source: libcfa/src/heap.cfa@ fa2e183

ADT ast-experimental
Last change on this file since fa2e183 was 88ac843e, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Moved lockfree containers to containers/lockfree.hfa.
Added poison_list, which is a lock-free bag with push and poison as only operations.
  • Property mode set to 100644
File size: 68.1 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.cfa --
8//
9// Author : Peter A. Buhr
10// Created On : Tue Dec 19 21:58:35 2017
11// Last Modified By : Peter A. Buhr
12// Last Modified On : Thu Oct 13 22:21:52 2022
13// Update Count : 1557
14//
15
16#include <stdio.h>
17#include <string.h> // memset, memcpy
18#include <limits.h> // ULONG_MAX
19#include <stdlib.h> // EXIT_FAILURE
20#include <errno.h> // errno, ENOMEM, EINVAL
21#include <unistd.h> // STDERR_FILENO, sbrk, sysconf
22#include <malloc.h> // memalign, malloc_usable_size
23#include <sys/mman.h> // mmap, munmap
24extern "C" {
25#include <sys/sysinfo.h> // get_nprocs
26} // extern "C"
27
28#include "bits/align.hfa" // libAlign
29#include "bits/defs.hfa" // likely, unlikely
30#include "concurrency/kernel/fwd.hfa" // __POLL_PREEMPTION
31#include "startup.hfa" // STARTUP_PRIORITY_MEMORY
32#include "math.hfa" // ceiling, min
33#include "bitmanip.hfa" // is_pow2, ceiling2
34
35// supported mallopt options
36#ifndef M_MMAP_THRESHOLD
37#define M_MMAP_THRESHOLD (-1)
38#endif // M_MMAP_THRESHOLD
39
40#ifndef M_TOP_PAD
41#define M_TOP_PAD (-2)
42#endif // M_TOP_PAD
43
44#define FASTLOOKUP // use O(1) table lookup from allocation size to bucket size
45#define RETURNSPIN // toggle spinlock / lockfree stack
46#define OWNERSHIP // return freed memory to owner thread
47
48#define CACHE_ALIGN 64
49#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))
50
51#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
52
53//#define __STATISTICS__
54
55enum {
56 // The default extension heap amount in units of bytes. When the current heap reaches the brk address, the brk
57 // address is extended by the extension amount.
58 __CFA_DEFAULT_HEAP_EXPANSION__ = 10 * 1024 * 1024,
59
60 // The mmap crossover point during allocation. Allocations less than this amount are allocated from buckets; values
61 // greater than or equal to this value are mmap from the operating system.
62 __CFA_DEFAULT_MMAP_START__ = 512 * 1024 + 1,
63
64 // The default unfreed storage amount in units of bytes. When the uC++ program ends it subtracts this amount from
65 // the malloc/free counter to adjust for storage the program does not free.
66 __CFA_DEFAULT_HEAP_UNFREED__ = 0
67}; // enum
68
69
70//####################### Heap Trace/Print ####################
71
72
73static bool traceHeap = false;
74
75inline bool traceHeap() libcfa_public { return traceHeap; }
76
77bool traceHeapOn() libcfa_public {
78 bool temp = traceHeap;
79 traceHeap = true;
80 return temp;
81} // traceHeapOn
82
83bool traceHeapOff() libcfa_public {
84 bool temp = traceHeap;
85 traceHeap = false;
86 return temp;
87} // traceHeapOff
88
89bool traceHeapTerm() libcfa_public { return false; }
90
91
92static bool prtFree = false;
93
94bool prtFree() {
95 return prtFree;
96} // prtFree
97
98bool prtFreeOn() {
99 bool temp = prtFree;
100 prtFree = true;
101 return temp;
102} // prtFreeOn
103
104bool prtFreeOff() {
105 bool temp = prtFree;
106 prtFree = false;
107 return temp;
108} // prtFreeOff
109
110
111//######################### Spin Lock #########################
112
113
114// pause to prevent excess processor bus usage
115#if defined( __i386 ) || defined( __x86_64 )
116 #define Pause() __asm__ __volatile__ ( "pause" : : : )
117#elif defined(__ARM_ARCH)
118 #define Pause() __asm__ __volatile__ ( "YIELD" : : : )
119#else
120 #error unsupported architecture
121#endif
122
123typedef volatile uintptr_t SpinLock_t CALIGN; // aligned addressable word-size
124
125static inline __attribute__((always_inline)) void lock( volatile SpinLock_t & slock ) {
126 enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
127 unsigned int spin = SPIN_START;
128
129 for ( unsigned int i = 1;; i += 1 ) {
130 if ( slock == 0 && __atomic_test_and_set( &slock, __ATOMIC_SEQ_CST ) == 0 ) break; // Fence
131 for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
132 spin += spin; // powers of 2
133 //if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
134 if ( spin > SPIN_END ) spin = SPIN_END; // cap spinning
135 } // for
136} // spin_lock
137
138static inline __attribute__((always_inline)) void unlock( volatile SpinLock_t & slock ) {
139 __atomic_clear( &slock, __ATOMIC_SEQ_CST ); // Fence
140} // spin_unlock
141
142
143//####################### Heap Statistics ####################
144
145
146#ifdef __STATISTICS__
147enum { CntTriples = 12 }; // number of counter triples
148enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC, FREE };
149
150struct StatsOverlay { // overlay for iteration
151 unsigned int calls, calls_0;
152 unsigned long long int request, alloc;
153};
154
155// Heap statistics counters.
156union HeapStatistics {
157 struct { // minimum qualification
158 unsigned int malloc_calls, malloc_0_calls;
159 unsigned long long int malloc_storage_request, malloc_storage_alloc;
160 unsigned int aalloc_calls, aalloc_0_calls;
161 unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
162 unsigned int calloc_calls, calloc_0_calls;
163 unsigned long long int calloc_storage_request, calloc_storage_alloc;
164 unsigned int memalign_calls, memalign_0_calls;
165 unsigned long long int memalign_storage_request, memalign_storage_alloc;
166 unsigned int amemalign_calls, amemalign_0_calls;
167 unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
168 unsigned int cmemalign_calls, cmemalign_0_calls;
169 unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
170 unsigned int resize_calls, resize_0_calls;
171 unsigned long long int resize_storage_request, resize_storage_alloc;
172 unsigned int realloc_calls, realloc_0_calls;
173 unsigned long long int realloc_storage_request, realloc_storage_alloc;
174 unsigned int free_calls, free_null_calls;
175 unsigned long long int free_storage_request, free_storage_alloc;
176 unsigned int return_pulls, return_pushes;
177 unsigned long long int return_storage_request, return_storage_alloc;
178 unsigned int mmap_calls, mmap_0_calls; // no zero calls
179 unsigned long long int mmap_storage_request, mmap_storage_alloc;
180 unsigned int munmap_calls, munmap_0_calls; // no zero calls
181 unsigned long long int munmap_storage_request, munmap_storage_alloc;
182 };
183 struct StatsOverlay counters[CntTriples]; // overlay for iteration
184}; // HeapStatistics
185
186static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(StatsOverlay),
187 "Heap statistics counter-triplets does not match with array size" );
188
189static void HeapStatisticsCtor( HeapStatistics & stats ) {
190 memset( &stats, '\0', sizeof(stats) ); // very fast
191 // for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
192 // stats.counters[i].calls = stats.counters[i].calls_0 = stats.counters[i].request = stats.counters[i].alloc = 0;
193 // } // for
194} // HeapStatisticsCtor
195
196static HeapStatistics & ?+=?( HeapStatistics & lhs, const HeapStatistics & rhs ) {
197 for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
198 lhs.counters[i].calls += rhs.counters[i].calls;
199 lhs.counters[i].calls_0 += rhs.counters[i].calls_0;
200 lhs.counters[i].request += rhs.counters[i].request;
201 lhs.counters[i].alloc += rhs.counters[i].alloc;
202 } // for
203 return lhs;
204} // ?+=?
205#endif // __STATISTICS__
206
207
208#define SPINLOCK 0
209#define LOCKFREE 1
210#define BUCKETLOCK SPINLOCK
211#if BUCKETLOCK == SPINLOCK
212#elif BUCKETLOCK == LOCKFREE
213#include <containers/lockfree.hfa>
214#else
215 #error undefined lock type for bucket lock
216#endif // LOCKFREE
217
218// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
219// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
220enum { NoBucketSizes = 91 }; // number of buckets sizes
221
222struct Heap {
223 struct Storage {
224 struct Header { // header
225 union Kind {
226 struct RealHeader {
227 union {
228 struct { // 4-byte word => 8-byte header, 8-byte word => 16-byte header
229 union {
230 // 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
231 // FreeHeader * home; // allocated block points back to home locations (must overlay alignment)
232 void * home; // allocated block points back to home locations (must overlay alignment)
233 size_t blockSize; // size for munmap (must overlay alignment)
234 #if BUCKETLOCK == SPINLOCK
235 Storage * next; // freed block points to next freed block of same size
236 #endif // SPINLOCK
237 };
238 size_t size; // allocation size in bytes
239 };
240 #if BUCKETLOCK == LOCKFREE
241 Link(Storage) next; // freed block points next freed block of same size (double-wide)
242 #endif // LOCKFREE
243 };
244 } real; // RealHeader
245
246 struct FakeHeader {
247 uintptr_t alignment; // 1st low-order bit => fake header & alignment
248 uintptr_t offset;
249 } fake; // FakeHeader
250 } kind; // Kind
251 } header; // Header
252
253 char pad[libAlign() - sizeof( Header )];
254 char data[0]; // storage
255 }; // Storage
256
257 static_assert( libAlign() >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );
258
259 struct __attribute__(( aligned (8) )) FreeHeader {
260 size_t blockSize __attribute__(( aligned(8) )); // size of allocations on this list
261 #if BUCKETLOCK == SPINLOCK
262 #ifdef OWNERSHIP
263 #ifdef RETURNSPIN
264 SpinLock_t returnLock;
265 #endif // RETURNSPIN
266 Storage * returnList; // other thread return list
267 #endif // OWNERSHIP
268 Storage * freeList; // thread free list
269 #else
270 StackLF(Storage) freeList;
271 #endif // BUCKETLOCK
272 Heap * homeManager; // heap owner (free storage to bucket, from bucket to heap)
273 }; // FreeHeader
274
275 FreeHeader freeLists[NoBucketSizes]; // buckets for different allocation sizes
276 void * heapBuffer; // start of free storage in buffer
277 size_t heapReserve; // amount of remaining free storage in buffer
278
279 #if defined( __STATISTICS__ ) || defined( __CFA_DEBUG__ )
280 Heap * nextHeapManager; // intrusive link of existing heaps; traversed to collect statistics or check unfreed storage
281 #endif // __STATISTICS__ || __CFA_DEBUG__
282 Heap * nextFreeHeapManager; // intrusive link of free heaps from terminated threads; reused by new threads
283
284 #ifdef __CFA_DEBUG__
285 int64_t allocUnfreed; // running total of allocations minus frees; can be negative
286 #endif // __CFA_DEBUG__
287
288 #ifdef __STATISTICS__
289 HeapStatistics stats; // local statistic table for this heap
290 #endif // __STATISTICS__
291}; // Heap
292
293#if BUCKETLOCK == LOCKFREE
294inline __attribute__((always_inline))
295static {
296 Link(Heap.Storage) * ?`next( Heap.Storage * this ) { return &this->header.kind.real.next; }
297 void ?{}( Heap.FreeHeader & ) {}
298 void ^?{}( Heap.FreeHeader & ) {}
299} // distribution
300#endif // LOCKFREE
301
302
303struct HeapMaster {
304 SpinLock_t extLock; // protects allocation-buffer extension
305 SpinLock_t mgrLock; // protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd
306
307 void * heapBegin; // start of heap
308 void * heapEnd; // logical end of heap
309 size_t heapRemaining; // amount of storage not allocated in the current chunk
310 size_t pageSize; // architecture pagesize
311 size_t heapExpand; // sbrk advance
312 size_t mmapStart; // cross over point for mmap
313 unsigned int maxBucketsUsed; // maximum number of buckets in use
314
315 Heap * heapManagersList; // heap-list head
316 Heap * freeHeapManagersList; // free-list head
317
318 // Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
319 Heap * heapManagersStorage; // next heap to use in heap superblock
320 Heap * heapManagersStorageEnd; // logical heap outside of superblock's end
321
322 #ifdef __STATISTICS__
323 HeapStatistics stats; // global stats for thread-local heaps to add there counters when exiting
324 unsigned long int threads_started, threads_exited; // counts threads that have started and exited
325 unsigned long int reused_heap, new_heap; // counts reusability of heaps
326 unsigned int sbrk_calls;
327 unsigned long long int sbrk_storage;
328 int stats_fd;
329 #endif // __STATISTICS__
330}; // HeapMaster
331
332
333#ifdef FASTLOOKUP
334enum { LookupSizes = 65_536 + sizeof(Heap.Storage) }; // number of fast lookup sizes
335static unsigned char lookup[LookupSizes]; // O(1) lookup for small sizes
336#endif // FASTLOOKUP
337
338static volatile bool heapMasterBootFlag = false; // trigger for first heap
339static HeapMaster heapMaster @= {}; // program global
340
341static void heapMasterCtor();
342static void heapMasterDtor();
343static Heap * getHeap();
344
345
346// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
347// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
348// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
349static const unsigned int bucketSizes[] @= { // different bucket sizes
350 16 + sizeof(Heap.Storage), 32 + sizeof(Heap.Storage), 48 + sizeof(Heap.Storage), 64 + sizeof(Heap.Storage), // 4
351 96 + sizeof(Heap.Storage), 112 + sizeof(Heap.Storage), 128 + sizeof(Heap.Storage), // 3
352 160, 192, 224, 256 + sizeof(Heap.Storage), // 4
353 320, 384, 448, 512 + sizeof(Heap.Storage), // 4
354 640, 768, 896, 1_024 + sizeof(Heap.Storage), // 4
355 1_536, 2_048 + sizeof(Heap.Storage), // 2
356 2_560, 3_072, 3_584, 4_096 + sizeof(Heap.Storage), // 4
357 6_144, 8_192 + sizeof(Heap.Storage), // 2
358 9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(Heap.Storage), // 8
359 18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(Heap.Storage), // 8
360 36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(Heap.Storage), // 8
361 73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(Heap.Storage), // 8
362 147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(Heap.Storage), // 8
363 294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(Heap.Storage), // 8
364 655_360, 786_432, 917_504, 1_048_576 + sizeof(Heap.Storage), // 4
365 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
366 2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(Heap.Storage), // 4
367};
368
369static_assert( NoBucketSizes == sizeof(bucketSizes) / sizeof(bucketSizes[0] ), "size of bucket array wrong" );
370
371
372// extern visibility, used by runtime kernel
373libcfa_public size_t __page_size; // architecture pagesize
374libcfa_public int __map_prot; // common mmap/mprotect protection
375
376
377// Thread-local storage is allocated lazily when the storage is accessed.
378static __thread size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
379static __thread Heap * volatile heapManager CALIGN TLSMODEL;
380static __thread size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing
381
382
383// declare helper functions for HeapMaster
384void noMemory(); // forward, called by "builtin_new" when malloc returns 0
385
386
387// generic Bsearchl does not inline, so substitute with hand-coded binary-search.
388inline __attribute__((always_inline))
389static size_t Bsearchl( unsigned int key, const unsigned int vals[], size_t dim ) {
390 size_t l = 0, m, h = dim;
391 while ( l < h ) {
392 m = (l + h) / 2;
393 if ( (unsigned int &)(vals[m]) < key ) { // cast away const
394 l = m + 1;
395 } else {
396 h = m;
397 } // if
398 } // while
399 return l;
400} // Bsearchl
401
402
403void heapMasterCtor() with( heapMaster ) {
404 // Singleton pattern to initialize heap master
405
406 verify( bucketSizes[0] == (16 + sizeof(Heap.Storage)) );
407
408 __page_size = sysconf( _SC_PAGESIZE );
409 __map_prot = PROT_READ | PROT_WRITE | PROT_EXEC;
410
411 ?{}( extLock );
412 ?{}( mgrLock );
413
414 char * end = (char *)sbrk( 0 );
415 heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, libAlign() ) - end ); // move start of heap to multiple of alignment
416 heapRemaining = 0;
417 heapExpand = malloc_expansion();
418 mmapStart = malloc_mmap_start();
419
420 // find the closest bucket size less than or equal to the mmapStart size
421 maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
422
423 verify( (mmapStart >= pageSize) && (bucketSizes[NoBucketSizes - 1] >= mmapStart) );
424 verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
425 verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
426
427 heapManagersList = 0p;
428 freeHeapManagersList = 0p;
429
430 heapManagersStorage = 0p;
431 heapManagersStorageEnd = 0p;
432
433 #ifdef __STATISTICS__
434 HeapStatisticsCtor( stats ); // clear statistic counters
435 threads_started = threads_exited = 0;
436 reused_heap = new_heap = 0;
437 sbrk_calls = sbrk_storage = 0;
438 stats_fd = STDERR_FILENO;
439 #endif // __STATISTICS__
440
441 #ifdef FASTLOOKUP
442 for ( unsigned int i = 0, idx = 0; i < LookupSizes; i += 1 ) {
443 if ( i > bucketSizes[idx] ) idx += 1;
444 lookup[i] = idx;
445 verify( i <= bucketSizes[idx] );
446 verify( (i <= 32 && idx == 0) || (i > bucketSizes[idx - 1]) );
447 } // for
448 #endif // FASTLOOKUP
449
450 heapMasterBootFlag = true;
451} // heapMasterCtor
452
453
454#define NO_MEMORY_MSG "**** Error **** insufficient heap memory available to allocate %zd new bytes."
455
456Heap * getHeap() with( heapMaster ) {
457 Heap * heap;
458 if ( freeHeapManagersList ) { // free heap for reused ?
459 heap = freeHeapManagersList;
460 freeHeapManagersList = heap->nextFreeHeapManager;
461
462 #ifdef __STATISTICS__
463 reused_heap += 1;
464 #endif // __STATISTICS__
465 } else { // free heap not found, create new
466 // Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps *
467 // 12K ~= 120K byte superblock. Where 128-heap superblock handles a medium sized multi-processor server.
468 size_t remaining = heapManagersStorageEnd - heapManagersStorage; // remaining free heaps in superblock
469 if ( ! heapManagersStorage || remaining != 0 ) {
470 // Each block of heaps is a multiple of the number of cores on the computer.
471 int HeapDim = get_nprocs(); // get_nprocs_conf does not work
472 size_t size = HeapDim * sizeof( Heap );
473
474 heapManagersStorage = (Heap *)mmap( 0, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0 );
475 if ( unlikely( heapManagersStorage == (Heap *)MAP_FAILED ) ) { // failed ?
476 if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
477 // Do not call strerror( errno ) as it may call malloc.
478 abort( "**** Error **** attempt to allocate block of heaps of size %zu bytes and mmap failed with errno %d.", size, errno );
479 } // if
480 heapManagersStorageEnd = &heapManagersStorage[HeapDim]; // outside array
481 } // if
482
483 heap = heapManagersStorage;
484 heapManagersStorage = heapManagersStorage + 1; // bump next heap
485
486 #if defined( __STATISTICS__ ) || defined( __CFA_DEBUG__ )
487 heap->nextHeapManager = heapManagersList;
488 #endif // __STATISTICS__ || __CFA_DEBUG__
489 heapManagersList = heap;
490
491 #ifdef __STATISTICS__
492 new_heap += 1;
493 #endif // __STATISTICS__
494
495 with( *heap ) {
496 for ( unsigned int j = 0; j < NoBucketSizes; j += 1 ) { // initialize free lists
497 #ifdef OWNERSHIP
498 #ifdef RETURNSPIN
499 ?{}( freeLists[j].returnLock );
500 #endif // RETURNSPIN
501 freeLists[j].returnList = 0p;
502 #endif // OWNERSHIP
503 freeLists[j].freeList = 0p;
504 freeLists[j].homeManager = heap;
505 freeLists[j].blockSize = bucketSizes[j];
506 } // for
507
508 heapBuffer = 0p;
509 heapReserve = 0;
510 nextFreeHeapManager = 0p;
511 #ifdef __CFA_DEBUG__
512 allocUnfreed = 0;
513 #endif // __CFA_DEBUG__
514 } // with
515 } // if
516
517 return heap;
518} // getHeap
519
520
521void heapManagerCtor() libcfa_public {
522 if ( unlikely( ! heapMasterBootFlag ) ) heapMasterCtor();
523
524 lock( heapMaster.mgrLock ); // protect heapMaster counters
525
526 // get storage for heap manager
527
528 heapManager = getHeap();
529
530 #ifdef __STATISTICS__
531 HeapStatisticsCtor( heapManager->stats ); // heap local
532 heapMaster.threads_started += 1;
533 #endif // __STATISTICS__
534
535 unlock( heapMaster.mgrLock );
536} // heapManagerCtor
537
538
539void heapManagerDtor() libcfa_public {
540 lock( heapMaster.mgrLock );
541
542 // place heap on list of free heaps for reusability
543 heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
544 heapMaster.freeHeapManagersList = heapManager;
545
546 #ifdef __STATISTICS__
547 heapMaster.threads_exited += 1;
548 #endif // __STATISTICS__
549
550 // Do not set heapManager to NULL because it is used after Cforall is shutdown but before the program shuts down.
551
552 unlock( heapMaster.mgrLock );
553} // heapManagerDtor
554
555
556//####################### Memory Allocation Routines Helpers ####################
557
558
559extern int cfa_main_returned; // from interpose.cfa
560extern "C" {
561 void memory_startup( void ) {
562 if ( ! heapMasterBootFlag ) heapManagerCtor(); // sanity check
563 } // memory_startup
564
565 void memory_shutdown( void ) {
566 heapManagerDtor();
567 } // memory_shutdown
568
569 void heapAppStart() { // called by __cfaabi_appready_startup
570 verify( heapManager );
571 #ifdef __CFA_DEBUG__
572 heapManager->allocUnfreed = 0; // clear prior allocation counts
573 #endif // __CFA_DEBUG__
574
575 #ifdef __STATISTICS__
576 HeapStatisticsCtor( heapManager->stats ); // clear prior statistic counters
577 #endif // __STATISTICS__
578 } // heapAppStart
579
580 void heapAppStop() { // called by __cfaabi_appready_startdown
581 fclose( stdin ); fclose( stdout ); // free buffer storage
582 if ( ! cfa_main_returned ) return; // do not check unfreed storage if exit called
583
584 #ifdef __CFA_DEBUG__
585 // allocUnfreed is set to 0 when a heap is created and it accumulates any unfreed storage during its multiple thread
586 // usages. At the end, add up each heap allocUnfreed value across all heaps to get the total unfreed storage.
587 int64_t allocUnfreed = 0;
588 for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
589 allocUnfreed += heap->allocUnfreed;
590 } // for
591
592 allocUnfreed -= malloc_unfreed(); // subtract any user specified unfreed storage
593 if ( allocUnfreed > 0 ) {
594 // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
595 char helpText[512];
596 __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
597 "CFA warning (UNIX pid:%ld) : program terminating with %ju(0x%jx) bytes of storage allocated but not freed.\n"
598 "Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
599 (long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
600 } // if
601 #endif // __CFA_DEBUG__
602 } // heapAppStop
603} // extern "C"
604
605
606#ifdef __STATISTICS__
607static HeapStatistics stats; // zero filled
608
609#define prtFmt \
610 "\nHeap statistics: (storage request / allocation)\n" \
611 " malloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
612 " aalloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
613 " calloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
614 " memalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
615 " amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
616 " cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
617 " resize >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
618 " realloc >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n" \
619 " free !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n" \
620 " return pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n" \
621 " sbrk calls %'u; storage %'llu bytes\n" \
622 " mmap calls %'u; storage %'llu / %'llu bytes\n" \
623 " munmap calls %'u; storage %'llu / %'llu bytes\n" \
624 " threads started %'lu; exited %'lu\n" \
625 " heaps new %'lu; reused %'lu\n"
626
627// Use "write" because streams may be shutdown when calls are made.
628static int printStats( HeapStatistics & stats ) with( heapMaster, stats ) { // see malloc_stats
629 char helpText[sizeof(prtFmt) + 1024]; // space for message and values
630 return __cfaabi_bits_print_buffer( stats_fd, helpText, sizeof(helpText), prtFmt,
631 malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
632 aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
633 calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
634 memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
635 amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
636 cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
637 resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
638 realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
639 free_calls, free_null_calls, free_storage_request, free_storage_alloc,
640 return_pulls, return_pushes, return_storage_request, return_storage_alloc,
641 sbrk_calls, sbrk_storage,
642 mmap_calls, mmap_storage_request, mmap_storage_alloc,
643 munmap_calls, munmap_storage_request, munmap_storage_alloc,
644 threads_started, threads_exited,
645 new_heap, reused_heap
646 );
647} // printStats
648
649#define prtFmtXML \
650 "<malloc version=\"1\">\n" \
651 "<heap nr=\"0\">\n" \
652 "<sizes>\n" \
653 "</sizes>\n" \
654 "<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
655 "<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
656 "<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
657 "<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
658 "<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
659 "<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
660 "<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
661 "<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
662 "<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
663 "<total type=\"return\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
664 "<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n" \
665 "<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n" \
666 "<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n" \
667 "<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n" \
668 "<total type=\"heaps\" new=\"%'lu;\" reused=\"%'lu\"/>\n" \
669 "</malloc>"
670
671static int printStatsXML( HeapStatistics & stats, FILE * stream ) with( heapMaster, stats ) { // see malloc_info
672 char helpText[sizeof(prtFmtXML) + 1024]; // space for message and values
673 return __cfaabi_bits_print_buffer( fileno( stream ), helpText, sizeof(helpText), prtFmtXML,
674 malloc_calls, malloc_0_calls, malloc_storage_request, malloc_storage_alloc,
675 aalloc_calls, aalloc_0_calls, aalloc_storage_request, aalloc_storage_alloc,
676 calloc_calls, calloc_0_calls, calloc_storage_request, calloc_storage_alloc,
677 memalign_calls, memalign_0_calls, memalign_storage_request, memalign_storage_alloc,
678 amemalign_calls, amemalign_0_calls, amemalign_storage_request, amemalign_storage_alloc,
679 cmemalign_calls, cmemalign_0_calls, cmemalign_storage_request, cmemalign_storage_alloc,
680 resize_calls, resize_0_calls, resize_storage_request, resize_storage_alloc,
681 realloc_calls, realloc_0_calls, realloc_storage_request, realloc_storage_alloc,
682 free_calls, free_null_calls, free_storage_request, free_storage_alloc,
683 return_pulls, return_pushes, return_storage_request, return_storage_alloc,
684 sbrk_calls, sbrk_storage,
685 mmap_calls, mmap_storage_request, mmap_storage_alloc,
686 munmap_calls, munmap_storage_request, munmap_storage_alloc,
687 threads_started, threads_exited,
688 new_heap, reused_heap
689 );
690} // printStatsXML
691
692static HeapStatistics & collectStats( HeapStatistics & stats ) with( heapMaster ) {
693 lock( mgrLock );
694
695 stats += heapMaster.stats;
696 for ( Heap * heap = heapManagersList; heap; heap = heap->nextHeapManager ) {
697 stats += heap->stats;
698 } // for
699
700 unlock( mgrLock );
701 return stats;
702} // collectStats
703#endif // __STATISTICS__
704
705
706static bool setMmapStart( size_t value ) with( heapMaster ) { // true => mmapped, false => sbrk
707 if ( value < __page_size || bucketSizes[NoBucketSizes - 1] < value ) return false;
708 mmapStart = value; // set global
709
710 // find the closest bucket size less than or equal to the mmapStart size
711 maxBucketsUsed = Bsearchl( mmapStart, bucketSizes, NoBucketSizes ); // binary search
712 verify( maxBucketsUsed < NoBucketSizes ); // subscript failure ?
713 verify( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
714 return true;
715} // setMmapStart
716
717
718// <-------+----------------------------------------------------> bsize (bucket size)
719// |header |addr
720//==================================================================================
721// align/offset |
722// <-----------------<------------+-----------------------------> bsize (bucket size)
723// |fake-header | addr
724#define HeaderAddr( addr ) ((Heap.Storage.Header *)( (char *)addr - sizeof(Heap.Storage) ))
725#define RealHeader( header ) ((Heap.Storage.Header *)((char *)header - header->kind.fake.offset))
726
727// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
728// |header |addr
729//==================================================================================
730// align/offset |
731// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
732// |fake-header |addr
733#define DataStorage( bsize, addr, header ) (bsize - ( (char *)addr - (char *)header ))
734
735
736inline __attribute__((always_inline))
737static void checkAlign( size_t alignment ) {
738 if ( unlikely( alignment < libAlign() || ! is_pow2( alignment ) ) ) {
739 abort( "**** Error **** alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
740 } // if
741} // checkAlign
742
743
744inline __attribute__((always_inline))
745static void checkHeader( bool check, const char name[], void * addr ) {
746 if ( unlikely( check ) ) { // bad address ?
747 abort( "**** Error **** attempt to %s storage %p with address outside the heap.\n"
748 "Possible cause is duplicate free on same block or overwriting of memory.",
749 name, addr );
750 } // if
751} // checkHeader
752
753
754// Manipulate sticky bits stored in unused 3 low-order bits of an address.
755// bit0 => alignment => fake header
756// bit1 => zero filled (calloc)
757// bit2 => mapped allocation versus sbrk
758#define StickyBits( header ) (((header)->kind.real.blockSize & 0x7))
759#define ClearStickyBits( addr ) (typeof(addr))((uintptr_t)(addr) & ~7)
760#define MarkAlignmentBit( align ) ((align) | 1)
761#define AlignmentBit( header ) ((((header)->kind.fake.alignment) & 1))
762#define ClearAlignmentBit( header ) (((header)->kind.fake.alignment) & ~1)
763#define ZeroFillBit( header ) ((((header)->kind.real.blockSize) & 2))
764#define ClearZeroFillBit( header ) ((((header)->kind.real.blockSize) &= ~2))
765#define MarkZeroFilledBit( header ) ((header)->kind.real.blockSize |= 2)
766#define MmappedBit( header ) ((((header)->kind.real.blockSize) & 4))
767#define MarkMmappedBit( size ) ((size) | 4)
768
769
770inline __attribute__((always_inline))
771static void fakeHeader( Heap.Storage.Header *& header, size_t & alignment ) {
772 if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
773 alignment = ClearAlignmentBit( header ); // clear flag from value
774 #ifdef __CFA_DEBUG__
775 checkAlign( alignment ); // check alignment
776 #endif // __CFA_DEBUG__
777 header = RealHeader( header ); // backup from fake to real header
778 } else {
779 alignment = libAlign(); // => no fake header
780 } // if
781} // fakeHeader
782
783
784inline __attribute__((always_inline))
785static bool headers( const char name[] __attribute__(( unused )), void * addr, Heap.Storage.Header *& header,
786 Heap.FreeHeader *& freeHead, size_t & size, size_t & alignment ) with( heapMaster, *heapManager ) {
787 header = HeaderAddr( addr );
788
789 #ifdef __CFA_DEBUG__
790 checkHeader( header < (Heap.Storage.Header *)heapBegin, name, addr ); // bad low address ?
791 #endif // __CFA_DEBUG__
792
793 if ( likely( ! StickyBits( header ) ) ) { // no sticky bits ?
794 freeHead = (Heap.FreeHeader *)(header->kind.real.home);
795 alignment = libAlign();
796 } else {
797 fakeHeader( header, alignment );
798 if ( unlikely( MmappedBit( header ) ) ) { // mmapped ?
799 verify( addr < heapBegin || heapEnd < addr );
800 size = ClearStickyBits( header->kind.real.blockSize ); // mmap size
801 return true;
802 } // if
803
804 freeHead = (Heap.FreeHeader *)(ClearStickyBits( header->kind.real.home ));
805 } // if
806 size = freeHead->blockSize;
807
808 #ifdef __CFA_DEBUG__
809 checkHeader( header < (Heap.Storage.Header *)heapBegin || (Heap.Storage.Header *)heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
810
811 Heap * homeManager;
812 if ( unlikely( freeHead == 0p || // freed and only free-list node => null link
813 // freed and link points at another free block not to a bucket in the bucket array.
814 (homeManager = freeHead->homeManager, freeHead < &homeManager->freeLists[0] ||
815 &homeManager->freeLists[NoBucketSizes] <= freeHead ) ) ) {
816 abort( "**** Error **** attempt to %s storage %p with corrupted header.\n"
817 "Possible cause is duplicate free on same block or overwriting of header information.",
818 name, addr );
819 } // if
820 #endif // __CFA_DEBUG__
821
822 return false;
823} // headers
824
825
826static void * master_extend( size_t size ) with( heapMaster ) {
827 lock( extLock );
828
829 ptrdiff_t rem = heapRemaining - size;
830 if ( unlikely( rem < 0 ) ) {
831 // If the size requested is bigger than the current remaining storage, increase the size of the heap.
832
833 size_t increase = ceiling2( size > heapExpand ? size : heapExpand, libAlign() );
834 // Do not call abort or strerror( errno ) as they may call malloc.
835 if ( unlikely( sbrk( increase ) == (void *)-1 ) ) { // failed, no memory ?
836 unlock( extLock );
837 abort( NO_MEMORY_MSG, size ); // no memory
838 } // if
839
840 // Make storage executable for thunks.
841 if ( mprotect( (char *)heapEnd + heapRemaining, increase, __map_prot ) ) {
842 unlock( extLock );
843 abort( "**** Error **** attempt to make heap storage executable for thunks and mprotect failed with errno %d.", errno );
844 } // if
845
846 rem = heapRemaining + increase - size;
847
848 #ifdef __STATISTICS__
849 sbrk_calls += 1;
850 sbrk_storage += increase;
851 #endif // __STATISTICS__
852 } // if
853
854 Heap.Storage * block = (Heap.Storage *)heapEnd;
855 heapRemaining = rem;
856 heapEnd = (char *)heapEnd + size;
857
858 unlock( extLock );
859 return block;
860} // master_extend
861
862
863__attribute__(( noinline ))
864static void * manager_extend( size_t size ) with( *heapManager ) {
865 ptrdiff_t rem = heapReserve - size;
866
867 if ( unlikely( rem < 0 ) ) { // negative
868 // If the size requested is bigger than the current remaining reserve, use the current reserve to populate
869 // smaller freeLists, and increase the reserve.
870
871 rem = heapReserve; // positive
872
873 if ( rem >= bucketSizes[0] ) { // minimal size ? otherwise ignore
874 size_t bucket;
875 #ifdef FASTLOOKUP
876 if ( likely( rem < LookupSizes ) ) bucket = lookup[rem];
877 #endif // FASTLOOKUP
878 bucket = Bsearchl( rem, bucketSizes, heapMaster.maxBucketsUsed );
879 verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
880 Heap.FreeHeader * freeHead = &(freeLists[bucket]);
881
882 // The remaining storage many not be bucket size, whereas all other allocations are. Round down to previous
883 // bucket size in this case.
884 if ( unlikely( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
885 Heap.Storage * block = (Heap.Storage *)heapBuffer;
886
887 block->header.kind.real.next = freeHead->freeList; // push on stack
888 freeHead->freeList = block;
889 } // if
890
891 size_t increase = ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), libAlign() );
892 heapBuffer = master_extend( increase );
893 rem = increase - size;
894 } // if
895
896 Heap.Storage * block = (Heap.Storage *)heapBuffer;
897 heapReserve = rem;
898 heapBuffer = (char *)heapBuffer + size;
899
900 return block;
901} // manager_extend
902
903
904#define BOOT_HEAP_MANAGER \
905 if ( unlikely( ! heapMasterBootFlag ) ) { \
906 heapManagerCtor(); /* trigger for first heap */ \
907 } /* if */
908
909#ifdef __STATISTICS__
910#define STAT_NAME __counter
911#define STAT_PARM , unsigned int STAT_NAME
912#define STAT_ARG( name ) , name
913#define STAT_0_CNT( counter ) stats.counters[counter].calls_0 += 1
914#else
915#define STAT_NAME
916#define STAT_PARM
917#define STAT_ARG( name )
918#define STAT_0_CNT( counter )
919#endif // __STATISTICS__
920
921#define PROLOG( counter, ... ) \
922 BOOT_HEAP_MANAGER; \
923 if ( unlikely( size == 0 ) || /* 0 BYTE ALLOCATION RETURNS NULL POINTER */ \
924 unlikely( size > ULONG_MAX - sizeof(Heap.Storage) ) ) { /* error check */ \
925 STAT_0_CNT( counter ); \
926 __VA_ARGS__; \
927 return 0p; \
928 } /* if */
929
930
931#define SCRUB_SIZE 1024lu
932// Do not use '\xfe' for scrubbing because dereferencing an address composed of it causes a SIGSEGV *without* a valid IP
933// pointer in the interrupt frame.
934#define SCRUB '\xff'
935
936static void * doMalloc( size_t size STAT_PARM ) libcfa_nopreempt with( *heapManager ) {
937 PROLOG( STAT_NAME );
938
939 verify( heapManager );
940 Heap.Storage * block; // pointer to new block of storage
941
942 // Look up size in the size list. Make sure the user request includes space for the header that must be allocated
943 // along with the block and is a multiple of the alignment size.
944 size_t tsize = size + sizeof(Heap.Storage);
945
946 #ifdef __STATISTICS__
947 stats.counters[STAT_NAME].calls += 1;
948 stats.counters[STAT_NAME].request += size;
949 #endif // __STATISTICS__
950
951 #ifdef __CFA_DEBUG__
952 allocUnfreed += size;
953 #endif // __CFA_DEBUG__
954
955 if ( likely( tsize < heapMaster.mmapStart ) ) { // small size => sbrk
956 size_t bucket;
957 #ifdef FASTLOOKUP
958 if ( likely( tsize < LookupSizes ) ) bucket = lookup[tsize];
959 else
960 #endif // FASTLOOKUP
961 bucket = Bsearchl( tsize, bucketSizes, heapMaster.maxBucketsUsed );
962 verify( 0 <= bucket && bucket <= heapMaster.maxBucketsUsed );
963 Heap.FreeHeader * freeHead = &freeLists[bucket];
964
965 verify( freeHead <= &freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
966 verify( tsize <= freeHead->blockSize ); // search failure ?
967
968 tsize = freeHead->blockSize; // total space needed for request
969 #ifdef __STATISTICS__
970 stats.counters[STAT_NAME].alloc += tsize;
971 #endif // __STATISTICS__
972
973 // Spin until the lock is acquired for this particular size of block.
974
975 #if BUCKETLOCK == SPINLOCK
976 block = freeHead->freeList; // remove node from stack
977 #else
978 block = pop( freeHead->freeList );
979 #endif // BUCKETLOCK
980 if ( unlikely( block == 0p ) ) { // no free block ?
981 #ifdef OWNERSHIP
982 // Freelist for that size is empty, so carve it out of the heap, if there is enough left, or get some more
983 // and then carve it off.
984 #ifdef RETURNSPIN
985 #if BUCKETLOCK == SPINLOCK
986 lock( freeHead->returnLock );
987 block = freeHead->returnList;
988 freeHead->returnList = 0p;
989 unlock( freeHead->returnLock );
990 #else
991 block = __atomic_exchange_n( &freeHead->returnList, nullptr, __ATOMIC_SEQ_CST );
992 #endif // RETURNSPIN
993
994 if ( likely( block == 0p ) ) { // return list also empty?
995 #endif // OWNERSHIP
996 // Do not leave kernel thread as manager_extend accesses heapManager.
997 disable_interrupts();
998 block = (Heap.Storage *)manager_extend( tsize ); // mutual exclusion on call
999 enable_interrupts( false );
1000
1001 // OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1002
1003 #ifdef __CFA_DEBUG__
1004 // Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first 1024 bytes.
1005 memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1006 #endif // __CFA_DEBUG__
1007 #endif // BUCKETLOCK
1008 #ifdef OWNERSHIP
1009 } else { // merge returnList into freeHead
1010 #ifdef __STATISTICS__
1011 stats.return_pulls += 1;
1012 #endif // __STATISTICS__
1013
1014 // OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1015
1016 freeHead->freeList = block->header.kind.real.next;
1017 } // if
1018 #endif // OWNERSHIP
1019 } else {
1020 // Memory is scrubbed in doFree.
1021 freeHead->freeList = block->header.kind.real.next;
1022 } // if
1023
1024 block->header.kind.real.home = freeHead; // pointer back to free list of apropriate size
1025 } else { // large size => mmap
1026 if ( unlikely( size > ULONG_MAX - __page_size ) ) return 0p;
1027 tsize = ceiling2( tsize, __page_size ); // must be multiple of page size
1028 #ifdef __STATISTICS__
1029 stats.counters[STAT_NAME].alloc += tsize;
1030 stats.mmap_calls += 1;
1031 stats.mmap_storage_request += size;
1032 stats.mmap_storage_alloc += tsize;
1033 #endif // __STATISTICS__
1034
1035 disable_interrupts();
1036 block = (Heap.Storage *)mmap( 0, tsize, __map_prot, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0 );
1037 enable_interrupts( false );
1038
1039 // OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1040
1041 if ( unlikely( block == (Heap.Storage *)MAP_FAILED ) ) { // failed ?
1042 if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
1043 // Do not call strerror( errno ) as it may call malloc.
1044 abort( "**** Error **** attempt to allocate large object (> %zu) of size %zu bytes and mmap failed with errno %d.", size, heapMaster.mmapStart, errno );
1045 } // if
1046 block->header.kind.real.blockSize = MarkMmappedBit( tsize ); // storage size for munmap
1047
1048 #ifdef __CFA_DEBUG__
1049 // Scrub new memory so subsequent uninitialized usages might fail. Only scrub the first 1024 bytes. The rest of
1050 // the storage set to 0 by mmap.
1051 memset( block->data, SCRUB, min( SCRUB_SIZE, tsize - sizeof(Heap.Storage) ) );
1052 #endif // __CFA_DEBUG__
1053 } // if
1054
1055 block->header.kind.real.size = size; // store allocation size
1056 void * addr = &(block->data); // adjust off header to user bytes
1057 verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?
1058
1059 #ifdef __CFA_DEBUG__
1060 if ( traceHeap() ) {
1061 char helpText[64];
1062 __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1063 "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize ); // print debug/nodebug
1064 } // if
1065 #endif // __CFA_DEBUG__
1066
1067// poll_interrupts(); // call rollforward
1068
1069 return addr;
1070} // doMalloc
1071
1072
1073static void doFree( void * addr ) libcfa_nopreempt with( *heapManager ) {
1074 verify( addr );
1075
1076 // detect free after thread-local storage destruction and use global stats in that case
1077
1078 Heap.Storage.Header * header;
1079 Heap.FreeHeader * freeHead;
1080 size_t size, alignment;
1081
1082 bool mapped = headers( "free", addr, header, freeHead, size, alignment );
1083 #if defined( __STATISTICS__ ) || defined( __CFA_DEBUG__ )
1084 size_t rsize = header->kind.real.size; // optimization
1085 #endif // __STATISTICS__ || __CFA_DEBUG__
1086
1087 #ifdef __STATISTICS__
1088 stats.free_storage_request += rsize;
1089 stats.free_storage_alloc += size;
1090 #endif // __STATISTICS__
1091
1092 #ifdef __CFA_DEBUG__
1093 allocUnfreed -= rsize;
1094 #endif // __CFA_DEBUG__
1095
1096 if ( unlikely( mapped ) ) { // mmapped ?
1097 #ifdef __STATISTICS__
1098 stats.munmap_calls += 1;
1099 stats.munmap_storage_request += rsize;
1100 stats.munmap_storage_alloc += size;
1101 #endif // __STATISTICS__
1102
1103 // OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1104
1105 // Does not matter where this storage is freed.
1106 if ( unlikely( munmap( header, size ) == -1 ) ) {
1107 // Do not call strerror( errno ) as it may call malloc.
1108 abort( "**** Error **** attempt to deallocate large object %p and munmap failed with errno %d.\n"
1109 "Possible cause is invalid delete pointer: either not allocated or with corrupt header.",
1110 addr, errno );
1111 } // if
1112 } else {
1113 #ifdef __CFA_DEBUG__
1114 // memset is NOT always inlined!
1115 disable_interrupts();
1116 // Scrub old memory so subsequent usages might fail. Only scrub the first/last SCRUB_SIZE bytes.
1117 char * data = ((Heap.Storage *)header)->data; // data address
1118 size_t dsize = size - sizeof(Heap.Storage); // data size
1119 if ( dsize <= SCRUB_SIZE * 2 ) {
1120 memset( data, SCRUB, dsize ); // scrub all
1121 } else {
1122 memset( data, SCRUB, SCRUB_SIZE ); // scrub front
1123 memset( data + dsize - SCRUB_SIZE, SCRUB, SCRUB_SIZE ); // scrub back
1124 } // if
1125 enable_interrupts( false );
1126 #endif // __CFA_DEBUG__
1127
1128 if ( likely( heapManager == freeHead->homeManager ) ) { // belongs to this thread
1129 header->kind.real.next = freeHead->freeList; // push on stack
1130 freeHead->freeList = (Heap.Storage *)header;
1131 } else { // return to thread owner
1132 verify( heapManager );
1133
1134 #ifdef OWNERSHIP
1135 #ifdef RETURNSPIN
1136 lock( freeHead->returnLock );
1137 header->kind.real.next = freeHead->returnList; // push to bucket return list
1138 freeHead->returnList = (Heap.Storage *)header;
1139 unlock( freeHead->returnLock );
1140 #else // lock free
1141 header->kind.real.next = freeHead->returnList; // link new node to top node
1142 // CAS resets header->kind.real.next = freeHead->returnList on failure
1143 while ( ! __atomic_compare_exchange_n( &freeHead->returnList, &header->kind.real.next, header,
1144 false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
1145 #endif // RETURNSPIN
1146
1147 #else // no OWNERSHIP
1148
1149 freeHead = &heap->freeLists[ClearStickyBits( header->kind.real.home ) - &freeHead->homeManager->freeLists[0]];
1150 header->kind.real.next = freeHead->freeList; // push on stack
1151 freeHead->freeList = (Heap.Storage *)header;
1152 #endif // ! OWNERSHIP
1153
1154 #ifdef __U_STATISTICS__
1155 stats.return_pushes += 1;
1156 stats.return_storage_request += rsize;
1157 stats.return_storage_alloc += size;
1158 #endif // __U_STATISTICS__
1159
1160 // OK TO BE PREEMPTED HERE AS heapManager IS NO LONGER ACCESSED.
1161 } // if
1162 } // if
1163
1164 #ifdef __CFA_DEBUG__
1165 if ( traceHeap() ) {
1166 char helpText[64];
1167 __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
1168 "Free( %p ) size:%zu\n", addr, size ); // print debug/nodebug
1169 } // if
1170 #endif // __CFA_DEBUG__
1171
1172// poll_interrupts(); // call rollforward
1173} // doFree
1174
1175
1176size_t prtFree( Heap & manager ) with( manager ) {
1177 size_t total = 0;
1178 #ifdef __STATISTICS__
1179 __cfaabi_bits_acquire();
1180 __cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
1181 #endif // __STATISTICS__
1182 for ( unsigned int i = 0; i < heapMaster.maxBucketsUsed; i += 1 ) {
1183 size_t size = freeLists[i].blockSize;
1184 #ifdef __STATISTICS__
1185 unsigned int N = 0;
1186 #endif // __STATISTICS__
1187
1188 #if BUCKETLOCK == SPINLOCK
1189 for ( Heap.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
1190 #else
1191 for(;;) {
1192// for ( Heap.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
1193// for ( Heap.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
1194// Heap.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
1195// typeof(p) temp = (( p )`next)->top; // FIX ME: direct assignent fails, initialization works`
1196// p = temp;
1197 #endif // BUCKETLOCK
1198 total += size;
1199 #ifdef __STATISTICS__
1200 N += 1;
1201 #endif // __STATISTICS__
1202 } // for
1203
1204 #ifdef __STATISTICS__
1205 __cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u ", size, N );
1206 if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
1207 #endif // __STATISTICS__
1208 } // for
1209 #ifdef __STATISTICS__
1210 __cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
1211 __cfaabi_bits_release();
1212 #endif // __STATISTICS__
1213 return (char *)heapMaster.heapEnd - (char *)heapMaster.heapBegin - total;
1214} // prtFree
1215
1216
1217#ifdef __STATISTICS__
1218static void incCalls( intptr_t statName ) libcfa_nopreempt {
1219 heapManager->stats.counters[statName].calls += 1;
1220} // incCalls
1221
1222static void incZeroCalls( intptr_t statName ) libcfa_nopreempt {
1223 heapManager->stats.counters[statName].calls_0 += 1;
1224} // incZeroCalls
1225#endif // __STATISTICS__
1226
1227#ifdef __CFA_DEBUG__
1228static void incUnfreed( intptr_t offset ) libcfa_nopreempt {
1229 heapManager->allocUnfreed += offset;
1230} // incUnfreed
1231#endif // __CFA_DEBUG__
1232
1233
1234static void * memalignNoStats( size_t alignment, size_t size STAT_PARM ) {
1235 checkAlign( alignment ); // check alignment
1236
1237 // if alignment <= default alignment or size == 0, do normal malloc as two headers are unnecessary
1238 if ( unlikely( alignment <= libAlign() || size == 0 ) ) return doMalloc( size STAT_ARG( STAT_NAME ) );
1239
1240 // Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
1241 // administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
1242 // .-------------v-----------------v----------------v----------,
1243 // | Real Header | ... padding ... | Fake Header | data ... |
1244 // `-------------^-----------------^-+--------------^----------'
1245 // |<--------------------------------' offset/align |<-- alignment boundary
1246
1247 // subtract libAlign() because it is already the minimum alignment
1248 // add sizeof(Storage) for fake header
1249 size_t offset = alignment - libAlign() + sizeof(Heap.Storage);
1250 char * addr = (char *)doMalloc( size + offset STAT_ARG( STAT_NAME ) );
1251
1252 // address in the block of the "next" alignment address
1253 char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(Heap.Storage)), alignment );
1254
1255 // address of header from malloc
1256 Heap.Storage.Header * realHeader = HeaderAddr( addr );
1257 realHeader->kind.real.size = size; // correct size to eliminate above alignment offset
1258 #ifdef __CFA_DEBUG__
1259 incUnfreed( -offset ); // adjustment off the offset from call to doMalloc
1260 #endif // __CFA_DEBUG__
1261
1262 // address of fake header *before* the alignment location
1263 Heap.Storage.Header * fakeHeader = HeaderAddr( user );
1264
1265 // SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
1266 fakeHeader->kind.fake.offset = (char *)fakeHeader - (char *)realHeader;
1267 // SKULLDUGGERY: odd alignment implies fake header
1268 fakeHeader->kind.fake.alignment = MarkAlignmentBit( alignment );
1269
1270 return user;
1271} // memalignNoStats
1272
1273
1274//####################### Memory Allocation Routines ####################
1275
1276
1277extern "C" {
1278 // Allocates size bytes and returns a pointer to the allocated memory. The contents are undefined. If size is 0,
1279 // then malloc() returns a unique pointer value that can later be successfully passed to free().
1280 void * malloc( size_t size ) libcfa_public {
1281 return doMalloc( size STAT_ARG( MALLOC ) );
1282 } // malloc
1283
1284
1285 // Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
1286 void * aalloc( size_t dim, size_t elemSize ) libcfa_public {
1287 return doMalloc( dim * elemSize STAT_ARG( AALLOC ) );
1288 } // aalloc
1289
1290
1291 // Same as aalloc() with memory set to zero.
1292 void * calloc( size_t dim, size_t elemSize ) libcfa_public {
1293 size_t size = dim * elemSize;
1294 char * addr = (char *)doMalloc( size STAT_ARG( CALLOC ) );
1295
1296 if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1297
1298 Heap.Storage.Header * header;
1299 Heap.FreeHeader * freeHead;
1300 size_t bsize, alignment;
1301
1302 #ifndef __CFA_DEBUG__
1303 bool mapped =
1304 #endif // __CFA_DEBUG__
1305 headers( "calloc", addr, header, freeHead, bsize, alignment );
1306
1307 #ifndef __CFA_DEBUG__
1308 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1309 if ( likely( ! mapped ) )
1310 #endif // __CFA_DEBUG__
1311 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1312 // `-header`-addr `-size
1313 memset( addr, '\0', size ); // set to zeros
1314
1315 MarkZeroFilledBit( header ); // mark as zero fill
1316 return addr;
1317 } // calloc
1318
1319
1320 // Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined. If oaddr is
1321 // 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
1322 // not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, it must have been returned by an earlier
1323 // call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
1324 void * resize( void * oaddr, size_t size ) libcfa_public {
1325 if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1326 return doMalloc( size STAT_ARG( RESIZE ) );
1327 } // if
1328
1329 PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1330
1331 Heap.Storage.Header * header;
1332 Heap.FreeHeader * freeHead;
1333 size_t bsize, oalign;
1334 headers( "resize", oaddr, header, freeHead, bsize, oalign );
1335
1336 size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1337 // same size, DO NOT preserve STICKY PROPERTIES.
1338 if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
1339 ClearZeroFillBit( header ); // no alignment and turn off 0 fill
1340 #ifdef __CFA_DEBUG__
1341 incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1342 #endif // __CFA_DEBUG__
1343 header->kind.real.size = size; // reset allocation size
1344 #ifdef __STATISTICS__
1345 incCalls( RESIZE );
1346 #endif // __STATISTICS__
1347 return oaddr;
1348 } // if
1349
1350 // change size, DO NOT preserve STICKY PROPERTIES.
1351 doFree( oaddr ); // free previous storage
1352
1353 return doMalloc( size STAT_ARG( RESIZE ) ); // create new area
1354 } // resize
1355
1356
1357 // Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
1358 // the old and new sizes.
1359 void * realloc( void * oaddr, size_t size ) libcfa_public {
1360 if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1361 return doMalloc( size STAT_ARG( REALLOC ) );
1362 } // if
1363
1364 PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1365
1366 Heap.Storage.Header * header;
1367 Heap.FreeHeader * freeHead;
1368 size_t bsize, oalign;
1369 headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1370
1371 size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1372 size_t osize = header->kind.real.size; // old allocation size
1373 bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1374 if ( unlikely( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
1375 #ifdef __CFA_DEBUG__
1376 incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1377 #endif // __CFA_DEBUG__
1378 header->kind.real.size = size; // reset allocation size
1379 if ( unlikely( ozfill ) && size > osize ) { // previous request zero fill and larger ?
1380 memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
1381 } // if
1382 #ifdef __STATISTICS__
1383 incCalls( REALLOC );
1384 #endif // __STATISTICS__
1385 return oaddr;
1386 } // if
1387
1388 // change size and copy old content to new storage
1389
1390 void * naddr;
1391 if ( likely( oalign <= libAlign() ) ) { // previous request not aligned ?
1392 naddr = doMalloc( size STAT_ARG( REALLOC ) ); // create new area
1393 } else {
1394 naddr = memalignNoStats( oalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1395 } // if
1396
1397 headers( "realloc", naddr, header, freeHead, bsize, oalign );
1398 // To preserve prior fill, the entire bucket must be copied versus the size.
1399 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1400 doFree( oaddr ); // free previous storage
1401
1402 if ( unlikely( ozfill ) ) { // previous request zero fill ?
1403 MarkZeroFilledBit( header ); // mark new request as zero filled
1404 if ( size > osize ) { // previous request larger ?
1405 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1406 } // if
1407 } // if
1408 return naddr;
1409 } // realloc
1410
1411
1412 // Same as realloc() except the new allocation size is large enough for an array of nelem elements of size elsize.
1413 void * reallocarray( void * oaddr, size_t dim, size_t elemSize ) libcfa_public {
1414 return realloc( oaddr, dim * elemSize );
1415 } // reallocarray
1416
1417
1418 // Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
1419 void * memalign( size_t alignment, size_t size ) libcfa_public {
1420 return memalignNoStats( alignment, size STAT_ARG( MEMALIGN ) );
1421 } // memalign
1422
1423
1424 // Same as aalloc() with memory alignment.
1425 void * amemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1426 return memalignNoStats( alignment, dim * elemSize STAT_ARG( AMEMALIGN ) );
1427 } // amemalign
1428
1429
1430 // Same as calloc() with memory alignment.
1431 void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) libcfa_public {
1432 size_t size = dim * elemSize;
1433 char * addr = (char *)memalignNoStats( alignment, size STAT_ARG( CMEMALIGN ) );
1434
1435 if ( unlikely( addr == NULL ) ) return NULL; // stop further processing if 0p is returned
1436
1437 Heap.Storage.Header * header;
1438 Heap.FreeHeader * freeHead;
1439 size_t bsize;
1440
1441 #ifndef __CFA_DEBUG__
1442 bool mapped =
1443 #endif // __CFA_DEBUG__
1444 headers( "cmemalign", addr, header, freeHead, bsize, alignment );
1445
1446 // Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
1447 #ifndef __CFA_DEBUG__
1448 if ( ! mapped )
1449 #endif // __CFA_DEBUG__
1450 // <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
1451 // `-header`-addr `-size
1452 memset( addr, '\0', size ); // set to zeros
1453
1454 MarkZeroFilledBit( header ); // mark as zero filled
1455 return addr;
1456 } // cmemalign
1457
1458
1459 // Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
1460 // of alignment. This requirement is universally ignored.
1461 void * aligned_alloc( size_t alignment, size_t size ) libcfa_public {
1462 return memalign( alignment, size );
1463 } // aligned_alloc
1464
1465
1466 // Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
1467 // memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
1468 // is 0, then posix_memalign() returns either 0p, or a unique pointer value that can later be successfully passed to
1469 // free(3).
1470 int posix_memalign( void ** memptr, size_t alignment, size_t size ) libcfa_public {
1471 if ( unlikely( alignment < libAlign() || ! is_pow2( alignment ) ) ) return EINVAL; // check alignment
1472 *memptr = memalign( alignment, size );
1473 return 0;
1474 } // posix_memalign
1475
1476
1477 // Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
1478 // page size. It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
1479 void * valloc( size_t size ) libcfa_public {
1480 return memalign( __page_size, size );
1481 } // valloc
1482
1483
1484 // Same as valloc but rounds size to multiple of page size.
1485 void * pvalloc( size_t size ) libcfa_public {
1486 return memalign( __page_size, ceiling2( size, __page_size ) ); // round size to multiple of page size
1487 } // pvalloc
1488
1489
1490 // Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
1491 // or realloc(). Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
1492 // 0p, no operation is performed.
1493 void free( void * addr ) libcfa_public {
1494// verify( heapManager );
1495
1496 if ( unlikely( addr == 0p ) ) { // special case
1497 #ifdef __STATISTICS__
1498 if ( heapManager )
1499 incZeroCalls( FREE );
1500 #endif // __STATISTICS__
1501 return;
1502 } // if
1503
1504 #ifdef __STATISTICS__
1505 incCalls( FREE );
1506 #endif // __STATISTICS__
1507
1508 doFree( addr ); // handles heapManager == nullptr
1509 } // free
1510
1511
1512 // Returns the alignment of an allocation.
1513 size_t malloc_alignment( void * addr ) libcfa_public {
1514 if ( unlikely( addr == 0p ) ) return libAlign(); // minimum alignment
1515 Heap.Storage.Header * header = HeaderAddr( addr );
1516 if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1517 return ClearAlignmentBit( header ); // clear flag from value
1518 } else {
1519 return libAlign(); // minimum alignment
1520 } // if
1521 } // malloc_alignment
1522
1523
1524 // Returns true if the allocation is zero filled, e.g., allocated by calloc().
1525 bool malloc_zero_fill( void * addr ) libcfa_public {
1526 if ( unlikely( addr == 0p ) ) return false; // null allocation is not zero fill
1527 Heap.Storage.Header * header = HeaderAddr( addr );
1528 if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1529 header = RealHeader( header ); // backup from fake to real header
1530 } // if
1531 return ZeroFillBit( header ); // zero filled ?
1532 } // malloc_zero_fill
1533
1534
1535 // Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
1536 size_t malloc_size( void * addr ) libcfa_public {
1537 if ( unlikely( addr == 0p ) ) return 0; // null allocation has zero size
1538 Heap.Storage.Header * header = HeaderAddr( addr );
1539 if ( unlikely( AlignmentBit( header ) ) ) { // fake header ?
1540 header = RealHeader( header ); // backup from fake to real header
1541 } // if
1542 return header->kind.real.size;
1543 } // malloc_size
1544
1545
1546 // Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
1547 // malloc or a related function.
1548 size_t malloc_usable_size( void * addr ) libcfa_public {
1549 if ( unlikely( addr == 0p ) ) return 0; // null allocation has 0 size
1550 Heap.Storage.Header * header;
1551 Heap.FreeHeader * freeHead;
1552 size_t bsize, alignment;
1553
1554 headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
1555 return DataStorage( bsize, addr, header ); // data storage in bucket
1556 } // malloc_usable_size
1557
1558
1559 // Prints (on default standard error) statistics about memory allocated by malloc and related functions.
1560 void malloc_stats( void ) libcfa_public {
1561 #ifdef __STATISTICS__
1562 HeapStatistics stats;
1563 HeapStatisticsCtor( stats );
1564 if ( printStats( collectStats( stats ) ) == -1 ) {
1565 #else
1566 #define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
1567 if ( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ) == -1 ) {
1568 #endif // __STATISTICS__
1569 abort( "**** Error **** write failed in malloc_stats" );
1570 } // if
1571 } // malloc_stats
1572
1573
1574 // Changes the file descriptor where malloc_stats() writes statistics.
1575 int malloc_stats_fd( int fd __attribute__(( unused )) ) libcfa_public {
1576 #ifdef __STATISTICS__
1577 int temp = heapMaster.stats_fd;
1578 heapMaster.stats_fd = fd;
1579 return temp;
1580 #else
1581 return -1; // unsupported
1582 #endif // __STATISTICS__
1583 } // malloc_stats_fd
1584
1585
1586 // Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
1587 // The string is printed on the file stream stream. The exported string includes information about all arenas (see
1588 // malloc).
1589 int malloc_info( int options, FILE * stream __attribute__(( unused )) ) libcfa_public {
1590 if ( options != 0 ) { errno = EINVAL; return -1; }
1591 #ifdef __STATISTICS__
1592 HeapStatistics stats;
1593 HeapStatisticsCtor( stats );
1594 return printStatsXML( collectStats( stats ), stream ); // returns bytes written or -1
1595 #else
1596 return 0; // unsupported
1597 #endif // __STATISTICS__
1598 } // malloc_info
1599
1600
1601 // Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
1602 // specifies the parameter to be modified, and value specifies the new value for that parameter.
1603 int mallopt( int option, int value ) libcfa_public {
1604 if ( value < 0 ) return 0;
1605 choose( option ) {
1606 case M_TOP_PAD:
1607 heapMaster.heapExpand = ceiling2( value, __page_size );
1608 return 1;
1609 case M_MMAP_THRESHOLD:
1610 if ( setMmapStart( value ) ) return 1;
1611 } // choose
1612 return 0; // error, unsupported
1613 } // mallopt
1614
1615
1616 // Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
1617 int malloc_trim( size_t ) libcfa_public {
1618 return 0; // => impossible to release memory
1619 } // malloc_trim
1620
1621
1622 // Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
1623 // or the state of malloc_hook functions pointers). The state is recorded in a system-dependent opaque data
1624 // structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
1625 // result. (The caller must free this memory.)
1626 void * malloc_get_state( void ) libcfa_public {
1627 return 0p; // unsupported
1628 } // malloc_get_state
1629
1630
1631 // Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
1632 // structure pointed to by state.
1633 int malloc_set_state( void * ) libcfa_public {
1634 return 0; // unsupported
1635 } // malloc_set_state
1636
1637
1638 // Sets the amount (bytes) to extend the heap when there is insufficent free storage to service an allocation.
1639 __attribute__((weak)) size_t malloc_expansion() libcfa_public { return __CFA_DEFAULT_HEAP_EXPANSION__; }
1640
1641 // Sets the crossover point between allocations occuring in the sbrk area or separately mmapped.
1642 __attribute__((weak)) size_t malloc_mmap_start() libcfa_public { return __CFA_DEFAULT_MMAP_START__; }
1643
1644 // Amount subtracted to adjust for unfreed program storage (debug only).
1645 __attribute__((weak)) size_t malloc_unfreed() libcfa_public { return __CFA_DEFAULT_HEAP_UNFREED__; }
1646} // extern "C"
1647
1648
1649// Must have CFA linkage to overload with C linkage realloc.
1650void * resize( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1651 if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1652 return memalignNoStats( nalign, size STAT_ARG( RESIZE ) );
1653 } // if
1654
1655 PROLOG( RESIZE, doFree( oaddr ) ); // => free( oaddr )
1656
1657 // Attempt to reuse existing alignment.
1658 Heap.Storage.Header * header = HeaderAddr( oaddr );
1659 bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1660 size_t oalign;
1661
1662 if ( unlikely( isFakeHeader ) ) {
1663 checkAlign( nalign ); // check alignment
1664 oalign = ClearAlignmentBit( header ); // old alignment
1665 if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1666 && ( oalign <= nalign // going down
1667 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1668 ) ) {
1669 HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1670 Heap.FreeHeader * freeHead;
1671 size_t bsize, oalign;
1672 headers( "resize", oaddr, header, freeHead, bsize, oalign );
1673 size_t odsize = DataStorage( bsize, oaddr, header ); // data storage available in bucket
1674
1675 if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
1676 HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1677 ClearZeroFillBit( header ); // turn off 0 fill
1678 #ifdef __CFA_DEBUG__
1679 incUnfreed( size - header->kind.real.size ); // adjustment off the size difference
1680 #endif // __CFA_DEBUG__
1681 header->kind.real.size = size; // reset allocation size
1682 #ifdef __STATISTICS__
1683 incCalls( RESIZE );
1684 #endif // __STATISTICS__
1685 return oaddr;
1686 } // if
1687 } // if
1688 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1689 && nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1690 return resize( oaddr, size ); // duplicate special case checks
1691 } // if
1692
1693 // change size, DO NOT preserve STICKY PROPERTIES.
1694 doFree( oaddr ); // free previous storage
1695 return memalignNoStats( nalign, size STAT_ARG( RESIZE ) ); // create new aligned area
1696} // resize
1697
1698
1699void * realloc( void * oaddr, size_t nalign, size_t size ) libcfa_public {
1700 if ( unlikely( oaddr == 0p ) ) { // => malloc( size )
1701 return memalignNoStats( nalign, size STAT_ARG( REALLOC ) );
1702 } // if
1703
1704 PROLOG( REALLOC, doFree( oaddr ) ); // => free( oaddr )
1705
1706 // Attempt to reuse existing alignment.
1707 Heap.Storage.Header * header = HeaderAddr( oaddr );
1708 bool isFakeHeader = AlignmentBit( header ); // old fake header ?
1709 size_t oalign;
1710 if ( unlikely( isFakeHeader ) ) {
1711 checkAlign( nalign ); // check alignment
1712 oalign = ClearAlignmentBit( header ); // old alignment
1713 if ( unlikely( (uintptr_t)oaddr % nalign == 0 // lucky match ?
1714 && ( oalign <= nalign // going down
1715 || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
1716 ) ) {
1717 HeaderAddr( oaddr )->kind.fake.alignment = MarkAlignmentBit( nalign ); // update alignment (could be the same)
1718 return realloc( oaddr, size ); // duplicate special case checks
1719 } // if
1720 } else if ( ! isFakeHeader // old real header (aligned on libAlign) ?
1721 && nalign == libAlign() ) { // new alignment also on libAlign => no fake header needed
1722 return realloc( oaddr, size ); // duplicate special case checks
1723 } // if
1724
1725 Heap.FreeHeader * freeHead;
1726 size_t bsize;
1727 headers( "realloc", oaddr, header, freeHead, bsize, oalign );
1728
1729 // change size and copy old content to new storage
1730
1731 size_t osize = header->kind.real.size; // old allocation size
1732 bool ozfill = ZeroFillBit( header ); // old allocation zero filled
1733
1734 void * naddr = memalignNoStats( nalign, size STAT_ARG( REALLOC ) ); // create new aligned area
1735
1736 headers( "realloc", naddr, header, freeHead, bsize, oalign );
1737 memcpy( naddr, oaddr, min( osize, size ) ); // copy bytes
1738 doFree( oaddr ); // free previous storage
1739
1740 if ( unlikely( ozfill ) ) { // previous request zero fill ?
1741 MarkZeroFilledBit( header ); // mark new request as zero filled
1742 if ( size > osize ) { // previous request larger ?
1743 memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
1744 } // if
1745 } // if
1746 return naddr;
1747} // realloc
1748
1749
1750void * reallocarray( void * oaddr, size_t nalign, size_t dim, size_t elemSize ) __THROW {
1751 return realloc( oaddr, nalign, dim * elemSize );
1752} // reallocarray
1753
1754
1755// Local Variables: //
1756// tab-width: 4 //
1757// compile-command: "cfa -nodebug -O2 heap.cfa" //
1758// End: //
Note: See TracBrowser for help on using the repository browser.