2 * Copyright 2015 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
20 #ifndef FOLLY_MALLOC_H_
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>
47 using std::goodMallocSize;
48 using std::jemallocMinInPlaceExpandable;
49 using std::usingJEMalloc;
50 using std::smartRealloc;
51 using std::checkedMalloc;
52 using std::checkedCalloc;
53 using std::checkedRealloc;
56 #else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
58 #ifdef _LIBSTDCXX_FBSTRING
59 #pragma GCC system_header
62 * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
63 * jemalloc if we are using jemalloc, or will be NULL if we are using another
64 * malloc implementation.
66 extern "C" void* mallocx(size_t, int)
67 __attribute__((__weak__));
68 extern "C" void* rallocx(void*, size_t, int)
69 __attribute__((__weak__));
70 extern "C" size_t xallocx(void*, size_t, size_t, int)
71 __attribute__((__weak__));
72 extern "C" size_t sallocx(const void*, int)
73 __attribute__((__weak__));
74 extern "C" void dallocx(void*, int)
75 __attribute__((__weak__));
76 extern "C" size_t nallocx(size_t, int)
77 __attribute__((__weak__));
78 extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
79 __attribute__((__weak__));
80 extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
81 __attribute__((__weak__));
82 extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
84 __attribute__((__weak__));
86 #include <bits/functexcept.h>
87 #define FOLLY_HAVE_MALLOC_H 1
89 #include <folly/detail/Malloc.h> /* nolint */
92 // for malloc_usable_size
93 // NOTE: FreeBSD 9 doesn't have malloc.h. It's defitions
94 // are found in stdlib.h.
95 #if FOLLY_HAVE_MALLOC_H
109 #ifdef _LIBSTDCXX_FBSTRING
110 namespace std _GLIBCXX_VISIBILITY(default) {
111 _GLIBCXX_BEGIN_NAMESPACE_VERSION
116 // Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
118 #define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
120 #define FOLLY_MALLOC_NOINLINE
124 * Determine if we are using jemalloc or not.
126 inline bool usingJEMalloc() noexcept {
127 // Checking for rallocx != NULL is not sufficient; we may be in a dlopen()ed
128 // module that depends on libjemalloc, so rallocx is resolved, but the main
129 // program might be using a different memory allocator.
130 // How do we determine that we're using jemalloc? In the hackiest
131 // way possible. We allocate memory using malloc() and see if the
132 // per-thread counter of allocated memory increases. This makes me
133 // feel dirty inside. Also note that this requires jemalloc to have
134 // been compiled with --enable-stats.
135 static const bool result = [] () FOLLY_MALLOC_NOINLINE noexcept {
136 // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
137 // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
138 // (!!). http://goo.gl/xpmctm
139 if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
140 || sallocx == nullptr || dallocx == nullptr || nallocx == nullptr
141 || mallctl == nullptr || mallctlnametomib == nullptr
142 || mallctlbymib == nullptr) {
146 // "volatile" because gcc optimizes out the reads from *counter, because
147 // it "knows" malloc doesn't modify global state...
148 /* nolint */ volatile uint64_t* counter;
149 size_t counterLen = sizeof(uint64_t*);
151 if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
156 if (counterLen != sizeof(uint64_t*)) {
160 uint64_t origAllocated = *counter;
162 void* ptr = malloc(1);
164 // wtf, failing to allocate 1 byte
169 return (origAllocated != *counter);
176 * For jemalloc's size classes, see
177 * http://www.canonware.com/download/jemalloc/jemalloc-latest/doc/jemalloc.html
179 inline size_t goodMallocSize(size_t minSize) noexcept {
180 if (!usingJEMalloc()) {
181 // Not using jemalloc - no smarts
186 // Choose smallest allocation to be 64 bytes - no tripping over
187 // cache line boundaries, and small string optimization takes care
188 // of short strings anyway.
190 } else if (minSize <= 512) {
191 // Round up to the next multiple of 64; we don't want to trip over
192 // cache line boundaries.
193 goodSize = (minSize + 63) & ~size_t(63);
195 // Boundaries between size classes depend on numerious factors, some of
196 // which can even be modified at run-time. Determine the good allocation
197 // size by calling nallocx() directly.
198 goodSize = nallocx(minSize, 0);
200 assert(nallocx(goodSize, 0) == goodSize);
204 // We always request "good" sizes for allocation, so jemalloc can
205 // never grow in place small blocks; they're already occupied to the
206 // brim. Blocks larger than or equal to 4096 bytes can in fact be
207 // expanded in place, and this constant reflects that.
208 static const size_t jemallocMinInPlaceExpandable = 4096;
211 * Trivial wrappers around malloc, calloc, realloc that check for allocation
212 * failure and throw std::bad_alloc in that case.
214 inline void* checkedMalloc(size_t size) {
215 void* p = malloc(size);
216 if (!p) std::__throw_bad_alloc();
220 inline void* checkedCalloc(size_t n, size_t size) {
221 void* p = calloc(n, size);
222 if (!p) std::__throw_bad_alloc();
226 inline void* checkedRealloc(void* ptr, size_t size) {
227 void* p = realloc(ptr, size);
228 if (!p) std::__throw_bad_alloc();
233 * This function tries to reallocate a buffer of which only the first
234 * currentSize bytes are used. The problem with using realloc is that
235 * if currentSize is relatively small _and_ if realloc decides it
236 * needs to move the memory chunk to a new buffer, then realloc ends
237 * up copying data that is not used. It's impossible to hook into
238 * GNU's malloc to figure whether expansion will occur in-place or as
239 * a malloc-copy-free troika. (If an expand_in_place primitive would
240 * be available, smartRealloc would use it.) As things stand, this
241 * routine just tries to call realloc() (thus benefitting of potential
242 * copy-free coalescing) unless there's too much slack memory.
244 inline void* smartRealloc(void* p,
245 const size_t currentSize,
246 const size_t currentCapacity,
247 const size_t newCapacity) {
249 assert(currentSize <= currentCapacity &&
250 currentCapacity < newCapacity);
252 if (usingJEMalloc()) {
253 // using jemalloc's API. Don't forget that jemalloc can never grow
254 // in place blocks smaller than 4096 bytes.
256 // NB: newCapacity may not be precisely equal to a jemalloc size class,
257 // i.e. newCapacity is not guaranteed to be the result of a
258 // goodMallocSize() call, therefore xallocx() may return more than
259 // newCapacity bytes of space. Use >= rather than == to check whether
260 // xallocx() successfully expanded in place.
261 if (currentCapacity >= jemallocMinInPlaceExpandable &&
262 xallocx(p, newCapacity, 0, 0) >= newCapacity) {
263 // Managed to expand in place
266 // Cannot expand; must move
267 auto const result = checkedMalloc(newCapacity);
268 std::memcpy(result, p, currentSize);
273 // No jemalloc no honey
274 auto const slack = currentCapacity - currentSize;
275 if (slack * 2 > currentSize) {
276 // Too much slack, malloc-copy-free cycle:
277 auto const result = checkedMalloc(newCapacity);
278 std::memcpy(result, p, currentSize);
282 // If there's not too much slack, we realloc in hope of coalescing
283 return checkedRealloc(p, newCapacity);
286 #ifdef _LIBSTDCXX_FBSTRING
287 _GLIBCXX_END_NAMESPACE_VERSION
292 #endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
294 #endif // FOLLY_MALLOC_H_