alignment address
1114	char * user = (char *)Ceiling( (uintptr_t)(addr + sizeof(Heap::Storage)), alignment );
1115
1116	// address of header from malloc
1117	Heap::Storage::Header * realHeader = headerAddr( addr );
1118	realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1119	// address of fake header * before* the alignment location
1120	Heap::Storage::Header * fakeHeader = headerAddr( user );
1121	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1122	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
1123	// SKULLDUGGERY: odd alignment imples fake header
1124	fakeHeader->kind.fake.alignment = alignment \| 1;
1125
1126	ENABLE_INTERRUPTS;
1127	return user;
1128	} // memalignNoStats
1129
1130	// Operators new and new [] call malloc; delete calls free
1131
1132
1133	//####################### Memory Allocation Routines ####################
1134
1135
1136	extern "C" {
1137	// Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1138	// then malloc() returns a unique pointer value that can later be successfully passed to free().
1139	void * malloc( size_t size ) {
1140	return mallocNoStats( size
1141	#ifdef __STATISTICS__
1142	, HeapStatistics::MALLOC
1143	#endif // __STATISTICS__
1144	);
1145	} // malloc
1146
1147
1148	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1149	void * aalloc( size_t dim, size_t elemSize ) {
1150	return mallocNoStats( dim * elemSize
1151	#ifdef __STATISTICS__
1152	, HeapStatistics::AALLOC
1153	#endif // __STATISTICS__
1154	);
1155	} // aalloc
1156
1157
1158	// Same as aalloc() with memory set to zero.
1159	void * calloc( size_t dim, size_t elemSize ) {
1160	size_t size = dim * elemSize;
1161	char * addr = (char *)mallocNoStats( size
1162	#ifdef __STATISTICS__
1163	, HeapStatistics::CALLOC
1164	#endif // __STATISTICS__
1165	);
1166
1167	if ( UNLIKELY( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1168
1169	Heap::Storage::Header * header;
1170	Heap::FreeHeader * freeHead;
1171	size_t bsize, alignment;
1172
1173	#ifndef __DEBUG__
1174	bool mapped =
1175	#endif // __DEBUG__
1176	headers( "calloc", addr, header, freeHead, bsize, alignment );
1177
1178	#ifndef __DEBUG__
1179	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1180	if ( LIKELY( ! mapped ) )
1181	#endif // __DEBUG__
1182	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1183	// `-header`-addr `-size
1184	memset( addr, '\0', size ); // set to zeros
1185
1186	header->kind.real.blockSize \|= 2; // mark as zero filled
1187	return addr;
1188	} // calloc
1189
1190
1191	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1192	// nullptr, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1193	// not nullptr, then the call is equivalent to free(oaddr). Unless oaddr is nullptr, it must have been returned by an earlier
1194	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1195	void * resize( void * oaddr, size_t size ) {
1196	if ( UNLIKELY( oaddr == nullptr ) ) { // special cases
1197	return mallocNoStats( size
1198	#ifdef __STATISTICS__
1199	, HeapStatistics::RESIZE
1200	#endif // __STATISTICS__
1201	);
1202	} // if
1203
1204	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1205	if ( UNLIKELY( size == 0 ) ) { // special cases
1206	#ifdef __STATISTICS__
1207	heapManager->stats.resize_0_calls += 1;
1208	#endif // __STATISTICS__
1209	doFree( oaddr ); // free previous storage
1210	return nullptr;
1211	} // if
1212
1213	Heap::Storage::Header * header;
1214	Heap::FreeHeader * freeHead;
1215	size_t bsize, oalign;
1216	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1217
1218	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1219	// same size, DO NOT preserve STICKY PROPERTIES.
1220	if ( oalign == __ALIGN__ && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1221	#ifdef __STATISTICS__
1222	heapManager->stats.resize_calls += 1;
1223	#endif // __STATISTICS__
1224	header->kind.real.blockSize &= -2; // no alignment and turn off 0 fill
1225	header->kind.real.size = size; // reset allocation size
1226	return oaddr;
1227	} // if
1228
1229	// change size, DO NOT preserve STICKY PROPERTIES.
1230	doFree( oaddr ); // free previous storage
1231	return mallocNoStats( size // create new area
1232	#ifdef __STATISTICS__
1233	, HeapStatistics::RESIZE
1234	#endif // __STATISTICS__
1235	);
1236	} // resize
1237
1238
1239	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1240	// the old and new sizes.
1241	void * realloc( void * oaddr, size_t size ) {
1242	if ( UNLIKELY( oaddr == nullptr ) ) { // special cases
1243	return mallocNoStats( size
1244	#ifdef __STATISTICS__
1245	, HeapStatistics::REALLOC
1246	#endif // __STATISTICS__
1247	);
1248	} // if
1249
1250	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1251	if ( UNLIKELY( size == 0 ) ) { // special cases
1252	#ifdef __STATISTICS__
1253	heapManager->stats.realloc_0_calls += 1;
1254	#endif // __STATISTICS__
1255	doFree( oaddr ); // free previous storage
1256	return nullptr;
1257	} // if
1258
1259	Heap::Storage::Header * header;
1260	Heap::FreeHeader * freeHead;
1261	size_t bsize, oalign;
1262	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1263
1264	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1265	size_t osize = header->kind.real.size; // old allocation size
1266	bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1267	if ( UNLIKELY( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1268	#ifdef __STATISTICS__
1269	heapManager->stats.realloc_calls += 1;
1270	heapManager->stats.realloc_storage_request += size;
1271	#endif // __STATISTICS__
1272
1273	header->kind.real.size = size; // reset allocation size
1274	if ( UNLIKELY( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1275	memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1276	} // if
1277	return oaddr;
1278	} // if
1279
1280	// change size and copy old content to new storage
1281
1282	void * naddr;
1283	if ( UNLIKELY( oalign <= __ALIGN__ ) ) { // previous request not aligned ?
1284	naddr = mallocNoStats( size // create new area
1285	#ifdef __STATISTICS__
1286	, HeapStatistics::REALLOC
1287	#endif // __STATISTICS__
1288	);
1289	} else {
1290	naddr = memalignNoStats( oalign, size // create new aligned area
1291	#ifdef __STATISTICS__
1292	, HeapStatistics::REALLOC
1293	#endif // __STATISTICS__
1294	);
1295	} // if
1296
1297	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1298	// To preserve prior fill, the entire bucket must be copied versus the size.
1299	memcpy( naddr, oaddr, std::min( osize, size ) ); // copy bytes
1300	doFree( oaddr ); // free previous storage
1301
1302	if ( UNLIKELY( ozfill ) ) { // previous request zero fill ?
1303	header->kind.real.blockSize \|= 2; // mark new request as zero filled
1304	if ( size > osize ) { // previous request larger ?
1305	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1306	} // if
1307	} // if
1308	return naddr;
1309	} // realloc
1310
1311
1312	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1313	void * memalign( size_t alignment, size_t size ) {
1314	return memalignNoStats( alignment, size
1315	#ifdef __STATISTICS__
1316	, HeapStatistics::MEMALIGN
1317	#endif // __STATISTICS__
1318	);
1319	} // memalign
1320
1321
1322	// Same as aalloc() with memory alignment.
1323	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
1324	return memalignNoStats( alignment, dim * elemSize
1325	#ifdef __STATISTICS__
1326	, HeapStatistics::AMEMALIGN
1327	#endif // __STATISTICS__
1328	);
1329	} // amemalign
1330
1331
1332	// Same as calloc() with memory alignment.
1333	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
1334	size_t size = dim * elemSize;
1335	char * addr = (char *)memalignNoStats( alignment, size
1336	#ifdef __STATISTICS__
1337	, HeapStatistics::CMEMALIGN
1338	#endif // __STATISTICS__
1339	);
1340
1341	if ( UNLIKELY( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1342
1343	Heap::Storage::Header * header;
1344	Heap::FreeHeader * freeHead;
1345	size_t bsize;
1346
1347	#ifndef __DEBUG__
1348	bool mapped =
1349	#endif // __DEBUG__
1350	headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1351
1352	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1353	#ifndef __DEBUG__
1354	if ( LIKELY( ! mapped ) )
1355	#endif // __DEBUG__
1356	// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1357	// `-header`-addr `-size
1358	memset( addr, '\0', size ); // set to zeros
1359
1360	header->kind.real.blockSize \|= 2; // mark as zero filled
1361	return addr;
1362	} // cmemalign
1363
1364
1365	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1366	// of alignment. This requirement is universally ignored.
1367	void * aligned_alloc( size_t alignment, size_t size ) {
1368	return memalign( alignment, size );
1369	} // aligned_alloc
1370
1371
1372	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1373	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1374	// is 0, then posix_memalign() returns either nullptr, or a unique pointer value that can later be successfully passed to
1375	// free(3).
1376	int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
1377	if ( UNLIKELY( alignment < __ALIGN__ \|\| ! Pow2( alignment ) ) ) return EINVAL; // check alignment
1378	*memptr = memalign( alignment, size );
1379	return 0;
1380	} // posix_memalign
1381
1382
1383	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1384	// page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1385	void * valloc( size_t size ) {
1386	return memalign( __cfa_page_size, size );
1387	} // valloc
1388
1389
1390	// Same as valloc but rounds size to multiple of page size.
1391	void * pvalloc( size_t size ) { // round size to multiple of page size
1392	return memalign( __cfa_page_size, Ceiling( size, __cfa_page_size ) );
1393	} // pvalloc
1394
1395
1396	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1397	// or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1398	// nullptr, no operation is performed.
1399	void free( void * addr ) {
1400	// detect free after thread-local storage destruction and use global stats in that case
1401	#ifdef __STATISTICS__
1402	DISABLE_INTERRUPTS;
1403	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
1404	#ifdef __DEBUG__
1405	0
1406	#endif // __DEBUG__
1407	);
1408	#endif // __STATISTICS__
1409
1410	// detect free after thread-local storage destruction and use global stats in that case
1411	if ( UNLIKELY( addr == nullptr ) ) { // special case
1412	#ifdef __STATISTICS__
1413	if ( LIKELY( heapManager ) ) heapManager->stats.free_null_calls += 1;
1414	else AtomicFetchAdd( heapMaster.stats.free_null_calls, 1 );
1415	ENABLE_INTERRUPTS;
1416	#endif // __STATISTICS__
1417	return;
1418	} // fi
1419
1420	#ifdef __STATISTICS__
1421	if ( LIKELY( heapManager ) ) heapManager->stats.free_calls += 1;
1422	else AtomicFetchAdd( heapMaster.stats.free_calls, 1 );
1423	ENABLE_INTERRUPTS;
1424	#endif // __STATISTICS__
1425
1426	doFree( addr );
1427	} // free
1428
1429
1430	// Returns the alignment of an allocation.
1431	size_t malloc_alignment( void * addr ) {
1432	if ( UNLIKELY( addr == nullptr ) ) return __ALIGN__; // minimum alignment
1433	Heap::Storage::Header * header = headerAddr( addr );
1434	if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
1435	return header->kind.fake.alignment & -2; // remove flag from value
1436	} else {
1437	return __ALIGN__; // minimum alignment
1438	} // if
1439	} // malloc_alignment
1440
1441
1442	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
1443	bool malloc_zero_fill( void * addr ) {
1444	if ( UNLIKELY( addr == nullptr ) ) return false; // null allocation is not zero fill
1445	Heap::Storage::Header * header = headerAddr( addr );
1446	if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
1447	header = realHeader( header ); // backup from fake to real header
1448	} // if
1449	return (header->kind.real.blockSize & 2) != 0; // zero filled ?
1450	} // malloc_zero_fill
1451
1452
1453	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1454	size_t malloc_size( void * addr ) {
1455	if ( UNLIKELY( addr == nullptr ) ) return 0; // null allocation is not zero fill
1456	Heap::Storage::Header * header = headerAddr( addr );
1457	if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
1458	header = realHeader( header ); // backup from fake to real header
1459	} // if
1460	return header->kind.real.size;
1461	} // malloc_size
1462
1463
1464	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1465	// malloc or a related function.
1466	size_t malloc_usable_size( void * addr ) {
1467	if ( UNLIKELY( addr == nullptr ) ) return 0; // null allocation has 0 size
1468	Heap::Storage::Header * header;
1469	Heap::FreeHeader * freeHead;
1470	size_t bsize, alignment;
1471
1472	headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1473	return dataStorage( bsize, addr, header ); // data storage in bucket
1474	} // malloc_usable_size
1475
1476
1477	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1478	void malloc_stats() {
1479	#ifdef __STATISTICS__
1480	HeapStatistics stats;
1481	printStats( collectStats( stats ) );
1482	#else
1483	#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1484	NOWARNING( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ), unused-result );
1485	#endif // __STATISTICS__
1486	} // malloc_stats
1487
1488
1489	// Changes the file descriptor where malloc_stats() writes statistics.
1490	int malloc_stats_fd( int fd __attribute__(( unused )) ) {
1491	#ifdef __STATISTICS__
1492	int temp = heapMaster.stats_fd;
1493	heapMaster.stats_fd = fd;
1494	return temp;
1495	#else
1496	return -1; // unsupported
1497	#endif // __STATISTICS__
1498	} // malloc_stats_fd
1499
1500
1501	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1502	// The string is printed on the file stream stream. The exported string includes information about all arenas (see
1503	// malloc).
1504	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) {
1505	if ( options != 0 ) { errno = EINVAL; return -1; }
1506	#ifdef __STATISTICS__
1507	HeapStatistics stats;
1508	return printStatsXML( collectStats( stats ), stream );
1509	#else
1510	return 0; // unsupported
1511	#endif // __STATISTICS__
1512	} // malloc_info
1513
1514
1515	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1516	// specifies the parameter to be modified, and value specifies the new value for that parameter.
1517	int mallopt( int option, int value ) {
1518	if ( value < 0 ) return 0;
1519	switch( option ) {
1520	case M_TOP_PAD:
1521	heapMaster.heapExpand = Ceiling( value, __cfa_page_size );
1522	return 1;
1523	case M_MMAP_THRESHOLD:
1524	if ( setMmapStart( value ) ) return 1;
1525	break;
1526	} // switch
1527	return 0; // error, unsupported
1528	} // mallopt
1529
1530
1531	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1532	int malloc_trim( size_t ) {
1533	return 0; // => impossible to release memory
1534	} // malloc_trim
1535
1536
1537	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1538	// or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1539	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1540	// result. (The caller must free this memory.)
1541	void * malloc_get_state( void ) {
1542	return nullptr; // unsupported
1543	} // malloc_get_state
1544
1545
1546	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1547	// structure pointed to by state.
1548	int malloc_set_state( void * ) {
1549	return 0; // unsupported
1550	} // malloc_set_state
1551
1552	// Set the amount (bytes) to extend the heap size once all the current storage in the heap is allocated.
1553	size_t malloc_expansion() { return __DEFAULT_HEAP_EXPANSION__; }
1554
1555	// Set the crossover point between allocations occuring in the sbrk area or separately mmapped.
1556	size_t malloc_mmap_start() { return __DEFAULT_MMAP_START__; }
1557
1558	// Amount subtracted to adjust for unfreed program storage (debug only).
1559	size_t malloc_unfreed() { return __DEFAULT_HEAP_UNFREED__; }
1560	} // extern "C"
1561
1562
1563	#ifdef __cforall
1564	void * resize( void * oaddr, size_t nalign, size_t size )
1565	#else
1566	extern "C" {
1567	void * _X6resizeFPv_Pvmm__1( void * oaddr, size_t nalign, size_t size )
1568	#endif
1569	{
1570	if ( UNLIKELY( oaddr == nullptr ) ) {
1571	return memalignNoStats( nalign, size
1572	#ifdef __STATISTICS__
1573	, HeapStatistics::RESIZE
1574	#endif // __STATISTICS__
1575	);
1576	} // if
1577
1578	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1579	if ( UNLIKELY( size == 0 ) ) { // special cases
1580	#ifdef __STATISTICS__
1581	heapManager->stats.resize_0_calls += 1;
1582	#endif // __STATISTICS__
1583	doFree( oaddr ); // free previous storage
1584	return nullptr;
1585	} // if
1586
1587	#ifdef __DEBUG__
1588	checkAlign( nalign ); // check alignment
1589	#endif // __DEBUG__
1590
1591	// Attempt to reuse existing alignment.
1592	Heap::Storage::Header * header = headerAddr( oaddr );
1593	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1594	size_t oalign;
1595	if ( UNLIKELY( isFakeHeader ) ) {
1596	oalign = header->kind.fake.alignment & -2; // old alignment
1597	if ( UNLIKELY( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1598	&& ( oalign <= nalign // going down
1599	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1600	) ) {
1601	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1602	Heap::FreeHeader * freeHead;
1603	size_t bsize, oalign;
1604	headers( "resize", oaddr, header, freeHead, bsize, oalign );
1605	size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
1606
1607	if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1608	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1609
1610	header->kind.real.blockSize &= -2; // turn off 0 fill
1611	header->kind.real.size = size; // reset allocation size
1612	return oaddr;
1613	} // if
1614	} // if
1615	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1616	&& nalign == __ALIGN__ ) { // new alignment also on libAlign => no fake header needed
1617	return resize( oaddr, size ); // duplicate special case checks
1618	} // if
1619
1620	// change size, DO NOT preserve STICKY PROPERTIES.
1621	doFree( oaddr ); // free previous storage
1622	return memalignNoStats( nalign, size // create new aligned area
1623	#ifdef __STATISTICS__
1624	, HeapStatistics::RESIZE
1625	#endif // __STATISTICS__
1626	);
1627	} // resize
1628	#ifndef __cforall
1629	}
1630	#endif
1631
1632
1633	#ifdef __cforall
1634	void * realloc( void * oaddr, size_t nalign, size_t size )
1635	#else
1636	extern "C" {
1637	void * _X7reallocFPv_Pvmm__1( void * oaddr, size_t nalign, size_t size )
1638	#endif
1639	{
1640	if ( UNLIKELY( oaddr == nullptr ) ) {
1641	return memalignNoStats( nalign, size
1642	#ifdef __STATISTICS__
1643	, HeapStatistics::REALLOC
1644	#endif // __STATISTICS__
1645	);
1646	} // if
1647
1648	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
1649	if ( UNLIKELY( size == 0 ) ) { // special cases
1650	#ifdef __STATISTICS__
1651	heapManager->stats.realloc_0_calls += 1;
1652	#endif // __STATISTICS__
1653	doFree( oaddr ); // free previous storage
1654	return nullptr;
1655	} // if
1656
1657	#ifdef __DEBUG__
1658	checkAlign( nalign ); // check alignment
1659	#endif // __DEBUG__
1660
1661	// Attempt to reuse existing alignment.
1662	Heap::Storage::Header * header = headerAddr( oaddr );
1663	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
1664	size_t oalign;
1665	if ( UNLIKELY( isFakeHeader ) ) {
1666	oalign = header->kind.fake.alignment & -2; // old alignment
1667	if ( UNLIKELY( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1668	&& ( oalign <= nalign // going down
1669	\|\| (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1670	) ) {
1671	headerAddr( oaddr )->kind.fake.alignment = nalign \| 1; // update alignment (could be the same)
1672	return realloc( oaddr, size ); // duplicate special case checks
1673	} // if
1674	} else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1675	&& nalign == __ALIGN__ ) { // new alignment also on libAlign => no fake header needed
1676	return realloc( oaddr, size ); // duplicate special case checks
1677	} // if
1678
1679	Heap::FreeHeader * freeHead;
1680	size_t bsize;
1681	headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1682
1683	// change size and copy old content to new storage
1684
1685	size_t osize = header->kind.real.size; // old allocation size
1686	bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
1687
1688	void * naddr = memalignNoStats( nalign, size // create new aligned area
1689	#ifdef __STATISTICS__
1690	, HeapStatistics::REALLOC
1691	#endif // __STATISTICS__
1692	);
1693
1694	headers( "realloc", naddr, header, freeHead, bsize, oalign );
1695	memcpy( naddr, oaddr, std::min( osize, size ) ); // copy bytes
1696	doFree( oaddr ); // free previous storage
1697
1698	if ( UNLIKELY( ozfill ) ) { // previous request zero fill ?
1699	header->kind.real.blockSize \|= 2; // mark new request as zero filled
1700	if ( size > osize ) { // previous request larger ?
1701	memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1702	} // if
1703	} // if
1704	return naddr;
1705	} // realloc
1706	#ifndef __cforall
1707	}
1708	#endif
1709
1710	// zip -r HeapPerThread.zip heap/HeapPerThread.h heap/HeapPerThread.cc heap/Makefile heap/affinity.h heap/test.cc heap/away.cc
1711
1712	// g++-10 -Wall -Wextra -g -O3 -DNDEBUG -D__STATISTICS__ -DTLS HeapPerThread.cc -fPIC -shared -o HeapPerThread.so
1713
1714	// Local Variables: //
1715	// tab-width: 4 //
1716	// compile-command: "g++-10 -Wall -Wextra -g -O3 -DNDEBUG -D__STATISTICS__ HeapPerThread.cc -c" //
1717	// End: //

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

Download in other formats: