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 #include <folly/detail/MemoryIdler.h>
19 #include <folly/Logging.h>
20 #include <folly/Portability.h>
21 #include <folly/ScopeGuard.h>
22 #include <folly/concurrency/CacheLocality.h>
23 #include <folly/memory/MallctlHelper.h>
24 #include <folly/memory/Malloc.h>
25 #include <folly/portability/PThread.h>
26 #include <folly/portability/SysMman.h>
27 #include <folly/portability/Unistd.h>
34 namespace folly { namespace detail {
36 AtomicStruct<std::chrono::steady_clock::duration>
37 MemoryIdler::defaultIdleTimeout(std::chrono::seconds(5));
39 void MemoryIdler::flushLocalMallocCaches() {
40 if (!usingJEMalloc()) {
43 if (!mallctl || !mallctlnametomib || !mallctlbymib) {
44 FB_LOG_EVERY_MS(ERROR, 10000) << "mallctl* weak link failed";
49 // Not using mallctlCall as this will fail if tcache is disabled.
50 mallctl("thread.tcache.flush", nullptr, nullptr, nullptr, 0);
52 // By default jemalloc has 4 arenas per cpu, and then assigns each
53 // thread to one of those arenas. This means that in any service
54 // that doesn't perform a lot of context switching, the chances that
55 // another thread will be using the current thread's arena (and hence
56 // doing the appropriate dirty-page purging) are low. Some good
57 // tuned configurations (such as that used by hhvm) use fewer arenas
58 // and then pin threads to avoid contended access. In that case,
59 // purging the arenas is counter-productive. We use the heuristic
60 // that if narenas <= 2 * num_cpus then we shouldn't do anything here,
61 // which detects when the narenas has been reduced from the default
63 unsigned arenaForCurrent;
67 mallctlRead("opt.narenas", &narenas);
68 mallctlRead("thread.arena", &arenaForCurrent);
69 if (narenas > 2 * CacheLocality::system().numCpus &&
70 mallctlnametomib("arena.0.purge", mib, &miblen) == 0) {
71 mib[1] = static_cast<size_t>(arenaForCurrent);
72 mallctlbymib(mib, miblen, nullptr, nullptr, nullptr, 0);
74 } catch (const std::runtime_error& ex) {
75 FB_LOG_EVERY_MS(WARNING, 10000) << ex.what();
80 // Stack madvise isn't Linux or glibc specific, but the system calls
81 // and arithmetic (and bug compatibility) are not portable. The set of
82 // platforms could be increased if it was useful.
83 #if (FOLLY_X64 || FOLLY_PPC64) && defined(_GNU_SOURCE) && \
84 defined(__linux__) && !FOLLY_MOBILE && !FOLLY_SANITIZE_ADDRESS
86 static FOLLY_TLS uintptr_t tls_stackLimit;
87 static FOLLY_TLS size_t tls_stackSize;
89 static size_t pageSize() {
90 static const size_t s_pageSize = sysconf(_SC_PAGESIZE);
94 static void fetchStackLimits() {
96 pthread_getattr_np(pthread_self(), &attr);
97 SCOPE_EXIT { pthread_attr_destroy(&attr); };
102 if ((err = pthread_attr_getstack(&attr, &addr, &rawSize))) {
103 // unexpected, but it is better to continue in prod than do nothing
104 FB_LOG_EVERY_MS(ERROR, 10000) << "pthread_attr_getstack error " << err;
109 assert(addr != nullptr);
110 assert(rawSize >= PTHREAD_STACK_MIN);
112 // glibc subtracts guard page from stack size, even though pthread docs
113 // seem to imply the opposite
115 if (pthread_attr_getguardsize(&attr, &guardSize) != 0) {
118 assert(rawSize > guardSize);
120 // stack goes down, so guard page adds to the base addr
121 tls_stackLimit = reinterpret_cast<uintptr_t>(addr) + guardSize;
122 tls_stackSize = rawSize - guardSize;
124 assert((tls_stackLimit & (pageSize() - 1)) == 0);
127 FOLLY_NOINLINE static uintptr_t getStackPtr() {
129 auto rv = reinterpret_cast<uintptr_t>(&marker);
133 void MemoryIdler::unmapUnusedStack(size_t retain) {
134 if (tls_stackSize == 0) {
137 if (tls_stackSize <= std::max(static_cast<size_t>(1), retain)) {
138 // covers both missing stack info, and impossibly large retain
142 auto sp = getStackPtr();
143 assert(sp >= tls_stackLimit);
144 assert(sp - tls_stackLimit < tls_stackSize);
146 auto end = (sp - retain) & ~(pageSize() - 1);
147 if (end <= tls_stackLimit) {
148 // no pages are eligible for unmapping
152 size_t len = end - tls_stackLimit;
153 assert((len & (pageSize() - 1)) == 0);
154 if (madvise((void*)tls_stackLimit, len, MADV_DONTNEED) != 0) {
155 // It is likely that the stack vma hasn't been fully grown. In this
156 // case madvise will apply dontneed to the present vmas, then return
157 // errno of ENOMEM. We can also get an EAGAIN, theoretically.
158 // EINVAL means either an invalid alignment or length, or that some
159 // of the pages are locked or shared. Neither should occur.
160 assert(errno == EAGAIN || errno == ENOMEM);
166 void MemoryIdler::unmapUnusedStack(size_t /* retain */) {}
170 } // namespace detail