2 * Copyright 2014 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++, then just define stub code
36 // here to typedef the fbstring type into the folly namespace.
37 // This provides backwards compatibility for code that explicitly
38 // includes and uses fbstring.
39 #if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
41 #include <folly/detail/Malloc.h>
46 using std::goodMallocSize;
47 using std::jemallocMinInPlaceExpandable;
48 using std::usingJEMalloc;
49 using std::smartRealloc;
50 using std::checkedMalloc;
51 using std::checkedCalloc;
52 using std::checkedRealloc;
55 #else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
57 #ifdef _LIBSTDCXX_FBSTRING
58 #pragma GCC system_header
61 * Declare *allocx() and mallctl() as weak symbols. These will be provided by
62 * jemalloc if we are using jemalloc, or will be NULL if we are using another
63 * malloc implementation.
65 extern "C" void* mallocx(size_t, int)
66 __attribute__((__weak__));
67 extern "C" void* rallocx(void*, size_t, int)
68 __attribute__((__weak__));
69 extern "C" size_t xallocx(void*, size_t, size_t, int)
70 __attribute__((__weak__));
71 extern "C" size_t sallocx(const void*, int)
72 __attribute__((__weak__));
73 extern "C" void dallocx(void*, int)
74 __attribute__((__weak__));
75 extern "C" size_t nallocx(size_t, int)
76 __attribute__((__weak__));
77 extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
78 __attribute__((__weak__));
80 #include <bits/functexcept.h>
81 #define FOLLY_HAVE_MALLOC_H 1
83 #include <folly/detail/Malloc.h> /* nolint */
84 #include <folly/Portability.h>
87 #include <folly/ScopeGuard.h>
89 // for malloc_usable_size
90 // NOTE: FreeBSD 9 doesn't have malloc.h. It's defitions
91 // are found in stdlib.h.
92 #if FOLLY_HAVE_MALLOC_H
105 #ifdef _LIBSTDCXX_FBSTRING
106 namespace std _GLIBCXX_VISIBILITY(default) {
107 _GLIBCXX_BEGIN_NAMESPACE_VERSION
112 bool usingJEMallocSlow();
115 * Determine if we are using jemalloc or not.
117 inline bool usingJEMalloc() {
118 // Checking for rallocx != NULL is not sufficient; we may be in a dlopen()ed
119 // module that depends on libjemalloc, so rallocx is resolved, but the main
120 // program might be using a different memory allocator. Look at the
121 // implementation of usingJEMallocSlow() for the (hacky) details.
122 static const bool result = usingJEMallocSlow();
127 * For jemalloc's size classes, see
128 * http://www.canonware.com/download/jemalloc/jemalloc-latest/doc/jemalloc.html
130 inline size_t goodMallocSize(size_t minSize) noexcept {
131 if (!usingJEMalloc()) {
132 // Not using jemalloc - no smarts
136 SCOPE_EXIT { assert(nallocx(goodSize, 0) == goodSize); };
138 // Choose smallest allocation to be 64 bytes - no tripping over
139 // cache line boundaries, and small string optimization takes care
140 // of short strings anyway.
141 return goodSize = 64;
143 if (minSize <= 512) {
144 // Round up to the next multiple of 64; we don't want to trip over
145 // cache line boundaries.
146 return goodSize = (minSize + 63) & ~size_t(63);
148 if (minSize <= 3584) {
149 // Round up to the next multiple of 256. For some size classes jemalloc
150 // will additionally round up to the nearest multiple of 512, hence the
152 return goodSize = nallocx((minSize + 255) & ~size_t(255), 0);
154 if (minSize <= 4072 * 1024) {
155 // Round up to the next multiple of 4KB
156 return goodSize = (minSize + 4095) & ~size_t(4095);
159 // Round up to the next multiple of 4MB
160 return goodSize = (minSize + 4194303) & ~size_t(4194303);
163 // We always request "good" sizes for allocation, so jemalloc can
164 // never grow in place small blocks; they're already occupied to the
165 // brim. Blocks larger than or equal to 4096 bytes can in fact be
166 // expanded in place, and this constant reflects that.
167 static const size_t jemallocMinInPlaceExpandable = 4096;
170 * Trivial wrappers around malloc, calloc, realloc that check for allocation
171 * failure and throw std::bad_alloc in that case.
173 inline void* checkedMalloc(size_t size) {
174 void* p = malloc(size);
175 if (!p) std::__throw_bad_alloc();
179 inline void* checkedCalloc(size_t n, size_t size) {
180 void* p = calloc(n, size);
181 if (!p) std::__throw_bad_alloc();
185 inline void* checkedRealloc(void* ptr, size_t size) {
186 void* p = realloc(ptr, size);
187 if (!p) std::__throw_bad_alloc();
192 * This function tries to reallocate a buffer of which only the first
193 * currentSize bytes are used. The problem with using realloc is that
194 * if currentSize is relatively small _and_ if realloc decides it
195 * needs to move the memory chunk to a new buffer, then realloc ends
196 * up copying data that is not used. It's impossible to hook into
197 * GNU's malloc to figure whether expansion will occur in-place or as
198 * a malloc-copy-free troika. (If an expand_in_place primitive would
199 * be available, smartRealloc would use it.) As things stand, this
200 * routine just tries to call realloc() (thus benefitting of potential
201 * copy-free coalescing) unless there's too much slack memory.
203 inline void* smartRealloc(void* p,
204 const size_t currentSize,
205 const size_t currentCapacity,
206 const size_t newCapacity) {
208 assert(currentSize <= currentCapacity &&
209 currentCapacity < newCapacity);
211 if (usingJEMalloc()) {
212 // using jemalloc's API. Don't forget that jemalloc can never grow
213 // in place blocks smaller than 4096 bytes.
215 // NB: newCapacity may not be precisely equal to a jemalloc size class,
216 // i.e. newCapacity is not guaranteed to be the result of a
217 // goodMallocSize() call, therefore xallocx() may return more than
218 // newCapacity bytes of space. Use >= rather than == to check whether
219 // xallocx() successfully expanded in place.
220 if (currentCapacity >= jemallocMinInPlaceExpandable &&
221 xallocx(p, newCapacity, 0, 0) >= newCapacity) {
222 // Managed to expand in place
225 // Cannot expand; must move
226 auto const result = checkedMalloc(newCapacity);
227 std::memcpy(result, p, currentSize);
232 // No jemalloc no honey
233 auto const slack = currentCapacity - currentSize;
234 if (slack * 2 > currentSize) {
235 // Too much slack, malloc-copy-free cycle:
236 auto const result = checkedMalloc(newCapacity);
237 std::memcpy(result, p, currentSize);
241 // If there's not too much slack, we realloc in hope of coalescing
242 return checkedRealloc(p, newCapacity);
245 #ifdef _LIBSTDCXX_FBSTRING
246 _GLIBCXX_END_NAMESPACE_VERSION
251 #endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
253 #endif // FOLLY_MALLOC_H_