2 * Copyright 2017 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
22 #include <folly/CPortability.h>
23 #include <folly/portability/Config.h>
26 * Define various MALLOCX_* macros normally provided by jemalloc. We define
27 * them so that we don't have to include jemalloc.h, in case the program is
28 * built without jemalloc support.
30 #if (defined(USE_JEMALLOC) || defined(FOLLY_USE_JEMALLOC)) && !FOLLY_SANITIZE
31 // We have JEMalloc, so use it.
32 # include <jemalloc/jemalloc.h>
34 # ifndef MALLOCX_LG_ALIGN
35 # define MALLOCX_LG_ALIGN(la) (la)
38 # define MALLOCX_ZERO (static_cast<int>(0x40))
42 // If using fbstring from libstdc++ (see comment in FBString.h), then
43 // just define stub code here to typedef the fbstring type into the
45 // This provides backwards compatibility for code that explicitly
46 // includes and uses fbstring.
47 #if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
49 #include <folly/memory/detail/MallocImpl.h>
50 #include <folly/portability/BitsFunctexcept.h>
55 using std::goodMallocSize;
56 using std::jemallocMinInPlaceExpandable;
57 using std::usingJEMalloc;
58 using std::smartRealloc;
59 using std::checkedMalloc;
60 using std::checkedCalloc;
61 using std::checkedRealloc;
64 #else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
66 #ifdef _LIBSTDCXX_FBSTRING
67 #pragma GCC system_header
70 * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
71 * jemalloc if we are using jemalloc, or will be nullptr if we are using another
72 * malloc implementation.
74 extern "C" void* mallocx(size_t, int)
75 __attribute__((__weak__));
76 extern "C" void* rallocx(void*, size_t, int)
77 __attribute__((__weak__));
78 extern "C" size_t xallocx(void*, size_t, size_t, int)
79 __attribute__((__weak__));
80 extern "C" size_t sallocx(const void*, int)
81 __attribute__((__weak__));
82 extern "C" void dallocx(void*, int)
83 __attribute__((__weak__));
84 extern "C" void sdallocx(void*, size_t, int)
85 __attribute__((__weak__));
86 extern "C" size_t nallocx(size_t, int)
87 __attribute__((__weak__));
88 extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
89 __attribute__((__weak__));
90 extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
91 __attribute__((__weak__));
92 extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
94 __attribute__((__weak__));
96 #include <bits/functexcept.h>
98 #define FOLLY_HAVE_MALLOC_H 1
100 #else // !defined(_LIBSTDCXX_FBSTRING)
102 #include <folly/memory/detail/MallocImpl.h> /* nolint */
103 #include <folly/portability/BitsFunctexcept.h> /* nolint */
107 // for malloc_usable_size
108 // NOTE: FreeBSD 9 doesn't have malloc.h. Its definitions
109 // are found in stdlib.h.
110 #if FOLLY_HAVE_MALLOC_H
125 #ifdef _LIBSTDCXX_FBSTRING
126 namespace std _GLIBCXX_VISIBILITY(default) {
127 _GLIBCXX_BEGIN_NAMESPACE_VERSION
132 // Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
133 #if defined(__GNUC__)
134 #define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
135 #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL) >= 40900
136 // This is for checked malloc-like functions (returns non-null pointer
137 // which cannot alias any outstanding pointer).
138 #define FOLLY_MALLOC_CHECKED_MALLOC \
139 __attribute__((__returns_nonnull__, __malloc__))
141 #define FOLLY_MALLOC_CHECKED_MALLOC __attribute__((__malloc__))
144 #define FOLLY_MALLOC_NOINLINE
145 #define FOLLY_MALLOC_CHECKED_MALLOC
148 #include <folly/CPortability.h>
150 * Determine if we are using jemalloc or not.
152 #if defined(USE_JEMALLOC) && !FOLLY_SANITIZE
153 inline bool usingJEMalloc() noexcept {
157 FOLLY_MALLOC_NOINLINE inline bool usingJEMalloc() noexcept {
158 // Checking for rallocx != nullptr is not sufficient; we may be in a
159 // dlopen()ed module that depends on libjemalloc, so rallocx is resolved, but
160 // the main program might be using a different memory allocator.
161 // How do we determine that we're using jemalloc? In the hackiest
162 // way possible. We allocate memory using malloc() and see if the
163 // per-thread counter of allocated memory increases. This makes me
164 // feel dirty inside. Also note that this requires jemalloc to have
165 // been compiled with --enable-stats.
166 static const bool result = [] () noexcept {
167 // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
168 // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
169 // (!!). http://goo.gl/xpmctm
170 if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
171 || sallocx == nullptr || dallocx == nullptr || sdallocx == nullptr
172 || nallocx == nullptr || mallctl == nullptr
173 || mallctlnametomib == nullptr || mallctlbymib == nullptr) {
177 // "volatile" because gcc optimizes out the reads from *counter, because
178 // it "knows" malloc doesn't modify global state...
179 /* nolint */ volatile uint64_t* counter;
180 size_t counterLen = sizeof(uint64_t*);
182 if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
187 if (counterLen != sizeof(uint64_t*)) {
191 uint64_t origAllocated = *counter;
193 static const void* volatile ptr = malloc(1);
195 // wtf, failing to allocate 1 byte
199 return (origAllocated != *counter);
206 inline size_t goodMallocSize(size_t minSize) noexcept {
211 if (!usingJEMalloc()) {
212 // Not using jemalloc - no smarts
216 return nallocx(minSize, 0);
219 // We always request "good" sizes for allocation, so jemalloc can
220 // never grow in place small blocks; they're already occupied to the
221 // brim. Blocks larger than or equal to 4096 bytes can in fact be
222 // expanded in place, and this constant reflects that.
223 static const size_t jemallocMinInPlaceExpandable = 4096;
226 * Trivial wrappers around malloc, calloc, realloc that check for allocation
227 * failure and throw std::bad_alloc in that case.
229 inline void* checkedMalloc(size_t size) {
230 void* p = malloc(size);
232 std::__throw_bad_alloc();
237 inline void* checkedCalloc(size_t n, size_t size) {
238 void* p = calloc(n, size);
240 std::__throw_bad_alloc();
245 inline void* checkedRealloc(void* ptr, size_t size) {
246 void* p = realloc(ptr, size);
248 std::__throw_bad_alloc();
254 * This function tries to reallocate a buffer of which only the first
255 * currentSize bytes are used. The problem with using realloc is that
256 * if currentSize is relatively small _and_ if realloc decides it
257 * needs to move the memory chunk to a new buffer, then realloc ends
258 * up copying data that is not used. It's generally not a win to try
259 * to hook in to realloc() behavior to avoid copies - at least in
260 * jemalloc, realloc() almost always ends up doing a copy, because
261 * there is little fragmentation / slack space to take advantage of.
263 FOLLY_MALLOC_CHECKED_MALLOC FOLLY_MALLOC_NOINLINE inline void* smartRealloc(
265 const size_t currentSize,
266 const size_t currentCapacity,
267 const size_t newCapacity) {
269 assert(currentSize <= currentCapacity &&
270 currentCapacity < newCapacity);
272 auto const slack = currentCapacity - currentSize;
273 if (slack * 2 > currentSize) {
274 // Too much slack, malloc-copy-free cycle:
275 auto const result = checkedMalloc(newCapacity);
276 std::memcpy(result, p, currentSize);
280 // If there's not too much slack, we realloc in hope of coalescing
281 return checkedRealloc(p, newCapacity);
284 #ifdef _LIBSTDCXX_FBSTRING
285 _GLIBCXX_END_NAMESPACE_VERSION
290 #endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)