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/portability/Config.h>
25 * Define various MALLOCX_* macros normally provided by jemalloc. We define
26 * them so that we don't have to include jemalloc.h, in case the program is
27 * built without jemalloc support.
29 #if defined(USE_JEMALLOC) || defined(FOLLY_USE_JEMALLOC)
30 // We have JEMalloc, so use it.
31 # include <jemalloc/jemalloc.h>
33 # ifndef MALLOCX_LG_ALIGN
34 # define MALLOCX_LG_ALIGN(la) (la)
37 # define MALLOCX_ZERO (static_cast<int>(0x40))
41 // If using fbstring from libstdc++ (see comment in FBString.h), then
42 // just define stub code here to typedef the fbstring type into the
44 // This provides backwards compatibility for code that explicitly
45 // includes and uses fbstring.
46 #if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
48 #include <folly/detail/Malloc.h>
49 #include <folly/portability/BitsFunctexcept.h>
54 using std::goodMallocSize;
55 using std::jemallocMinInPlaceExpandable;
56 using std::usingJEMalloc;
57 using std::smartRealloc;
58 using std::checkedMalloc;
59 using std::checkedCalloc;
60 using std::checkedRealloc;
63 #else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
65 #ifdef _LIBSTDCXX_FBSTRING
66 #pragma GCC system_header
69 * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
70 * jemalloc if we are using jemalloc, or will be NULL if we are using another
71 * malloc implementation.
73 extern "C" void* mallocx(size_t, int)
74 __attribute__((__weak__));
75 extern "C" void* rallocx(void*, size_t, int)
76 __attribute__((__weak__));
77 extern "C" size_t xallocx(void*, size_t, size_t, int)
78 __attribute__((__weak__));
79 extern "C" size_t sallocx(const void*, int)
80 __attribute__((__weak__));
81 extern "C" void dallocx(void*, int)
82 __attribute__((__weak__));
83 extern "C" void sdallocx(void*, size_t, int)
84 __attribute__((__weak__));
85 extern "C" size_t nallocx(size_t, int)
86 __attribute__((__weak__));
87 extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
88 __attribute__((__weak__));
89 extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
90 __attribute__((__weak__));
91 extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
93 __attribute__((__weak__));
95 #include <bits/functexcept.h>
97 #define FOLLY_HAVE_MALLOC_H 1
99 #else // !defined(_LIBSTDCXX_FBSTRING)
101 #include <folly/detail/Malloc.h> /* nolint */
102 #include <folly/portability/BitsFunctexcept.h> /* nolint */
106 // for malloc_usable_size
107 // NOTE: FreeBSD 9 doesn't have malloc.h. Its definitions
108 // are found in stdlib.h.
109 #if FOLLY_HAVE_MALLOC_H
124 #ifdef _LIBSTDCXX_FBSTRING
125 namespace std _GLIBCXX_VISIBILITY(default) {
126 _GLIBCXX_BEGIN_NAMESPACE_VERSION
131 // Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
132 #if defined(__GNUC__)
133 #define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
134 #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL) >= 40900
135 // This is for checked malloc-like functions (returns non-null pointer
136 // which cannot alias any outstanding pointer).
137 #define FOLLY_MALLOC_CHECKED_MALLOC \
138 __attribute__((__returns_nonnull__, __malloc__))
140 #define FOLLY_MALLOC_CHECKED_MALLOC __attribute__((__malloc__))
143 #define FOLLY_MALLOC_NOINLINE
144 #define FOLLY_MALLOC_CHECKED_MALLOC
148 * Determine if we are using jemalloc or not.
150 FOLLY_MALLOC_NOINLINE inline bool usingJEMalloc() noexcept {
151 // Checking for rallocx != NULL is not sufficient; we may be in a dlopen()ed
152 // module that depends on libjemalloc, so rallocx is resolved, but the main
153 // program might be using a different memory allocator.
154 // How do we determine that we're using jemalloc? In the hackiest
155 // way possible. We allocate memory using malloc() and see if the
156 // per-thread counter of allocated memory increases. This makes me
157 // feel dirty inside. Also note that this requires jemalloc to have
158 // been compiled with --enable-stats.
159 static const bool result = [] () noexcept {
160 // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
161 // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
162 // (!!). http://goo.gl/xpmctm
163 if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
164 || sallocx == nullptr || dallocx == nullptr || sdallocx == nullptr
165 || nallocx == nullptr || mallctl == nullptr
166 || mallctlnametomib == nullptr || mallctlbymib == nullptr) {
170 // "volatile" because gcc optimizes out the reads from *counter, because
171 // it "knows" malloc doesn't modify global state...
172 /* nolint */ volatile uint64_t* counter;
173 size_t counterLen = sizeof(uint64_t*);
175 if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
180 if (counterLen != sizeof(uint64_t*)) {
184 uint64_t origAllocated = *counter;
186 // Static because otherwise clever compilers will find out that
187 // the ptr is not used and does not escape the scope, so they will
188 // just optimize away the malloc.
189 static const void* ptr = malloc(1);
191 // wtf, failing to allocate 1 byte
195 return (origAllocated != *counter);
201 inline size_t goodMallocSize(size_t minSize) noexcept {
206 if (!usingJEMalloc()) {
207 // Not using jemalloc - no smarts
211 return nallocx(minSize, 0);
214 // We always request "good" sizes for allocation, so jemalloc can
215 // never grow in place small blocks; they're already occupied to the
216 // brim. Blocks larger than or equal to 4096 bytes can in fact be
217 // expanded in place, and this constant reflects that.
218 static const size_t jemallocMinInPlaceExpandable = 4096;
221 * Trivial wrappers around malloc, calloc, realloc that check for allocation
222 * failure and throw std::bad_alloc in that case.
224 inline void* checkedMalloc(size_t size) {
225 void* p = malloc(size);
226 if (!p) std::__throw_bad_alloc();
230 inline void* checkedCalloc(size_t n, size_t size) {
231 void* p = calloc(n, size);
232 if (!p) std::__throw_bad_alloc();
236 inline void* checkedRealloc(void* ptr, size_t size) {
237 void* p = realloc(ptr, size);
238 if (!p) std::__throw_bad_alloc();
243 * This function tries to reallocate a buffer of which only the first
244 * currentSize bytes are used. The problem with using realloc is that
245 * if currentSize is relatively small _and_ if realloc decides it
246 * needs to move the memory chunk to a new buffer, then realloc ends
247 * up copying data that is not used. It's impossible to hook into
248 * GNU's malloc to figure whether expansion will occur in-place or as
249 * a malloc-copy-free troika. (If an expand_in_place primitive would
250 * be available, smartRealloc would use it.) As things stand, this
251 * routine just tries to call realloc() (thus benefitting of potential
252 * copy-free coalescing) unless there's too much slack memory.
254 FOLLY_MALLOC_CHECKED_MALLOC FOLLY_MALLOC_NOINLINE inline void* smartRealloc(
256 const size_t currentSize,
257 const size_t currentCapacity,
258 const size_t newCapacity) {
260 assert(currentSize <= currentCapacity &&
261 currentCapacity < newCapacity);
263 if (usingJEMalloc()) {
264 // using jemalloc's API. Don't forget that jemalloc can never grow
265 // in place blocks smaller than 4096 bytes.
267 // NB: newCapacity may not be precisely equal to a jemalloc size class,
268 // i.e. newCapacity is not guaranteed to be the result of a
269 // goodMallocSize() call, therefore xallocx() may return more than
270 // newCapacity bytes of space. Use >= rather than == to check whether
271 // xallocx() successfully expanded in place.
272 if (currentCapacity >= jemallocMinInPlaceExpandable &&
273 xallocx(p, newCapacity, 0, 0) >= newCapacity) {
274 // Managed to expand in place
277 // Cannot expand; must move
278 auto const result = checkedMalloc(newCapacity);
279 std::memcpy(result, p, currentSize);
284 // No jemalloc no honey
285 auto const slack = currentCapacity - currentSize;
286 if (slack * 2 > currentSize) {
287 // Too much slack, malloc-copy-free cycle:
288 auto const result = checkedMalloc(newCapacity);
289 std::memcpy(result, p, currentSize);
293 // If there's not too much slack, we realloc in hope of coalescing
294 return checkedRealloc(p, newCapacity);
297 #ifdef _LIBSTDCXX_FBSTRING
298 _GLIBCXX_END_NAMESPACE_VERSION
303 #endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)