1 // Copyright (c) 2005, 2007, Google Inc.
2 // All rights reserved.
3 // Copyright (C) 2005, 2006, 2007, 2008, 2009, 2011 Apple Inc. All rights reserved.
4 //
5 // Redistribution and use in source and binary forms, with or without
6 // modification, are permitted provided that the following conditions are
7 // met:
8 //
9 // * Redistributions of source code must retain the above copyright
10 // notice, this list of conditions and the following disclaimer.
11 // * Redistributions in binary form must reproduce the above
12 // copyright notice, this list of conditions and the following disclaimer
13 // in the documentation and/or other materials provided with the
14 // distribution.
15 // * Neither the name of Google Inc. nor the names of its
16 // contributors may be used to endorse or promote products derived from
17 // this software without specific prior written permission.
18 //
19 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
25 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30
31 // ---
32 // Author: Sanjay Ghemawat <opensource@google.com>
33 //
34 // A malloc that uses a per-thread cache to satisfy small malloc requests.
35 // (The time for malloc/free of a small object drops from 300 ns to 50 ns.)
36 //
37 // See doc/tcmalloc.html for a high-level
38 // description of how this malloc works.
39 //
40 // SYNCHRONIZATION
41 // 1. The thread-specific lists are accessed without acquiring any locks.
42 // This is safe because each such list is only accessed by one thread.
43 // 2. We have a lock per central free-list, and hold it while manipulating
44 // the central free list for a particular size.
45 // 3. The central page allocator is protected by "pageheap_lock".
46 // 4. The pagemap (which maps from page-number to descriptor),
47 // can be read without holding any locks, and written while holding
48 // the "pageheap_lock".
49 // 5. To improve performance, a subset of the information one can get
50 // from the pagemap is cached in a data structure, pagemap_cache_,
51 // that atomically reads and writes its entries. This cache can be
52 // read and written without locking.
53 //
54 // This multi-threaded access to the pagemap is safe for fairly
55 // subtle reasons. We basically assume that when an object X is
56 // allocated by thread A and deallocated by thread B, there must
57 // have been appropriate synchronization in the handoff of object
58 // X from thread A to thread B. The same logic applies to pagemap_cache_.
59 //
60 // THE PAGEID-TO-SIZECLASS CACHE
61 // Hot PageID-to-sizeclass mappings are held by pagemap_cache_. If this cache
62 // returns 0 for a particular PageID then that means "no information," not that
63 // the sizeclass is 0. The cache may have stale information for pages that do
64 // not hold the beginning of any free()'able object. Staleness is eliminated
65 // in Populate() for pages with sizeclass > 0 objects, and in do_malloc() and
66 // do_memalign() for all other relevant pages.
67 //
68 // TODO: Bias reclamation to larger addresses
69 // TODO: implement mallinfo/mallopt
70 // TODO: Better testing
71 //
72 // 9/28/2003 (new page-level allocator replaces ptmalloc2):
73 // * malloc/free of small objects goes from ~300 ns to ~50 ns.
74 // * allocation of a reasonably complicated struct
75 // goes from about 1100 ns to about 300 ns.
76
77 #include "config.h"
78 #include "FastMalloc.h"
79
80 #include "Assertions.h"
81 #include <limits>
82 #if ENABLE(JSC_MULTIPLE_THREADS)
83 #include <pthread.h>
84 #endif
85 #include <wtf/StdLibExtras.h>
86
87 #ifndef NO_TCMALLOC_SAMPLES
88 #ifdef WTF_CHANGES
89 #define NO_TCMALLOC_SAMPLES
90 #endif
91 #endif
92
93 #if !(defined(USE_SYSTEM_MALLOC) && USE_SYSTEM_MALLOC) && defined(NDEBUG)
94 #define FORCE_SYSTEM_MALLOC 0
95 #else
96 #define FORCE_SYSTEM_MALLOC 1
97 #endif
98
99 // Use a background thread to periodically scavenge memory to release back to the system
100 // https://bugs.webkit.org/show_bug.cgi?id=27900: don't turn this on for Tiger until we have figured out why it caused a crash.
101 #if defined(BUILDING_ON_TIGER)
102 #define USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY 0
103 #else
104 #define USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY 1
105 #endif
106
107 #ifndef NDEBUG
108 namespace WTF {
109
110 #if ENABLE(JSC_MULTIPLE_THREADS)
111 static pthread_key_t isForbiddenKey;
112 static pthread_once_t isForbiddenKeyOnce = PTHREAD_ONCE_INIT;
initializeIsForbiddenKey()113 static void initializeIsForbiddenKey()
114 {
115 pthread_key_create(&isForbiddenKey, 0);
116 }
117
118 #if !ASSERT_DISABLED
isForbidden()119 static bool isForbidden()
120 {
121 pthread_once(&isForbiddenKeyOnce, initializeIsForbiddenKey);
122 return !!pthread_getspecific(isForbiddenKey);
123 }
124 #endif
125
fastMallocForbid()126 void fastMallocForbid()
127 {
128 pthread_once(&isForbiddenKeyOnce, initializeIsForbiddenKey);
129 pthread_setspecific(isForbiddenKey, &isForbiddenKey);
130 }
131
fastMallocAllow()132 void fastMallocAllow()
133 {
134 pthread_once(&isForbiddenKeyOnce, initializeIsForbiddenKey);
135 pthread_setspecific(isForbiddenKey, 0);
136 }
137
138 #else
139
140 static bool staticIsForbidden;
141 static bool isForbidden()
142 {
143 return staticIsForbidden;
144 }
145
146 void fastMallocForbid()
147 {
148 staticIsForbidden = true;
149 }
150
151 void fastMallocAllow()
152 {
153 staticIsForbidden = false;
154 }
155 #endif // ENABLE(JSC_MULTIPLE_THREADS)
156
157 } // namespace WTF
158 #endif // NDEBUG
159
160 #include <string.h>
161
162 namespace WTF {
163
164 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
165
166 namespace Internal {
167
fastMallocMatchFailed(void *)168 void fastMallocMatchFailed(void*)
169 {
170 CRASH();
171 }
172
173 } // namespace Internal
174
175 #endif
176
fastZeroedMalloc(size_t n)177 void* fastZeroedMalloc(size_t n)
178 {
179 void* result = fastMalloc(n);
180 memset(result, 0, n);
181 return result;
182 }
183
fastStrDup(const char * src)184 char* fastStrDup(const char* src)
185 {
186 size_t len = strlen(src) + 1;
187 char* dup = static_cast<char*>(fastMalloc(len));
188 memcpy(dup, src, len);
189 return dup;
190 }
191
tryFastZeroedMalloc(size_t n)192 TryMallocReturnValue tryFastZeroedMalloc(size_t n)
193 {
194 void* result;
195 if (!tryFastMalloc(n).getValue(result))
196 return 0;
197 memset(result, 0, n);
198 return result;
199 }
200
201 } // namespace WTF
202
203 #if FORCE_SYSTEM_MALLOC
204
205 #if PLATFORM(BREWMP)
206 #include "brew/SystemMallocBrew.h"
207 #endif
208
209 #if OS(DARWIN)
210 #include <malloc/malloc.h>
211 #elif OS(WINDOWS)
212 #include <malloc.h>
213 #endif
214
215 namespace WTF {
216
tryFastMalloc(size_t n)217 TryMallocReturnValue tryFastMalloc(size_t n)
218 {
219 ASSERT(!isForbidden());
220
221 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
222 if (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= n) // If overflow would occur...
223 return 0;
224
225 void* result = malloc(n + sizeof(AllocAlignmentInteger));
226 if (!result)
227 return 0;
228
229 *static_cast<AllocAlignmentInteger*>(result) = Internal::AllocTypeMalloc;
230 result = static_cast<AllocAlignmentInteger*>(result) + 1;
231
232 return result;
233 #else
234 return malloc(n);
235 #endif
236 }
237
fastMalloc(size_t n)238 void* fastMalloc(size_t n)
239 {
240 ASSERT(!isForbidden());
241
242 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
243 TryMallocReturnValue returnValue = tryFastMalloc(n);
244 void* result;
245 if (!returnValue.getValue(result))
246 CRASH();
247 #else
248 void* result = malloc(n);
249 #endif
250
251 if (!result) {
252 #if PLATFORM(BREWMP)
253 // The behavior of malloc(0) is implementation defined.
254 // To make sure that fastMalloc never returns 0, retry with fastMalloc(1).
255 if (!n)
256 return fastMalloc(1);
257 #endif
258 CRASH();
259 }
260
261 return result;
262 }
263
tryFastCalloc(size_t n_elements,size_t element_size)264 TryMallocReturnValue tryFastCalloc(size_t n_elements, size_t element_size)
265 {
266 ASSERT(!isForbidden());
267
268 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
269 size_t totalBytes = n_elements * element_size;
270 if (n_elements > 1 && element_size && (totalBytes / element_size) != n_elements || (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= totalBytes))
271 return 0;
272
273 totalBytes += sizeof(AllocAlignmentInteger);
274 void* result = malloc(totalBytes);
275 if (!result)
276 return 0;
277
278 memset(result, 0, totalBytes);
279 *static_cast<AllocAlignmentInteger*>(result) = Internal::AllocTypeMalloc;
280 result = static_cast<AllocAlignmentInteger*>(result) + 1;
281 return result;
282 #else
283 return calloc(n_elements, element_size);
284 #endif
285 }
286
fastCalloc(size_t n_elements,size_t element_size)287 void* fastCalloc(size_t n_elements, size_t element_size)
288 {
289 ASSERT(!isForbidden());
290
291 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
292 TryMallocReturnValue returnValue = tryFastCalloc(n_elements, element_size);
293 void* result;
294 if (!returnValue.getValue(result))
295 CRASH();
296 #else
297 void* result = calloc(n_elements, element_size);
298 #endif
299
300 if (!result) {
301 #if PLATFORM(BREWMP)
302 // If either n_elements or element_size is 0, the behavior of calloc is implementation defined.
303 // To make sure that fastCalloc never returns 0, retry with fastCalloc(1, 1).
304 if (!n_elements || !element_size)
305 return fastCalloc(1, 1);
306 #endif
307 CRASH();
308 }
309
310 return result;
311 }
312
fastFree(void * p)313 void fastFree(void* p)
314 {
315 ASSERT(!isForbidden());
316
317 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
318 if (!p)
319 return;
320
321 AllocAlignmentInteger* header = Internal::fastMallocMatchValidationValue(p);
322 if (*header != Internal::AllocTypeMalloc)
323 Internal::fastMallocMatchFailed(p);
324 free(header);
325 #else
326 free(p);
327 #endif
328 }
329
tryFastRealloc(void * p,size_t n)330 TryMallocReturnValue tryFastRealloc(void* p, size_t n)
331 {
332 ASSERT(!isForbidden());
333
334 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
335 if (p) {
336 if (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= n) // If overflow would occur...
337 return 0;
338 AllocAlignmentInteger* header = Internal::fastMallocMatchValidationValue(p);
339 if (*header != Internal::AllocTypeMalloc)
340 Internal::fastMallocMatchFailed(p);
341 void* result = realloc(header, n + sizeof(AllocAlignmentInteger));
342 if (!result)
343 return 0;
344
345 // This should not be needed because the value is already there:
346 // *static_cast<AllocAlignmentInteger*>(result) = Internal::AllocTypeMalloc;
347 result = static_cast<AllocAlignmentInteger*>(result) + 1;
348 return result;
349 } else {
350 return fastMalloc(n);
351 }
352 #else
353 return realloc(p, n);
354 #endif
355 }
356
fastRealloc(void * p,size_t n)357 void* fastRealloc(void* p, size_t n)
358 {
359 ASSERT(!isForbidden());
360
361 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
362 TryMallocReturnValue returnValue = tryFastRealloc(p, n);
363 void* result;
364 if (!returnValue.getValue(result))
365 CRASH();
366 #else
367 void* result = realloc(p, n);
368 #endif
369
370 if (!result)
371 CRASH();
372 return result;
373 }
374
releaseFastMallocFreeMemory()375 void releaseFastMallocFreeMemory() { }
376
fastMallocStatistics()377 FastMallocStatistics fastMallocStatistics()
378 {
379 FastMallocStatistics statistics = { 0, 0, 0 };
380 return statistics;
381 }
382
fastMallocSize(const void * p)383 size_t fastMallocSize(const void* p)
384 {
385 #if OS(DARWIN)
386 return malloc_size(p);
387 #elif OS(WINDOWS) && !PLATFORM(BREWMP)
388 // Brew MP uses its own memory allocator, so _msize does not work on the Brew MP simulator.
389 return _msize(const_cast<void*>(p));
390 #else
391 return 1;
392 #endif
393 }
394
395 } // namespace WTF
396
397 #if OS(DARWIN)
398 // This symbol is present in the JavaScriptCore exports file even when FastMalloc is disabled.
399 // It will never be used in this case, so it's type and value are less interesting than its presence.
400 extern "C" const int jscore_fastmalloc_introspection = 0;
401 #endif
402
403 #else // FORCE_SYSTEM_MALLOC
404
405 #if HAVE(STDINT_H)
406 #include <stdint.h>
407 #elif HAVE(INTTYPES_H)
408 #include <inttypes.h>
409 #else
410 #include <sys/types.h>
411 #endif
412
413 #include "AlwaysInline.h"
414 #include "Assertions.h"
415 #include "TCPackedCache.h"
416 #include "TCPageMap.h"
417 #include "TCSpinLock.h"
418 #include "TCSystemAlloc.h"
419 #include <algorithm>
420 #include <limits>
421 #include <pthread.h>
422 #include <stdarg.h>
423 #include <stddef.h>
424 #include <stdio.h>
425 #if HAVE(ERRNO_H)
426 #include <errno.h>
427 #endif
428 #if OS(UNIX)
429 #include <unistd.h>
430 #endif
431 #if OS(WINDOWS)
432 #ifndef WIN32_LEAN_AND_MEAN
433 #define WIN32_LEAN_AND_MEAN
434 #endif
435 #include <windows.h>
436 #endif
437
438 #ifdef WTF_CHANGES
439
440 #if OS(DARWIN)
441 #include "MallocZoneSupport.h"
442 #include <wtf/HashSet.h>
443 #include <wtf/Vector.h>
444 #endif
445
446 #if HAVE(HEADER_DETECTION_H)
447 #include "HeaderDetection.h"
448 #endif
449
450 #if HAVE(DISPATCH_H)
451 #include <dispatch/dispatch.h>
452 #endif
453
454 #if HAVE(PTHREAD_MACHDEP_H)
455 #include <System/pthread_machdep.h>
456
457 #if defined(__PTK_FRAMEWORK_JAVASCRIPTCORE_KEY0)
458 #define WTF_USE_PTHREAD_GETSPECIFIC_DIRECT 1
459 #endif
460 #endif
461
462 #ifndef PRIuS
463 #define PRIuS "zu"
464 #endif
465
466 // Calling pthread_getspecific through a global function pointer is faster than a normal
467 // call to the function on Mac OS X, and it's used in performance-critical code. So we
468 // use a function pointer. But that's not necessarily faster on other platforms, and we had
469 // problems with this technique on Windows, so we'll do this only on Mac OS X.
470 #if OS(DARWIN)
471 #if !USE(PTHREAD_GETSPECIFIC_DIRECT)
472 static void* (*pthread_getspecific_function_pointer)(pthread_key_t) = pthread_getspecific;
473 #define pthread_getspecific(key) pthread_getspecific_function_pointer(key)
474 #else
475 #define pthread_getspecific(key) _pthread_getspecific_direct(key)
476 #define pthread_setspecific(key, val) _pthread_setspecific_direct(key, (val))
477 #endif
478 #endif
479
480 #define DEFINE_VARIABLE(type, name, value, meaning) \
481 namespace FLAG__namespace_do_not_use_directly_use_DECLARE_##type##_instead { \
482 type FLAGS_##name(value); \
483 char FLAGS_no##name; \
484 } \
485 using FLAG__namespace_do_not_use_directly_use_DECLARE_##type##_instead::FLAGS_##name
486
487 #define DEFINE_int64(name, value, meaning) \
488 DEFINE_VARIABLE(int64_t, name, value, meaning)
489
490 #define DEFINE_double(name, value, meaning) \
491 DEFINE_VARIABLE(double, name, value, meaning)
492
493 namespace WTF {
494
495 #define malloc fastMalloc
496 #define calloc fastCalloc
497 #define free fastFree
498 #define realloc fastRealloc
499
500 #define MESSAGE LOG_ERROR
501 #define CHECK_CONDITION ASSERT
502
503 #if OS(DARWIN)
504 struct Span;
505 class TCMalloc_Central_FreeListPadded;
506 class TCMalloc_PageHeap;
507 class TCMalloc_ThreadCache;
508 template <typename T> class PageHeapAllocator;
509
510 class FastMallocZone {
511 public:
512 static void init();
513
514 static kern_return_t enumerate(task_t, void*, unsigned typeMmask, vm_address_t zoneAddress, memory_reader_t, vm_range_recorder_t);
goodSize(malloc_zone_t *,size_t size)515 static size_t goodSize(malloc_zone_t*, size_t size) { return size; }
check(malloc_zone_t *)516 static boolean_t check(malloc_zone_t*) { return true; }
print(malloc_zone_t *,boolean_t)517 static void print(malloc_zone_t*, boolean_t) { }
log(malloc_zone_t *,void *)518 static void log(malloc_zone_t*, void*) { }
forceLock(malloc_zone_t *)519 static void forceLock(malloc_zone_t*) { }
forceUnlock(malloc_zone_t *)520 static void forceUnlock(malloc_zone_t*) { }
statistics(malloc_zone_t *,malloc_statistics_t * stats)521 static void statistics(malloc_zone_t*, malloc_statistics_t* stats) { memset(stats, 0, sizeof(malloc_statistics_t)); }
522
523 private:
524 FastMallocZone(TCMalloc_PageHeap*, TCMalloc_ThreadCache**, TCMalloc_Central_FreeListPadded*, PageHeapAllocator<Span>*, PageHeapAllocator<TCMalloc_ThreadCache>*);
525 static size_t size(malloc_zone_t*, const void*);
526 static void* zoneMalloc(malloc_zone_t*, size_t);
527 static void* zoneCalloc(malloc_zone_t*, size_t numItems, size_t size);
528 static void zoneFree(malloc_zone_t*, void*);
529 static void* zoneRealloc(malloc_zone_t*, void*, size_t);
zoneValloc(malloc_zone_t *,size_t)530 static void* zoneValloc(malloc_zone_t*, size_t) { LOG_ERROR("valloc is not supported"); return 0; }
zoneDestroy(malloc_zone_t *)531 static void zoneDestroy(malloc_zone_t*) { }
532
533 malloc_zone_t m_zone;
534 TCMalloc_PageHeap* m_pageHeap;
535 TCMalloc_ThreadCache** m_threadHeaps;
536 TCMalloc_Central_FreeListPadded* m_centralCaches;
537 PageHeapAllocator<Span>* m_spanAllocator;
538 PageHeapAllocator<TCMalloc_ThreadCache>* m_pageHeapAllocator;
539 };
540
541 #endif
542
543 #endif
544
545 #ifndef WTF_CHANGES
546 // This #ifdef should almost never be set. Set NO_TCMALLOC_SAMPLES if
547 // you're porting to a system where you really can't get a stacktrace.
548 #ifdef NO_TCMALLOC_SAMPLES
549 // We use #define so code compiles even if you #include stacktrace.h somehow.
550 # define GetStackTrace(stack, depth, skip) (0)
551 #else
552 # include <google/stacktrace.h>
553 #endif
554 #endif
555
556 // Even if we have support for thread-local storage in the compiler
557 // and linker, the OS may not support it. We need to check that at
558 // runtime. Right now, we have to keep a manual set of "bad" OSes.
559 #if defined(HAVE_TLS)
560 static bool kernel_supports_tls = false; // be conservative
KernelSupportsTLS()561 static inline bool KernelSupportsTLS() {
562 return kernel_supports_tls;
563 }
564 # if !HAVE_DECL_UNAME // if too old for uname, probably too old for TLS
CheckIfKernelSupportsTLS()565 static void CheckIfKernelSupportsTLS() {
566 kernel_supports_tls = false;
567 }
568 # else
569 # include <sys/utsname.h> // DECL_UNAME checked for <sys/utsname.h> too
CheckIfKernelSupportsTLS()570 static void CheckIfKernelSupportsTLS() {
571 struct utsname buf;
572 if (uname(&buf) != 0) { // should be impossible
573 MESSAGE("uname failed assuming no TLS support (errno=%d)\n", errno);
574 kernel_supports_tls = false;
575 } else if (strcasecmp(buf.sysname, "linux") == 0) {
576 // The linux case: the first kernel to support TLS was 2.6.0
577 if (buf.release[0] < '2' && buf.release[1] == '.') // 0.x or 1.x
578 kernel_supports_tls = false;
579 else if (buf.release[0] == '2' && buf.release[1] == '.' &&
580 buf.release[2] >= '0' && buf.release[2] < '6' &&
581 buf.release[3] == '.') // 2.0 - 2.5
582 kernel_supports_tls = false;
583 else
584 kernel_supports_tls = true;
585 } else { // some other kernel, we'll be optimisitic
586 kernel_supports_tls = true;
587 }
588 // TODO(csilvers): VLOG(1) the tls status once we support RAW_VLOG
589 }
590 # endif // HAVE_DECL_UNAME
591 #endif // HAVE_TLS
592
593 // __THROW is defined in glibc systems. It means, counter-intuitively,
594 // "This function will never throw an exception." It's an optional
595 // optimization tool, but we may need to use it to match glibc prototypes.
596 #ifndef __THROW // I guess we're not on a glibc system
597 # define __THROW // __THROW is just an optimization, so ok to make it ""
598 #endif
599
600 //-------------------------------------------------------------------
601 // Configuration
602 //-------------------------------------------------------------------
603
604 // Not all possible combinations of the following parameters make
605 // sense. In particular, if kMaxSize increases, you may have to
606 // increase kNumClasses as well.
607 static const size_t kPageShift = 12;
608 static const size_t kPageSize = 1 << kPageShift;
609 static const size_t kMaxSize = 8u * kPageSize;
610 static const size_t kAlignShift = 3;
611 static const size_t kAlignment = 1 << kAlignShift;
612 static const size_t kNumClasses = 68;
613
614 // Allocates a big block of memory for the pagemap once we reach more than
615 // 128MB
616 static const size_t kPageMapBigAllocationThreshold = 128 << 20;
617
618 // Minimum number of pages to fetch from system at a time. Must be
619 // significantly bigger than kPageSize to amortize system-call
620 // overhead, and also to reduce external fragementation. Also, we
621 // should keep this value big because various incarnations of Linux
622 // have small limits on the number of mmap() regions per
623 // address-space.
624 static const size_t kMinSystemAlloc = 1 << (20 - kPageShift);
625
626 // Number of objects to move between a per-thread list and a central
627 // list in one shot. We want this to be not too small so we can
628 // amortize the lock overhead for accessing the central list. Making
629 // it too big may temporarily cause unnecessary memory wastage in the
630 // per-thread free list until the scavenger cleans up the list.
631 static int num_objects_to_move[kNumClasses];
632
633 // Maximum length we allow a per-thread free-list to have before we
634 // move objects from it into the corresponding central free-list. We
635 // want this big to avoid locking the central free-list too often. It
636 // should not hurt to make this list somewhat big because the
637 // scavenging code will shrink it down when its contents are not in use.
638 static const int kMaxFreeListLength = 256;
639
640 // Lower and upper bounds on the per-thread cache sizes
641 static const size_t kMinThreadCacheSize = kMaxSize * 2;
642 static const size_t kMaxThreadCacheSize = 2 << 20;
643
644 // Default bound on the total amount of thread caches
645 static const size_t kDefaultOverallThreadCacheSize = 16 << 20;
646
647 // For all span-lengths < kMaxPages we keep an exact-size list.
648 // REQUIRED: kMaxPages >= kMinSystemAlloc;
649 static const size_t kMaxPages = kMinSystemAlloc;
650
651 /* The smallest prime > 2^n */
652 static int primes_list[] = {
653 // Small values might cause high rates of sampling
654 // and hence commented out.
655 // 2, 5, 11, 17, 37, 67, 131, 257,
656 // 521, 1031, 2053, 4099, 8209, 16411,
657 32771, 65537, 131101, 262147, 524309, 1048583,
658 2097169, 4194319, 8388617, 16777259, 33554467 };
659
660 // Twice the approximate gap between sampling actions.
661 // I.e., we take one sample approximately once every
662 // tcmalloc_sample_parameter/2
663 // bytes of allocation, i.e., ~ once every 128KB.
664 // Must be a prime number.
665 #ifdef NO_TCMALLOC_SAMPLES
666 DEFINE_int64(tcmalloc_sample_parameter, 0,
667 "Unused: code is compiled with NO_TCMALLOC_SAMPLES");
668 static size_t sample_period = 0;
669 #else
670 DEFINE_int64(tcmalloc_sample_parameter, 262147,
671 "Twice the approximate gap between sampling actions."
672 " Must be a prime number. Otherwise will be rounded up to a "
673 " larger prime number");
674 static size_t sample_period = 262147;
675 #endif
676
677 // Protects sample_period above
678 static SpinLock sample_period_lock = SPINLOCK_INITIALIZER;
679
680 // Parameters for controlling how fast memory is returned to the OS.
681
682 DEFINE_double(tcmalloc_release_rate, 1,
683 "Rate at which we release unused memory to the system. "
684 "Zero means we never release memory back to the system. "
685 "Increase this flag to return memory faster; decrease it "
686 "to return memory slower. Reasonable rates are in the "
687 "range [0,10]");
688
689 //-------------------------------------------------------------------
690 // Mapping from size to size_class and vice versa
691 //-------------------------------------------------------------------
692
693 // Sizes <= 1024 have an alignment >= 8. So for such sizes we have an
694 // array indexed by ceil(size/8). Sizes > 1024 have an alignment >= 128.
695 // So for these larger sizes we have an array indexed by ceil(size/128).
696 //
697 // We flatten both logical arrays into one physical array and use
698 // arithmetic to compute an appropriate index. The constants used by
699 // ClassIndex() were selected to make the flattening work.
700 //
701 // Examples:
702 // Size Expression Index
703 // -------------------------------------------------------
704 // 0 (0 + 7) / 8 0
705 // 1 (1 + 7) / 8 1
706 // ...
707 // 1024 (1024 + 7) / 8 128
708 // 1025 (1025 + 127 + (120<<7)) / 128 129
709 // ...
710 // 32768 (32768 + 127 + (120<<7)) / 128 376
711 static const size_t kMaxSmallSize = 1024;
712 static const int shift_amount[2] = { 3, 7 }; // For divides by 8 or 128
713 static const int add_amount[2] = { 7, 127 + (120 << 7) };
714 static unsigned char class_array[377];
715
716 // Compute index of the class_array[] entry for a given size
ClassIndex(size_t s)717 static inline int ClassIndex(size_t s) {
718 const int i = (s > kMaxSmallSize);
719 return static_cast<int>((s + add_amount[i]) >> shift_amount[i]);
720 }
721
722 // Mapping from size class to max size storable in that class
723 static size_t class_to_size[kNumClasses];
724
725 // Mapping from size class to number of pages to allocate at a time
726 static size_t class_to_pages[kNumClasses];
727
728 // TransferCache is used to cache transfers of num_objects_to_move[size_class]
729 // back and forth between thread caches and the central cache for a given size
730 // class.
731 struct TCEntry {
732 void *head; // Head of chain of objects.
733 void *tail; // Tail of chain of objects.
734 };
735 // A central cache freelist can have anywhere from 0 to kNumTransferEntries
736 // slots to put link list chains into. To keep memory usage bounded the total
737 // number of TCEntries across size classes is fixed. Currently each size
738 // class is initially given one TCEntry which also means that the maximum any
739 // one class can have is kNumClasses.
740 static const int kNumTransferEntries = kNumClasses;
741
742 // Note: the following only works for "n"s that fit in 32-bits, but
743 // that is fine since we only use it for small sizes.
LgFloor(size_t n)744 static inline int LgFloor(size_t n) {
745 int log = 0;
746 for (int i = 4; i >= 0; --i) {
747 int shift = (1 << i);
748 size_t x = n >> shift;
749 if (x != 0) {
750 n = x;
751 log += shift;
752 }
753 }
754 ASSERT(n == 1);
755 return log;
756 }
757
758 // Some very basic linked list functions for dealing with using void * as
759 // storage.
760
SLL_Next(void * t)761 static inline void *SLL_Next(void *t) {
762 return *(reinterpret_cast<void**>(t));
763 }
764
SLL_SetNext(void * t,void * n)765 static inline void SLL_SetNext(void *t, void *n) {
766 *(reinterpret_cast<void**>(t)) = n;
767 }
768
SLL_Push(void ** list,void * element)769 static inline void SLL_Push(void **list, void *element) {
770 SLL_SetNext(element, *list);
771 *list = element;
772 }
773
SLL_Pop(void ** list)774 static inline void *SLL_Pop(void **list) {
775 void *result = *list;
776 *list = SLL_Next(*list);
777 return result;
778 }
779
780
781 // Remove N elements from a linked list to which head points. head will be
782 // modified to point to the new head. start and end will point to the first
783 // and last nodes of the range. Note that end will point to NULL after this
784 // function is called.
SLL_PopRange(void ** head,int N,void ** start,void ** end)785 static inline void SLL_PopRange(void **head, int N, void **start, void **end) {
786 if (N == 0) {
787 *start = NULL;
788 *end = NULL;
789 return;
790 }
791
792 void *tmp = *head;
793 for (int i = 1; i < N; ++i) {
794 tmp = SLL_Next(tmp);
795 }
796
797 *start = *head;
798 *end = tmp;
799 *head = SLL_Next(tmp);
800 // Unlink range from list.
801 SLL_SetNext(tmp, NULL);
802 }
803
SLL_PushRange(void ** head,void * start,void * end)804 static inline void SLL_PushRange(void **head, void *start, void *end) {
805 if (!start) return;
806 SLL_SetNext(end, *head);
807 *head = start;
808 }
809
SLL_Size(void * head)810 static inline size_t SLL_Size(void *head) {
811 int count = 0;
812 while (head) {
813 count++;
814 head = SLL_Next(head);
815 }
816 return count;
817 }
818
819 // Setup helper functions.
820
SizeClass(size_t size)821 static ALWAYS_INLINE size_t SizeClass(size_t size) {
822 return class_array[ClassIndex(size)];
823 }
824
825 // Get the byte-size for a specified class
ByteSizeForClass(size_t cl)826 static ALWAYS_INLINE size_t ByteSizeForClass(size_t cl) {
827 return class_to_size[cl];
828 }
NumMoveSize(size_t size)829 static int NumMoveSize(size_t size) {
830 if (size == 0) return 0;
831 // Use approx 64k transfers between thread and central caches.
832 int num = static_cast<int>(64.0 * 1024.0 / size);
833 if (num < 2) num = 2;
834 // Clamp well below kMaxFreeListLength to avoid ping pong between central
835 // and thread caches.
836 if (num > static_cast<int>(0.8 * kMaxFreeListLength))
837 num = static_cast<int>(0.8 * kMaxFreeListLength);
838
839 // Also, avoid bringing in too many objects into small object free
840 // lists. There are lots of such lists, and if we allow each one to
841 // fetch too many at a time, we end up having to scavenge too often
842 // (especially when there are lots of threads and each thread gets a
843 // small allowance for its thread cache).
844 //
845 // TODO: Make thread cache free list sizes dynamic so that we do not
846 // have to equally divide a fixed resource amongst lots of threads.
847 if (num > 32) num = 32;
848
849 return num;
850 }
851
852 // Initialize the mapping arrays
InitSizeClasses()853 static void InitSizeClasses() {
854 // Do some sanity checking on add_amount[]/shift_amount[]/class_array[]
855 if (ClassIndex(0) < 0) {
856 MESSAGE("Invalid class index %d for size 0\n", ClassIndex(0));
857 CRASH();
858 }
859 if (static_cast<size_t>(ClassIndex(kMaxSize)) >= sizeof(class_array)) {
860 MESSAGE("Invalid class index %d for kMaxSize\n", ClassIndex(kMaxSize));
861 CRASH();
862 }
863
864 // Compute the size classes we want to use
865 size_t sc = 1; // Next size class to assign
866 unsigned char alignshift = kAlignShift;
867 int last_lg = -1;
868 for (size_t size = kAlignment; size <= kMaxSize; size += (1 << alignshift)) {
869 int lg = LgFloor(size);
870 if (lg > last_lg) {
871 // Increase alignment every so often.
872 //
873 // Since we double the alignment every time size doubles and
874 // size >= 128, this means that space wasted due to alignment is
875 // at most 16/128 i.e., 12.5%. Plus we cap the alignment at 256
876 // bytes, so the space wasted as a percentage starts falling for
877 // sizes > 2K.
878 if ((lg >= 7) && (alignshift < 8)) {
879 alignshift++;
880 }
881 last_lg = lg;
882 }
883
884 // Allocate enough pages so leftover is less than 1/8 of total.
885 // This bounds wasted space to at most 12.5%.
886 size_t psize = kPageSize;
887 while ((psize % size) > (psize >> 3)) {
888 psize += kPageSize;
889 }
890 const size_t my_pages = psize >> kPageShift;
891
892 if (sc > 1 && my_pages == class_to_pages[sc-1]) {
893 // See if we can merge this into the previous class without
894 // increasing the fragmentation of the previous class.
895 const size_t my_objects = (my_pages << kPageShift) / size;
896 const size_t prev_objects = (class_to_pages[sc-1] << kPageShift)
897 / class_to_size[sc-1];
898 if (my_objects == prev_objects) {
899 // Adjust last class to include this size
900 class_to_size[sc-1] = size;
901 continue;
902 }
903 }
904
905 // Add new class
906 class_to_pages[sc] = my_pages;
907 class_to_size[sc] = size;
908 sc++;
909 }
910 if (sc != kNumClasses) {
911 MESSAGE("wrong number of size classes: found %" PRIuS " instead of %d\n",
912 sc, int(kNumClasses));
913 CRASH();
914 }
915
916 // Initialize the mapping arrays
917 int next_size = 0;
918 for (unsigned char c = 1; c < kNumClasses; c++) {
919 const size_t max_size_in_class = class_to_size[c];
920 for (size_t s = next_size; s <= max_size_in_class; s += kAlignment) {
921 class_array[ClassIndex(s)] = c;
922 }
923 next_size = static_cast<int>(max_size_in_class + kAlignment);
924 }
925
926 // Double-check sizes just to be safe
927 for (size_t size = 0; size <= kMaxSize; size++) {
928 const size_t sc = SizeClass(size);
929 if (sc == 0) {
930 MESSAGE("Bad size class %" PRIuS " for %" PRIuS "\n", sc, size);
931 CRASH();
932 }
933 if (sc > 1 && size <= class_to_size[sc-1]) {
934 MESSAGE("Allocating unnecessarily large class %" PRIuS " for %" PRIuS
935 "\n", sc, size);
936 CRASH();
937 }
938 if (sc >= kNumClasses) {
939 MESSAGE("Bad size class %" PRIuS " for %" PRIuS "\n", sc, size);
940 CRASH();
941 }
942 const size_t s = class_to_size[sc];
943 if (size > s) {
944 MESSAGE("Bad size %" PRIuS " for %" PRIuS " (sc = %" PRIuS ")\n", s, size, sc);
945 CRASH();
946 }
947 if (s == 0) {
948 MESSAGE("Bad size %" PRIuS " for %" PRIuS " (sc = %" PRIuS ")\n", s, size, sc);
949 CRASH();
950 }
951 }
952
953 // Initialize the num_objects_to_move array.
954 for (size_t cl = 1; cl < kNumClasses; ++cl) {
955 num_objects_to_move[cl] = NumMoveSize(ByteSizeForClass(cl));
956 }
957
958 #ifndef WTF_CHANGES
959 if (false) {
960 // Dump class sizes and maximum external wastage per size class
961 for (size_t cl = 1; cl < kNumClasses; ++cl) {
962 const int alloc_size = class_to_pages[cl] << kPageShift;
963 const int alloc_objs = alloc_size / class_to_size[cl];
964 const int min_used = (class_to_size[cl-1] + 1) * alloc_objs;
965 const int max_waste = alloc_size - min_used;
966 MESSAGE("SC %3d [ %8d .. %8d ] from %8d ; %2.0f%% maxwaste\n",
967 int(cl),
968 int(class_to_size[cl-1] + 1),
969 int(class_to_size[cl]),
970 int(class_to_pages[cl] << kPageShift),
971 max_waste * 100.0 / alloc_size
972 );
973 }
974 }
975 #endif
976 }
977
978 // -------------------------------------------------------------------------
979 // Simple allocator for objects of a specified type. External locking
980 // is required before accessing one of these objects.
981 // -------------------------------------------------------------------------
982
983 // Metadata allocator -- keeps stats about how many bytes allocated
984 static uint64_t metadata_system_bytes = 0;
MetaDataAlloc(size_t bytes)985 static void* MetaDataAlloc(size_t bytes) {
986 void* result = TCMalloc_SystemAlloc(bytes, 0);
987 if (result != NULL) {
988 metadata_system_bytes += bytes;
989 }
990 return result;
991 }
992
993 template <class T>
994 class PageHeapAllocator {
995 private:
996 // How much to allocate from system at a time
997 static const size_t kAllocIncrement = 32 << 10;
998
999 // Aligned size of T
1000 static const size_t kAlignedSize
1001 = (((sizeof(T) + kAlignment - 1) / kAlignment) * kAlignment);
1002
1003 // Free area from which to carve new objects
1004 char* free_area_;
1005 size_t free_avail_;
1006
1007 // Linked list of all regions allocated by this allocator
1008 void* allocated_regions_;
1009
1010 // Free list of already carved objects
1011 void* free_list_;
1012
1013 // Number of allocated but unfreed objects
1014 int inuse_;
1015
1016 public:
Init()1017 void Init() {
1018 ASSERT(kAlignedSize <= kAllocIncrement);
1019 inuse_ = 0;
1020 allocated_regions_ = 0;
1021 free_area_ = NULL;
1022 free_avail_ = 0;
1023 free_list_ = NULL;
1024 }
1025
New()1026 T* New() {
1027 // Consult free list
1028 void* result;
1029 if (free_list_ != NULL) {
1030 result = free_list_;
1031 free_list_ = *(reinterpret_cast<void**>(result));
1032 } else {
1033 if (free_avail_ < kAlignedSize) {
1034 // Need more room
1035 char* new_allocation = reinterpret_cast<char*>(MetaDataAlloc(kAllocIncrement));
1036 if (!new_allocation)
1037 CRASH();
1038
1039 *reinterpret_cast_ptr<void**>(new_allocation) = allocated_regions_;
1040 allocated_regions_ = new_allocation;
1041 free_area_ = new_allocation + kAlignedSize;
1042 free_avail_ = kAllocIncrement - kAlignedSize;
1043 }
1044 result = free_area_;
1045 free_area_ += kAlignedSize;
1046 free_avail_ -= kAlignedSize;
1047 }
1048 inuse_++;
1049 return reinterpret_cast<T*>(result);
1050 }
1051
Delete(T * p)1052 void Delete(T* p) {
1053 *(reinterpret_cast<void**>(p)) = free_list_;
1054 free_list_ = p;
1055 inuse_--;
1056 }
1057
inuse() const1058 int inuse() const { return inuse_; }
1059
1060 #if defined(WTF_CHANGES) && OS(DARWIN)
1061 template <class Recorder>
recordAdministrativeRegions(Recorder & recorder,const RemoteMemoryReader & reader)1062 void recordAdministrativeRegions(Recorder& recorder, const RemoteMemoryReader& reader)
1063 {
1064 for (void* adminAllocation = allocated_regions_; adminAllocation; adminAllocation = reader.nextEntryInLinkedList(reinterpret_cast<void**>(adminAllocation)))
1065 recorder.recordRegion(reinterpret_cast<vm_address_t>(adminAllocation), kAllocIncrement);
1066 }
1067 #endif
1068 };
1069
1070 // -------------------------------------------------------------------------
1071 // Span - a contiguous run of pages
1072 // -------------------------------------------------------------------------
1073
1074 // Type that can hold a page number
1075 typedef uintptr_t PageID;
1076
1077 // Type that can hold the length of a run of pages
1078 typedef uintptr_t Length;
1079
1080 static const Length kMaxValidPages = (~static_cast<Length>(0)) >> kPageShift;
1081
1082 // Convert byte size into pages. This won't overflow, but may return
1083 // an unreasonably large value if bytes is huge enough.
pages(size_t bytes)1084 static inline Length pages(size_t bytes) {
1085 return (bytes >> kPageShift) +
1086 ((bytes & (kPageSize - 1)) > 0 ? 1 : 0);
1087 }
1088
1089 // Convert a user size into the number of bytes that will actually be
1090 // allocated
AllocationSize(size_t bytes)1091 static size_t AllocationSize(size_t bytes) {
1092 if (bytes > kMaxSize) {
1093 // Large object: we allocate an integral number of pages
1094 ASSERT(bytes <= (kMaxValidPages << kPageShift));
1095 return pages(bytes) << kPageShift;
1096 } else {
1097 // Small object: find the size class to which it belongs
1098 return ByteSizeForClass(SizeClass(bytes));
1099 }
1100 }
1101
1102 // Information kept for a span (a contiguous run of pages).
1103 struct Span {
1104 PageID start; // Starting page number
1105 Length length; // Number of pages in span
1106 Span* next; // Used when in link list
1107 Span* prev; // Used when in link list
1108 void* objects; // Linked list of free objects
1109 unsigned int free : 1; // Is the span free
1110 #ifndef NO_TCMALLOC_SAMPLES
1111 unsigned int sample : 1; // Sampled object?
1112 #endif
1113 unsigned int sizeclass : 8; // Size-class for small objects (or 0)
1114 unsigned int refcount : 11; // Number of non-free objects
1115 bool decommitted : 1;
1116
1117 #undef SPAN_HISTORY
1118 #ifdef SPAN_HISTORY
1119 // For debugging, we can keep a log events per span
1120 int nexthistory;
1121 char history[64];
1122 int value[64];
1123 #endif
1124 };
1125
1126 #define ASSERT_SPAN_COMMITTED(span) ASSERT(!span->decommitted)
1127
1128 #ifdef SPAN_HISTORY
Event(Span * span,char op,int v=0)1129 void Event(Span* span, char op, int v = 0) {
1130 span->history[span->nexthistory] = op;
1131 span->value[span->nexthistory] = v;
1132 span->nexthistory++;
1133 if (span->nexthistory == sizeof(span->history)) span->nexthistory = 0;
1134 }
1135 #else
1136 #define Event(s,o,v) ((void) 0)
1137 #endif
1138
1139 // Allocator/deallocator for spans
1140 static PageHeapAllocator<Span> span_allocator;
NewSpan(PageID p,Length len)1141 static Span* NewSpan(PageID p, Length len) {
1142 Span* result = span_allocator.New();
1143 memset(result, 0, sizeof(*result));
1144 result->start = p;
1145 result->length = len;
1146 #ifdef SPAN_HISTORY
1147 result->nexthistory = 0;
1148 #endif
1149 return result;
1150 }
1151
DeleteSpan(Span * span)1152 static inline void DeleteSpan(Span* span) {
1153 #ifndef NDEBUG
1154 // In debug mode, trash the contents of deleted Spans
1155 memset(span, 0x3f, sizeof(*span));
1156 #endif
1157 span_allocator.Delete(span);
1158 }
1159
1160 // -------------------------------------------------------------------------
1161 // Doubly linked list of spans.
1162 // -------------------------------------------------------------------------
1163
DLL_Init(Span * list)1164 static inline void DLL_Init(Span* list) {
1165 list->next = list;
1166 list->prev = list;
1167 }
1168
DLL_Remove(Span * span)1169 static inline void DLL_Remove(Span* span) {
1170 span->prev->next = span->next;
1171 span->next->prev = span->prev;
1172 span->prev = NULL;
1173 span->next = NULL;
1174 }
1175
DLL_IsEmpty(const Span * list)1176 static ALWAYS_INLINE bool DLL_IsEmpty(const Span* list) {
1177 return list->next == list;
1178 }
1179
DLL_Length(const Span * list)1180 static int DLL_Length(const Span* list) {
1181 int result = 0;
1182 for (Span* s = list->next; s != list; s = s->next) {
1183 result++;
1184 }
1185 return result;
1186 }
1187
1188 #if 0 /* Not needed at the moment -- causes compiler warnings if not used */
1189 static void DLL_Print(const char* label, const Span* list) {
1190 MESSAGE("%-10s %p:", label, list);
1191 for (const Span* s = list->next; s != list; s = s->next) {
1192 MESSAGE(" <%p,%u,%u>", s, s->start, s->length);
1193 }
1194 MESSAGE("\n");
1195 }
1196 #endif
1197
DLL_Prepend(Span * list,Span * span)1198 static inline void DLL_Prepend(Span* list, Span* span) {
1199 ASSERT(span->next == NULL);
1200 ASSERT(span->prev == NULL);
1201 span->next = list->next;
1202 span->prev = list;
1203 list->next->prev = span;
1204 list->next = span;
1205 }
1206
1207 // -------------------------------------------------------------------------
1208 // Stack traces kept for sampled allocations
1209 // The following state is protected by pageheap_lock_.
1210 // -------------------------------------------------------------------------
1211
1212 // size/depth are made the same size as a pointer so that some generic
1213 // code below can conveniently cast them back and forth to void*.
1214 static const int kMaxStackDepth = 31;
1215 struct StackTrace {
1216 uintptr_t size; // Size of object
1217 uintptr_t depth; // Number of PC values stored in array below
1218 void* stack[kMaxStackDepth];
1219 };
1220 static PageHeapAllocator<StackTrace> stacktrace_allocator;
1221 static Span sampled_objects;
1222
1223 // -------------------------------------------------------------------------
1224 // Map from page-id to per-page data
1225 // -------------------------------------------------------------------------
1226
1227 // We use PageMap2<> for 32-bit and PageMap3<> for 64-bit machines.
1228 // We also use a simple one-level cache for hot PageID-to-sizeclass mappings,
1229 // because sometimes the sizeclass is all the information we need.
1230
1231 // Selector class -- general selector uses 3-level map
1232 template <int BITS> class MapSelector {
1233 public:
1234 typedef TCMalloc_PageMap3<BITS-kPageShift> Type;
1235 typedef PackedCache<BITS, uint64_t> CacheType;
1236 };
1237
1238 #if defined(WTF_CHANGES)
1239 #if CPU(X86_64)
1240 // On all known X86-64 platforms, the upper 16 bits are always unused and therefore
1241 // can be excluded from the PageMap key.
1242 // See http://en.wikipedia.org/wiki/X86-64#Virtual_address_space_details
1243
1244 static const size_t kBitsUnusedOn64Bit = 16;
1245 #else
1246 static const size_t kBitsUnusedOn64Bit = 0;
1247 #endif
1248
1249 // A three-level map for 64-bit machines
1250 template <> class MapSelector<64> {
1251 public:
1252 typedef TCMalloc_PageMap3<64 - kPageShift - kBitsUnusedOn64Bit> Type;
1253 typedef PackedCache<64, uint64_t> CacheType;
1254 };
1255 #endif
1256
1257 // A two-level map for 32-bit machines
1258 template <> class MapSelector<32> {
1259 public:
1260 typedef TCMalloc_PageMap2<32 - kPageShift> Type;
1261 typedef PackedCache<32 - kPageShift, uint16_t> CacheType;
1262 };
1263
1264 // -------------------------------------------------------------------------
1265 // Page-level allocator
1266 // * Eager coalescing
1267 //
1268 // Heap for page-level allocation. We allow allocating and freeing a
1269 // contiguous runs of pages (called a "span").
1270 // -------------------------------------------------------------------------
1271
1272 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1273 // The page heap maintains a free list for spans that are no longer in use by
1274 // the central cache or any thread caches. We use a background thread to
1275 // periodically scan the free list and release a percentage of it back to the OS.
1276
1277 // If free_committed_pages_ exceeds kMinimumFreeCommittedPageCount, the
1278 // background thread:
1279 // - wakes up
1280 // - pauses for kScavengeDelayInSeconds
1281 // - returns to the OS a percentage of the memory that remained unused during
1282 // that pause (kScavengePercentage * min_free_committed_pages_since_last_scavenge_)
1283 // The goal of this strategy is to reduce memory pressure in a timely fashion
1284 // while avoiding thrashing the OS allocator.
1285
1286 // Time delay before the page heap scavenger will consider returning pages to
1287 // the OS.
1288 static const int kScavengeDelayInSeconds = 2;
1289
1290 // Approximate percentage of free committed pages to return to the OS in one
1291 // scavenge.
1292 static const float kScavengePercentage = .5f;
1293
1294 // number of span lists to keep spans in when memory is returned.
1295 static const int kMinSpanListsWithSpans = 32;
1296
1297 // Number of free committed pages that we want to keep around. The minimum number of pages used when there
1298 // is 1 span in each of the first kMinSpanListsWithSpans spanlists. Currently 528 pages.
1299 static const size_t kMinimumFreeCommittedPageCount = kMinSpanListsWithSpans * ((1.0f+kMinSpanListsWithSpans) / 2.0f);
1300
1301 #endif
1302
1303 class TCMalloc_PageHeap {
1304 public:
1305 void init();
1306
1307 // Allocate a run of "n" pages. Returns zero if out of memory.
1308 Span* New(Length n);
1309
1310 // Delete the span "[p, p+n-1]".
1311 // REQUIRES: span was returned by earlier call to New() and
1312 // has not yet been deleted.
1313 void Delete(Span* span);
1314
1315 // Mark an allocated span as being used for small objects of the
1316 // specified size-class.
1317 // REQUIRES: span was returned by an earlier call to New()
1318 // and has not yet been deleted.
1319 void RegisterSizeClass(Span* span, size_t sc);
1320
1321 // Split an allocated span into two spans: one of length "n" pages
1322 // followed by another span of length "span->length - n" pages.
1323 // Modifies "*span" to point to the first span of length "n" pages.
1324 // Returns a pointer to the second span.
1325 //
1326 // REQUIRES: "0 < n < span->length"
1327 // REQUIRES: !span->free
1328 // REQUIRES: span->sizeclass == 0
1329 Span* Split(Span* span, Length n);
1330
1331 // Return the descriptor for the specified page.
GetDescriptor(PageID p) const1332 inline Span* GetDescriptor(PageID p) const {
1333 return reinterpret_cast<Span*>(pagemap_.get(p));
1334 }
1335
1336 #ifdef WTF_CHANGES
GetDescriptorEnsureSafe(PageID p)1337 inline Span* GetDescriptorEnsureSafe(PageID p)
1338 {
1339 pagemap_.Ensure(p, 1);
1340 return GetDescriptor(p);
1341 }
1342
1343 size_t ReturnedBytes() const;
1344 #endif
1345
1346 // Dump state to stderr
1347 #ifndef WTF_CHANGES
1348 void Dump(TCMalloc_Printer* out);
1349 #endif
1350
1351 // Return number of bytes allocated from system
SystemBytes() const1352 inline uint64_t SystemBytes() const { return system_bytes_; }
1353
1354 // Return number of free bytes in heap
FreeBytes() const1355 uint64_t FreeBytes() const {
1356 return (static_cast<uint64_t>(free_pages_) << kPageShift);
1357 }
1358
1359 bool Check();
1360 bool CheckList(Span* list, Length min_pages, Length max_pages);
1361
1362 // Release all pages on the free list for reuse by the OS:
1363 void ReleaseFreePages();
1364
1365 // Return 0 if we have no information, or else the correct sizeclass for p.
1366 // Reads and writes to pagemap_cache_ do not require locking.
1367 // The entries are 64 bits on 64-bit hardware and 16 bits on
1368 // 32-bit hardware, and we don't mind raciness as long as each read of
1369 // an entry yields a valid entry, not a partially updated entry.
GetSizeClassIfCached(PageID p) const1370 size_t GetSizeClassIfCached(PageID p) const {
1371 return pagemap_cache_.GetOrDefault(p, 0);
1372 }
CacheSizeClass(PageID p,size_t cl) const1373 void CacheSizeClass(PageID p, size_t cl) const { pagemap_cache_.Put(p, cl); }
1374
1375 private:
1376 // Pick the appropriate map and cache types based on pointer size
1377 typedef MapSelector<8*sizeof(uintptr_t)>::Type PageMap;
1378 typedef MapSelector<8*sizeof(uintptr_t)>::CacheType PageMapCache;
1379 PageMap pagemap_;
1380 mutable PageMapCache pagemap_cache_;
1381
1382 // We segregate spans of a given size into two circular linked
1383 // lists: one for normal spans, and one for spans whose memory
1384 // has been returned to the system.
1385 struct SpanList {
1386 Span normal;
1387 Span returned;
1388 };
1389
1390 // List of free spans of length >= kMaxPages
1391 SpanList large_;
1392
1393 // Array mapping from span length to a doubly linked list of free spans
1394 SpanList free_[kMaxPages];
1395
1396 // Number of pages kept in free lists
1397 uintptr_t free_pages_;
1398
1399 // Bytes allocated from system
1400 uint64_t system_bytes_;
1401
1402 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1403 // Number of pages kept in free lists that are still committed.
1404 Length free_committed_pages_;
1405
1406 // Minimum number of free committed pages since last scavenge. (Can be 0 if
1407 // we've committed new pages since the last scavenge.)
1408 Length min_free_committed_pages_since_last_scavenge_;
1409 #endif
1410
1411 bool GrowHeap(Length n);
1412
1413 // REQUIRES span->length >= n
1414 // Remove span from its free list, and move any leftover part of
1415 // span into appropriate free lists. Also update "span" to have
1416 // length exactly "n" and mark it as non-free so it can be returned
1417 // to the client.
1418 //
1419 // "released" is true iff "span" was found on a "returned" list.
1420 void Carve(Span* span, Length n, bool released);
1421
RecordSpan(Span * span)1422 void RecordSpan(Span* span) {
1423 pagemap_.set(span->start, span);
1424 if (span->length > 1) {
1425 pagemap_.set(span->start + span->length - 1, span);
1426 }
1427 }
1428
1429 // Allocate a large span of length == n. If successful, returns a
1430 // span of exactly the specified length. Else, returns NULL.
1431 Span* AllocLarge(Length n);
1432
1433 #if !USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1434 // Incrementally release some memory to the system.
1435 // IncrementalScavenge(n) is called whenever n pages are freed.
1436 void IncrementalScavenge(Length n);
1437 #endif
1438
1439 // Number of pages to deallocate before doing more scavenging
1440 int64_t scavenge_counter_;
1441
1442 // Index of last free list we scavenged
1443 size_t scavenge_index_;
1444
1445 #if defined(WTF_CHANGES) && OS(DARWIN)
1446 friend class FastMallocZone;
1447 #endif
1448
1449 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1450 void initializeScavenger();
1451 ALWAYS_INLINE void signalScavenger();
1452 void scavenge();
1453 ALWAYS_INLINE bool shouldScavenge() const;
1454
1455 #if HAVE(DISPATCH_H) || OS(WINDOWS)
1456 void periodicScavenge();
1457 ALWAYS_INLINE bool isScavengerSuspended();
1458 ALWAYS_INLINE void scheduleScavenger();
1459 ALWAYS_INLINE void rescheduleScavenger();
1460 ALWAYS_INLINE void suspendScavenger();
1461 #endif
1462
1463 #if HAVE(DISPATCH_H)
1464 dispatch_queue_t m_scavengeQueue;
1465 dispatch_source_t m_scavengeTimer;
1466 bool m_scavengingSuspended;
1467 #elif OS(WINDOWS)
1468 static void CALLBACK scavengerTimerFired(void*, BOOLEAN);
1469 HANDLE m_scavengeQueueTimer;
1470 #else
1471 static NO_RETURN_WITH_VALUE void* runScavengerThread(void*);
1472 NO_RETURN void scavengerThread();
1473
1474 // Keeps track of whether the background thread is actively scavenging memory every kScavengeDelayInSeconds, or
1475 // it's blocked waiting for more pages to be deleted.
1476 bool m_scavengeThreadActive;
1477
1478 pthread_mutex_t m_scavengeMutex;
1479 pthread_cond_t m_scavengeCondition;
1480 #endif
1481
1482 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1483 };
1484
init()1485 void TCMalloc_PageHeap::init()
1486 {
1487 pagemap_.init(MetaDataAlloc);
1488 pagemap_cache_ = PageMapCache(0);
1489 free_pages_ = 0;
1490 system_bytes_ = 0;
1491
1492 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1493 free_committed_pages_ = 0;
1494 min_free_committed_pages_since_last_scavenge_ = 0;
1495 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1496
1497 scavenge_counter_ = 0;
1498 // Start scavenging at kMaxPages list
1499 scavenge_index_ = kMaxPages-1;
1500 COMPILE_ASSERT(kNumClasses <= (1 << PageMapCache::kValuebits), valuebits);
1501 DLL_Init(&large_.normal);
1502 DLL_Init(&large_.returned);
1503 for (size_t i = 0; i < kMaxPages; i++) {
1504 DLL_Init(&free_[i].normal);
1505 DLL_Init(&free_[i].returned);
1506 }
1507
1508 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1509 initializeScavenger();
1510 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1511 }
1512
1513 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1514
1515 #if HAVE(DISPATCH_H)
1516
initializeScavenger()1517 void TCMalloc_PageHeap::initializeScavenger()
1518 {
1519 m_scavengeQueue = dispatch_queue_create("com.apple.JavaScriptCore.FastMallocSavenger", NULL);
1520 m_scavengeTimer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, m_scavengeQueue);
1521 dispatch_time_t startTime = dispatch_time(DISPATCH_TIME_NOW, kScavengeDelayInSeconds * NSEC_PER_SEC);
1522 dispatch_source_set_timer(m_scavengeTimer, startTime, kScavengeDelayInSeconds * NSEC_PER_SEC, 1000 * NSEC_PER_USEC);
1523 dispatch_source_set_event_handler(m_scavengeTimer, ^{ periodicScavenge(); });
1524 m_scavengingSuspended = true;
1525 }
1526
isScavengerSuspended()1527 ALWAYS_INLINE bool TCMalloc_PageHeap::isScavengerSuspended()
1528 {
1529 ASSERT(IsHeld(pageheap_lock));
1530 return m_scavengingSuspended;
1531 }
1532
scheduleScavenger()1533 ALWAYS_INLINE void TCMalloc_PageHeap::scheduleScavenger()
1534 {
1535 ASSERT(IsHeld(pageheap_lock));
1536 m_scavengingSuspended = false;
1537 dispatch_resume(m_scavengeTimer);
1538 }
1539
rescheduleScavenger()1540 ALWAYS_INLINE void TCMalloc_PageHeap::rescheduleScavenger()
1541 {
1542 // Nothing to do here for libdispatch.
1543 }
1544
suspendScavenger()1545 ALWAYS_INLINE void TCMalloc_PageHeap::suspendScavenger()
1546 {
1547 ASSERT(IsHeld(pageheap_lock));
1548 m_scavengingSuspended = true;
1549 dispatch_suspend(m_scavengeTimer);
1550 }
1551
1552 #elif OS(WINDOWS)
1553
scavengerTimerFired(void * context,BOOLEAN)1554 void TCMalloc_PageHeap::scavengerTimerFired(void* context, BOOLEAN)
1555 {
1556 static_cast<TCMalloc_PageHeap*>(context)->periodicScavenge();
1557 }
1558
initializeScavenger()1559 void TCMalloc_PageHeap::initializeScavenger()
1560 {
1561 m_scavengeQueueTimer = 0;
1562 }
1563
isScavengerSuspended()1564 ALWAYS_INLINE bool TCMalloc_PageHeap::isScavengerSuspended()
1565 {
1566 ASSERT(IsHeld(pageheap_lock));
1567 return !m_scavengeQueueTimer;
1568 }
1569
scheduleScavenger()1570 ALWAYS_INLINE void TCMalloc_PageHeap::scheduleScavenger()
1571 {
1572 // We need to use WT_EXECUTEONLYONCE here and reschedule the timer, because
1573 // Windows will fire the timer event even when the function is already running.
1574 ASSERT(IsHeld(pageheap_lock));
1575 CreateTimerQueueTimer(&m_scavengeQueueTimer, 0, scavengerTimerFired, this, kScavengeDelayInSeconds * 1000, 0, WT_EXECUTEONLYONCE);
1576 }
1577
rescheduleScavenger()1578 ALWAYS_INLINE void TCMalloc_PageHeap::rescheduleScavenger()
1579 {
1580 // We must delete the timer and create it again, because it is not possible to retrigger a timer on Windows.
1581 suspendScavenger();
1582 scheduleScavenger();
1583 }
1584
suspendScavenger()1585 ALWAYS_INLINE void TCMalloc_PageHeap::suspendScavenger()
1586 {
1587 ASSERT(IsHeld(pageheap_lock));
1588 HANDLE scavengeQueueTimer = m_scavengeQueueTimer;
1589 m_scavengeQueueTimer = 0;
1590 DeleteTimerQueueTimer(0, scavengeQueueTimer, 0);
1591 }
1592
1593 #else
1594
initializeScavenger()1595 void TCMalloc_PageHeap::initializeScavenger()
1596 {
1597 // Create a non-recursive mutex.
1598 #if !defined(PTHREAD_MUTEX_NORMAL) || PTHREAD_MUTEX_NORMAL == PTHREAD_MUTEX_DEFAULT
1599 pthread_mutex_init(&m_scavengeMutex, 0);
1600 #else
1601 pthread_mutexattr_t attr;
1602 pthread_mutexattr_init(&attr);
1603 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_NORMAL);
1604
1605 pthread_mutex_init(&m_scavengeMutex, &attr);
1606
1607 pthread_mutexattr_destroy(&attr);
1608 #endif
1609
1610 pthread_cond_init(&m_scavengeCondition, 0);
1611 m_scavengeThreadActive = true;
1612 pthread_t thread;
1613 pthread_create(&thread, 0, runScavengerThread, this);
1614 }
1615
runScavengerThread(void * context)1616 void* TCMalloc_PageHeap::runScavengerThread(void* context)
1617 {
1618 static_cast<TCMalloc_PageHeap*>(context)->scavengerThread();
1619 #if (COMPILER(MSVC) || COMPILER(SUNCC))
1620 // Without this, Visual Studio and Sun Studio will complain that this method does not return a value.
1621 return 0;
1622 #endif
1623 }
1624
signalScavenger()1625 ALWAYS_INLINE void TCMalloc_PageHeap::signalScavenger()
1626 {
1627 // m_scavengeMutex should be held before accessing m_scavengeThreadActive.
1628 ASSERT(pthread_mutex_trylock(m_scavengeMutex));
1629 if (!m_scavengeThreadActive && shouldScavenge())
1630 pthread_cond_signal(&m_scavengeCondition);
1631 }
1632
1633 #endif
1634
scavenge()1635 void TCMalloc_PageHeap::scavenge()
1636 {
1637 size_t pagesToRelease = min_free_committed_pages_since_last_scavenge_ * kScavengePercentage;
1638 size_t targetPageCount = std::max<size_t>(kMinimumFreeCommittedPageCount, free_committed_pages_ - pagesToRelease);
1639
1640 while (free_committed_pages_ > targetPageCount) {
1641 for (int i = kMaxPages; i > 0 && free_committed_pages_ >= targetPageCount; i--) {
1642 SpanList* slist = (static_cast<size_t>(i) == kMaxPages) ? &large_ : &free_[i];
1643 // If the span size is bigger than kMinSpanListsWithSpans pages return all the spans in the list, else return all but 1 span.
1644 // Return only 50% of a spanlist at a time so spans of size 1 are not the only ones left.
1645 size_t length = DLL_Length(&slist->normal);
1646 size_t numSpansToReturn = (i > kMinSpanListsWithSpans) ? length : length / 2;
1647 for (int j = 0; static_cast<size_t>(j) < numSpansToReturn && !DLL_IsEmpty(&slist->normal) && free_committed_pages_ > targetPageCount; j++) {
1648 Span* s = slist->normal.prev;
1649 DLL_Remove(s);
1650 ASSERT(!s->decommitted);
1651 if (!s->decommitted) {
1652 TCMalloc_SystemRelease(reinterpret_cast<void*>(s->start << kPageShift),
1653 static_cast<size_t>(s->length << kPageShift));
1654 ASSERT(free_committed_pages_ >= s->length);
1655 free_committed_pages_ -= s->length;
1656 s->decommitted = true;
1657 }
1658 DLL_Prepend(&slist->returned, s);
1659 }
1660 }
1661 }
1662
1663 min_free_committed_pages_since_last_scavenge_ = free_committed_pages_;
1664 }
1665
shouldScavenge() const1666 ALWAYS_INLINE bool TCMalloc_PageHeap::shouldScavenge() const
1667 {
1668 return free_committed_pages_ > kMinimumFreeCommittedPageCount;
1669 }
1670
1671 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1672
New(Length n)1673 inline Span* TCMalloc_PageHeap::New(Length n) {
1674 ASSERT(Check());
1675 ASSERT(n > 0);
1676
1677 // Find first size >= n that has a non-empty list
1678 for (Length s = n; s < kMaxPages; s++) {
1679 Span* ll = NULL;
1680 bool released = false;
1681 if (!DLL_IsEmpty(&free_[s].normal)) {
1682 // Found normal span
1683 ll = &free_[s].normal;
1684 } else if (!DLL_IsEmpty(&free_[s].returned)) {
1685 // Found returned span; reallocate it
1686 ll = &free_[s].returned;
1687 released = true;
1688 } else {
1689 // Keep looking in larger classes
1690 continue;
1691 }
1692
1693 Span* result = ll->next;
1694 Carve(result, n, released);
1695 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1696 // The newly allocated memory is from a span that's in the normal span list (already committed). Update the
1697 // free committed pages count.
1698 ASSERT(free_committed_pages_ >= n);
1699 free_committed_pages_ -= n;
1700 if (free_committed_pages_ < min_free_committed_pages_since_last_scavenge_)
1701 min_free_committed_pages_since_last_scavenge_ = free_committed_pages_;
1702 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1703 ASSERT(Check());
1704 free_pages_ -= n;
1705 return result;
1706 }
1707
1708 Span* result = AllocLarge(n);
1709 if (result != NULL) {
1710 ASSERT_SPAN_COMMITTED(result);
1711 return result;
1712 }
1713
1714 // Grow the heap and try again
1715 if (!GrowHeap(n)) {
1716 ASSERT(Check());
1717 return NULL;
1718 }
1719
1720 return AllocLarge(n);
1721 }
1722
AllocLarge(Length n)1723 Span* TCMalloc_PageHeap::AllocLarge(Length n) {
1724 // find the best span (closest to n in size).
1725 // The following loops implements address-ordered best-fit.
1726 bool from_released = false;
1727 Span *best = NULL;
1728
1729 // Search through normal list
1730 for (Span* span = large_.normal.next;
1731 span != &large_.normal;
1732 span = span->next) {
1733 if (span->length >= n) {
1734 if ((best == NULL)
1735 || (span->length < best->length)
1736 || ((span->length == best->length) && (span->start < best->start))) {
1737 best = span;
1738 from_released = false;
1739 }
1740 }
1741 }
1742
1743 // Search through released list in case it has a better fit
1744 for (Span* span = large_.returned.next;
1745 span != &large_.returned;
1746 span = span->next) {
1747 if (span->length >= n) {
1748 if ((best == NULL)
1749 || (span->length < best->length)
1750 || ((span->length == best->length) && (span->start < best->start))) {
1751 best = span;
1752 from_released = true;
1753 }
1754 }
1755 }
1756
1757 if (best != NULL) {
1758 Carve(best, n, from_released);
1759 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1760 // The newly allocated memory is from a span that's in the normal span list (already committed). Update the
1761 // free committed pages count.
1762 ASSERT(free_committed_pages_ >= n);
1763 free_committed_pages_ -= n;
1764 if (free_committed_pages_ < min_free_committed_pages_since_last_scavenge_)
1765 min_free_committed_pages_since_last_scavenge_ = free_committed_pages_;
1766 #endif // USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1767 ASSERT(Check());
1768 free_pages_ -= n;
1769 return best;
1770 }
1771 return NULL;
1772 }
1773
Split(Span * span,Length n)1774 Span* TCMalloc_PageHeap::Split(Span* span, Length n) {
1775 ASSERT(0 < n);
1776 ASSERT(n < span->length);
1777 ASSERT(!span->free);
1778 ASSERT(span->sizeclass == 0);
1779 Event(span, 'T', n);
1780
1781 const Length extra = span->length - n;
1782 Span* leftover = NewSpan(span->start + n, extra);
1783 Event(leftover, 'U', extra);
1784 RecordSpan(leftover);
1785 pagemap_.set(span->start + n - 1, span); // Update map from pageid to span
1786 span->length = n;
1787
1788 return leftover;
1789 }
1790
Carve(Span * span,Length n,bool released)1791 inline void TCMalloc_PageHeap::Carve(Span* span, Length n, bool released) {
1792 ASSERT(n > 0);
1793 DLL_Remove(span);
1794 span->free = 0;
1795 Event(span, 'A', n);
1796
1797 if (released) {
1798 // If the span chosen to carve from is decommited, commit the entire span at once to avoid committing spans 1 page at a time.
1799 ASSERT(span->decommitted);
1800 TCMalloc_SystemCommit(reinterpret_cast<void*>(span->start << kPageShift), static_cast<size_t>(span->length << kPageShift));
1801 span->decommitted = false;
1802 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1803 free_committed_pages_ += span->length;
1804 #endif
1805 }
1806
1807 const int extra = static_cast<int>(span->length - n);
1808 ASSERT(extra >= 0);
1809 if (extra > 0) {
1810 Span* leftover = NewSpan(span->start + n, extra);
1811 leftover->free = 1;
1812 leftover->decommitted = false;
1813 Event(leftover, 'S', extra);
1814 RecordSpan(leftover);
1815
1816 // Place leftover span on appropriate free list
1817 SpanList* listpair = (static_cast<size_t>(extra) < kMaxPages) ? &free_[extra] : &large_;
1818 Span* dst = &listpair->normal;
1819 DLL_Prepend(dst, leftover);
1820
1821 span->length = n;
1822 pagemap_.set(span->start + n - 1, span);
1823 }
1824 }
1825
mergeDecommittedStates(Span * destination,Span * other)1826 static ALWAYS_INLINE void mergeDecommittedStates(Span* destination, Span* other)
1827 {
1828 if (destination->decommitted && !other->decommitted) {
1829 TCMalloc_SystemRelease(reinterpret_cast<void*>(other->start << kPageShift),
1830 static_cast<size_t>(other->length << kPageShift));
1831 } else if (other->decommitted && !destination->decommitted) {
1832 TCMalloc_SystemRelease(reinterpret_cast<void*>(destination->start << kPageShift),
1833 static_cast<size_t>(destination->length << kPageShift));
1834 destination->decommitted = true;
1835 }
1836 }
1837
Delete(Span * span)1838 inline void TCMalloc_PageHeap::Delete(Span* span) {
1839 ASSERT(Check());
1840 ASSERT(!span->free);
1841 ASSERT(span->length > 0);
1842 ASSERT(GetDescriptor(span->start) == span);
1843 ASSERT(GetDescriptor(span->start + span->length - 1) == span);
1844 span->sizeclass = 0;
1845 #ifndef NO_TCMALLOC_SAMPLES
1846 span->sample = 0;
1847 #endif
1848
1849 // Coalesce -- we guarantee that "p" != 0, so no bounds checking
1850 // necessary. We do not bother resetting the stale pagemap
1851 // entries for the pieces we are merging together because we only
1852 // care about the pagemap entries for the boundaries.
1853 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1854 // Track the total size of the neighboring free spans that are committed.
1855 Length neighboringCommittedSpansLength = 0;
1856 #endif
1857 const PageID p = span->start;
1858 const Length n = span->length;
1859 Span* prev = GetDescriptor(p-1);
1860 if (prev != NULL && prev->free) {
1861 // Merge preceding span into this span
1862 ASSERT(prev->start + prev->length == p);
1863 const Length len = prev->length;
1864 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1865 if (!prev->decommitted)
1866 neighboringCommittedSpansLength += len;
1867 #endif
1868 mergeDecommittedStates(span, prev);
1869 DLL_Remove(prev);
1870 DeleteSpan(prev);
1871 span->start -= len;
1872 span->length += len;
1873 pagemap_.set(span->start, span);
1874 Event(span, 'L', len);
1875 }
1876 Span* next = GetDescriptor(p+n);
1877 if (next != NULL && next->free) {
1878 // Merge next span into this span
1879 ASSERT(next->start == p+n);
1880 const Length len = next->length;
1881 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1882 if (!next->decommitted)
1883 neighboringCommittedSpansLength += len;
1884 #endif
1885 mergeDecommittedStates(span, next);
1886 DLL_Remove(next);
1887 DeleteSpan(next);
1888 span->length += len;
1889 pagemap_.set(span->start + span->length - 1, span);
1890 Event(span, 'R', len);
1891 }
1892
1893 Event(span, 'D', span->length);
1894 span->free = 1;
1895 if (span->decommitted) {
1896 if (span->length < kMaxPages)
1897 DLL_Prepend(&free_[span->length].returned, span);
1898 else
1899 DLL_Prepend(&large_.returned, span);
1900 } else {
1901 if (span->length < kMaxPages)
1902 DLL_Prepend(&free_[span->length].normal, span);
1903 else
1904 DLL_Prepend(&large_.normal, span);
1905 }
1906 free_pages_ += n;
1907
1908 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
1909 if (span->decommitted) {
1910 // If the merged span is decommitted, that means we decommitted any neighboring spans that were
1911 // committed. Update the free committed pages count.
1912 free_committed_pages_ -= neighboringCommittedSpansLength;
1913 if (free_committed_pages_ < min_free_committed_pages_since_last_scavenge_)
1914 min_free_committed_pages_since_last_scavenge_ = free_committed_pages_;
1915 } else {
1916 // If the merged span remains committed, add the deleted span's size to the free committed pages count.
1917 free_committed_pages_ += n;
1918 }
1919
1920 // Make sure the scavenge thread becomes active if we have enough freed pages to release some back to the system.
1921 signalScavenger();
1922 #else
1923 IncrementalScavenge(n);
1924 #endif
1925
1926 ASSERT(Check());
1927 }
1928
1929 #if !USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
IncrementalScavenge(Length n)1930 void TCMalloc_PageHeap::IncrementalScavenge(Length n) {
1931 // Fast path; not yet time to release memory
1932 scavenge_counter_ -= n;
1933 if (scavenge_counter_ >= 0) return; // Not yet time to scavenge
1934
1935 // If there is nothing to release, wait for so many pages before
1936 // scavenging again. With 4K pages, this comes to 16MB of memory.
1937 static const size_t kDefaultReleaseDelay = 1 << 8;
1938
1939 // Find index of free list to scavenge
1940 size_t index = scavenge_index_ + 1;
1941 for (size_t i = 0; i < kMaxPages+1; i++) {
1942 if (index > kMaxPages) index = 0;
1943 SpanList* slist = (index == kMaxPages) ? &large_ : &free_[index];
1944 if (!DLL_IsEmpty(&slist->normal)) {
1945 // Release the last span on the normal portion of this list
1946 Span* s = slist->normal.prev;
1947 DLL_Remove(s);
1948 TCMalloc_SystemRelease(reinterpret_cast<void*>(s->start << kPageShift),
1949 static_cast<size_t>(s->length << kPageShift));
1950 s->decommitted = true;
1951 DLL_Prepend(&slist->returned, s);
1952
1953 scavenge_counter_ = std::max<size_t>(64UL, std::min<size_t>(kDefaultReleaseDelay, kDefaultReleaseDelay - (free_pages_ / kDefaultReleaseDelay)));
1954
1955 if (index == kMaxPages && !DLL_IsEmpty(&slist->normal))
1956 scavenge_index_ = index - 1;
1957 else
1958 scavenge_index_ = index;
1959 return;
1960 }
1961 index++;
1962 }
1963
1964 // Nothing to scavenge, delay for a while
1965 scavenge_counter_ = kDefaultReleaseDelay;
1966 }
1967 #endif
1968
RegisterSizeClass(Span * span,size_t sc)1969 void TCMalloc_PageHeap::RegisterSizeClass(Span* span, size_t sc) {
1970 // Associate span object with all interior pages as well
1971 ASSERT(!span->free);
1972 ASSERT(GetDescriptor(span->start) == span);
1973 ASSERT(GetDescriptor(span->start+span->length-1) == span);
1974 Event(span, 'C', sc);
1975 span->sizeclass = static_cast<unsigned int>(sc);
1976 for (Length i = 1; i < span->length-1; i++) {
1977 pagemap_.set(span->start+i, span);
1978 }
1979 }
1980
1981 #ifdef WTF_CHANGES
ReturnedBytes() const1982 size_t TCMalloc_PageHeap::ReturnedBytes() const {
1983 size_t result = 0;
1984 for (unsigned s = 0; s < kMaxPages; s++) {
1985 const int r_length = DLL_Length(&free_[s].returned);
1986 unsigned r_pages = s * r_length;
1987 result += r_pages << kPageShift;
1988 }
1989
1990 for (Span* s = large_.returned.next; s != &large_.returned; s = s->next)
1991 result += s->length << kPageShift;
1992 return result;
1993 }
1994 #endif
1995
1996 #ifndef WTF_CHANGES
PagesToMB(uint64_t pages)1997 static double PagesToMB(uint64_t pages) {
1998 return (pages << kPageShift) / 1048576.0;
1999 }
2000
Dump(TCMalloc_Printer * out)2001 void TCMalloc_PageHeap::Dump(TCMalloc_Printer* out) {
2002 int nonempty_sizes = 0;
2003 for (int s = 0; s < kMaxPages; s++) {
2004 if (!DLL_IsEmpty(&free_[s].normal) || !DLL_IsEmpty(&free_[s].returned)) {
2005 nonempty_sizes++;
2006 }
2007 }
2008 out->printf("------------------------------------------------\n");
2009 out->printf("PageHeap: %d sizes; %6.1f MB free\n",
2010 nonempty_sizes, PagesToMB(free_pages_));
2011 out->printf("------------------------------------------------\n");
2012 uint64_t total_normal = 0;
2013 uint64_t total_returned = 0;
2014 for (int s = 0; s < kMaxPages; s++) {
2015 const int n_length = DLL_Length(&free_[s].normal);
2016 const int r_length = DLL_Length(&free_[s].returned);
2017 if (n_length + r_length > 0) {
2018 uint64_t n_pages = s * n_length;
2019 uint64_t r_pages = s * r_length;
2020 total_normal += n_pages;
2021 total_returned += r_pages;
2022 out->printf("%6u pages * %6u spans ~ %6.1f MB; %6.1f MB cum"
2023 "; unmapped: %6.1f MB; %6.1f MB cum\n",
2024 s,
2025 (n_length + r_length),
2026 PagesToMB(n_pages + r_pages),
2027 PagesToMB(total_normal + total_returned),
2028 PagesToMB(r_pages),
2029 PagesToMB(total_returned));
2030 }
2031 }
2032
2033 uint64_t n_pages = 0;
2034 uint64_t r_pages = 0;
2035 int n_spans = 0;
2036 int r_spans = 0;
2037 out->printf("Normal large spans:\n");
2038 for (Span* s = large_.normal.next; s != &large_.normal; s = s->next) {
2039 out->printf(" [ %6" PRIuS " pages ] %6.1f MB\n",
2040 s->length, PagesToMB(s->length));
2041 n_pages += s->length;
2042 n_spans++;
2043 }
2044 out->printf("Unmapped large spans:\n");
2045 for (Span* s = large_.returned.next; s != &large_.returned; s = s->next) {
2046 out->printf(" [ %6" PRIuS " pages ] %6.1f MB\n",
2047 s->length, PagesToMB(s->length));
2048 r_pages += s->length;
2049 r_spans++;
2050 }
2051 total_normal += n_pages;
2052 total_returned += r_pages;
2053 out->printf(">255 large * %6u spans ~ %6.1f MB; %6.1f MB cum"
2054 "; unmapped: %6.1f MB; %6.1f MB cum\n",
2055 (n_spans + r_spans),
2056 PagesToMB(n_pages + r_pages),
2057 PagesToMB(total_normal + total_returned),
2058 PagesToMB(r_pages),
2059 PagesToMB(total_returned));
2060 }
2061 #endif
2062
GrowHeap(Length n)2063 bool TCMalloc_PageHeap::GrowHeap(Length n) {
2064 ASSERT(kMaxPages >= kMinSystemAlloc);
2065 if (n > kMaxValidPages) return false;
2066 Length ask = (n>kMinSystemAlloc) ? n : static_cast<Length>(kMinSystemAlloc);
2067 size_t actual_size;
2068 void* ptr = TCMalloc_SystemAlloc(ask << kPageShift, &actual_size, kPageSize);
2069 if (ptr == NULL) {
2070 if (n < ask) {
2071 // Try growing just "n" pages
2072 ask = n;
2073 ptr = TCMalloc_SystemAlloc(ask << kPageShift, &actual_size, kPageSize);
2074 }
2075 if (ptr == NULL) return false;
2076 }
2077 ask = actual_size >> kPageShift;
2078
2079 uint64_t old_system_bytes = system_bytes_;
2080 system_bytes_ += (ask << kPageShift);
2081 const PageID p = reinterpret_cast<uintptr_t>(ptr) >> kPageShift;
2082 ASSERT(p > 0);
2083
2084 // If we have already a lot of pages allocated, just pre allocate a bunch of
2085 // memory for the page map. This prevents fragmentation by pagemap metadata
2086 // when a program keeps allocating and freeing large blocks.
2087
2088 if (old_system_bytes < kPageMapBigAllocationThreshold
2089 && system_bytes_ >= kPageMapBigAllocationThreshold) {
2090 pagemap_.PreallocateMoreMemory();
2091 }
2092
2093 // Make sure pagemap_ has entries for all of the new pages.
2094 // Plus ensure one before and one after so coalescing code
2095 // does not need bounds-checking.
2096 if (pagemap_.Ensure(p-1, ask+2)) {
2097 // Pretend the new area is allocated and then Delete() it to
2098 // cause any necessary coalescing to occur.
2099 //
2100 // We do not adjust free_pages_ here since Delete() will do it for us.
2101 Span* span = NewSpan(p, ask);
2102 RecordSpan(span);
2103 Delete(span);
2104 ASSERT(Check());
2105 return true;
2106 } else {
2107 // We could not allocate memory within "pagemap_"
2108 // TODO: Once we can return memory to the system, return the new span
2109 return false;
2110 }
2111 }
2112
Check()2113 bool TCMalloc_PageHeap::Check() {
2114 ASSERT(free_[0].normal.next == &free_[0].normal);
2115 ASSERT(free_[0].returned.next == &free_[0].returned);
2116 CheckList(&large_.normal, kMaxPages, 1000000000);
2117 CheckList(&large_.returned, kMaxPages, 1000000000);
2118 for (Length s = 1; s < kMaxPages; s++) {
2119 CheckList(&free_[s].normal, s, s);
2120 CheckList(&free_[s].returned, s, s);
2121 }
2122 return true;
2123 }
2124
2125 #if ASSERT_DISABLED
CheckList(Span *,Length,Length)2126 bool TCMalloc_PageHeap::CheckList(Span*, Length, Length) {
2127 return true;
2128 }
2129 #else
CheckList(Span * list,Length min_pages,Length max_pages)2130 bool TCMalloc_PageHeap::CheckList(Span* list, Length min_pages, Length max_pages) {
2131 for (Span* s = list->next; s != list; s = s->next) {
2132 CHECK_CONDITION(s->free);
2133 CHECK_CONDITION(s->length >= min_pages);
2134 CHECK_CONDITION(s->length <= max_pages);
2135 CHECK_CONDITION(GetDescriptor(s->start) == s);
2136 CHECK_CONDITION(GetDescriptor(s->start+s->length-1) == s);
2137 }
2138 return true;
2139 }
2140 #endif
2141
ReleaseFreeList(Span * list,Span * returned)2142 static void ReleaseFreeList(Span* list, Span* returned) {
2143 // Walk backwards through list so that when we push these
2144 // spans on the "returned" list, we preserve the order.
2145 while (!DLL_IsEmpty(list)) {
2146 Span* s = list->prev;
2147 DLL_Remove(s);
2148 DLL_Prepend(returned, s);
2149 TCMalloc_SystemRelease(reinterpret_cast<void*>(s->start << kPageShift),
2150 static_cast<size_t>(s->length << kPageShift));
2151 }
2152 }
2153
ReleaseFreePages()2154 void TCMalloc_PageHeap::ReleaseFreePages() {
2155 for (Length s = 0; s < kMaxPages; s++) {
2156 ReleaseFreeList(&free_[s].normal, &free_[s].returned);
2157 }
2158 ReleaseFreeList(&large_.normal, &large_.returned);
2159 ASSERT(Check());
2160 }
2161
2162 //-------------------------------------------------------------------
2163 // Free list
2164 //-------------------------------------------------------------------
2165
2166 class TCMalloc_ThreadCache_FreeList {
2167 private:
2168 void* list_; // Linked list of nodes
2169 uint16_t length_; // Current length
2170 uint16_t lowater_; // Low water mark for list length
2171
2172 public:
Init()2173 void Init() {
2174 list_ = NULL;
2175 length_ = 0;
2176 lowater_ = 0;
2177 }
2178
2179 // Return current length of list
length() const2180 int length() const {
2181 return length_;
2182 }
2183
2184 // Is list empty?
empty() const2185 bool empty() const {
2186 return list_ == NULL;
2187 }
2188
2189 // Low-water mark management
lowwatermark() const2190 int lowwatermark() const { return lowater_; }
clear_lowwatermark()2191 void clear_lowwatermark() { lowater_ = length_; }
2192
Push(void * ptr)2193 ALWAYS_INLINE void Push(void* ptr) {
2194 SLL_Push(&list_, ptr);
2195 length_++;
2196 }
2197
PushRange(int N,void * start,void * end)2198 void PushRange(int N, void *start, void *end) {
2199 SLL_PushRange(&list_, start, end);
2200 length_ = length_ + static_cast<uint16_t>(N);
2201 }
2202
PopRange(int N,void ** start,void ** end)2203 void PopRange(int N, void **start, void **end) {
2204 SLL_PopRange(&list_, N, start, end);
2205 ASSERT(length_ >= N);
2206 length_ = length_ - static_cast<uint16_t>(N);
2207 if (length_ < lowater_) lowater_ = length_;
2208 }
2209
Pop()2210 ALWAYS_INLINE void* Pop() {
2211 ASSERT(list_ != NULL);
2212 length_--;
2213 if (length_ < lowater_) lowater_ = length_;
2214 return SLL_Pop(&list_);
2215 }
2216
2217 #ifdef WTF_CHANGES
2218 template <class Finder, class Reader>
enumerateFreeObjects(Finder & finder,const Reader & reader)2219 void enumerateFreeObjects(Finder& finder, const Reader& reader)
2220 {
2221 for (void* nextObject = list_; nextObject; nextObject = reader.nextEntryInLinkedList(reinterpret_cast<void**>(nextObject)))
2222 finder.visit(nextObject);
2223 }
2224 #endif
2225 };
2226
2227 //-------------------------------------------------------------------
2228 // Data kept per thread
2229 //-------------------------------------------------------------------
2230
2231 class TCMalloc_ThreadCache {
2232 private:
2233 typedef TCMalloc_ThreadCache_FreeList FreeList;
2234 #if OS(WINDOWS)
2235 typedef DWORD ThreadIdentifier;
2236 #else
2237 typedef pthread_t ThreadIdentifier;
2238 #endif
2239
2240 size_t size_; // Combined size of data
2241 ThreadIdentifier tid_; // Which thread owns it
2242 bool in_setspecific_; // Called pthread_setspecific?
2243 FreeList list_[kNumClasses]; // Array indexed by size-class
2244
2245 // We sample allocations, biased by the size of the allocation
2246 uint32_t rnd_; // Cheap random number generator
2247 size_t bytes_until_sample_; // Bytes until we sample next
2248
2249 // Allocate a new heap. REQUIRES: pageheap_lock is held.
2250 static inline TCMalloc_ThreadCache* NewHeap(ThreadIdentifier tid);
2251
2252 // Use only as pthread thread-specific destructor function.
2253 static void DestroyThreadCache(void* ptr);
2254 public:
2255 // All ThreadCache objects are kept in a linked list (for stats collection)
2256 TCMalloc_ThreadCache* next_;
2257 TCMalloc_ThreadCache* prev_;
2258
2259 void Init(ThreadIdentifier tid);
2260 void Cleanup();
2261
2262 // Accessors (mostly just for printing stats)
freelist_length(size_t cl) const2263 int freelist_length(size_t cl) const { return list_[cl].length(); }
2264
2265 // Total byte size in cache
Size() const2266 size_t Size() const { return size_; }
2267
2268 ALWAYS_INLINE void* Allocate(size_t size);
2269 void Deallocate(void* ptr, size_t size_class);
2270
2271 ALWAYS_INLINE void FetchFromCentralCache(size_t cl, size_t allocationSize);
2272 void ReleaseToCentralCache(size_t cl, int N);
2273 void Scavenge();
2274 void Print() const;
2275
2276 // Record allocation of "k" bytes. Return true iff allocation
2277 // should be sampled
2278 bool SampleAllocation(size_t k);
2279
2280 // Pick next sampling point
2281 void PickNextSample(size_t k);
2282
2283 static void InitModule();
2284 static void InitTSD();
2285 static TCMalloc_ThreadCache* GetThreadHeap();
2286 static TCMalloc_ThreadCache* GetCache();
2287 static TCMalloc_ThreadCache* GetCacheIfPresent();
2288 static TCMalloc_ThreadCache* CreateCacheIfNecessary();
2289 static void DeleteCache(TCMalloc_ThreadCache* heap);
2290 static void BecomeIdle();
2291 static void RecomputeThreadCacheSize();
2292
2293 #ifdef WTF_CHANGES
2294 template <class Finder, class Reader>
enumerateFreeObjects(Finder & finder,const Reader & reader)2295 void enumerateFreeObjects(Finder& finder, const Reader& reader)
2296 {
2297 for (unsigned sizeClass = 0; sizeClass < kNumClasses; sizeClass++)
2298 list_[sizeClass].enumerateFreeObjects(finder, reader);
2299 }
2300 #endif
2301 };
2302
2303 //-------------------------------------------------------------------
2304 // Data kept per size-class in central cache
2305 //-------------------------------------------------------------------
2306
2307 class TCMalloc_Central_FreeList {
2308 public:
2309 void Init(size_t cl);
2310
2311 // These methods all do internal locking.
2312
2313 // Insert the specified range into the central freelist. N is the number of
2314 // elements in the range.
2315 void InsertRange(void *start, void *end, int N);
2316
2317 // Returns the actual number of fetched elements into N.
2318 void RemoveRange(void **start, void **end, int *N);
2319
2320 // Returns the number of free objects in cache.
length()2321 size_t length() {
2322 SpinLockHolder h(&lock_);
2323 return counter_;
2324 }
2325
2326 // Returns the number of free objects in the transfer cache.
tc_length()2327 int tc_length() {
2328 SpinLockHolder h(&lock_);
2329 return used_slots_ * num_objects_to_move[size_class_];
2330 }
2331
2332 #ifdef WTF_CHANGES
2333 template <class Finder, class Reader>
enumerateFreeObjects(Finder & finder,const Reader & reader,TCMalloc_Central_FreeList * remoteCentralFreeList)2334 void enumerateFreeObjects(Finder& finder, const Reader& reader, TCMalloc_Central_FreeList* remoteCentralFreeList)
2335 {
2336 for (Span* span = &empty_; span && span != &empty_; span = (span->next ? reader(span->next) : 0))
2337 ASSERT(!span->objects);
2338
2339 ASSERT(!nonempty_.objects);
2340 static const ptrdiff_t nonemptyOffset = reinterpret_cast<const char*>(&nonempty_) - reinterpret_cast<const char*>(this);
2341
2342 Span* remoteNonempty = reinterpret_cast<Span*>(reinterpret_cast<char*>(remoteCentralFreeList) + nonemptyOffset);
2343 Span* remoteSpan = nonempty_.next;
2344
2345 for (Span* span = reader(remoteSpan); span && remoteSpan != remoteNonempty; remoteSpan = span->next, span = (span->next ? reader(span->next) : 0)) {
2346 for (void* nextObject = span->objects; nextObject; nextObject = reader.nextEntryInLinkedList(reinterpret_cast<void**>(nextObject)))
2347 finder.visit(nextObject);
2348 }
2349 }
2350 #endif
2351
2352 private:
2353 // REQUIRES: lock_ is held
2354 // Remove object from cache and return.
2355 // Return NULL if no free entries in cache.
2356 void* FetchFromSpans();
2357
2358 // REQUIRES: lock_ is held
2359 // Remove object from cache and return. Fetches
2360 // from pageheap if cache is empty. Only returns
2361 // NULL on allocation failure.
2362 void* FetchFromSpansSafe();
2363
2364 // REQUIRES: lock_ is held
2365 // Release a linked list of objects to spans.
2366 // May temporarily release lock_.
2367 void ReleaseListToSpans(void *start);
2368
2369 // REQUIRES: lock_ is held
2370 // Release an object to spans.
2371 // May temporarily release lock_.
2372 ALWAYS_INLINE void ReleaseToSpans(void* object);
2373
2374 // REQUIRES: lock_ is held
2375 // Populate cache by fetching from the page heap.
2376 // May temporarily release lock_.
2377 ALWAYS_INLINE void Populate();
2378
2379 // REQUIRES: lock is held.
2380 // Tries to make room for a TCEntry. If the cache is full it will try to
2381 // expand it at the cost of some other cache size. Return false if there is
2382 // no space.
2383 bool MakeCacheSpace();
2384
2385 // REQUIRES: lock_ for locked_size_class is held.
2386 // Picks a "random" size class to steal TCEntry slot from. In reality it
2387 // just iterates over the sizeclasses but does so without taking a lock.
2388 // Returns true on success.
2389 // May temporarily lock a "random" size class.
2390 static ALWAYS_INLINE bool EvictRandomSizeClass(size_t locked_size_class, bool force);
2391
2392 // REQUIRES: lock_ is *not* held.
2393 // Tries to shrink the Cache. If force is true it will relase objects to
2394 // spans if it allows it to shrink the cache. Return false if it failed to
2395 // shrink the cache. Decrements cache_size_ on succeess.
2396 // May temporarily take lock_. If it takes lock_, the locked_size_class
2397 // lock is released to the thread from holding two size class locks
2398 // concurrently which could lead to a deadlock.
2399 bool ShrinkCache(int locked_size_class, bool force);
2400
2401 // This lock protects all the data members. cached_entries and cache_size_
2402 // may be looked at without holding the lock.
2403 SpinLock lock_;
2404
2405 // We keep linked lists of empty and non-empty spans.
2406 size_t size_class_; // My size class
2407 Span empty_; // Dummy header for list of empty spans
2408 Span nonempty_; // Dummy header for list of non-empty spans
2409 size_t counter_; // Number of free objects in cache entry
2410
2411 // Here we reserve space for TCEntry cache slots. Since one size class can
2412 // end up getting all the TCEntries quota in the system we just preallocate
2413 // sufficient number of entries here.
2414 TCEntry tc_slots_[kNumTransferEntries];
2415
2416 // Number of currently used cached entries in tc_slots_. This variable is
2417 // updated under a lock but can be read without one.
2418 int32_t used_slots_;
2419 // The current number of slots for this size class. This is an
2420 // adaptive value that is increased if there is lots of traffic
2421 // on a given size class.
2422 int32_t cache_size_;
2423 };
2424
2425 // Pad each CentralCache object to multiple of 64 bytes
2426 class TCMalloc_Central_FreeListPadded : public TCMalloc_Central_FreeList {
2427 private:
2428 char pad_[(64 - (sizeof(TCMalloc_Central_FreeList) % 64)) % 64];
2429 };
2430
2431 //-------------------------------------------------------------------
2432 // Global variables
2433 //-------------------------------------------------------------------
2434
2435 // Central cache -- a collection of free-lists, one per size-class.
2436 // We have a separate lock per free-list to reduce contention.
2437 static TCMalloc_Central_FreeListPadded central_cache[kNumClasses];
2438
2439 // Page-level allocator
2440 static SpinLock pageheap_lock = SPINLOCK_INITIALIZER;
2441 static AllocAlignmentInteger pageheap_memory[(sizeof(TCMalloc_PageHeap) + sizeof(AllocAlignmentInteger) - 1) / sizeof(AllocAlignmentInteger)];
2442 static bool phinited = false;
2443
2444 // Avoid extra level of indirection by making "pageheap" be just an alias
2445 // of pageheap_memory.
2446 typedef union {
2447 void* m_memory;
2448 TCMalloc_PageHeap* m_pageHeap;
2449 } PageHeapUnion;
2450
getPageHeap()2451 static inline TCMalloc_PageHeap* getPageHeap()
2452 {
2453 PageHeapUnion u = { &pageheap_memory[0] };
2454 return u.m_pageHeap;
2455 }
2456
2457 #define pageheap getPageHeap()
2458
2459 #if USE_BACKGROUND_THREAD_TO_SCAVENGE_MEMORY
2460
2461 #if HAVE(DISPATCH_H) || OS(WINDOWS)
2462
periodicScavenge()2463 void TCMalloc_PageHeap::periodicScavenge()
2464 {
2465 SpinLockHolder h(&pageheap_lock);
2466 pageheap->scavenge();
2467
2468 if (shouldScavenge()) {
2469 rescheduleScavenger();
2470 return;
2471 }
2472
2473 suspendScavenger();
2474 }
2475
signalScavenger()2476 ALWAYS_INLINE void TCMalloc_PageHeap::signalScavenger()
2477 {
2478 ASSERT(IsHeld(pageheap_lock));
2479 if (isScavengerSuspended() && shouldScavenge())
2480 scheduleScavenger();
2481 }
2482
2483 #else
2484
scavengerThread()2485 void TCMalloc_PageHeap::scavengerThread()
2486 {
2487 #if HAVE(PTHREAD_SETNAME_NP)
2488 pthread_setname_np("JavaScriptCore: FastMalloc scavenger");
2489 #endif
2490
2491 while (1) {
2492 if (!shouldScavenge()) {
2493 pthread_mutex_lock(&m_scavengeMutex);
2494 m_scavengeThreadActive = false;
2495 // Block until there are enough free committed pages to release back to the system.
2496 pthread_cond_wait(&m_scavengeCondition, &m_scavengeMutex);
2497 m_scavengeThreadActive = true;
2498 pthread_mutex_unlock(&m_scavengeMutex);
2499 }
2500 sleep(kScavengeDelayInSeconds);
2501 {
2502 SpinLockHolder h(&pageheap_lock);
2503 pageheap->scavenge();
2504 }
2505 }
2506 }
2507
2508 #endif
2509
2510 #endif
2511
2512 // If TLS is available, we also store a copy
2513 // of the per-thread object in a __thread variable
2514 // since __thread variables are faster to read
2515 // than pthread_getspecific(). We still need
2516 // pthread_setspecific() because __thread
2517 // variables provide no way to run cleanup
2518 // code when a thread is destroyed.
2519 #ifdef HAVE_TLS
2520 static __thread TCMalloc_ThreadCache *threadlocal_heap;
2521 #endif
2522 // Thread-specific key. Initialization here is somewhat tricky
2523 // because some Linux startup code invokes malloc() before it
2524 // is in a good enough state to handle pthread_keycreate().
2525 // Therefore, we use TSD keys only after tsd_inited is set to true.
2526 // Until then, we use a slow path to get the heap object.
2527 static bool tsd_inited = false;
2528 #if USE(PTHREAD_GETSPECIFIC_DIRECT)
2529 static const pthread_key_t heap_key = __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY0;
2530 #else
2531 static pthread_key_t heap_key;
2532 #endif
2533 #if OS(WINDOWS)
2534 DWORD tlsIndex = TLS_OUT_OF_INDEXES;
2535 #endif
2536
setThreadHeap(TCMalloc_ThreadCache * heap)2537 static ALWAYS_INLINE void setThreadHeap(TCMalloc_ThreadCache* heap)
2538 {
2539 #if USE(PTHREAD_GETSPECIFIC_DIRECT)
2540 // Can't have two libraries both doing this in the same process,
2541 // so check and make this crash right away.
2542 if (pthread_getspecific(heap_key))
2543 CRASH();
2544 #endif
2545
2546 // Still do pthread_setspecific even if there's an alternate form
2547 // of thread-local storage in use, to benefit from the delete callback.
2548 pthread_setspecific(heap_key, heap);
2549
2550 #if OS(WINDOWS)
2551 TlsSetValue(tlsIndex, heap);
2552 #endif
2553 }
2554
2555 // Allocator for thread heaps
2556 static PageHeapAllocator<TCMalloc_ThreadCache> threadheap_allocator;
2557
2558 // Linked list of heap objects. Protected by pageheap_lock.
2559 static TCMalloc_ThreadCache* thread_heaps = NULL;
2560 static int thread_heap_count = 0;
2561
2562 // Overall thread cache size. Protected by pageheap_lock.
2563 static size_t overall_thread_cache_size = kDefaultOverallThreadCacheSize;
2564
2565 // Global per-thread cache size. Writes are protected by
2566 // pageheap_lock. Reads are done without any locking, which should be
2567 // fine as long as size_t can be written atomically and we don't place
2568 // invariants between this variable and other pieces of state.
2569 static volatile size_t per_thread_cache_size = kMaxThreadCacheSize;
2570
2571 //-------------------------------------------------------------------
2572 // Central cache implementation
2573 //-------------------------------------------------------------------
2574
Init(size_t cl)2575 void TCMalloc_Central_FreeList::Init(size_t cl) {
2576 lock_.Init();
2577 size_class_ = cl;
2578 DLL_Init(&empty_);
2579 DLL_Init(&nonempty_);
2580 counter_ = 0;
2581
2582 cache_size_ = 1;
2583 used_slots_ = 0;
2584 ASSERT(cache_size_ <= kNumTransferEntries);
2585 }
2586
ReleaseListToSpans(void * start)2587 void TCMalloc_Central_FreeList::ReleaseListToSpans(void* start) {
2588 while (start) {
2589 void *next = SLL_Next(start);
2590 ReleaseToSpans(start);
2591 start = next;
2592 }
2593 }
2594
ReleaseToSpans(void * object)2595 ALWAYS_INLINE void TCMalloc_Central_FreeList::ReleaseToSpans(void* object) {
2596 const PageID p = reinterpret_cast<uintptr_t>(object) >> kPageShift;
2597 Span* span = pageheap->GetDescriptor(p);
2598 ASSERT(span != NULL);
2599 ASSERT(span->refcount > 0);
2600
2601 // If span is empty, move it to non-empty list
2602 if (span->objects == NULL) {
2603 DLL_Remove(span);
2604 DLL_Prepend(&nonempty_, span);
2605 Event(span, 'N', 0);
2606 }
2607
2608 // The following check is expensive, so it is disabled by default
2609 if (false) {
2610 // Check that object does not occur in list
2611 unsigned got = 0;
2612 for (void* p = span->objects; p != NULL; p = *((void**) p)) {
2613 ASSERT(p != object);
2614 got++;
2615 }
2616 ASSERT(got + span->refcount ==
2617 (span->length<<kPageShift)/ByteSizeForClass(span->sizeclass));
2618 }
2619
2620 counter_++;
2621 span->refcount--;
2622 if (span->refcount == 0) {
2623 Event(span, '#', 0);
2624 counter_ -= (span->length<<kPageShift) / ByteSizeForClass(span->sizeclass);
2625 DLL_Remove(span);
2626
2627 // Release central list lock while operating on pageheap
2628 lock_.Unlock();
2629 {
2630 SpinLockHolder h(&pageheap_lock);
2631 pageheap->Delete(span);
2632 }
2633 lock_.Lock();
2634 } else {
2635 *(reinterpret_cast<void**>(object)) = span->objects;
2636 span->objects = object;
2637 }
2638 }
2639
EvictRandomSizeClass(size_t locked_size_class,bool force)2640 ALWAYS_INLINE bool TCMalloc_Central_FreeList::EvictRandomSizeClass(
2641 size_t locked_size_class, bool force) {
2642 static int race_counter = 0;
2643 int t = race_counter++; // Updated without a lock, but who cares.
2644 if (t >= static_cast<int>(kNumClasses)) {
2645 while (t >= static_cast<int>(kNumClasses)) {
2646 t -= kNumClasses;
2647 }
2648 race_counter = t;
2649 }
2650 ASSERT(t >= 0);
2651 ASSERT(t < static_cast<int>(kNumClasses));
2652 if (t == static_cast<int>(locked_size_class)) return false;
2653 return central_cache[t].ShrinkCache(static_cast<int>(locked_size_class), force);
2654 }
2655
MakeCacheSpace()2656 bool TCMalloc_Central_FreeList::MakeCacheSpace() {
2657 // Is there room in the cache?
2658 if (used_slots_ < cache_size_) return true;
2659 // Check if we can expand this cache?
2660 if (cache_size_ == kNumTransferEntries) return false;
2661 // Ok, we'll try to grab an entry from some other size class.
2662 if (EvictRandomSizeClass(size_class_, false) ||
2663 EvictRandomSizeClass(size_class_, true)) {
2664 // Succeeded in evicting, we're going to make our cache larger.
2665 cache_size_++;
2666 return true;
2667 }
2668 return false;
2669 }
2670
2671
2672 namespace {
2673 class LockInverter {
2674 private:
2675 SpinLock *held_, *temp_;
2676 public:
LockInverter(SpinLock * held,SpinLock * temp)2677 inline explicit LockInverter(SpinLock* held, SpinLock *temp)
2678 : held_(held), temp_(temp) { held_->Unlock(); temp_->Lock(); }
~LockInverter()2679 inline ~LockInverter() { temp_->Unlock(); held_->Lock(); }
2680 };
2681 }
2682
ShrinkCache(int locked_size_class,bool force)2683 bool TCMalloc_Central_FreeList::ShrinkCache(int locked_size_class, bool force) {
2684 // Start with a quick check without taking a lock.
2685 if (cache_size_ == 0) return false;
2686 // We don't evict from a full cache unless we are 'forcing'.
2687 if (force == false && used_slots_ == cache_size_) return false;
2688
2689 // Grab lock, but first release the other lock held by this thread. We use
2690 // the lock inverter to ensure that we never hold two size class locks
2691 // concurrently. That can create a deadlock because there is no well
2692 // defined nesting order.
2693 LockInverter li(¢ral_cache[locked_size_class].lock_, &lock_);
2694 ASSERT(used_slots_ <= cache_size_);
2695 ASSERT(0 <= cache_size_);
2696 if (cache_size_ == 0) return false;
2697 if (used_slots_ == cache_size_) {
2698 if (force == false) return false;
2699 // ReleaseListToSpans releases the lock, so we have to make all the
2700 // updates to the central list before calling it.
2701 cache_size_--;
2702 used_slots_--;
2703 ReleaseListToSpans(tc_slots_[used_slots_].head);
2704 return true;
2705 }
2706 cache_size_--;
2707 return true;
2708 }
2709
InsertRange(void * start,void * end,int N)2710 void TCMalloc_Central_FreeList::InsertRange(void *start, void *end, int N) {
2711 SpinLockHolder h(&lock_);
2712 if (N == num_objects_to_move[size_class_] &&
2713 MakeCacheSpace()) {
2714 int slot = used_slots_++;
2715 ASSERT(slot >=0);
2716 ASSERT(slot < kNumTransferEntries);
2717 TCEntry *entry = &tc_slots_[slot];
2718 entry->head = start;
2719 entry->tail = end;
2720 return;
2721 }
2722 ReleaseListToSpans(start);
2723 }
2724
RemoveRange(void ** start,void ** end,int * N)2725 void TCMalloc_Central_FreeList::RemoveRange(void **start, void **end, int *N) {
2726 int num = *N;
2727 ASSERT(num > 0);
2728
2729 SpinLockHolder h(&lock_);
2730 if (num == num_objects_to_move[size_class_] && used_slots_ > 0) {
2731 int slot = --used_slots_;
2732 ASSERT(slot >= 0);
2733 TCEntry *entry = &tc_slots_[slot];
2734 *start = entry->head;
2735 *end = entry->tail;
2736 return;
2737 }
2738
2739 // TODO: Prefetch multiple TCEntries?
2740 void *tail = FetchFromSpansSafe();
2741 if (!tail) {
2742 // We are completely out of memory.
2743 *start = *end = NULL;
2744 *N = 0;
2745 return;
2746 }
2747
2748 SLL_SetNext(tail, NULL);
2749 void *head = tail;
2750 int count = 1;
2751 while (count < num) {
2752 void *t = FetchFromSpans();
2753 if (!t) break;
2754 SLL_Push(&head, t);
2755 count++;
2756 }
2757 *start = head;
2758 *end = tail;
2759 *N = count;
2760 }
2761
2762
FetchFromSpansSafe()2763 void* TCMalloc_Central_FreeList::FetchFromSpansSafe() {
2764 void *t = FetchFromSpans();
2765 if (!t) {
2766 Populate();
2767 t = FetchFromSpans();
2768 }
2769 return t;
2770 }
2771
FetchFromSpans()2772 void* TCMalloc_Central_FreeList::FetchFromSpans() {
2773 if (DLL_IsEmpty(&nonempty_)) return NULL;
2774 Span* span = nonempty_.next;
2775
2776 ASSERT(span->objects != NULL);
2777 ASSERT_SPAN_COMMITTED(span);
2778 span->refcount++;
2779 void* result = span->objects;
2780 span->objects = *(reinterpret_cast<void**>(result));
2781 if (span->objects == NULL) {
2782 // Move to empty list
2783 DLL_Remove(span);
2784 DLL_Prepend(&empty_, span);
2785 Event(span, 'E', 0);
2786 }
2787 counter_--;
2788 return result;
2789 }
2790
2791 // Fetch memory from the system and add to the central cache freelist.
Populate()2792 ALWAYS_INLINE void TCMalloc_Central_FreeList::Populate() {
2793 // Release central list lock while operating on pageheap
2794 lock_.Unlock();
2795 const size_t npages = class_to_pages[size_class_];
2796
2797 Span* span;
2798 {
2799 SpinLockHolder h(&pageheap_lock);
2800 span = pageheap->New(npages);
2801 if (span) pageheap->RegisterSizeClass(span, size_class_);
2802 }
2803 if (span == NULL) {
2804 #if HAVE(ERRNO_H)
2805 MESSAGE("allocation failed: %d\n", errno);
2806 #elif OS(WINDOWS)
2807 MESSAGE("allocation failed: %d\n", ::GetLastError());
2808 #else
2809 MESSAGE("allocation failed\n");
2810 #endif
2811 lock_.Lock();
2812 return;
2813 }
2814 ASSERT_SPAN_COMMITTED(span);
2815 ASSERT(span->length == npages);
2816 // Cache sizeclass info eagerly. Locking is not necessary.
2817 // (Instead of being eager, we could just replace any stale info
2818 // about this span, but that seems to be no better in practice.)
2819 for (size_t i = 0; i < npages; i++) {
2820 pageheap->CacheSizeClass(span->start + i, size_class_);
2821 }
2822
2823 // Split the block into pieces and add to the free-list
2824 // TODO: coloring of objects to avoid cache conflicts?
2825 void** tail = &span->objects;
2826 char* ptr = reinterpret_cast<char*>(span->start << kPageShift);
2827 char* limit = ptr + (npages << kPageShift);
2828 const size_t size = ByteSizeForClass(size_class_);
2829 int num = 0;
2830 char* nptr;
2831 while ((nptr = ptr + size) <= limit) {
2832 *tail = ptr;
2833 tail = reinterpret_cast_ptr<void**>(ptr);
2834 ptr = nptr;
2835 num++;
2836 }
2837 ASSERT(ptr <= limit);
2838 *tail = NULL;
2839 span->refcount = 0; // No sub-object in use yet
2840
2841 // Add span to list of non-empty spans
2842 lock_.Lock();
2843 DLL_Prepend(&nonempty_, span);
2844 counter_ += num;
2845 }
2846
2847 //-------------------------------------------------------------------
2848 // TCMalloc_ThreadCache implementation
2849 //-------------------------------------------------------------------
2850
SampleAllocation(size_t k)2851 inline bool TCMalloc_ThreadCache::SampleAllocation(size_t k) {
2852 if (bytes_until_sample_ < k) {
2853 PickNextSample(k);
2854 return true;
2855 } else {
2856 bytes_until_sample_ -= k;
2857 return false;
2858 }
2859 }
2860
Init(ThreadIdentifier tid)2861 void TCMalloc_ThreadCache::Init(ThreadIdentifier tid) {
2862 size_ = 0;
2863 next_ = NULL;
2864 prev_ = NULL;
2865 tid_ = tid;
2866 in_setspecific_ = false;
2867 for (size_t cl = 0; cl < kNumClasses; ++cl) {
2868 list_[cl].Init();
2869 }
2870
2871 // Initialize RNG -- run it for a bit to get to good values
2872 bytes_until_sample_ = 0;
2873 rnd_ = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(this));
2874 for (int i = 0; i < 100; i++) {
2875 PickNextSample(static_cast<size_t>(FLAGS_tcmalloc_sample_parameter * 2));
2876 }
2877 }
2878
Cleanup()2879 void TCMalloc_ThreadCache::Cleanup() {
2880 // Put unused memory back into central cache
2881 for (size_t cl = 0; cl < kNumClasses; ++cl) {
2882 if (list_[cl].length() > 0) {
2883 ReleaseToCentralCache(cl, list_[cl].length());
2884 }
2885 }
2886 }
2887
Allocate(size_t size)2888 ALWAYS_INLINE void* TCMalloc_ThreadCache::Allocate(size_t size) {
2889 ASSERT(size <= kMaxSize);
2890 const size_t cl = SizeClass(size);
2891 FreeList* list = &list_[cl];
2892 size_t allocationSize = ByteSizeForClass(cl);
2893 if (list->empty()) {
2894 FetchFromCentralCache(cl, allocationSize);
2895 if (list->empty()) return NULL;
2896 }
2897 size_ -= allocationSize;
2898 return list->Pop();
2899 }
2900
Deallocate(void * ptr,size_t cl)2901 inline void TCMalloc_ThreadCache::Deallocate(void* ptr, size_t cl) {
2902 size_ += ByteSizeForClass(cl);
2903 FreeList* list = &list_[cl];
2904 list->Push(ptr);
2905 // If enough data is free, put back into central cache
2906 if (list->length() > kMaxFreeListLength) {
2907 ReleaseToCentralCache(cl, num_objects_to_move[cl]);
2908 }
2909 if (size_ >= per_thread_cache_size) Scavenge();
2910 }
2911
2912 // Remove some objects of class "cl" from central cache and add to thread heap
FetchFromCentralCache(size_t cl,size_t allocationSize)2913 ALWAYS_INLINE void TCMalloc_ThreadCache::FetchFromCentralCache(size_t cl, size_t allocationSize) {
2914 int fetch_count = num_objects_to_move[cl];
2915 void *start, *end;
2916 central_cache[cl].RemoveRange(&start, &end, &fetch_count);
2917 list_[cl].PushRange(fetch_count, start, end);
2918 size_ += allocationSize * fetch_count;
2919 }
2920
2921 // Remove some objects of class "cl" from thread heap and add to central cache
ReleaseToCentralCache(size_t cl,int N)2922 inline void TCMalloc_ThreadCache::ReleaseToCentralCache(size_t cl, int N) {
2923 ASSERT(N > 0);
2924 FreeList* src = &list_[cl];
2925 if (N > src->length()) N = src->length();
2926 size_ -= N*ByteSizeForClass(cl);
2927
2928 // We return prepackaged chains of the correct size to the central cache.
2929 // TODO: Use the same format internally in the thread caches?
2930 int batch_size = num_objects_to_move[cl];
2931 while (N > batch_size) {
2932 void *tail, *head;
2933 src->PopRange(batch_size, &head, &tail);
2934 central_cache[cl].InsertRange(head, tail, batch_size);
2935 N -= batch_size;
2936 }
2937 void *tail, *head;
2938 src->PopRange(N, &head, &tail);
2939 central_cache[cl].InsertRange(head, tail, N);
2940 }
2941
2942 // Release idle memory to the central cache
Scavenge()2943 inline void TCMalloc_ThreadCache::Scavenge() {
2944 // If the low-water mark for the free list is L, it means we would
2945 // not have had to allocate anything from the central cache even if
2946 // we had reduced the free list size by L. We aim to get closer to
2947 // that situation by dropping L/2 nodes from the free list. This
2948 // may not release much memory, but if so we will call scavenge again
2949 // pretty soon and the low-water marks will be high on that call.
2950 //int64 start = CycleClock::Now();
2951
2952 for (size_t cl = 0; cl < kNumClasses; cl++) {
2953 FreeList* list = &list_[cl];
2954 const int lowmark = list->lowwatermark();
2955 if (lowmark > 0) {
2956 const int drop = (lowmark > 1) ? lowmark/2 : 1;
2957 ReleaseToCentralCache(cl, drop);
2958 }
2959 list->clear_lowwatermark();
2960 }
2961
2962 //int64 finish = CycleClock::Now();
2963 //CycleTimer ct;
2964 //MESSAGE("GC: %.0f ns\n", ct.CyclesToUsec(finish-start)*1000.0);
2965 }
2966
PickNextSample(size_t k)2967 void TCMalloc_ThreadCache::PickNextSample(size_t k) {
2968 // Make next "random" number
2969 // x^32+x^22+x^2+x^1+1 is a primitive polynomial for random numbers
2970 static const uint32_t kPoly = (1 << 22) | (1 << 2) | (1 << 1) | (1 << 0);
2971 uint32_t r = rnd_;
2972 rnd_ = (r << 1) ^ ((static_cast<int32_t>(r) >> 31) & kPoly);
2973
2974 // Next point is "rnd_ % (sample_period)". I.e., average
2975 // increment is "sample_period/2".
2976 const int flag_value = static_cast<int>(FLAGS_tcmalloc_sample_parameter);
2977 static int last_flag_value = -1;
2978
2979 if (flag_value != last_flag_value) {
2980 SpinLockHolder h(&sample_period_lock);
2981 int i;
2982 for (i = 0; i < (static_cast<int>(sizeof(primes_list)/sizeof(primes_list[0])) - 1); i++) {
2983 if (primes_list[i] >= flag_value) {
2984 break;
2985 }
2986 }
2987 sample_period = primes_list[i];
2988 last_flag_value = flag_value;
2989 }
2990
2991 bytes_until_sample_ += rnd_ % sample_period;
2992
2993 if (k > (static_cast<size_t>(-1) >> 2)) {
2994 // If the user has asked for a huge allocation then it is possible
2995 // for the code below to loop infinitely. Just return (note that
2996 // this throws off the sampling accuracy somewhat, but a user who
2997 // is allocating more than 1G of memory at a time can live with a
2998 // minor inaccuracy in profiling of small allocations, and also
2999 // would rather not wait for the loop below to terminate).
3000 return;
3001 }
3002
3003 while (bytes_until_sample_ < k) {
3004 // Increase bytes_until_sample_ by enough average sampling periods
3005 // (sample_period >> 1) to allow us to sample past the current
3006 // allocation.
3007 bytes_until_sample_ += (sample_period >> 1);
3008 }
3009
3010 bytes_until_sample_ -= k;
3011 }
3012
InitModule()3013 void TCMalloc_ThreadCache::InitModule() {
3014 // There is a slight potential race here because of double-checked
3015 // locking idiom. However, as long as the program does a small
3016 // allocation before switching to multi-threaded mode, we will be
3017 // fine. We increase the chances of doing such a small allocation
3018 // by doing one in the constructor of the module_enter_exit_hook
3019 // object declared below.
3020 SpinLockHolder h(&pageheap_lock);
3021 if (!phinited) {
3022 #ifdef WTF_CHANGES
3023 InitTSD();
3024 #endif
3025 InitSizeClasses();
3026 threadheap_allocator.Init();
3027 span_allocator.Init();
3028 span_allocator.New(); // Reduce cache conflicts
3029 span_allocator.New(); // Reduce cache conflicts
3030 stacktrace_allocator.Init();
3031 DLL_Init(&sampled_objects);
3032 for (size_t i = 0; i < kNumClasses; ++i) {
3033 central_cache[i].Init(i);
3034 }
3035 pageheap->init();
3036 phinited = 1;
3037 #if defined(WTF_CHANGES) && OS(DARWIN)
3038 FastMallocZone::init();
3039 #endif
3040 }
3041 }
3042
NewHeap(ThreadIdentifier tid)3043 inline TCMalloc_ThreadCache* TCMalloc_ThreadCache::NewHeap(ThreadIdentifier tid) {
3044 // Create the heap and add it to the linked list
3045 TCMalloc_ThreadCache *heap = threadheap_allocator.New();
3046 heap->Init(tid);
3047 heap->next_ = thread_heaps;
3048 heap->prev_ = NULL;
3049 if (thread_heaps != NULL) thread_heaps->prev_ = heap;
3050 thread_heaps = heap;
3051 thread_heap_count++;
3052 RecomputeThreadCacheSize();
3053 return heap;
3054 }
3055
GetThreadHeap()3056 inline TCMalloc_ThreadCache* TCMalloc_ThreadCache::GetThreadHeap() {
3057 #ifdef HAVE_TLS
3058 // __thread is faster, but only when the kernel supports it
3059 if (KernelSupportsTLS())
3060 return threadlocal_heap;
3061 #elif OS(WINDOWS)
3062 return static_cast<TCMalloc_ThreadCache*>(TlsGetValue(tlsIndex));
3063 #else
3064 return static_cast<TCMalloc_ThreadCache*>(pthread_getspecific(heap_key));
3065 #endif
3066 }
3067
GetCache()3068 inline TCMalloc_ThreadCache* TCMalloc_ThreadCache::GetCache() {
3069 TCMalloc_ThreadCache* ptr = NULL;
3070 if (!tsd_inited) {
3071 InitModule();
3072 } else {
3073 ptr = GetThreadHeap();
3074 }
3075 if (ptr == NULL) ptr = CreateCacheIfNecessary();
3076 return ptr;
3077 }
3078
3079 // In deletion paths, we do not try to create a thread-cache. This is
3080 // because we may be in the thread destruction code and may have
3081 // already cleaned up the cache for this thread.
GetCacheIfPresent()3082 inline TCMalloc_ThreadCache* TCMalloc_ThreadCache::GetCacheIfPresent() {
3083 if (!tsd_inited) return NULL;
3084 void* const p = GetThreadHeap();
3085 return reinterpret_cast<TCMalloc_ThreadCache*>(p);
3086 }
3087
InitTSD()3088 void TCMalloc_ThreadCache::InitTSD() {
3089 ASSERT(!tsd_inited);
3090 #if USE(PTHREAD_GETSPECIFIC_DIRECT)
3091 pthread_key_init_np(heap_key, DestroyThreadCache);
3092 #else
3093 pthread_key_create(&heap_key, DestroyThreadCache);
3094 #endif
3095 #if OS(WINDOWS)
3096 tlsIndex = TlsAlloc();
3097 #endif
3098 tsd_inited = true;
3099
3100 #if !OS(WINDOWS)
3101 // We may have used a fake pthread_t for the main thread. Fix it.
3102 pthread_t zero;
3103 memset(&zero, 0, sizeof(zero));
3104 #endif
3105 #ifndef WTF_CHANGES
3106 SpinLockHolder h(&pageheap_lock);
3107 #else
3108 ASSERT(pageheap_lock.IsHeld());
3109 #endif
3110 for (TCMalloc_ThreadCache* h = thread_heaps; h != NULL; h = h->next_) {
3111 #if OS(WINDOWS)
3112 if (h->tid_ == 0) {
3113 h->tid_ = GetCurrentThreadId();
3114 }
3115 #else
3116 if (pthread_equal(h->tid_, zero)) {
3117 h->tid_ = pthread_self();
3118 }
3119 #endif
3120 }
3121 }
3122
CreateCacheIfNecessary()3123 TCMalloc_ThreadCache* TCMalloc_ThreadCache::CreateCacheIfNecessary() {
3124 // Initialize per-thread data if necessary
3125 TCMalloc_ThreadCache* heap = NULL;
3126 {
3127 SpinLockHolder h(&pageheap_lock);
3128
3129 #if OS(WINDOWS)
3130 DWORD me;
3131 if (!tsd_inited) {
3132 me = 0;
3133 } else {
3134 me = GetCurrentThreadId();
3135 }
3136 #else
3137 // Early on in glibc's life, we cannot even call pthread_self()
3138 pthread_t me;
3139 if (!tsd_inited) {
3140 memset(&me, 0, sizeof(me));
3141 } else {
3142 me = pthread_self();
3143 }
3144 #endif
3145
3146 // This may be a recursive malloc call from pthread_setspecific()
3147 // In that case, the heap for this thread has already been created
3148 // and added to the linked list. So we search for that first.
3149 for (TCMalloc_ThreadCache* h = thread_heaps; h != NULL; h = h->next_) {
3150 #if OS(WINDOWS)
3151 if (h->tid_ == me) {
3152 #else
3153 if (pthread_equal(h->tid_, me)) {
3154 #endif
3155 heap = h;
3156 break;
3157 }
3158 }
3159
3160 if (heap == NULL) heap = NewHeap(me);
3161 }
3162
3163 // We call pthread_setspecific() outside the lock because it may
3164 // call malloc() recursively. The recursive call will never get
3165 // here again because it will find the already allocated heap in the
3166 // linked list of heaps.
3167 if (!heap->in_setspecific_ && tsd_inited) {
3168 heap->in_setspecific_ = true;
3169 setThreadHeap(heap);
3170 }
3171 return heap;
3172 }
3173
3174 void TCMalloc_ThreadCache::BecomeIdle() {
3175 if (!tsd_inited) return; // No caches yet
3176 TCMalloc_ThreadCache* heap = GetThreadHeap();
3177 if (heap == NULL) return; // No thread cache to remove
3178 if (heap->in_setspecific_) return; // Do not disturb the active caller
3179
3180 heap->in_setspecific_ = true;
3181 setThreadHeap(NULL);
3182 #ifdef HAVE_TLS
3183 // Also update the copy in __thread
3184 threadlocal_heap = NULL;
3185 #endif
3186 heap->in_setspecific_ = false;
3187 if (GetThreadHeap() == heap) {
3188 // Somehow heap got reinstated by a recursive call to malloc
3189 // from pthread_setspecific. We give up in this case.
3190 return;
3191 }
3192
3193 // We can now get rid of the heap
3194 DeleteCache(heap);
3195 }
3196
3197 void TCMalloc_ThreadCache::DestroyThreadCache(void* ptr) {
3198 // Note that "ptr" cannot be NULL since pthread promises not
3199 // to invoke the destructor on NULL values, but for safety,
3200 // we check anyway.
3201 if (ptr == NULL) return;
3202 #ifdef HAVE_TLS
3203 // Prevent fast path of GetThreadHeap() from returning heap.
3204 threadlocal_heap = NULL;
3205 #endif
3206 DeleteCache(reinterpret_cast<TCMalloc_ThreadCache*>(ptr));
3207 }
3208
3209 void TCMalloc_ThreadCache::DeleteCache(TCMalloc_ThreadCache* heap) {
3210 // Remove all memory from heap
3211 heap->Cleanup();
3212
3213 // Remove from linked list
3214 SpinLockHolder h(&pageheap_lock);
3215 if (heap->next_ != NULL) heap->next_->prev_ = heap->prev_;
3216 if (heap->prev_ != NULL) heap->prev_->next_ = heap->next_;
3217 if (thread_heaps == heap) thread_heaps = heap->next_;
3218 thread_heap_count--;
3219 RecomputeThreadCacheSize();
3220
3221 threadheap_allocator.Delete(heap);
3222 }
3223
3224 void TCMalloc_ThreadCache::RecomputeThreadCacheSize() {
3225 // Divide available space across threads
3226 int n = thread_heap_count > 0 ? thread_heap_count : 1;
3227 size_t space = overall_thread_cache_size / n;
3228
3229 // Limit to allowed range
3230 if (space < kMinThreadCacheSize) space = kMinThreadCacheSize;
3231 if (space > kMaxThreadCacheSize) space = kMaxThreadCacheSize;
3232
3233 per_thread_cache_size = space;
3234 }
3235
3236 void TCMalloc_ThreadCache::Print() const {
3237 for (size_t cl = 0; cl < kNumClasses; ++cl) {
3238 MESSAGE(" %5" PRIuS " : %4d len; %4d lo\n",
3239 ByteSizeForClass(cl),
3240 list_[cl].length(),
3241 list_[cl].lowwatermark());
3242 }
3243 }
3244
3245 // Extract interesting stats
3246 struct TCMallocStats {
3247 uint64_t system_bytes; // Bytes alloced from system
3248 uint64_t thread_bytes; // Bytes in thread caches
3249 uint64_t central_bytes; // Bytes in central cache
3250 uint64_t transfer_bytes; // Bytes in central transfer cache
3251 uint64_t pageheap_bytes; // Bytes in page heap
3252 uint64_t metadata_bytes; // Bytes alloced for metadata
3253 };
3254
3255 #ifndef WTF_CHANGES
3256 // Get stats into "r". Also get per-size-class counts if class_count != NULL
3257 static void ExtractStats(TCMallocStats* r, uint64_t* class_count) {
3258 r->central_bytes = 0;
3259 r->transfer_bytes = 0;
3260 for (int cl = 0; cl < kNumClasses; ++cl) {
3261 const int length = central_cache[cl].length();
3262 const int tc_length = central_cache[cl].tc_length();
3263 r->central_bytes += static_cast<uint64_t>(ByteSizeForClass(cl)) * length;
3264 r->transfer_bytes +=
3265 static_cast<uint64_t>(ByteSizeForClass(cl)) * tc_length;
3266 if (class_count) class_count[cl] = length + tc_length;
3267 }
3268
3269 // Add stats from per-thread heaps
3270 r->thread_bytes = 0;
3271 { // scope
3272 SpinLockHolder h(&pageheap_lock);
3273 for (TCMalloc_ThreadCache* h = thread_heaps; h != NULL; h = h->next_) {
3274 r->thread_bytes += h->Size();
3275 if (class_count) {
3276 for (size_t cl = 0; cl < kNumClasses; ++cl) {
3277 class_count[cl] += h->freelist_length(cl);
3278 }
3279 }
3280 }
3281 }
3282
3283 { //scope
3284 SpinLockHolder h(&pageheap_lock);
3285 r->system_bytes = pageheap->SystemBytes();
3286 r->metadata_bytes = metadata_system_bytes;
3287 r->pageheap_bytes = pageheap->FreeBytes();
3288 }
3289 }
3290 #endif
3291
3292 #ifndef WTF_CHANGES
3293 // WRITE stats to "out"
3294 static void DumpStats(TCMalloc_Printer* out, int level) {
3295 TCMallocStats stats;
3296 uint64_t class_count[kNumClasses];
3297 ExtractStats(&stats, (level >= 2 ? class_count : NULL));
3298
3299 if (level >= 2) {
3300 out->printf("------------------------------------------------\n");
3301 uint64_t cumulative = 0;
3302 for (int cl = 0; cl < kNumClasses; ++cl) {
3303 if (class_count[cl] > 0) {
3304 uint64_t class_bytes = class_count[cl] * ByteSizeForClass(cl);
3305 cumulative += class_bytes;
3306 out->printf("class %3d [ %8" PRIuS " bytes ] : "
3307 "%8" PRIu64 " objs; %5.1f MB; %5.1f cum MB\n",
3308 cl, ByteSizeForClass(cl),
3309 class_count[cl],
3310 class_bytes / 1048576.0,
3311 cumulative / 1048576.0);
3312 }
3313 }
3314
3315 SpinLockHolder h(&pageheap_lock);
3316 pageheap->Dump(out);
3317 }
3318
3319 const uint64_t bytes_in_use = stats.system_bytes
3320 - stats.pageheap_bytes
3321 - stats.central_bytes
3322 - stats.transfer_bytes
3323 - stats.thread_bytes;
3324
3325 out->printf("------------------------------------------------\n"
3326 "MALLOC: %12" PRIu64 " Heap size\n"
3327 "MALLOC: %12" PRIu64 " Bytes in use by application\n"
3328 "MALLOC: %12" PRIu64 " Bytes free in page heap\n"
3329 "MALLOC: %12" PRIu64 " Bytes free in central cache\n"
3330 "MALLOC: %12" PRIu64 " Bytes free in transfer cache\n"
3331 "MALLOC: %12" PRIu64 " Bytes free in thread caches\n"
3332 "MALLOC: %12" PRIu64 " Spans in use\n"
3333 "MALLOC: %12" PRIu64 " Thread heaps in use\n"
3334 "MALLOC: %12" PRIu64 " Metadata allocated\n"
3335 "------------------------------------------------\n",
3336 stats.system_bytes,
3337 bytes_in_use,
3338 stats.pageheap_bytes,
3339 stats.central_bytes,
3340 stats.transfer_bytes,
3341 stats.thread_bytes,
3342 uint64_t(span_allocator.inuse()),
3343 uint64_t(threadheap_allocator.inuse()),
3344 stats.metadata_bytes);
3345 }
3346
3347 static void PrintStats(int level) {
3348 const int kBufferSize = 16 << 10;
3349 char* buffer = new char[kBufferSize];
3350 TCMalloc_Printer printer(buffer, kBufferSize);
3351 DumpStats(&printer, level);
3352 write(STDERR_FILENO, buffer, strlen(buffer));
3353 delete[] buffer;
3354 }
3355
3356 static void** DumpStackTraces() {
3357 // Count how much space we need
3358 int needed_slots = 0;
3359 {
3360 SpinLockHolder h(&pageheap_lock);
3361 for (Span* s = sampled_objects.next; s != &sampled_objects; s = s->next) {
3362 StackTrace* stack = reinterpret_cast<StackTrace*>(s->objects);
3363 needed_slots += 3 + stack->depth;
3364 }
3365 needed_slots += 100; // Slop in case sample grows
3366 needed_slots += needed_slots/8; // An extra 12.5% slop
3367 }
3368
3369 void** result = new void*[needed_slots];
3370 if (result == NULL) {
3371 MESSAGE("tcmalloc: could not allocate %d slots for stack traces\n",
3372 needed_slots);
3373 return NULL;
3374 }
3375
3376 SpinLockHolder h(&pageheap_lock);
3377 int used_slots = 0;
3378 for (Span* s = sampled_objects.next; s != &sampled_objects; s = s->next) {
3379 ASSERT(used_slots < needed_slots); // Need to leave room for terminator
3380 StackTrace* stack = reinterpret_cast<StackTrace*>(s->objects);
3381 if (used_slots + 3 + stack->depth >= needed_slots) {
3382 // No more room
3383 break;
3384 }
3385
3386 result[used_slots+0] = reinterpret_cast<void*>(static_cast<uintptr_t>(1));
3387 result[used_slots+1] = reinterpret_cast<void*>(stack->size);
3388 result[used_slots+2] = reinterpret_cast<void*>(stack->depth);
3389 for (int d = 0; d < stack->depth; d++) {
3390 result[used_slots+3+d] = stack->stack[d];
3391 }
3392 used_slots += 3 + stack->depth;
3393 }
3394 result[used_slots] = reinterpret_cast<void*>(static_cast<uintptr_t>(0));
3395 return result;
3396 }
3397 #endif
3398
3399 #ifndef WTF_CHANGES
3400
3401 // TCMalloc's support for extra malloc interfaces
3402 class TCMallocImplementation : public MallocExtension {
3403 public:
3404 virtual void GetStats(char* buffer, int buffer_length) {
3405 ASSERT(buffer_length > 0);
3406 TCMalloc_Printer printer(buffer, buffer_length);
3407
3408 // Print level one stats unless lots of space is available
3409 if (buffer_length < 10000) {
3410 DumpStats(&printer, 1);
3411 } else {
3412 DumpStats(&printer, 2);
3413 }
3414 }
3415
3416 virtual void** ReadStackTraces() {
3417 return DumpStackTraces();
3418 }
3419
3420 virtual bool GetNumericProperty(const char* name, size_t* value) {
3421 ASSERT(name != NULL);
3422
3423 if (strcmp(name, "generic.current_allocated_bytes") == 0) {
3424 TCMallocStats stats;
3425 ExtractStats(&stats, NULL);
3426 *value = stats.system_bytes
3427 - stats.thread_bytes
3428 - stats.central_bytes
3429 - stats.pageheap_bytes;
3430 return true;
3431 }
3432
3433 if (strcmp(name, "generic.heap_size") == 0) {
3434 TCMallocStats stats;
3435 ExtractStats(&stats, NULL);
3436 *value = stats.system_bytes;
3437 return true;
3438 }
3439
3440 if (strcmp(name, "tcmalloc.slack_bytes") == 0) {
3441 // We assume that bytes in the page heap are not fragmented too
3442 // badly, and are therefore available for allocation.
3443 SpinLockHolder l(&pageheap_lock);
3444 *value = pageheap->FreeBytes();
3445 return true;
3446 }
3447
3448 if (strcmp(name, "tcmalloc.max_total_thread_cache_bytes") == 0) {
3449 SpinLockHolder l(&pageheap_lock);
3450 *value = overall_thread_cache_size;
3451 return true;
3452 }
3453
3454 if (strcmp(name, "tcmalloc.current_total_thread_cache_bytes") == 0) {
3455 TCMallocStats stats;
3456 ExtractStats(&stats, NULL);
3457 *value = stats.thread_bytes;
3458 return true;
3459 }
3460
3461 return false;
3462 }
3463
3464 virtual bool SetNumericProperty(const char* name, size_t value) {
3465 ASSERT(name != NULL);
3466
3467 if (strcmp(name, "tcmalloc.max_total_thread_cache_bytes") == 0) {
3468 // Clip the value to a reasonable range
3469 if (value < kMinThreadCacheSize) value = kMinThreadCacheSize;
3470 if (value > (1<<30)) value = (1<<30); // Limit to 1GB
3471
3472 SpinLockHolder l(&pageheap_lock);
3473 overall_thread_cache_size = static_cast<size_t>(value);
3474 TCMalloc_ThreadCache::RecomputeThreadCacheSize();
3475 return true;
3476 }
3477
3478 return false;
3479 }
3480
3481 virtual void MarkThreadIdle() {
3482 TCMalloc_ThreadCache::BecomeIdle();
3483 }
3484
3485 virtual void ReleaseFreeMemory() {
3486 SpinLockHolder h(&pageheap_lock);
3487 pageheap->ReleaseFreePages();
3488 }
3489 };
3490 #endif
3491
3492 // The constructor allocates an object to ensure that initialization
3493 // runs before main(), and therefore we do not have a chance to become
3494 // multi-threaded before initialization. We also create the TSD key
3495 // here. Presumably by the time this constructor runs, glibc is in
3496 // good enough shape to handle pthread_key_create().
3497 //
3498 // The constructor also takes the opportunity to tell STL to use
3499 // tcmalloc. We want to do this early, before construct time, so
3500 // all user STL allocations go through tcmalloc (which works really
3501 // well for STL).
3502 //
3503 // The destructor prints stats when the program exits.
3504 class TCMallocGuard {
3505 public:
3506
3507 TCMallocGuard() {
3508 #ifdef HAVE_TLS // this is true if the cc/ld/libc combo support TLS
3509 // Check whether the kernel also supports TLS (needs to happen at runtime)
3510 CheckIfKernelSupportsTLS();
3511 #endif
3512 #ifndef WTF_CHANGES
3513 #ifdef WIN32 // patch the windows VirtualAlloc, etc.
3514 PatchWindowsFunctions(); // defined in windows/patch_functions.cc
3515 #endif
3516 #endif
3517 free(malloc(1));
3518 TCMalloc_ThreadCache::InitTSD();
3519 free(malloc(1));
3520 #ifndef WTF_CHANGES
3521 MallocExtension::Register(new TCMallocImplementation);
3522 #endif
3523 }
3524
3525 #ifndef WTF_CHANGES
3526 ~TCMallocGuard() {
3527 const char* env = getenv("MALLOCSTATS");
3528 if (env != NULL) {
3529 int level = atoi(env);
3530 if (level < 1) level = 1;
3531 PrintStats(level);
3532 }
3533 #ifdef WIN32
3534 UnpatchWindowsFunctions();
3535 #endif
3536 }
3537 #endif
3538 };
3539
3540 #ifndef WTF_CHANGES
3541 static TCMallocGuard module_enter_exit_hook;
3542 #endif
3543
3544
3545 //-------------------------------------------------------------------
3546 // Helpers for the exported routines below
3547 //-------------------------------------------------------------------
3548
3549 #ifndef WTF_CHANGES
3550
3551 static Span* DoSampledAllocation(size_t size) {
3552
3553 // Grab the stack trace outside the heap lock
3554 StackTrace tmp;
3555 tmp.depth = GetStackTrace(tmp.stack, kMaxStackDepth, 1);
3556 tmp.size = size;
3557
3558 SpinLockHolder h(&pageheap_lock);
3559 // Allocate span
3560 Span *span = pageheap->New(pages(size == 0 ? 1 : size));
3561 if (span == NULL) {
3562 return NULL;
3563 }
3564
3565 // Allocate stack trace
3566 StackTrace *stack = stacktrace_allocator.New();
3567 if (stack == NULL) {
3568 // Sampling failed because of lack of memory
3569 return span;
3570 }
3571
3572 *stack = tmp;
3573 span->sample = 1;
3574 span->objects = stack;
3575 DLL_Prepend(&sampled_objects, span);
3576
3577 return span;
3578 }
3579 #endif
3580
3581 static inline bool CheckCachedSizeClass(void *ptr) {
3582 PageID p = reinterpret_cast<uintptr_t>(ptr) >> kPageShift;
3583 size_t cached_value = pageheap->GetSizeClassIfCached(p);
3584 return cached_value == 0 ||
3585 cached_value == pageheap->GetDescriptor(p)->sizeclass;
3586 }
3587
3588 static inline void* CheckedMallocResult(void *result)
3589 {
3590 ASSERT(result == 0 || CheckCachedSizeClass(result));
3591 return result;
3592 }
3593
3594 static inline void* SpanToMallocResult(Span *span) {
3595 ASSERT_SPAN_COMMITTED(span);
3596 pageheap->CacheSizeClass(span->start, 0);
3597 return
3598 CheckedMallocResult(reinterpret_cast<void*>(span->start << kPageShift));
3599 }
3600
3601 #ifdef WTF_CHANGES
3602 template <bool crashOnFailure>
3603 #endif
3604 static ALWAYS_INLINE void* do_malloc(size_t size) {
3605 void* ret = NULL;
3606
3607 #ifdef WTF_CHANGES
3608 ASSERT(!isForbidden());
3609 #endif
3610
3611 // The following call forces module initialization
3612 TCMalloc_ThreadCache* heap = TCMalloc_ThreadCache::GetCache();
3613 #ifndef WTF_CHANGES
3614 if ((FLAGS_tcmalloc_sample_parameter > 0) && heap->SampleAllocation(size)) {
3615 Span* span = DoSampledAllocation(size);
3616 if (span != NULL) {
3617 ret = SpanToMallocResult(span);
3618 }
3619 } else
3620 #endif
3621 if (size > kMaxSize) {
3622 // Use page-level allocator
3623 SpinLockHolder h(&pageheap_lock);
3624 Span* span = pageheap->New(pages(size));
3625 if (span != NULL) {
3626 ret = SpanToMallocResult(span);
3627 }
3628 } else {
3629 // The common case, and also the simplest. This just pops the
3630 // size-appropriate freelist, afer replenishing it if it's empty.
3631 ret = CheckedMallocResult(heap->Allocate(size));
3632 }
3633 if (!ret) {
3634 #ifdef WTF_CHANGES
3635 if (crashOnFailure) // This branch should be optimized out by the compiler.
3636 CRASH();
3637 #else
3638 errno = ENOMEM;
3639 #endif
3640 }
3641 return ret;
3642 }
3643
3644 static ALWAYS_INLINE void do_free(void* ptr) {
3645 if (ptr == NULL) return;
3646 ASSERT(pageheap != NULL); // Should not call free() before malloc()
3647 const PageID p = reinterpret_cast<uintptr_t>(ptr) >> kPageShift;
3648 Span* span = NULL;
3649 size_t cl = pageheap->GetSizeClassIfCached(p);
3650
3651 if (cl == 0) {
3652 span = pageheap->GetDescriptor(p);
3653 cl = span->sizeclass;
3654 pageheap->CacheSizeClass(p, cl);
3655 }
3656 if (cl != 0) {
3657 #ifndef NO_TCMALLOC_SAMPLES
3658 ASSERT(!pageheap->GetDescriptor(p)->sample);
3659 #endif
3660 TCMalloc_ThreadCache* heap = TCMalloc_ThreadCache::GetCacheIfPresent();
3661 if (heap != NULL) {
3662 heap->Deallocate(ptr, cl);
3663 } else {
3664 // Delete directly into central cache
3665 SLL_SetNext(ptr, NULL);
3666 central_cache[cl].InsertRange(ptr, ptr, 1);
3667 }
3668 } else {
3669 SpinLockHolder h(&pageheap_lock);
3670 ASSERT(reinterpret_cast<uintptr_t>(ptr) % kPageSize == 0);
3671 ASSERT(span != NULL && span->start == p);
3672 #ifndef NO_TCMALLOC_SAMPLES
3673 if (span->sample) {
3674 DLL_Remove(span);
3675 stacktrace_allocator.Delete(reinterpret_cast<StackTrace*>(span->objects));
3676 span->objects = NULL;
3677 }
3678 #endif
3679 pageheap->Delete(span);
3680 }
3681 }
3682
3683 #ifndef WTF_CHANGES
3684 // For use by exported routines below that want specific alignments
3685 //
3686 // Note: this code can be slow, and can significantly fragment memory.
3687 // The expectation is that memalign/posix_memalign/valloc/pvalloc will
3688 // not be invoked very often. This requirement simplifies our
3689 // implementation and allows us to tune for expected allocation
3690 // patterns.
3691 static void* do_memalign(size_t align, size_t size) {
3692 ASSERT((align & (align - 1)) == 0);
3693 ASSERT(align > 0);
3694 if (pageheap == NULL) TCMalloc_ThreadCache::InitModule();
3695
3696 // Allocate at least one byte to avoid boundary conditions below
3697 if (size == 0) size = 1;
3698
3699 if (size <= kMaxSize && align < kPageSize) {
3700 // Search through acceptable size classes looking for one with
3701 // enough alignment. This depends on the fact that
3702 // InitSizeClasses() currently produces several size classes that
3703 // are aligned at powers of two. We will waste time and space if
3704 // we miss in the size class array, but that is deemed acceptable
3705 // since memalign() should be used rarely.
3706 size_t cl = SizeClass(size);
3707 while (cl < kNumClasses && ((class_to_size[cl] & (align - 1)) != 0)) {
3708 cl++;
3709 }
3710 if (cl < kNumClasses) {
3711 TCMalloc_ThreadCache* heap = TCMalloc_ThreadCache::GetCache();
3712 return CheckedMallocResult(heap->Allocate(class_to_size[cl]));
3713 }
3714 }
3715
3716 // We will allocate directly from the page heap
3717 SpinLockHolder h(&pageheap_lock);
3718
3719 if (align <= kPageSize) {
3720 // Any page-level allocation will be fine
3721 // TODO: We could put the rest of this page in the appropriate
3722 // TODO: cache but it does not seem worth it.
3723 Span* span = pageheap->New(pages(size));
3724 return span == NULL ? NULL : SpanToMallocResult(span);
3725 }
3726
3727 // Allocate extra pages and carve off an aligned portion
3728 const Length alloc = pages(size + align);
3729 Span* span = pageheap->New(alloc);
3730 if (span == NULL) return NULL;
3731
3732 // Skip starting portion so that we end up aligned
3733 Length skip = 0;
3734 while ((((span->start+skip) << kPageShift) & (align - 1)) != 0) {
3735 skip++;
3736 }
3737 ASSERT(skip < alloc);
3738 if (skip > 0) {
3739 Span* rest = pageheap->Split(span, skip);
3740 pageheap->Delete(span);
3741 span = rest;
3742 }
3743
3744 // Skip trailing portion that we do not need to return
3745 const Length needed = pages(size);
3746 ASSERT(span->length >= needed);
3747 if (span->length > needed) {
3748 Span* trailer = pageheap->Split(span, needed);
3749 pageheap->Delete(trailer);
3750 }
3751 return SpanToMallocResult(span);
3752 }
3753 #endif
3754
3755 // Helpers for use by exported routines below:
3756
3757 #ifndef WTF_CHANGES
3758 static inline void do_malloc_stats() {
3759 PrintStats(1);
3760 }
3761 #endif
3762
3763 static inline int do_mallopt(int, int) {
3764 return 1; // Indicates error
3765 }
3766
3767 #ifdef HAVE_STRUCT_MALLINFO // mallinfo isn't defined on freebsd, for instance
3768 static inline struct mallinfo do_mallinfo() {
3769 TCMallocStats stats;
3770 ExtractStats(&stats, NULL);
3771
3772 // Just some of the fields are filled in.
3773 struct mallinfo info;
3774 memset(&info, 0, sizeof(info));
3775
3776 // Unfortunately, the struct contains "int" field, so some of the
3777 // size values will be truncated.
3778 info.arena = static_cast<int>(stats.system_bytes);
3779 info.fsmblks = static_cast<int>(stats.thread_bytes
3780 + stats.central_bytes
3781 + stats.transfer_bytes);
3782 info.fordblks = static_cast<int>(stats.pageheap_bytes);
3783 info.uordblks = static_cast<int>(stats.system_bytes
3784 - stats.thread_bytes
3785 - stats.central_bytes
3786 - stats.transfer_bytes
3787 - stats.pageheap_bytes);
3788
3789 return info;
3790 }
3791 #endif
3792
3793 //-------------------------------------------------------------------
3794 // Exported routines
3795 //-------------------------------------------------------------------
3796
3797 // CAVEAT: The code structure below ensures that MallocHook methods are always
3798 // called from the stack frame of the invoked allocation function.
3799 // heap-checker.cc depends on this to start a stack trace from
3800 // the call to the (de)allocation function.
3801
3802 #ifndef WTF_CHANGES
3803 extern "C"
3804 #else
3805 #define do_malloc do_malloc<crashOnFailure>
3806
3807 template <bool crashOnFailure>
3808 ALWAYS_INLINE void* malloc(size_t);
3809
3810 void* fastMalloc(size_t size)
3811 {
3812 return malloc<true>(size);
3813 }
3814
3815 TryMallocReturnValue tryFastMalloc(size_t size)
3816 {
3817 return malloc<false>(size);
3818 }
3819
3820 template <bool crashOnFailure>
3821 ALWAYS_INLINE
3822 #endif
3823 void* malloc(size_t size) {
3824 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
3825 if (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= size) // If overflow would occur...
3826 return 0;
3827 size += sizeof(AllocAlignmentInteger);
3828 void* result = do_malloc(size);
3829 if (!result)
3830 return 0;
3831
3832 *static_cast<AllocAlignmentInteger*>(result) = Internal::AllocTypeMalloc;
3833 result = static_cast<AllocAlignmentInteger*>(result) + 1;
3834 #else
3835 void* result = do_malloc(size);
3836 #endif
3837
3838 #ifndef WTF_CHANGES
3839 MallocHook::InvokeNewHook(result, size);
3840 #endif
3841 return result;
3842 }
3843
3844 #ifndef WTF_CHANGES
3845 extern "C"
3846 #endif
3847 void free(void* ptr) {
3848 #ifndef WTF_CHANGES
3849 MallocHook::InvokeDeleteHook(ptr);
3850 #endif
3851
3852 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
3853 if (!ptr)
3854 return;
3855
3856 AllocAlignmentInteger* header = Internal::fastMallocMatchValidationValue(ptr);
3857 if (*header != Internal::AllocTypeMalloc)
3858 Internal::fastMallocMatchFailed(ptr);
3859 do_free(header);
3860 #else
3861 do_free(ptr);
3862 #endif
3863 }
3864
3865 #ifndef WTF_CHANGES
3866 extern "C"
3867 #else
3868 template <bool crashOnFailure>
3869 ALWAYS_INLINE void* calloc(size_t, size_t);
3870
3871 void* fastCalloc(size_t n, size_t elem_size)
3872 {
3873 return calloc<true>(n, elem_size);
3874 }
3875
3876 TryMallocReturnValue tryFastCalloc(size_t n, size_t elem_size)
3877 {
3878 return calloc<false>(n, elem_size);
3879 }
3880
3881 template <bool crashOnFailure>
3882 ALWAYS_INLINE
3883 #endif
3884 void* calloc(size_t n, size_t elem_size) {
3885 size_t totalBytes = n * elem_size;
3886
3887 // Protect against overflow
3888 if (n > 1 && elem_size && (totalBytes / elem_size) != n)
3889 return 0;
3890
3891 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
3892 if (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= totalBytes) // If overflow would occur...
3893 return 0;
3894
3895 totalBytes += sizeof(AllocAlignmentInteger);
3896 void* result = do_malloc(totalBytes);
3897 if (!result)
3898 return 0;
3899
3900 memset(result, 0, totalBytes);
3901 *static_cast<AllocAlignmentInteger*>(result) = Internal::AllocTypeMalloc;
3902 result = static_cast<AllocAlignmentInteger*>(result) + 1;
3903 #else
3904 void* result = do_malloc(totalBytes);
3905 if (result != NULL) {
3906 memset(result, 0, totalBytes);
3907 }
3908 #endif
3909
3910 #ifndef WTF_CHANGES
3911 MallocHook::InvokeNewHook(result, totalBytes);
3912 #endif
3913 return result;
3914 }
3915
3916 // Since cfree isn't used anywhere, we don't compile it in.
3917 #ifndef WTF_CHANGES
3918 #ifndef WTF_CHANGES
3919 extern "C"
3920 #endif
3921 void cfree(void* ptr) {
3922 #ifndef WTF_CHANGES
3923 MallocHook::InvokeDeleteHook(ptr);
3924 #endif
3925 do_free(ptr);
3926 }
3927 #endif
3928
3929 #ifndef WTF_CHANGES
3930 extern "C"
3931 #else
3932 template <bool crashOnFailure>
3933 ALWAYS_INLINE void* realloc(void*, size_t);
3934
3935 void* fastRealloc(void* old_ptr, size_t new_size)
3936 {
3937 return realloc<true>(old_ptr, new_size);
3938 }
3939
3940 TryMallocReturnValue tryFastRealloc(void* old_ptr, size_t new_size)
3941 {
3942 return realloc<false>(old_ptr, new_size);
3943 }
3944
3945 template <bool crashOnFailure>
3946 ALWAYS_INLINE
3947 #endif
3948 void* realloc(void* old_ptr, size_t new_size) {
3949 if (old_ptr == NULL) {
3950 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
3951 void* result = malloc(new_size);
3952 #else
3953 void* result = do_malloc(new_size);
3954 #ifndef WTF_CHANGES
3955 MallocHook::InvokeNewHook(result, new_size);
3956 #endif
3957 #endif
3958 return result;
3959 }
3960 if (new_size == 0) {
3961 #ifndef WTF_CHANGES
3962 MallocHook::InvokeDeleteHook(old_ptr);
3963 #endif
3964 free(old_ptr);
3965 return NULL;
3966 }
3967
3968 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
3969 if (std::numeric_limits<size_t>::max() - sizeof(AllocAlignmentInteger) <= new_size) // If overflow would occur...
3970 return 0;
3971 new_size += sizeof(AllocAlignmentInteger);
3972 AllocAlignmentInteger* header = Internal::fastMallocMatchValidationValue(old_ptr);
3973 if (*header != Internal::AllocTypeMalloc)
3974 Internal::fastMallocMatchFailed(old_ptr);
3975 old_ptr = header;
3976 #endif
3977
3978 // Get the size of the old entry
3979 const PageID p = reinterpret_cast<uintptr_t>(old_ptr) >> kPageShift;
3980 size_t cl = pageheap->GetSizeClassIfCached(p);
3981 Span *span = NULL;
3982 size_t old_size;
3983 if (cl == 0) {
3984 span = pageheap->GetDescriptor(p);
3985 cl = span->sizeclass;
3986 pageheap->CacheSizeClass(p, cl);
3987 }
3988 if (cl != 0) {
3989 old_size = ByteSizeForClass(cl);
3990 } else {
3991 ASSERT(span != NULL);
3992 old_size = span->length << kPageShift;
3993 }
3994
3995 // Reallocate if the new size is larger than the old size,
3996 // or if the new size is significantly smaller than the old size.
3997 if ((new_size > old_size) || (AllocationSize(new_size) < old_size)) {
3998 // Need to reallocate
3999 void* new_ptr = do_malloc(new_size);
4000 if (new_ptr == NULL) {
4001 return NULL;
4002 }
4003 #ifndef WTF_CHANGES
4004 MallocHook::InvokeNewHook(new_ptr, new_size);
4005 #endif
4006 memcpy(new_ptr, old_ptr, ((old_size < new_size) ? old_size : new_size));
4007 #ifndef WTF_CHANGES
4008 MallocHook::InvokeDeleteHook(old_ptr);
4009 #endif
4010 // We could use a variant of do_free() that leverages the fact
4011 // that we already know the sizeclass of old_ptr. The benefit
4012 // would be small, so don't bother.
4013 do_free(old_ptr);
4014 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
4015 new_ptr = static_cast<AllocAlignmentInteger*>(new_ptr) + 1;
4016 #endif
4017 return new_ptr;
4018 } else {
4019 #if ENABLE(FAST_MALLOC_MATCH_VALIDATION)
4020 old_ptr = static_cast<AllocAlignmentInteger*>(old_ptr) + 1; // Set old_ptr back to the user pointer.
4021 #endif
4022 return old_ptr;
4023 }
4024 }
4025
4026 #ifdef WTF_CHANGES
4027 #undef do_malloc
4028 #else
4029
4030 static SpinLock set_new_handler_lock = SPINLOCK_INITIALIZER;
4031
4032 static inline void* cpp_alloc(size_t size, bool nothrow) {
4033 for (;;) {
4034 void* p = do_malloc(size);
4035 #ifdef PREANSINEW
4036 return p;
4037 #else
4038 if (p == NULL) { // allocation failed
4039 // Get the current new handler. NB: this function is not
4040 // thread-safe. We make a feeble stab at making it so here, but
4041 // this lock only protects against tcmalloc interfering with
4042 // itself, not with other libraries calling set_new_handler.
4043 std::new_handler nh;
4044 {
4045 SpinLockHolder h(&set_new_handler_lock);
4046 nh = std::set_new_handler(0);
4047 (void) std::set_new_handler(nh);
4048 }
4049 // If no new_handler is established, the allocation failed.
4050 if (!nh) {
4051 if (nothrow) return 0;
4052 throw std::bad_alloc();
4053 }
4054 // Otherwise, try the new_handler. If it returns, retry the
4055 // allocation. If it throws std::bad_alloc, fail the allocation.
4056 // if it throws something else, don't interfere.
4057 try {
4058 (*nh)();
4059 } catch (const std::bad_alloc&) {
4060 if (!nothrow) throw;
4061 return p;
4062 }
4063 } else { // allocation success
4064 return p;
4065 }
4066 #endif
4067 }
4068 }
4069
4070 #if ENABLE(GLOBAL_FASTMALLOC_NEW)
4071
4072 void* operator new(size_t size) {
4073 void* p = cpp_alloc(size, false);
4074 // We keep this next instruction out of cpp_alloc for a reason: when
4075 // it's in, and new just calls cpp_alloc, the optimizer may fold the
4076 // new call into cpp_alloc, which messes up our whole section-based
4077 // stacktracing (see ATTRIBUTE_SECTION, above). This ensures cpp_alloc
4078 // isn't the last thing this fn calls, and prevents the folding.
4079 MallocHook::InvokeNewHook(p, size);
4080 return p;
4081 }
4082
4083 void* operator new(size_t size, const std::nothrow_t&) __THROW {
4084 void* p = cpp_alloc(size, true);
4085 MallocHook::InvokeNewHook(p, size);
4086 return p;
4087 }
4088
4089 void operator delete(void* p) __THROW {
4090 MallocHook::InvokeDeleteHook(p);
4091 do_free(p);
4092 }
4093
4094 void operator delete(void* p, const std::nothrow_t&) __THROW {
4095 MallocHook::InvokeDeleteHook(p);
4096 do_free(p);
4097 }
4098
4099 void* operator new[](size_t size) {
4100 void* p = cpp_alloc(size, false);
4101 // We keep this next instruction out of cpp_alloc for a reason: when
4102 // it's in, and new just calls cpp_alloc, the optimizer may fold the
4103 // new call into cpp_alloc, which messes up our whole section-based
4104 // stacktracing (see ATTRIBUTE_SECTION, above). This ensures cpp_alloc
4105 // isn't the last thing this fn calls, and prevents the folding.
4106 MallocHook::InvokeNewHook(p, size);
4107 return p;
4108 }
4109
4110 void* operator new[](size_t size, const std::nothrow_t&) __THROW {
4111 void* p = cpp_alloc(size, true);
4112 MallocHook::InvokeNewHook(p, size);
4113 return p;
4114 }
4115
4116 void operator delete[](void* p) __THROW {
4117 MallocHook::InvokeDeleteHook(p);
4118 do_free(p);
4119 }
4120
4121 void operator delete[](void* p, const std::nothrow_t&) __THROW {
4122 MallocHook::InvokeDeleteHook(p);
4123 do_free(p);
4124 }
4125
4126 #endif
4127
4128 extern "C" void* memalign(size_t align, size_t size) __THROW {
4129 void* result = do_memalign(align, size);
4130 MallocHook::InvokeNewHook(result, size);
4131 return result;
4132 }
4133
4134 extern "C" int posix_memalign(void** result_ptr, size_t align, size_t size)
4135 __THROW {
4136 if (((align % sizeof(void*)) != 0) ||
4137 ((align & (align - 1)) != 0) ||
4138 (align == 0)) {
4139 return EINVAL;
4140 }
4141
4142 void* result = do_memalign(align, size);
4143 MallocHook::InvokeNewHook(result, size);
4144 if (result == NULL) {
4145 return ENOMEM;
4146 } else {
4147 *result_ptr = result;
4148 return 0;
4149 }
4150 }
4151
4152 static size_t pagesize = 0;
4153
4154 extern "C" void* valloc(size_t size) __THROW {
4155 // Allocate page-aligned object of length >= size bytes
4156 if (pagesize == 0) pagesize = getpagesize();
4157 void* result = do_memalign(pagesize, size);
4158 MallocHook::InvokeNewHook(result, size);
4159 return result;
4160 }
4161
4162 extern "C" void* pvalloc(size_t size) __THROW {
4163 // Round up size to a multiple of pagesize
4164 if (pagesize == 0) pagesize = getpagesize();
4165 size = (size + pagesize - 1) & ~(pagesize - 1);
4166 void* result = do_memalign(pagesize, size);
4167 MallocHook::InvokeNewHook(result, size);
4168 return result;
4169 }
4170
4171 extern "C" void malloc_stats(void) {
4172 do_malloc_stats();
4173 }
4174
4175 extern "C" int mallopt(int cmd, int value) {
4176 return do_mallopt(cmd, value);
4177 }
4178
4179 #ifdef HAVE_STRUCT_MALLINFO
4180 extern "C" struct mallinfo mallinfo(void) {
4181 return do_mallinfo();
4182 }
4183 #endif
4184
4185 //-------------------------------------------------------------------
4186 // Some library routines on RedHat 9 allocate memory using malloc()
4187 // and free it using __libc_free() (or vice-versa). Since we provide
4188 // our own implementations of malloc/free, we need to make sure that
4189 // the __libc_XXX variants (defined as part of glibc) also point to
4190 // the same implementations.
4191 //-------------------------------------------------------------------
4192
4193 #if defined(__GLIBC__)
4194 extern "C" {
4195 #if COMPILER(GCC) && !defined(__MACH__) && defined(HAVE___ATTRIBUTE__)
4196 // Potentially faster variants that use the gcc alias extension.
4197 // Mach-O (Darwin) does not support weak aliases, hence the __MACH__ check.
4198 # define ALIAS(x) __attribute__ ((weak, alias (x)))
4199 void* __libc_malloc(size_t size) ALIAS("malloc");
4200 void __libc_free(void* ptr) ALIAS("free");
4201 void* __libc_realloc(void* ptr, size_t size) ALIAS("realloc");
4202 void* __libc_calloc(size_t n, size_t size) ALIAS("calloc");
4203 void __libc_cfree(void* ptr) ALIAS("cfree");
4204 void* __libc_memalign(size_t align, size_t s) ALIAS("memalign");
4205 void* __libc_valloc(size_t size) ALIAS("valloc");
4206 void* __libc_pvalloc(size_t size) ALIAS("pvalloc");
4207 int __posix_memalign(void** r, size_t a, size_t s) ALIAS("posix_memalign");
4208 # undef ALIAS
4209 # else /* not __GNUC__ */
4210 // Portable wrappers
4211 void* __libc_malloc(size_t size) { return malloc(size); }
4212 void __libc_free(void* ptr) { free(ptr); }
4213 void* __libc_realloc(void* ptr, size_t size) { return realloc(ptr, size); }
4214 void* __libc_calloc(size_t n, size_t size) { return calloc(n, size); }
4215 void __libc_cfree(void* ptr) { cfree(ptr); }
4216 void* __libc_memalign(size_t align, size_t s) { return memalign(align, s); }
4217 void* __libc_valloc(size_t size) { return valloc(size); }
4218 void* __libc_pvalloc(size_t size) { return pvalloc(size); }
4219 int __posix_memalign(void** r, size_t a, size_t s) {
4220 return posix_memalign(r, a, s);
4221 }
4222 # endif /* __GNUC__ */
4223 }
4224 #endif /* __GLIBC__ */
4225
4226 // Override __libc_memalign in libc on linux boxes specially.
4227 // They have a bug in libc that causes them to (very rarely) allocate
4228 // with __libc_memalign() yet deallocate with free() and the
4229 // definitions above don't catch it.
4230 // This function is an exception to the rule of calling MallocHook method
4231 // from the stack frame of the allocation function;
4232 // heap-checker handles this special case explicitly.
4233 static void *MemalignOverride(size_t align, size_t size, const void *caller)
4234 __THROW {
4235 void* result = do_memalign(align, size);
4236 MallocHook::InvokeNewHook(result, size);
4237 return result;
4238 }
4239 void *(*__memalign_hook)(size_t, size_t, const void *) = MemalignOverride;
4240
4241 #endif
4242
4243 #ifdef WTF_CHANGES
4244 void releaseFastMallocFreeMemory()
4245 {
4246 // Flush free pages in the current thread cache back to the page heap.
4247 // Low watermark mechanism in Scavenge() prevents full return on the first pass.
4248 // The second pass flushes everything.
4249 if (TCMalloc_ThreadCache* threadCache = TCMalloc_ThreadCache::GetCacheIfPresent()) {
4250 threadCache->Scavenge();
4251 threadCache->Scavenge();
4252 }
4253
4254 SpinLockHolder h(&pageheap_lock);
4255 pageheap->ReleaseFreePages();
4256 }
4257
4258 FastMallocStatistics fastMallocStatistics()
4259 {
4260 FastMallocStatistics statistics;
4261
4262 SpinLockHolder lockHolder(&pageheap_lock);
4263 statistics.reservedVMBytes = static_cast<size_t>(pageheap->SystemBytes());
4264 statistics.committedVMBytes = statistics.reservedVMBytes - pageheap->ReturnedBytes();
4265
4266 statistics.freeListBytes = 0;
4267 for (unsigned cl = 0; cl < kNumClasses; ++cl) {
4268 const int length = central_cache[cl].length();
4269 const int tc_length = central_cache[cl].tc_length();
4270
4271 statistics.freeListBytes += ByteSizeForClass(cl) * (length + tc_length);
4272 }
4273 for (TCMalloc_ThreadCache* threadCache = thread_heaps; threadCache ; threadCache = threadCache->next_)
4274 statistics.freeListBytes += threadCache->Size();
4275
4276 return statistics;
4277 }
4278
4279 size_t fastMallocSize(const void* ptr)
4280 {
4281 const PageID p = reinterpret_cast<uintptr_t>(ptr) >> kPageShift;
4282 Span* span = pageheap->GetDescriptorEnsureSafe(p);
4283
4284 if (!span || span->free)
4285 return 0;
4286
4287 for (void* free = span->objects; free != NULL; free = *((void**) free)) {
4288 if (ptr == free)
4289 return 0;
4290 }
4291
4292 if (size_t cl = span->sizeclass)
4293 return ByteSizeForClass(cl);
4294
4295 return span->length << kPageShift;
4296 }
4297
4298 #if OS(DARWIN)
4299
4300 class FreeObjectFinder {
4301 const RemoteMemoryReader& m_reader;
4302 HashSet<void*> m_freeObjects;
4303
4304 public:
4305 FreeObjectFinder(const RemoteMemoryReader& reader) : m_reader(reader) { }
4306
4307 void visit(void* ptr) { m_freeObjects.add(ptr); }
4308 bool isFreeObject(void* ptr) const { return m_freeObjects.contains(ptr); }
4309 bool isFreeObject(vm_address_t ptr) const { return isFreeObject(reinterpret_cast<void*>(ptr)); }
4310 size_t freeObjectCount() const { return m_freeObjects.size(); }
4311
4312 void findFreeObjects(TCMalloc_ThreadCache* threadCache)
4313 {
4314 for (; threadCache; threadCache = (threadCache->next_ ? m_reader(threadCache->next_) : 0))
4315 threadCache->enumerateFreeObjects(*this, m_reader);
4316 }
4317
4318 void findFreeObjects(TCMalloc_Central_FreeListPadded* centralFreeList, size_t numSizes, TCMalloc_Central_FreeListPadded* remoteCentralFreeList)
4319 {
4320 for (unsigned i = 0; i < numSizes; i++)
4321 centralFreeList[i].enumerateFreeObjects(*this, m_reader, remoteCentralFreeList + i);
4322 }
4323 };
4324
4325 class PageMapFreeObjectFinder {
4326 const RemoteMemoryReader& m_reader;
4327 FreeObjectFinder& m_freeObjectFinder;
4328
4329 public:
4330 PageMapFreeObjectFinder(const RemoteMemoryReader& reader, FreeObjectFinder& freeObjectFinder)
4331 : m_reader(reader)
4332 , m_freeObjectFinder(freeObjectFinder)
4333 { }
4334
4335 int visit(void* ptr) const
4336 {
4337 if (!ptr)
4338 return 1;
4339
4340 Span* span = m_reader(reinterpret_cast<Span*>(ptr));
4341 if (!span)
4342 return 1;
4343
4344 if (span->free) {
4345 void* ptr = reinterpret_cast<void*>(span->start << kPageShift);
4346 m_freeObjectFinder.visit(ptr);
4347 } else if (span->sizeclass) {
4348 // Walk the free list of the small-object span, keeping track of each object seen
4349 for (void* nextObject = span->objects; nextObject; nextObject = m_reader.nextEntryInLinkedList(reinterpret_cast<void**>(nextObject)))
4350 m_freeObjectFinder.visit(nextObject);
4351 }
4352 return span->length;
4353 }
4354 };
4355
4356 class PageMapMemoryUsageRecorder {
4357 task_t m_task;
4358 void* m_context;
4359 unsigned m_typeMask;
4360 vm_range_recorder_t* m_recorder;
4361 const RemoteMemoryReader& m_reader;
4362 const FreeObjectFinder& m_freeObjectFinder;
4363
4364 HashSet<void*> m_seenPointers;
4365 Vector<Span*> m_coalescedSpans;
4366
4367 public:
4368 PageMapMemoryUsageRecorder(task_t task, void* context, unsigned typeMask, vm_range_recorder_t* recorder, const RemoteMemoryReader& reader, const FreeObjectFinder& freeObjectFinder)
4369 : m_task(task)
4370 , m_context(context)
4371 , m_typeMask(typeMask)
4372 , m_recorder(recorder)
4373 , m_reader(reader)
4374 , m_freeObjectFinder(freeObjectFinder)
4375 { }
4376
4377 ~PageMapMemoryUsageRecorder()
4378 {
4379 ASSERT(!m_coalescedSpans.size());
4380 }
4381
4382 void recordPendingRegions()
4383 {
4384 Span* lastSpan = m_coalescedSpans[m_coalescedSpans.size() - 1];
4385 vm_range_t ptrRange = { m_coalescedSpans[0]->start << kPageShift, 0 };
4386 ptrRange.size = (lastSpan->start << kPageShift) - ptrRange.address + (lastSpan->length * kPageSize);
4387
4388 // Mark the memory region the spans represent as a candidate for containing pointers
4389 if (m_typeMask & MALLOC_PTR_REGION_RANGE_TYPE)
4390 (*m_recorder)(m_task, m_context, MALLOC_PTR_REGION_RANGE_TYPE, &ptrRange, 1);
4391
4392 if (!(m_typeMask & MALLOC_PTR_IN_USE_RANGE_TYPE)) {
4393 m_coalescedSpans.clear();
4394 return;
4395 }
4396
4397 Vector<vm_range_t, 1024> allocatedPointers;
4398 for (size_t i = 0; i < m_coalescedSpans.size(); ++i) {
4399 Span *theSpan = m_coalescedSpans[i];
4400 if (theSpan->free)
4401 continue;
4402
4403 vm_address_t spanStartAddress = theSpan->start << kPageShift;
4404 vm_size_t spanSizeInBytes = theSpan->length * kPageSize;
4405
4406 if (!theSpan->sizeclass) {
4407 // If it's an allocated large object span, mark it as in use
4408 if (!m_freeObjectFinder.isFreeObject(spanStartAddress))
4409 allocatedPointers.append((vm_range_t){spanStartAddress, spanSizeInBytes});
4410 } else {
4411 const size_t objectSize = ByteSizeForClass(theSpan->sizeclass);
4412
4413 // Mark each allocated small object within the span as in use
4414 const vm_address_t endOfSpan = spanStartAddress + spanSizeInBytes;
4415 for (vm_address_t object = spanStartAddress; object + objectSize <= endOfSpan; object += objectSize) {
4416 if (!m_freeObjectFinder.isFreeObject(object))
4417 allocatedPointers.append((vm_range_t){object, objectSize});
4418 }
4419 }
4420 }
4421
4422 (*m_recorder)(m_task, m_context, MALLOC_PTR_IN_USE_RANGE_TYPE, allocatedPointers.data(), allocatedPointers.size());
4423
4424 m_coalescedSpans.clear();
4425 }
4426
4427 int visit(void* ptr)
4428 {
4429 if (!ptr)
4430 return 1;
4431
4432 Span* span = m_reader(reinterpret_cast<Span*>(ptr));
4433 if (!span || !span->start)
4434 return 1;
4435
4436 if (m_seenPointers.contains(ptr))
4437 return span->length;
4438 m_seenPointers.add(ptr);
4439
4440 if (!m_coalescedSpans.size()) {
4441 m_coalescedSpans.append(span);
4442 return span->length;
4443 }
4444
4445 Span* previousSpan = m_coalescedSpans[m_coalescedSpans.size() - 1];
4446 vm_address_t previousSpanStartAddress = previousSpan->start << kPageShift;
4447 vm_size_t previousSpanSizeInBytes = previousSpan->length * kPageSize;
4448
4449 // If the new span is adjacent to the previous span, do nothing for now.
4450 vm_address_t spanStartAddress = span->start << kPageShift;
4451 if (spanStartAddress == previousSpanStartAddress + previousSpanSizeInBytes) {
4452 m_coalescedSpans.append(span);
4453 return span->length;
4454 }
4455
4456 // New span is not adjacent to previous span, so record the spans coalesced so far.
4457 recordPendingRegions();
4458 m_coalescedSpans.append(span);
4459
4460 return span->length;
4461 }
4462 };
4463
4464 class AdminRegionRecorder {
4465 task_t m_task;
4466 void* m_context;
4467 unsigned m_typeMask;
4468 vm_range_recorder_t* m_recorder;
4469 const RemoteMemoryReader& m_reader;
4470
4471 Vector<vm_range_t, 1024> m_pendingRegions;
4472
4473 public:
4474 AdminRegionRecorder(task_t task, void* context, unsigned typeMask, vm_range_recorder_t* recorder, const RemoteMemoryReader& reader)
4475 : m_task(task)
4476 , m_context(context)
4477 , m_typeMask(typeMask)
4478 , m_recorder(recorder)
4479 , m_reader(reader)
4480 { }
4481
4482 void recordRegion(vm_address_t ptr, size_t size)
4483 {
4484 if (m_typeMask & MALLOC_ADMIN_REGION_RANGE_TYPE)
4485 m_pendingRegions.append((vm_range_t){ ptr, size });
4486 }
4487
4488 void visit(void *ptr, size_t size)
4489 {
4490 recordRegion(reinterpret_cast<vm_address_t>(ptr), size);
4491 }
4492
4493 void recordPendingRegions()
4494 {
4495 if (m_pendingRegions.size()) {
4496 (*m_recorder)(m_task, m_context, MALLOC_ADMIN_REGION_RANGE_TYPE, m_pendingRegions.data(), m_pendingRegions.size());
4497 m_pendingRegions.clear();
4498 }
4499 }
4500
4501 ~AdminRegionRecorder()
4502 {
4503 ASSERT(!m_pendingRegions.size());
4504 }
4505 };
4506
4507 kern_return_t FastMallocZone::enumerate(task_t task, void* context, unsigned typeMask, vm_address_t zoneAddress, memory_reader_t reader, vm_range_recorder_t recorder)
4508 {
4509 RemoteMemoryReader memoryReader(task, reader);
4510
4511 InitSizeClasses();
4512
4513 FastMallocZone* mzone = memoryReader(reinterpret_cast<FastMallocZone*>(zoneAddress));
4514 TCMalloc_PageHeap* pageHeap = memoryReader(mzone->m_pageHeap);
4515 TCMalloc_ThreadCache** threadHeapsPointer = memoryReader(mzone->m_threadHeaps);
4516 TCMalloc_ThreadCache* threadHeaps = memoryReader(*threadHeapsPointer);
4517
4518 TCMalloc_Central_FreeListPadded* centralCaches = memoryReader(mzone->m_centralCaches, sizeof(TCMalloc_Central_FreeListPadded) * kNumClasses);
4519
4520 FreeObjectFinder finder(memoryReader);
4521 finder.findFreeObjects(threadHeaps);
4522 finder.findFreeObjects(centralCaches, kNumClasses, mzone->m_centralCaches);
4523
4524 TCMalloc_PageHeap::PageMap* pageMap = &pageHeap->pagemap_;
4525 PageMapFreeObjectFinder pageMapFinder(memoryReader, finder);
4526 pageMap->visitValues(pageMapFinder, memoryReader);
4527
4528 PageMapMemoryUsageRecorder usageRecorder(task, context, typeMask, recorder, memoryReader, finder);
4529 pageMap->visitValues(usageRecorder, memoryReader);
4530 usageRecorder.recordPendingRegions();
4531
4532 AdminRegionRecorder adminRegionRecorder(task, context, typeMask, recorder, memoryReader);
4533 pageMap->visitAllocations(adminRegionRecorder, memoryReader);
4534
4535 PageHeapAllocator<Span>* spanAllocator = memoryReader(mzone->m_spanAllocator);
4536 PageHeapAllocator<TCMalloc_ThreadCache>* pageHeapAllocator = memoryReader(mzone->m_pageHeapAllocator);
4537
4538 spanAllocator->recordAdministrativeRegions(adminRegionRecorder, memoryReader);
4539 pageHeapAllocator->recordAdministrativeRegions(adminRegionRecorder, memoryReader);
4540
4541 adminRegionRecorder.recordPendingRegions();
4542
4543 return 0;
4544 }
4545
4546 size_t FastMallocZone::size(malloc_zone_t*, const void*)
4547 {
4548 return 0;
4549 }
4550
4551 void* FastMallocZone::zoneMalloc(malloc_zone_t*, size_t)
4552 {
4553 return 0;
4554 }
4555
4556 void* FastMallocZone::zoneCalloc(malloc_zone_t*, size_t, size_t)
4557 {
4558 return 0;
4559 }
4560
4561 void FastMallocZone::zoneFree(malloc_zone_t*, void* ptr)
4562 {
4563 // Due to <rdar://problem/5671357> zoneFree may be called by the system free even if the pointer
4564 // is not in this zone. When this happens, the pointer being freed was not allocated by any
4565 // zone so we need to print a useful error for the application developer.
4566 malloc_printf("*** error for object %p: pointer being freed was not allocated\n", ptr);
4567 }
4568
4569 void* FastMallocZone::zoneRealloc(malloc_zone_t*, void*, size_t)
4570 {
4571 return 0;
4572 }
4573
4574
4575 #undef malloc
4576 #undef free
4577 #undef realloc
4578 #undef calloc
4579
4580 extern "C" {
4581 malloc_introspection_t jscore_fastmalloc_introspection = { &FastMallocZone::enumerate, &FastMallocZone::goodSize, &FastMallocZone::check, &FastMallocZone::print,
4582 &FastMallocZone::log, &FastMallocZone::forceLock, &FastMallocZone::forceUnlock, &FastMallocZone::statistics
4583
4584 #if !defined(BUILDING_ON_TIGER) && !defined(BUILDING_ON_LEOPARD)
4585 , 0 // zone_locked will not be called on the zone unless it advertises itself as version five or higher.
4586 #endif
4587 #if !defined(BUILDING_ON_TIGER) && !defined(BUILDING_ON_LEOPARD) && !defined(BUILDING_ON_SNOW_LEOPARD)
4588 , 0, 0, 0, 0 // These members will not be used unless the zone advertises itself as version seven or higher.
4589 #endif
4590
4591 };
4592 }
4593
4594 FastMallocZone::FastMallocZone(TCMalloc_PageHeap* pageHeap, TCMalloc_ThreadCache** threadHeaps, TCMalloc_Central_FreeListPadded* centralCaches, PageHeapAllocator<Span>* spanAllocator, PageHeapAllocator<TCMalloc_ThreadCache>* pageHeapAllocator)
4595 : m_pageHeap(pageHeap)
4596 , m_threadHeaps(threadHeaps)
4597 , m_centralCaches(centralCaches)
4598 , m_spanAllocator(spanAllocator)
4599 , m_pageHeapAllocator(pageHeapAllocator)
4600 {
4601 memset(&m_zone, 0, sizeof(m_zone));
4602 m_zone.version = 4;
4603 m_zone.zone_name = "JavaScriptCore FastMalloc";
4604 m_zone.size = &FastMallocZone::size;
4605 m_zone.malloc = &FastMallocZone::zoneMalloc;
4606 m_zone.calloc = &FastMallocZone::zoneCalloc;
4607 m_zone.realloc = &FastMallocZone::zoneRealloc;
4608 m_zone.free = &FastMallocZone::zoneFree;
4609 m_zone.valloc = &FastMallocZone::zoneValloc;
4610 m_zone.destroy = &FastMallocZone::zoneDestroy;
4611 m_zone.introspect = &jscore_fastmalloc_introspection;
4612 malloc_zone_register(&m_zone);
4613 }
4614
4615
4616 void FastMallocZone::init()
4617 {
4618 static FastMallocZone zone(pageheap, &thread_heaps, static_cast<TCMalloc_Central_FreeListPadded*>(central_cache), &span_allocator, &threadheap_allocator);
4619 }
4620
4621 #endif // OS(DARWIN)
4622
4623 } // namespace WTF
4624 #endif // WTF_CHANGES
4625
4626 #endif // FORCE_SYSTEM_MALLOC
4627