2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 // Functions to provide smarter use of jemalloc, if jemalloc is being used.
18 // http://www.canonware.com/download/jemalloc/jemalloc-latest/doc/jemalloc.html
21 #define FOLLY_MALLOC_H_
24 * Define various MALLOCX_* macros normally provided by jemalloc. We define
25 * them so that we don't have to include jemalloc.h, in case the program is
26 * built without jemalloc support.
28 #ifndef MALLOCX_LG_ALIGN
29 #define MALLOCX_LG_ALIGN(la) (la)
32 #define MALLOCX_ZERO (static_cast<int>(0x40))
35 // If using fbstring from libstdc++ (see comment in FBString.h), then
36 // just define stub code here to typedef the fbstring type into the
38 // This provides backwards compatibility for code that explicitly
39 // includes and uses fbstring.
40 #if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
42 #include <folly/detail/Malloc.h>
43 #include <folly/portability/BitsFunctexcept.h>
48 using std::goodMallocSize;
49 using std::jemallocMinInPlaceExpandable;
50 using std::usingJEMalloc;
51 using std::smartRealloc;
52 using std::checkedMalloc;
53 using std::checkedCalloc;
54 using std::checkedRealloc;
57 #else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
59 #ifdef _LIBSTDCXX_FBSTRING
60 #pragma GCC system_header
63 * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
64 * jemalloc if we are using jemalloc, or will be NULL if we are using another
65 * malloc implementation.
67 extern "C" void* mallocx(size_t, int)
68 __attribute__((__weak__));
69 extern "C" void* rallocx(void*, size_t, int)
70 __attribute__((__weak__));
71 extern "C" size_t xallocx(void*, size_t, size_t, int)
72 __attribute__((__weak__));
73 extern "C" size_t sallocx(const void*, int)
74 __attribute__((__weak__));
75 extern "C" void dallocx(void*, int)
76 __attribute__((__weak__));
77 extern "C" void sdallocx(void*, size_t, int)
78 __attribute__((__weak__));
79 extern "C" size_t nallocx(size_t, int)
80 __attribute__((__weak__));
81 extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
82 __attribute__((__weak__));
83 extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
84 __attribute__((__weak__));
85 extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
87 __attribute__((__weak__));
89 #include <bits/functexcept.h>
91 #define FOLLY_HAVE_MALLOC_H 1
93 #else // !defined(_LIBSTDCXX_FBSTRING)
95 #include <folly/detail/Malloc.h> /* nolint */
96 #include <folly/portability/BitsFunctexcept.h> /* nolint */
100 // for malloc_usable_size
101 // NOTE: FreeBSD 9 doesn't have malloc.h. Its definitions
102 // are found in stdlib.h.
103 #if FOLLY_HAVE_MALLOC_H
117 #ifdef _LIBSTDCXX_FBSTRING
118 namespace std _GLIBCXX_VISIBILITY(default) {
119 _GLIBCXX_BEGIN_NAMESPACE_VERSION
124 // Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
125 // Disabled for nvcc because it fails on attributes on lambdas.
126 #if defined(__GNUC__) && !defined(__NVCC__)
127 #define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
129 #define FOLLY_MALLOC_NOINLINE
133 * Determine if we are using jemalloc or not.
135 inline bool usingJEMalloc() noexcept {
136 // Checking for rallocx != NULL is not sufficient; we may be in a dlopen()ed
137 // module that depends on libjemalloc, so rallocx is resolved, but the main
138 // program might be using a different memory allocator.
139 // How do we determine that we're using jemalloc? In the hackiest
140 // way possible. We allocate memory using malloc() and see if the
141 // per-thread counter of allocated memory increases. This makes me
142 // feel dirty inside. Also note that this requires jemalloc to have
143 // been compiled with --enable-stats.
144 static const bool result = [] () FOLLY_MALLOC_NOINLINE noexcept {
145 // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
146 // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
147 // (!!). http://goo.gl/xpmctm
148 if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
149 || sallocx == nullptr || dallocx == nullptr || sdallocx == nullptr
150 || nallocx == nullptr || mallctl == nullptr
151 || mallctlnametomib == nullptr || mallctlbymib == nullptr) {
155 // "volatile" because gcc optimizes out the reads from *counter, because
156 // it "knows" malloc doesn't modify global state...
157 /* nolint */ volatile uint64_t* counter;
158 size_t counterLen = sizeof(uint64_t*);
160 if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
165 if (counterLen != sizeof(uint64_t*)) {
169 uint64_t origAllocated = *counter;
171 // Static because otherwise clever compilers will find out that
172 // the ptr is not used and does not escape the scope, so they will
173 // just optimize away the malloc.
174 static void* ptr = malloc(1);
176 // wtf, failing to allocate 1 byte
180 return (origAllocated != *counter);
186 inline size_t goodMallocSize(size_t minSize) noexcept {
191 if (!usingJEMalloc()) {
192 // Not using jemalloc - no smarts
196 return nallocx(minSize, 0);
199 // We always request "good" sizes for allocation, so jemalloc can
200 // never grow in place small blocks; they're already occupied to the
201 // brim. Blocks larger than or equal to 4096 bytes can in fact be
202 // expanded in place, and this constant reflects that.
203 static const size_t jemallocMinInPlaceExpandable = 4096;
206 * Trivial wrappers around malloc, calloc, realloc that check for allocation
207 * failure and throw std::bad_alloc in that case.
209 inline void* checkedMalloc(size_t size) {
210 void* p = malloc(size);
211 if (!p) std::__throw_bad_alloc();
215 inline void* checkedCalloc(size_t n, size_t size) {
216 void* p = calloc(n, size);
217 if (!p) std::__throw_bad_alloc();
221 inline void* checkedRealloc(void* ptr, size_t size) {
222 void* p = realloc(ptr, size);
223 if (!p) std::__throw_bad_alloc();
228 * This function tries to reallocate a buffer of which only the first
229 * currentSize bytes are used. The problem with using realloc is that
230 * if currentSize is relatively small _and_ if realloc decides it
231 * needs to move the memory chunk to a new buffer, then realloc ends
232 * up copying data that is not used. It's impossible to hook into
233 * GNU's malloc to figure whether expansion will occur in-place or as
234 * a malloc-copy-free troika. (If an expand_in_place primitive would
235 * be available, smartRealloc would use it.) As things stand, this
236 * routine just tries to call realloc() (thus benefitting of potential
237 * copy-free coalescing) unless there's too much slack memory.
239 inline void* smartRealloc(void* p,
240 const size_t currentSize,
241 const size_t currentCapacity,
242 const size_t newCapacity) {
244 assert(currentSize <= currentCapacity &&
245 currentCapacity < newCapacity);
247 if (usingJEMalloc()) {
248 // using jemalloc's API. Don't forget that jemalloc can never grow
249 // in place blocks smaller than 4096 bytes.
251 // NB: newCapacity may not be precisely equal to a jemalloc size class,
252 // i.e. newCapacity is not guaranteed to be the result of a
253 // goodMallocSize() call, therefore xallocx() may return more than
254 // newCapacity bytes of space. Use >= rather than == to check whether
255 // xallocx() successfully expanded in place.
256 if (currentCapacity >= jemallocMinInPlaceExpandable &&
257 xallocx(p, newCapacity, 0, 0) >= newCapacity) {
258 // Managed to expand in place
261 // Cannot expand; must move
262 auto const result = checkedMalloc(newCapacity);
263 std::memcpy(result, p, currentSize);
268 // No jemalloc no honey
269 auto const slack = currentCapacity - currentSize;
270 if (slack * 2 > currentSize) {
271 // Too much slack, malloc-copy-free cycle:
272 auto const result = checkedMalloc(newCapacity);
273 std::memcpy(result, p, currentSize);
277 // If there's not too much slack, we realloc in hope of coalescing
278 return checkedRealloc(p, newCapacity);
281 #ifdef _LIBSTDCXX_FBSTRING
282 _GLIBCXX_END_NAMESPACE_VERSION
287 #endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)