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 #include <folly/detail/MemoryIdler.h>
18 #include <folly/Logging.h>
19 #include <folly/Malloc.h>
20 #include <folly/ScopeGuard.h>
21 #include <folly/detail/CacheLocality.h>
31 namespace folly { namespace detail {
33 AtomicStruct<std::chrono::steady_clock::duration>
34 MemoryIdler::defaultIdleTimeout(std::chrono::seconds(5));
37 // Calls mallctl, optionally reading a value of type <T> if out is
38 // non-null. Logs on error.
40 static int mallctlRead(const char* cmd, T* out) {
41 size_t outLen = sizeof(T);
42 int err = mallctl(cmd,
43 out, out ? &outLen : nullptr,
46 FB_LOG_EVERY_MS(WARNING, 10000)
47 << "mallctl " << cmd << ": " << strerror(err) << " (" << err << ")";
52 static int mallctlCall(const char* cmd) {
53 // Use <unsigned> rather than <void> to avoid sizeof(void).
54 return mallctlRead<unsigned>(cmd, nullptr);
57 void MemoryIdler::flushLocalMallocCaches() {
58 if (usingJEMalloc()) {
59 if (!mallctl || !mallctlnametomib || !mallctlbymib) {
60 FB_LOG_EVERY_MS(ERROR, 10000) << "mallctl* weak link failed";
64 // "tcache.flush" was renamed to "thread.tcache.flush" in jemalloc 3
65 mallctlCall("thread.tcache.flush");
67 // By default jemalloc has 4 arenas per cpu, and then assigns each
68 // thread to one of those arenas. This means that in any service
69 // that doesn't perform a lot of context switching, the chances that
70 // another thread will be using the current thread's arena (and hence
71 // doing the appropriate dirty-page purging) are low. Some good
72 // tuned configurations (such as that used by hhvm) use fewer arenas
73 // and then pin threads to avoid contended access. In that case,
74 // purging the arenas is counter-productive. We use the heuristic
75 // that if narenas <= 2 * num_cpus then we shouldn't do anything here,
76 // which detects when the narenas has been reduced from the default
78 unsigned arenaForCurrent;
81 if (mallctlRead<size_t>("opt.narenas", &narenas) == 0 &&
82 narenas > 2 * CacheLocality::system().numCpus &&
83 mallctlRead<unsigned>("thread.arena", &arenaForCurrent) == 0 &&
84 mallctlnametomib("arena.0.purge", mib, &miblen) == 0) {
85 mib[1] = size_t(arenaForCurrent);
86 mallctlbymib(mib, miblen, nullptr, nullptr, nullptr, 0);
92 // Stack madvise isn't Linux or glibc specific, but the system calls
93 // and arithmetic (and bug compatibility) are not portable. The set of
94 // platforms could be increased if it was useful.
95 #if (FOLLY_X64 || FOLLY_PPC64 ) && defined(_GNU_SOURCE) && defined(__linux__)
97 static FOLLY_TLS uintptr_t tls_stackLimit;
98 static FOLLY_TLS size_t tls_stackSize;
100 static size_t pageSize() {
101 static const size_t s_pageSize = sysconf(_SC_PAGESIZE);
105 static void fetchStackLimits() {
107 pthread_getattr_np(pthread_self(), &attr);
108 SCOPE_EXIT { pthread_attr_destroy(&attr); };
113 if ((err = pthread_attr_getstack(&attr, &addr, &rawSize))) {
114 // unexpected, but it is better to continue in prod than do nothing
115 FB_LOG_EVERY_MS(ERROR, 10000) << "pthread_attr_getstack error " << err;
120 assert(addr != nullptr);
121 assert(rawSize >= PTHREAD_STACK_MIN);
123 // glibc subtracts guard page from stack size, even though pthread docs
124 // seem to imply the opposite
126 if (pthread_attr_getguardsize(&attr, &guardSize) != 0) {
129 assert(rawSize > guardSize);
131 // stack goes down, so guard page adds to the base addr
132 tls_stackLimit = uintptr_t(addr) + guardSize;
133 tls_stackSize = rawSize - guardSize;
135 assert((tls_stackLimit & (pageSize() - 1)) == 0);
138 FOLLY_NOINLINE static uintptr_t getStackPtr() {
140 auto rv = uintptr_t(&marker);
144 void MemoryIdler::unmapUnusedStack(size_t retain) {
145 if (tls_stackSize == 0) {
148 if (tls_stackSize <= std::max(size_t(1), retain)) {
149 // covers both missing stack info, and impossibly large retain
153 auto sp = getStackPtr();
154 assert(sp >= tls_stackLimit);
155 assert(sp - tls_stackLimit < tls_stackSize);
157 auto end = (sp - retain) & ~(pageSize() - 1);
158 if (end <= tls_stackLimit) {
159 // no pages are eligible for unmapping
163 size_t len = end - tls_stackLimit;
164 assert((len & (pageSize() - 1)) == 0);
165 if (madvise((void*)tls_stackLimit, len, MADV_DONTNEED) != 0) {
166 // It is likely that the stack vma hasn't been fully grown. In this
167 // case madvise will apply dontneed to the present vmas, then return
168 // errno of ENOMEM. We can also get an EAGAIN, theoretically.
169 // EINVAL means either an invalid alignment or length, or that some
170 // of the pages are locked or shared. Neither should occur.
171 assert(errno == EAGAIN || errno == ENOMEM);
177 void MemoryIdler::unmapUnusedStack(size_t retain) {