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/CacheLocality.h>
20 #define _GNU_SOURCE 1 // for RTLD_NOLOAD
25 #include <folly/Conv.h>
26 #include <folly/Exception.h>
27 #include <folly/FileUtil.h>
28 #include <folly/Format.h>
29 #include <folly/ScopeGuard.h>
34 ///////////// CacheLocality
36 /// Returns the best real CacheLocality information available
37 static CacheLocality getSystemLocalityInfo() {
40 return CacheLocality::readFromSysfs();
46 long numCpus = sysconf(_SC_NPROCESSORS_CONF);
48 // This shouldn't happen, but if it does we should try to keep
49 // going. We are probably not going to be able to parse /sys on
50 // this box either (although we will try), which means we are going
51 // to fall back to the SequentialThreadId splitter. On my 16 core
52 // (x hyperthreading) dev box 16 stripes is enough to get pretty good
53 // contention avoidance with SequentialThreadId, and there is little
54 // improvement from going from 32 to 64. This default gives us some
58 return CacheLocality::uniform(numCpus);
62 const CacheLocality& CacheLocality::system<std::atomic>() {
63 static CacheLocality cache(getSystemLocalityInfo());
67 // Each level of cache has sharing sets, which are the set of cpus
68 // that share a common cache at that level. These are available in a
69 // hex bitset form (/sys/devices/system/cpu/cpu0/index0/shared_cpu_map,
70 // for example). They are also available in a human-readable list form,
71 // as in /sys/devices/system/cpu/cpu0/index0/shared_cpu_list. The list
72 // is a comma-separated list of numbers and ranges, where the ranges are
73 // a pair of decimal numbers separated by a '-'.
75 // To sort the cpus for optimum locality we don't really need to parse
76 // the sharing sets, we just need a unique representative from the
77 // equivalence class. The smallest value works fine, and happens to be
78 // the first decimal number in the file. We load all of the equivalence
79 // class information from all of the cpu*/index* directories, order the
80 // cpus first by increasing last-level cache equivalence class, then by
81 // the smaller caches. Finally, we break ties with the cpu number itself.
83 /// Returns the first decimal number in the string, or throws an exception
84 /// if the string does not start with a number terminated by ',', '-',
86 static size_t parseLeadingNumber(const std::string& line) {
87 auto raw = line.c_str();
89 unsigned long val = strtoul(raw, &end, 10);
90 if (end == raw || (*end != ',' && *end != '-' && *end != '\n' && *end != 0)) {
91 throw std::runtime_error(
92 to<std::string>("error parsing list '", line, "'").c_str());
97 CacheLocality CacheLocality::readFromSysfsTree(
98 const std::function<std::string(std::string)>& mapping) {
99 // number of equivalence classes per level
100 std::vector<size_t> numCachesByLevel;
102 // the list of cache equivalence classes, where equivalance classes
103 // are named by the smallest cpu in the class
104 std::vector<std::vector<size_t>> equivClassesByCpu;
106 std::vector<size_t> cpus;
109 auto cpu = cpus.size();
110 std::vector<size_t> levels;
111 for (size_t index = 0;; ++index) {
113 format("/sys/devices/system/cpu/cpu{}/cache/index{}/", cpu, index)
115 auto cacheType = mapping(dir + "type");
116 auto equivStr = mapping(dir + "shared_cpu_list");
117 if (cacheType.size() == 0 || equivStr.size() == 0) {
121 if (cacheType[0] == 'I') {
122 // cacheType in { "Data", "Instruction", "Unified" }. skip icache
125 auto equiv = parseLeadingNumber(equivStr);
126 auto level = levels.size();
127 levels.push_back(equiv);
130 // we only want to count the equiv classes once, so we do it when
131 // we first encounter them
132 while (numCachesByLevel.size() <= level) {
133 numCachesByLevel.push_back(0);
135 numCachesByLevel[level]++;
139 if (levels.size() == 0) {
140 // no levels at all for this cpu, we must be done
143 equivClassesByCpu.emplace_back(std::move(levels));
147 if (cpus.size() == 0) {
148 throw std::runtime_error("unable to load cache sharing info");
151 std::sort(cpus.begin(),
153 [&](size_t lhs, size_t rhs) -> bool {
154 // sort first by equiv class of cache with highest index,
155 // direction doesn't matter. If different cpus have
156 // different numbers of caches then this code might produce
157 // a sub-optimal ordering, but it won't crash
158 auto& lhsEquiv = equivClassesByCpu[lhs];
159 auto& rhsEquiv = equivClassesByCpu[rhs];
160 for (int i = std::min(lhsEquiv.size(), rhsEquiv.size()) - 1;
163 if (lhsEquiv[i] != rhsEquiv[i]) {
164 return lhsEquiv[i] < rhsEquiv[i];
168 // break ties deterministically by cpu
172 // the cpus are now sorted by locality, with neighboring entries closer
173 // to each other than entries that are far away. For striping we want
174 // the inverse map, since we are starting with the cpu
175 std::vector<size_t> indexes(cpus.size());
176 for (size_t i = 0; i < cpus.size(); ++i) {
177 indexes[cpus[i]] = i;
180 return CacheLocality{
181 cpus.size(), std::move(numCachesByLevel), std::move(indexes)};
184 CacheLocality CacheLocality::readFromSysfs() {
185 return readFromSysfsTree([](std::string name) {
186 std::ifstream xi(name.c_str());
188 std::getline(xi, rv);
193 CacheLocality CacheLocality::uniform(size_t numCpus) {
196 rv.numCpus = numCpus;
198 // one cache shared by all cpus
199 rv.numCachesByLevel.push_back(numCpus);
201 // no permutations in locality index mapping
202 for (size_t cpu = 0; cpu < numCpus; ++cpu) {
203 rv.localityIndexByCpu.push_back(cpu);
209 ////////////// Getcpu
211 /// Resolves the dynamically loaded symbol __vdso_getcpu, returning null
213 static Getcpu::Func loadVdsoGetcpu() {
214 #if defined(_MSC_VER) || defined(__BIONIC__)
217 void* h = dlopen("linux-vdso.so.1", RTLD_LAZY | RTLD_LOCAL | RTLD_NOLOAD);
222 auto func = Getcpu::Func(dlsym(h, "__vdso_getcpu"));
223 if (func == nullptr) {
224 // technically a null result could either be a failure or a successful
225 // lookup of a symbol with the null value, but the second can't actually
226 // happen for this symbol. No point holding the handle forever if
227 // we don't need the code
235 Getcpu::Func Getcpu::vdsoFunc() {
236 static Func func = loadVdsoGetcpu();
241 /////////////// SequentialThreadId
244 std::atomic<size_t> SequentialThreadId<std::atomic>::prevId(0);
247 FOLLY_TLS size_t SequentialThreadId<std::atomic>::currentId(0);
250 /////////////// AccessSpreader
253 const AccessSpreader<std::atomic> AccessSpreader<std::atomic>::stripeByCore(
254 CacheLocality::system<>().numCachesByLevel.front());
257 const AccessSpreader<std::atomic> AccessSpreader<std::atomic>::stripeByChip(
258 CacheLocality::system<>().numCachesByLevel.back());
261 AccessSpreaderArray<std::atomic, 128>
262 AccessSpreaderArray<std::atomic, 128>::sharedInstance = {};
264 /// Always claims to be on CPU zero, node zero
265 static int degenerateGetcpu(unsigned* cpu, unsigned* node, void* /* unused */) {
266 if (cpu != nullptr) {
269 if (node != nullptr) {
276 Getcpu::Func AccessSpreader<std::atomic>::pickGetcpuFunc(size_t numStripes) {
277 if (numStripes == 1) {
278 // there's no need to call getcpu if there is only one stripe.
279 // This should not be common, so we don't want to waste a test and
280 // branch in the main code path, but we might as well use a faster
282 return °enerateGetcpu;
284 auto best = Getcpu::vdsoFunc();
285 return best ? best : &FallbackGetcpuType::getcpu;
289 } // namespace folly::detail