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 #include <folly/detail/MemoryIdler.h>
19 #include <folly/Logging.h>
20 #include <folly/Malloc.h>
21 #include <folly/Portability.h>
22 #include <folly/ScopeGuard.h>
23 #include <folly/detail/CacheLocality.h>
24 #include <folly/portability/SysMman.h>
25 #include <folly/portability/Unistd.h>
33 namespace folly { namespace detail {
35 AtomicStruct<std::chrono::steady_clock::duration>
36 MemoryIdler::defaultIdleTimeout(std::chrono::seconds(5));
39 // Calls mallctl, optionally reading a value of type <T> if out is
40 // non-null. Logs on error.
42 static int mallctlRead(const char* cmd, T* out) {
43 size_t outLen = sizeof(T);
44 int err = mallctl(cmd,
45 out, out ? &outLen : nullptr,
48 FB_LOG_EVERY_MS(WARNING, 10000)
49 << "mallctl " << cmd << ": " << strerror(err) << " (" << err << ")";
54 static int mallctlCall(const char* cmd) {
55 // Use <unsigned> rather than <void> to avoid sizeof(void).
56 return mallctlRead<unsigned>(cmd, nullptr);
59 void MemoryIdler::flushLocalMallocCaches() {
60 if (usingJEMalloc()) {
61 if (!mallctl || !mallctlnametomib || !mallctlbymib) {
62 FB_LOG_EVERY_MS(ERROR, 10000) << "mallctl* weak link failed";
66 // "tcache.flush" was renamed to "thread.tcache.flush" in jemalloc 3
67 mallctlCall("thread.tcache.flush");
69 // By default jemalloc has 4 arenas per cpu, and then assigns each
70 // thread to one of those arenas. This means that in any service
71 // that doesn't perform a lot of context switching, the chances that
72 // another thread will be using the current thread's arena (and hence
73 // doing the appropriate dirty-page purging) are low. Some good
74 // tuned configurations (such as that used by hhvm) use fewer arenas
75 // and then pin threads to avoid contended access. In that case,
76 // purging the arenas is counter-productive. We use the heuristic
77 // that if narenas <= 2 * num_cpus then we shouldn't do anything here,
78 // which detects when the narenas has been reduced from the default
80 unsigned arenaForCurrent;
83 if (mallctlRead<unsigned>("opt.narenas", &narenas) == 0 &&
84 narenas > 2 * CacheLocality::system().numCpus &&
85 mallctlRead<unsigned>("thread.arena", &arenaForCurrent) == 0 &&
86 mallctlnametomib("arena.0.purge", mib, &miblen) == 0) {
87 mib[1] = size_t(arenaForCurrent);
88 mallctlbymib(mib, miblen, nullptr, nullptr, nullptr, 0);
94 // Stack madvise isn't Linux or glibc specific, but the system calls
95 // and arithmetic (and bug compatibility) are not portable. The set of
96 // platforms could be increased if it was useful.
97 #if (FOLLY_X64 || FOLLY_PPC64) && defined(_GNU_SOURCE) && \
98 defined(__linux__) && !FOLLY_MOBILE
100 static FOLLY_TLS uintptr_t tls_stackLimit;
101 static FOLLY_TLS size_t tls_stackSize;
103 static size_t pageSize() {
104 static const size_t s_pageSize = sysconf(_SC_PAGESIZE);
108 static void fetchStackLimits() {
110 pthread_getattr_np(pthread_self(), &attr);
111 SCOPE_EXIT { pthread_attr_destroy(&attr); };
116 if ((err = pthread_attr_getstack(&attr, &addr, &rawSize))) {
117 // unexpected, but it is better to continue in prod than do nothing
118 FB_LOG_EVERY_MS(ERROR, 10000) << "pthread_attr_getstack error " << err;
123 assert(addr != nullptr);
124 assert(rawSize >= PTHREAD_STACK_MIN);
126 // glibc subtracts guard page from stack size, even though pthread docs
127 // seem to imply the opposite
129 if (pthread_attr_getguardsize(&attr, &guardSize) != 0) {
132 assert(rawSize > guardSize);
134 // stack goes down, so guard page adds to the base addr
135 tls_stackLimit = uintptr_t(addr) + guardSize;
136 tls_stackSize = rawSize - guardSize;
138 assert((tls_stackLimit & (pageSize() - 1)) == 0);
141 FOLLY_NOINLINE static uintptr_t getStackPtr() {
143 auto rv = uintptr_t(&marker);
147 void MemoryIdler::unmapUnusedStack(size_t retain) {
148 if (tls_stackSize == 0) {
151 if (tls_stackSize <= std::max(size_t(1), retain)) {
152 // covers both missing stack info, and impossibly large retain
156 auto sp = getStackPtr();
157 assert(sp >= tls_stackLimit);
158 assert(sp - tls_stackLimit < tls_stackSize);
160 auto end = (sp - retain) & ~(pageSize() - 1);
161 if (end <= tls_stackLimit) {
162 // no pages are eligible for unmapping
166 size_t len = end - tls_stackLimit;
167 assert((len & (pageSize() - 1)) == 0);
168 if (madvise((void*)tls_stackLimit, len, MADV_DONTNEED) != 0) {
169 // It is likely that the stack vma hasn't been fully grown. In this
170 // case madvise will apply dontneed to the present vmas, then return
171 // errno of ENOMEM. We can also get an EAGAIN, theoretically.
172 // EINVAL means either an invalid alignment or length, or that some
173 // of the pages are locked or shared. Neither should occur.
174 assert(errno == EAGAIN || errno == ENOMEM);
180 void MemoryIdler::unmapUnusedStack(size_t retain) {