Context Navigation

heap.cfa @ 0e0f128c

ADTaaron-thesisarm-ehast-experimentalcleanup-dtorsdeferred_resndemanglerenumforall-pointer-decayjacob/cs343-translationjenkins-sandboxnew-astnew-ast-unique-exprno_listpersistent-indexerpthread-emulationqualifiedEnum

Last change on this file since 0e0f128c was 0e0f128c, checked in by Thierry Delisle <tdelisle@…>, 6 years ago
Merge branch 'master' of plg.uwaterloo.ca:software/cfa/cfa-cc
Property mode set to `100644`
File size: 34.1 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.c --
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 : Sat Aug 11 08:22:16 2018
13	// Update Count : 495
14	//
15
16	#include <unistd.h> // sbrk, sysconf
17	#include <stdbool.h> // true, false
18	#include <stdio.h> // snprintf, fileno
19	#include <errno.h> // errno
20	extern "C" {
21	#include <sys/mman.h> // mmap, munmap
22	} // extern "C"
23
24	#include "bits/align.hfa" // libPow2
25	#include "bits/defs.hfa" // likely, unlikely
26	#include "bits/locks.hfa" // __spinlock_t
27	#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
28	#include "stdlib.hfa" // bsearchl
29	#include "malloc.h"
30
31
32	enum {
33	__CFA_DEFAULT_MMAP_START__ = (512 * 1024 + 1),
34	__CFA_DEFAULT_HEAP_EXPANSION__ = (1 * 1024 * 1024),
35	};
36
37	size_t default_mmap_start() __attribute__(( weak )) {
38	return __CFA_DEFAULT_MMAP_START__;
39	} // default_mmap_start
40
41	size_t default_heap_expansion() __attribute__(( weak )) {
42	return __CFA_DEFAULT_HEAP_EXPANSION__;
43	} // default_heap_expansion
44
45
46	// supported mallopt options
47	#ifndef M_MMAP_THRESHOLD
48	#define M_MMAP_THRESHOLD (-1)
49	#endif // M_TOP_PAD
50	#ifndef M_TOP_PAD
51	#define M_TOP_PAD (-2)
52	#endif // M_TOP_PAD
53
54	#define FASTLOOKUP
55	#define __STATISTICS__
56
57	#define SPINLOCK 0
58	#define LOCKFREE 1
59	#define BUCKETLOCK SPINLOCK
60	#if BUCKETLOCK == LOCKFREE
61	#include <uStackLF.h>
62	#endif // LOCKFREE
63
64	#define ALIGN 16
65
66	// enum { NoBucketSizes = 93, // number of buckets sizes
67	// #ifdef FASTLOOKUP
68	// LookupSizes = 65536, // number of fast lookup sizes
69	// #endif // FASTLOOKUP
70	// };
71	#define NoBucketSizes 93 // number of buckets sizes
72	#ifdef FASTLOOKUP
73	#define LookupSizes 65536 // number of fast lookup sizes
74	#endif // FASTLOOKUP
75
76
77	static bool traceHeap = false;
78
79	inline bool traceHeap() {
80	return traceHeap;
81	} // traceHeap
82
83	bool traceHeapOn() {
84	bool temp = traceHeap;
85	traceHeap = true;
86	return temp;
87	} // traceHeapOn
88
89	bool traceHeapOff() {
90	bool temp = traceHeap;
91	traceHeap = false;
92	return temp;
93	} // traceHeapOff
94
95
96	static bool checkFree = false;
97
98	inline bool checkFree() {
99	return checkFree;
100	} // checkFree
101
102	bool checkFreeOn() {
103	bool temp = checkFree;
104	checkFree = true;
105	return temp;
106	} // checkFreeOn
107
108	bool checkFreeOff() {
109	bool temp = checkFree;
110	checkFree = false;
111	return temp;
112	} // checkFreeOff
113
114
115	// static bool traceHeapTerm = false;
116
117	// inline bool traceHeapTerm() {
118	// return traceHeapTerm;
119	// } // traceHeapTerm
120
121	// bool traceHeapTermOn() {
122	// bool temp = traceHeapTerm;
123	// traceHeapTerm = true;
124	// return temp;
125	// } // traceHeapTermOn
126
127	// bool traceHeapTermOff() {
128	// bool temp = traceHeapTerm;
129	// traceHeapTerm = false;
130	// return temp;
131	// } // traceHeapTermOff
132
133
134	#ifdef __CFA_DEBUG__
135	static unsigned int allocFree; // running total of allocations minus frees
136
137	static void checkUnfreed() {
138	if ( allocFree != 0 ) {
139	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
140	// char helpText[512];
141	// int len = snprintf( helpText, sizeof(helpText), "CFA warning (UNIX pid:%ld) : program terminating with %u(0x%x) bytes of storage allocated but not freed.\n"
142	// "Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
143	// (long int)getpid(), allocFree, allocFree ); // always print the UNIX pid
144	// __cfaabi_dbg_bits_write( helpText, len );
145	} // if
146	} // checkUnfreed
147
148	extern "C" {
149	void heapAppStart() { // called by __cfaabi_appready_startup
150	allocFree = 0;
151	} // heapAppStart
152
153	void heapAppStop() { // called by __cfaabi_appready_startdown
154	fclose( stdin ); fclose( stdout );
155	checkUnfreed();
156	} // heapAppStop
157	} // extern "C"
158	#endif // __CFA_DEBUG__
159
160
161	struct HeapManager {
162	// struct FreeHeader; // forward declaration
163
164	struct Storage {
165	struct Header { // header
166	union Kind {
167	struct RealHeader {
168	union {
169	struct { // 32-bit word => 64-bit header, 64-bit word => 128-bit header
170	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
171	uint32_t padding; // unused, force home/blocksize to overlay alignment in fake header
172	#endif // __ORDER_BIG_ENDIAN__ && __U_WORDSIZE__ == 32
173
174	union {
175	// FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
176	void * home; // allocated block points back to home locations (must overlay alignment)
177	size_t blockSize; // size for munmap (must overlay alignment)
178	#if BUCKLOCK == SPINLOCK
179	Storage * next; // freed block points next freed block of same size
180	#endif // SPINLOCK
181	};
182
183	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
184	uint32_t padding; // unused, force home/blocksize to overlay alignment in fake header
185	#endif // __ORDER_LITTLE_ENDIAN__ && __U_WORDSIZE__ == 32
186
187	};
188	#if BUCKLOCK == LOCKFREE
189	Stack<Storage>::Link next; // freed block points next freed block of same size (double-wide)
190	#endif // LOCKFREE
191	};
192	} real; // RealHeader
193	struct FakeHeader {
194	#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
195	uint32_t alignment; // low-order bits of home/blockSize used for tricks
196	#endif // __ORDER_LITTLE_ENDIAN__
197
198	uint32_t offset;
199
200	#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
201	uint32_t alignment; // low-order bits of home/blockSize used for tricks
202	#endif // __ORDER_BIG_ENDIAN__
203	} fake; // FakeHeader
204	} kind; // Kind
205	} header; // Header
206	char pad[ALIGN - sizeof( Header )];
207	char data[0]; // storage
208	}; // Storage
209
210	static_assert( ALIGN >= sizeof( Storage ), "ALIGN < sizeof( Storage )" );
211
212	struct FreeHeader {
213	#if BUCKLOCK == SPINLOCK
214	__spinlock_t lock; // must be first field for alignment
215	Storage * freeList;
216	#elif BUCKLOCK == LOCKFREE
217	StackLF<Storage> freeList;
218	#else
219	#error undefined lock type for bucket lock
220	#endif // SPINLOCK
221	size_t blockSize; // size of allocations on this list
222	}; // FreeHeader
223
224	// must be first fields for alignment
225	__spinlock_t extlock; // protects allocation-buffer extension
226	FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
227
228	void * heapBegin; // start of heap
229	void * heapEnd; // logical end of heap
230	size_t heapRemaining; // amount of storage not allocated in the current chunk
231	}; // HeapManager
232
233	static inline size_t getKey( const HeapManager.FreeHeader & freeheader ) { return freeheader.blockSize; }
234	// statically allocated variables => zero filled.
235
236
237	static size_t pageSize; // architecture pagesize
238	static size_t heapExpand; // sbrk advance
239	static size_t mmapStart; // cross over point for mmap
240	static unsigned int maxBucketsUsed; // maximum number of buckets in use
241
242	// Powers of 2 are common allocation sizes, so make powers of 2 generate the minimum required size.
243	static unsigned int bucketSizes[NoBucketSizes] @= { // different bucket sizes
244	16, 32, 48, 64,
245	64 + sizeof(HeapManager.Storage), 96, 112, 128, 128 + sizeof(HeapManager.Storage), 160, 192, 224,
246	256 + sizeof(HeapManager.Storage), 320, 384, 448, 512 + sizeof(HeapManager.Storage), 640, 768, 896,
247	1_024 + sizeof(HeapManager.Storage), 1_536, 2_048 + sizeof(HeapManager.Storage), 2_560, 3_072, 3_584, 4_096 + sizeof(HeapManager.Storage), 6_144,
248	8_192 + sizeof(HeapManager.Storage), 9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360,
249	16_384 + sizeof(HeapManager.Storage), 18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720,
250	32_768 + sizeof(HeapManager.Storage), 36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440,
251	65_536 + sizeof(HeapManager.Storage), 73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880,
252	131_072 + sizeof(HeapManager.Storage), 147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760,
253	262_144 + sizeof(HeapManager.Storage), 294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520,
254	524_288 + sizeof(HeapManager.Storage), 655_360, 786_432, 917_504, 1_048_576 + sizeof(HeapManager.Storage), 1_179_648, 1_310_720, 1_441_792,
255	1_572_864, 1_703_936, 1_835_008, 1_966_080, 2_097_152 + sizeof(HeapManager.Storage), 2_621_440, 3_145_728, 3_670_016,
256	4_194_304 + sizeof(HeapManager.Storage)
257	};
258	#ifdef FASTLOOKUP
259	static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
260	#endif // FASTLOOKUP
261	static int mmapFd = -1; // fake or actual fd for anonymous file
262
263
264	#ifdef __CFA_DEBUG__
265	static bool heapBoot = 0; // detect recursion during boot
266	#endif // __CFA_DEBUG__
267	static HeapManager heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing
268
269
270	static inline bool setMmapStart( size_t value ) {
271	if ( value < pageSize \|\| bucketSizes[NoBucketSizes - 1] < value ) return true;
272	mmapStart = value; // set global
273
274	// find the closest bucket size less than or equal to the mmapStart size
275	maxBucketsUsed = bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
276	assert( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
277	assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
278	return false;
279	} // setMmapStart
280
281
282	static void ?{}( HeapManager & manager ) with ( manager ) {
283	pageSize = sysconf( _SC_PAGESIZE );
284
285	for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
286	freeLists[i].blockSize = bucketSizes[i];
287	} // for
288
289	#ifdef FASTLOOKUP
290	unsigned int idx = 0;
291	for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
292	if ( i > bucketSizes[idx] ) idx += 1;
293	lookup[i] = idx;
294	} // for
295	#endif // FASTLOOKUP
296
297	if ( setMmapStart( default_mmap_start() ) ) {
298	abort( "HeapManager : internal error, mmap start initialization failure." );
299	} // if
300	heapExpand = default_heap_expansion();
301
302	char * End = (char *)sbrk( 0 );
303	sbrk( (char *)libCeiling( (long unsigned int)End, libAlign() ) - End ); // move start of heap to multiple of alignment
304	heapBegin = heapEnd = sbrk( 0 ); // get new start point
305	} // HeapManager
306
307
308	static void ^?{}( HeapManager & ) {
309	#ifdef __STATISTICS__
310	// if ( traceHeapTerm() ) {
311	// printStats();
312	// if ( checkfree() ) checkFree( heapManager, true );
313	// } // if
314	#endif // __STATISTICS__
315	} // ~HeapManager
316
317
318	static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
319	void memory_startup( void ) {
320	#ifdef __CFA_DEBUG__
321	if ( unlikely( heapBoot ) ) { // check for recursion during system boot
322	// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
323	abort( "boot() : internal error, recursively invoked during system boot." );
324	} // if
325	heapBoot = true;
326	#endif // __CFA_DEBUG__
327
328	assert( heapManager.heapBegin == 0 );
329	heapManager{};
330	} // memory_startup
331
332	static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
333	void memory_shutdown( void ) {
334	^heapManager{};
335	} // memory_shutdown
336
337
338	#ifdef __STATISTICS__
339	static unsigned long long int mmap_storage; // heap statistics counters
340	static unsigned int mmap_calls;
341	static unsigned long long int munmap_storage;
342	static unsigned int munmap_calls;
343	static unsigned long long int sbrk_storage;
344	static unsigned int sbrk_calls;
345	static unsigned long long int malloc_storage;
346	static unsigned int malloc_calls;
347	static unsigned long long int free_storage;
348	static unsigned int free_calls;
349	static unsigned long long int calloc_storage;
350	static unsigned int calloc_calls;
351	static unsigned long long int memalign_storage;
352	static unsigned int memalign_calls;
353	static unsigned long long int cmemalign_storage;
354	static unsigned int cmemalign_calls;
355	static unsigned long long int realloc_storage;
356	static unsigned int realloc_calls;
357
358	static int statfd; // statistics file descriptor (changed by malloc_stats_fd)
359
360
361	// Use "write" because streams may be shutdown when calls are made.
362	static void printStats() {
363	char helpText[512];
364	__cfaabi_dbg_bits_print_buffer( helpText, sizeof(helpText),
365	"\nHeap statistics:\n"
366	" malloc: calls %u / storage %llu\n"
367	" calloc: calls %u / storage %llu\n"
368	" memalign: calls %u / storage %llu\n"
369	" cmemalign: calls %u / storage %llu\n"
370	" realloc: calls %u / storage %llu\n"
371	" free: calls %u / storage %llu\n"
372	" mmap: calls %u / storage %llu\n"
373	" munmap: calls %u / storage %llu\n"
374	" sbrk: calls %u / storage %llu\n",
375	malloc_calls, malloc_storage,
376	calloc_calls, calloc_storage,
377	memalign_calls, memalign_storage,
378	cmemalign_calls, cmemalign_storage,
379	realloc_calls, realloc_storage,
380	free_calls, free_storage,
381	mmap_calls, mmap_storage,
382	munmap_calls, munmap_storage,
383	sbrk_calls, sbrk_storage
384	);
385	} // printStats
386
387
388	static int printStatsXML( FILE * stream ) {
389	char helpText[512];
390	int len = snprintf( helpText, sizeof(helpText),
391	"<malloc version=\"1\">\n"
392	"<heap nr=\"0\">\n"
393	"<sizes>\n"
394	"</sizes>\n"
395	"<total type=\"malloc\" count=\"%u\" size=\"%llu\"/>\n"
396	"<total type=\"calloc\" count=\"%u\" size=\"%llu\"/>\n"
397	"<total type=\"memalign\" count=\"%u\" size=\"%llu\"/>\n"
398	"<total type=\"cmemalign\" count=\"%u\" size=\"%llu\"/>\n"
399	"<total type=\"realloc\" count=\"%u\" size=\"%llu\"/>\n"
400	"<total type=\"free\" count=\"%u\" size=\"%llu\"/>\n"
401	"<total type=\"mmap\" count=\"%u\" size=\"%llu\"/>\n"
402	"<total type=\"munmap\" count=\"%u\" size=\"%llu\"/>\n"
403	"<total type=\"sbrk\" count=\"%u\" size=\"%llu\"/>\n"
404	"</malloc>",
405	malloc_calls, malloc_storage,
406	calloc_calls, calloc_storage,
407	memalign_calls, memalign_storage,
408	cmemalign_calls, cmemalign_storage,
409	realloc_calls, realloc_storage,
410	free_calls, free_storage,
411	mmap_calls, mmap_storage,
412	munmap_calls, munmap_storage,
413	sbrk_calls, sbrk_storage
414	);
415	return write( fileno( stream ), helpText, len ); // -1 => error
416	} // printStatsXML
417	#endif // __STATISTICS__
418
419
420	static inline void noMemory() {
421	abort( "Heap memory exhausted at %zu bytes.\n"
422	"Possible cause is very large memory allocation and/or large amount of unfreed storage allocated by the program or system/library routines.",
423	((char )(sbrk( 0 )) - (char )(heapManager.heapBegin)) );
424	} // noMemory
425
426
427	static inline void checkAlign( size_t alignment ) {
428	if ( alignment < sizeof(void *) \|\| ! libPow2( alignment ) ) {
429	abort( "Alignment %zu for memory allocation is less than sizeof(void *) and/or not a power of 2.", alignment );
430	} // if
431	} // checkAlign
432
433
434	static inline bool setHeapExpand( size_t value ) {
435	if ( heapExpand < pageSize ) return true;
436	heapExpand = value;
437	return false;
438	} // setHeapExpand
439
440
441	static inline void checkHeader( bool check, const char * name, void * addr ) {
442	if ( unlikely( check ) ) { // bad address ?
443	abort( "Attempt to %s storage %p with address outside the heap.\n"
444	"Possible cause is duplicate free on same block or overwriting of memory.",
445	name, addr );
446	} // if
447	} // checkHeader
448
449
450	static inline void fakeHeader( HeapManager.Storage.Header *& header, size_t & size, size_t & alignment ) {
451	if ( unlikely( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
452	size_t offset = header->kind.fake.offset;
453	alignment = header->kind.fake.alignment & -2; // remove flag from value
454	#ifdef __CFA_DEBUG__
455	checkAlign( alignment ); // check alignment
456	#endif // __CFA_DEBUG__
457	header = (HeapManager.Storage.Header )((char )header - offset);
458	} // if
459	} // fakeHeader
460
461
462	#define headerAddr( addr ) ((HeapManager.Storage.Header )( (char )addr - sizeof(HeapManager.Storage) ))
463
464	static inline bool headers( const char * name, void * addr, HeapManager.Storage.Header & header, HeapManager.FreeHeader & freeElem, size_t & size, size_t & alignment ) with ( heapManager ) {
465	header = headerAddr( addr );
466
467	if ( unlikely( heapEnd < addr ) ) { // mmapped ?
468	fakeHeader( header, size, alignment );
469	size = header->kind.real.blockSize & -3; // mmap size
470	return true;
471	} // if
472
473	#ifdef __CFA_DEBUG__
474	checkHeader( addr < heapBegin \|\| header < (HeapManager.Storage.Header *)heapBegin, name, addr ); // bad low address ?
475	#endif // __CFA_DEBUG__
476	// header may be safe to dereference
477	fakeHeader( header, size, alignment );
478	#ifdef __CFA_DEBUG__
479	checkHeader( header < (HeapManager.Storage.Header )heapBegin \|\| (HeapManager.Storage.Header )heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
480	#endif // __CFA_DEBUG__
481
482	freeElem = (HeapManager.FreeHeader *)((size_t)header->kind.real.home & -3);
483	#ifdef __CFA_DEBUG__
484	if ( freeElem < &freeLists[0] \|\| &freeLists[NoBucketSizes] <= freeElem ) {
485	abort( "Attempt to %s storage %p with corrupted header.\n"
486	"Possible cause is duplicate free on same block or overwriting of header information.",
487	name, addr );
488	} // if
489	#endif // __CFA_DEBUG__
490	size = freeElem->blockSize;
491	return false;
492	} // headers
493
494
495	static inline void * extend( size_t size ) with ( heapManager ) {
496	lock( extlock __cfaabi_dbg_ctx2 );
497	ptrdiff_t rem = heapRemaining - size;
498	if ( rem < 0 ) {
499	// If the size requested is bigger than the current remaining storage, increase the size of the heap.
500
501	size_t increase = libCeiling( size > heapExpand ? size : heapExpand, libAlign() );
502	if ( sbrk( increase ) == (void *)-1 ) {
503	unlock( extlock );
504	errno = ENOMEM;
505	return 0;
506	} // if
507	#ifdef __STATISTICS__
508	sbrk_calls += 1;
509	sbrk_storage += increase;
510	#endif // __STATISTICS__
511	#ifdef __CFA_DEBUG__
512	// Set new memory to garbage so subsequent uninitialized usages might fail.
513	memset( (char *)heapEnd + heapRemaining, '\377', increase );
514	#endif // __CFA_DEBUG__
515	rem = heapRemaining + increase - size;
516	} // if
517
518	HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
519	heapRemaining = rem;
520	heapEnd = (char *)heapEnd + size;
521	unlock( extlock );
522	return block;
523	} // extend
524
525
526	static inline void * doMalloc( size_t size ) with ( heapManager ) {
527	HeapManager.Storage * block;
528
529	// Look up size in the size list. Make sure the user request includes space for the header that must be allocated
530	// along with the block and is a multiple of the alignment size.
531
532	size_t tsize = size + sizeof(HeapManager.Storage);
533	if ( likely( tsize < mmapStart ) ) { // small size => sbrk
534	HeapManager.FreeHeader * freeElem =
535	#ifdef FASTLOOKUP
536	tsize < LookupSizes ? &freeLists[lookup[tsize]] :
537	#endif // FASTLOOKUP
538	bsearchl( tsize, freeLists, (size_t)maxBucketsUsed ); // binary search
539	assert( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
540	assert( tsize <= freeElem->blockSize ); // search failure ?
541	tsize = freeElem->blockSize; // total space needed for request
542
543	// Spin until the lock is acquired for this particular size of block.
544
545	#if defined( SPINLOCK )
546	lock( freeElem->lock __cfaabi_dbg_ctx2 );
547	block = freeElem->freeList; // remove node from stack
548	#else
549	block = freeElem->freeList.pop();
550	#endif // SPINLOCK
551	if ( unlikely( block == 0 ) ) { // no free block ?
552	#if defined( SPINLOCK )
553	unlock( freeElem->lock );
554	#endif // SPINLOCK
555	// Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
556	// and then carve it off.
557
558	block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
559	if ( unlikely( block == 0 ) ) return 0;
560	#if defined( SPINLOCK )
561	} else {
562	freeElem->freeList = block->header.kind.real.next;
563	unlock( freeElem->lock );
564	#endif // SPINLOCK
565	} // if
566
567	block->header.kind.real.home = freeElem; // pointer back to free list of apropriate size
568	} else { // large size => mmap
569	tsize = libCeiling( tsize, pageSize ); // must be multiple of page size
570	#ifdef __STATISTICS__
571	__atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
572	__atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
573	#endif // __STATISTICS__
574	block = (HeapManager.Storage *)mmap( 0, tsize, PROT_READ \| PROT_WRITE, MAP_PRIVATE \| MAP_ANONYMOUS, mmapFd, 0 );
575	if ( block == (HeapManager.Storage *)MAP_FAILED ) {
576	// Do not call strerror( errno ) as it may call malloc.
577	abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu error:%d.", &heapManager, tsize, errno );
578	} // if
579	#ifdef __CFA_DEBUG__
580	// Set new memory to garbage so subsequent uninitialized usages might fail.
581	memset( block, '\377', tsize );
582	#endif // __CFA_DEBUG__
583	block->header.kind.real.blockSize = tsize; // storage size for munmap
584	} // if
585
586	void * area = &(block->data); // adjust off header to user bytes
587
588	#ifdef __CFA_DEBUG__
589	assert( ((uintptr_t)area & (libAlign() - 1)) == 0 ); // minimum alignment ?
590	__atomic_add_fetch( &allocFree, tsize, __ATOMIC_SEQ_CST );
591	if ( traceHeap() ) {
592	enum { BufferSize = 64 };
593	char helpText[BufferSize];
594	int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", area, size, tsize );
595	// int len = snprintf( helpText, BufferSize, "Malloc %p %zu\n", area, size );
596	__cfaabi_dbg_bits_write( helpText, len );
597	} // if
598	#endif // __CFA_DEBUG__
599
600	return area;
601	} // doMalloc
602
603
604	static inline void doFree( void * addr ) with ( heapManager ) {
605	#ifdef __CFA_DEBUG__
606	if ( unlikely( heapManager.heapBegin == 0 ) ) {
607	abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
608	} // if
609	#endif // __CFA_DEBUG__
610
611	HeapManager.Storage.Header * header;
612	HeapManager.FreeHeader * freeElem;
613	size_t size, alignment; // not used (see realloc)
614
615	if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
616	#ifdef __STATISTICS__
617	__atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
618	__atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
619	#endif // __STATISTICS__
620	if ( munmap( header, size ) == -1 ) {
621	#ifdef __CFA_DEBUG__
622	abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
623	"Possible cause is invalid pointer.",
624	addr );
625	#endif // __CFA_DEBUG__
626	} // if
627	} else {
628	#ifdef __CFA_DEBUG__
629	// Set free memory to garbage so subsequent usages might fail.
630	memset( ((HeapManager.Storage *)header)->data, '\377', freeElem->blockSize - sizeof( HeapManager.Storage ) );
631	#endif // __CFA_DEBUG__
632
633	#ifdef __STATISTICS__
634	free_storage += size;
635	#endif // __STATISTICS__
636	#if defined( SPINLOCK )
637	lock( freeElem->lock __cfaabi_dbg_ctx2 ); // acquire spin lock
638	header->kind.real.next = freeElem->freeList; // push on stack
639	freeElem->freeList = (HeapManager.Storage *)header;
640	unlock( freeElem->lock ); // release spin lock
641	#else
642	freeElem->freeList.push( (HeapManager.Storage )header );
643	#endif // SPINLOCK
644	} // if
645
646	#ifdef __CFA_DEBUG__
647	__atomic_add_fetch( &allocFree, -size, __ATOMIC_SEQ_CST );
648	if ( traceHeap() ) {
649	char helpText[64];
650	int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
651	__cfaabi_dbg_bits_write( helpText, len );
652	} // if
653	#endif // __CFA_DEBUG__
654	} // doFree
655
656
657	size_t checkFree( HeapManager & manager ) with ( manager ) {
658	size_t total = 0;
659	#ifdef __STATISTICS__
660	__cfaabi_dbg_bits_acquire();
661	__cfaabi_dbg_bits_print_nolock( "\nBin lists (bin size : free blocks on list)\n" );
662	#endif // __STATISTICS__
663	for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
664	size_t size = freeLists[i].blockSize;
665	#ifdef __STATISTICS__
666	unsigned int N = 0;
667	#endif // __STATISTICS__
668	#if defined( SPINLOCK )
669	for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0; p = p->header.kind.real.next ) {
670	#else
671	for ( HeapManager.Storage * p = freeLists[i].freeList.top(); p != 0; p = p->header.kind.real.next.top ) {
672	#endif // SPINLOCK
673	total += size;
674	#ifdef __STATISTICS__
675	N += 1;
676	#endif // __STATISTICS__
677	} // for
678	#ifdef __STATISTICS__
679	__cfaabi_dbg_bits_print_nolock( "%7zu, %-7u ", size, N );
680	if ( (i + 1) % 8 == 0 ) __cfaabi_dbg_bits_print_nolock( "\n" );
681	#endif // __STATISTICS__
682	} // for
683	#ifdef __STATISTICS__
684	__cfaabi_dbg_bits_print_nolock( "\ntotal free blocks:%zu\n", total );
685	__cfaabi_dbg_bits_release();
686	#endif // __STATISTICS__
687	return (char )heapEnd - (char )heapBegin - total;
688	} // checkFree
689
690
691	static inline void * malloc2( size_t size ) { // necessary for malloc statistics
692	assert( heapManager.heapBegin != 0 );
693	void * area = doMalloc( size );
694	if ( unlikely( area == 0 ) ) errno = ENOMEM; // POSIX
695	return area;
696	} // malloc2
697
698
699	static inline void * memalign2( size_t alignment, size_t size ) { // necessary for malloc statistics
700	#ifdef __CFA_DEBUG__
701	checkAlign( alignment ); // check alignment
702	#endif // __CFA_DEBUG__
703
704	// if alignment <= default alignment, do normal malloc as two headers are unnecessary
705	if ( unlikely( alignment <= libAlign() ) ) return malloc2( size );
706
707	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
708	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
709	// .-------------v-----------------v----------------v----------,
710	// \| Real Header \| ... padding ... \| Fake Header \| data ... \|
711	// `-------------^-----------------^-+--------------^----------'
712	// \|<--------------------------------' offset/align \|<-- alignment boundary
713
714	// subtract libAlign() because it is already the minimum alignment
715	// add sizeof(Storage) for fake header
716	char * area = (char *)doMalloc( size + alignment - libAlign() + sizeof(HeapManager.Storage) );
717	if ( unlikely( area == 0 ) ) return area;
718
719	// address in the block of the "next" alignment address
720	char * user = (char *)libCeiling( (uintptr_t)(area + sizeof(HeapManager.Storage)), alignment );
721
722	// address of header from malloc
723	HeapManager.Storage.Header * realHeader = headerAddr( area );
724	// address of fake header * before* the alignment location
725	HeapManager.Storage.Header * fakeHeader = headerAddr( user );
726	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
727	fakeHeader->kind.fake.offset = (char )fakeHeader - (char )realHeader;
728	// SKULLDUGGERY: odd alignment imples fake header
729	fakeHeader->kind.fake.alignment = alignment \| 1;
730
731	return user;
732	} // memalign2
733
734
735	extern "C" {
736	void * malloc( size_t size ) {
737	#ifdef __STATISTICS__
738	__atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
739	__atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
740	#endif // __STATISTICS__
741
742	return malloc2( size );
743	} // malloc
744
745
746	void * calloc( size_t noOfElems, size_t elemSize ) {
747	size_t size = noOfElems * elemSize;
748	#ifdef __STATISTICS__
749	__atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
750	__atomic_add_fetch( &calloc_storage, size, __ATOMIC_SEQ_CST );
751	#endif // __STATISTICS__
752
753	char * area = (char *)malloc2( size );
754	if ( unlikely( area == 0 ) ) return 0;
755	HeapManager.Storage.Header * header;
756	HeapManager.FreeHeader * freeElem;
757	size_t asize, alignment;
758	bool mapped __attribute__(( unused )) = headers( "calloc", area, header, freeElem, asize, alignment );
759	#ifndef __CFA_DEBUG__
760	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
761	if ( ! mapped )
762	#endif // __CFA_DEBUG__
763	memset( area, '\0', asize - sizeof(HeapManager.Storage) ); // set to zeros
764	header->kind.real.blockSize \|= 2; // mark as zero filled
765	return area;
766	} // calloc
767
768
769	void * cmemalign( size_t alignment, size_t noOfElems, size_t elemSize ) {
770	size_t size = noOfElems * elemSize;
771	#ifdef __STATISTICS__
772	__atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
773	__atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
774	#endif // __STATISTICS__
775
776	char * area = (char *)memalign2( alignment, size );
777	if ( unlikely( area == 0 ) ) return 0;
778	HeapManager.Storage.Header * header;
779	HeapManager.FreeHeader * freeElem;
780	size_t asize;
781	bool mapped __attribute__(( unused )) = headers( "cmemalign", area, header, freeElem, asize, alignment );
782	#ifndef __CFA_DEBUG__
783	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
784	if ( ! mapped )
785	#endif // __CFA_DEBUG__
786	memset( area, '\0', asize - ( (char )area - (char )header ) ); // set to zeros
787	header->kind.real.blockSize \|= 2; // mark as zero filled
788
789	return area;
790	} // cmemalign
791
792
793	void * realloc( void * addr, size_t size ) {
794	#ifdef __STATISTICS__
795	__atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
796	#endif // __STATISTICS__
797
798	if ( unlikely( addr == 0 ) ) return malloc2( size ); // special cases
799	if ( unlikely( size == 0 ) ) { free( addr ); return 0; }
800
801	HeapManager.Storage.Header * header;
802	HeapManager.FreeHeader * freeElem;
803	size_t asize, alignment = 0;
804	headers( "realloc", addr, header, freeElem, asize, alignment );
805
806	size_t usize = asize - ( (char )addr - (char )header ); // compute the amount of user storage in the block
807	if ( usize >= size ) { // already sufficient storage
808	// This case does not result in a new profiler entry because the previous one still exists and it must match with
809	// the free for this memory. Hence, this realloc does not appear in the profiler output.
810	return addr;
811	} // if
812
813	#ifdef __STATISTICS__
814	__atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
815	#endif // __STATISTICS__
816
817	void * area;
818	if ( unlikely( alignment != 0 ) ) { // previous request memalign?
819	area = memalign( alignment, size ); // create new area
820	} else {
821	area = malloc2( size ); // create new area
822	} // if
823	if ( unlikely( area == 0 ) ) return 0;
824	if ( unlikely( header->kind.real.blockSize & 2 ) ) { // previous request zero fill (calloc/cmemalign) ?
825	assert( (header->kind.real.blockSize & 1) == 0 );
826	bool mapped __attribute__(( unused )) = headers( "realloc", area, header, freeElem, asize, alignment );
827	#ifndef __CFA_DEBUG__
828	// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
829	if ( ! mapped )
830	#endif // __CFA_DEBUG__
831	memset( (char )area + usize, '\0', asize - ( (char )area - (char *)header ) - usize ); // zero-fill back part
832	header->kind.real.blockSize \|= 2; // mark new request as zero fill
833	} // if
834	memcpy( area, addr, usize ); // copy bytes
835	free( addr );
836	return area;
837	} // realloc
838
839
840	void * memalign( size_t alignment, size_t size ) {
841	#ifdef __STATISTICS__
842	__atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
843	__atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
844	#endif // __STATISTICS__
845
846	void * area = memalign2( alignment, size );
847
848	return area;
849	} // memalign
850
851
852	void * aligned_alloc( size_t alignment, size_t size ) {
853	return memalign( alignment, size );
854	} // aligned_alloc
855
856
857	int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
858	if ( alignment < sizeof(void *) \|\| ! libPow2( alignment ) ) return EINVAL; // check alignment
859	* memptr = memalign( alignment, size );
860	if ( unlikely( * memptr == 0 ) ) return ENOMEM;
861	return 0;
862	} // posix_memalign
863
864
865	void * valloc( size_t size ) {
866	return memalign( pageSize, size );
867	} // valloc
868
869
870	void free( void * addr ) {
871	#ifdef __STATISTICS__
872	__atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
873	#endif // __STATISTICS__
874
875	if ( unlikely( addr == 0 ) ) { // special case
876	#ifdef __CFA_DEBUG__
877	if ( traceHeap() ) {
878	#define nullmsg "Free( 0x0 ) size:0\n"
879	// Do not debug print free( 0 ), as it can cause recursive entry from sprintf.
880	__cfaabi_dbg_bits_write( nullmsg, sizeof(nullmsg) - 1 );
881	} // if
882	#endif // __CFA_DEBUG__
883	return;
884	} // exit
885
886	doFree( addr );
887	} // free
888
889
890	int mallopt( int option, int value ) {
891	choose( option ) {
892	case M_TOP_PAD:
893	if ( setHeapExpand( value ) ) fallthru default;
894	case M_MMAP_THRESHOLD:
895	if ( setMmapStart( value ) ) fallthru default;
896	default:
897	return 1; // success, or unsupported
898	} // switch
899	return 0; // error
900	} // mallopt
901
902
903	int malloc_trim( size_t ) {
904	return 0; // => impossible to release memory
905	} // malloc_trim
906
907	size_t malloc_usable_size( void * addr ) {
908	if ( unlikely( addr == 0 ) ) return 0; // null allocation has 0 size
909	HeapManager.Storage.Header * header;
910	HeapManager.FreeHeader * freeElem;
911	size_t size, alignment;
912
913	headers( "malloc_usable_size", addr, header, freeElem, size, alignment );
914	size_t usize = size - ( (char )addr - (char )header ); // compute the amount of user storage in the block
915	return usize;
916	} // malloc_usable_size
917
918
919	size_t malloc_alignment( void * addr ) {
920	if ( unlikely( addr == 0 ) ) return libAlign(); // minimum alignment
921	HeapManager.Storage.Header * header = (HeapManager.Storage.Header )( (char )addr - sizeof(HeapManager.Storage) );
922	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
923	return header->kind.fake.alignment & -2; // remove flag from value
924	} else {
925	return libAlign (); // minimum alignment
926	} // if
927	} // malloc_alignment
928
929
930	bool malloc_zero_fill( void * addr ) {
931	if ( unlikely( addr == 0 ) ) return false; // null allocation is not zero fill
932	HeapManager.Storage.Header * header = (HeapManager.Storage.Header )( (char )addr - sizeof(HeapManager.Storage) );
933	if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
934	header = (HeapManager.Storage.Header )((char )header - header->kind.fake.offset);
935	} // if
936	return (header->kind.real.blockSize & 2) != 0; // zero filled (calloc/cmemalign) ?
937	} // malloc_zero_fill
938
939
940	void malloc_stats( void ) {
941	#ifdef __STATISTICS__
942	printStats();
943	if ( checkFree() ) checkFree( heapManager );
944	#endif // __STATISTICS__
945	} // malloc_stats
946
947
948	int malloc_stats_fd( int fd ) {
949	#ifdef __STATISTICS__
950	int temp = statfd;
951	statfd = fd;
952	return temp;
953	#else
954	return -1;
955	#endif // __STATISTICS__
956	} // malloc_stats_fd
957
958
959	int malloc_info( int options, FILE * stream ) {
960	return printStatsXML( stream );
961	} // malloc_info
962
963
964	void * malloc_get_state( void ) {
965	return 0;
966	} // malloc_get_state
967
968
969	int malloc_set_state( void * ptr ) {
970	return 0;
971	} // malloc_set_state
972	} // extern "C"
973
974
975	// Local Variables: //
976	// tab-width: 4 //
977	// compile-command: "cfa -nodebug -O2 heap.c" //
978	// End: //

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

Download in other formats:

Original Format