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/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>
31 namespace folly { namespace detail {
33 ///////////// CacheLocality
35 /// Returns the best real CacheLocality information available
36 static CacheLocality getSystemLocalityInfo() {
39 return CacheLocality::readFromSysfs();
45 long numCpus = sysconf(_SC_NPROCESSORS_CONF);
47 // This shouldn't happen, but if it does we should try to keep
48 // going. We are probably not going to be able to parse /sys on
49 // this box either (although we will try), which means we are going
50 // to fall back to the SequentialThreadId splitter. On my 16 core
51 // (x hyperthreading) dev box 16 stripes is enough to get pretty good
52 // contention avoidance with SequentialThreadId, and there is little
53 // improvement from going from 32 to 64. This default gives us some
57 return CacheLocality::uniform(numCpus);
61 const CacheLocality& CacheLocality::system<std::atomic>() {
62 static CacheLocality cache(getSystemLocalityInfo());
66 // Each level of cache has sharing sets, which are the set of cpus
67 // that share a common cache at that level. These are available in a
68 // hex bitset form (/sys/devices/system/cpu/cpu0/index0/shared_cpu_map,
69 // for example). They are also available in a human-readable list form,
70 // as in /sys/devices/system/cpu/cpu0/index0/shared_cpu_list. The list
71 // is a comma-separated list of numbers and ranges, where the ranges are
72 // a pair of decimal numbers separated by a '-'.
74 // To sort the cpus for optimum locality we don't really need to parse
75 // the sharing sets, we just need a unique representative from the
76 // equivalence class. The smallest value works fine, and happens to be
77 // the first decimal number in the file. We load all of the equivalence
78 // class information from all of the cpu*/index* directories, order the
79 // cpus first by increasing last-level cache equivalence class, then by
80 // the smaller caches. Finally, we break ties with the cpu number itself.
82 /// Returns the first decimal number in the string, or throws an exception
83 /// if the string does not start with a number terminated by ',', '-',
85 static size_t parseLeadingNumber(const std::string& line) {
86 auto raw = line.c_str();
88 unsigned long val = strtoul(raw, &end, 10);
89 if (end == raw || (*end != ',' && *end != '-' && *end != '\n' && *end != 0)) {
90 throw std::runtime_error(to<std::string>(
91 "error parsing list '", line, "'").c_str());
96 CacheLocality CacheLocality::readFromSysfsTree(
97 const std::function<std::string(std::string)>& mapping) {
98 // number of equivalence classes per level
99 std::vector<size_t> numCachesByLevel;
101 // the list of cache equivalence classes, where equivalance classes
102 // are named by the smallest cpu in the class
103 std::vector<std::vector<size_t>> equivClassesByCpu;
105 std::vector<size_t> cpus;
108 auto cpu = cpus.size();
109 std::vector<size_t> levels;
110 for (size_t index = 0; ; ++index) {
111 auto dir = format("/sys/devices/system/cpu/cpu{}/cache/index{}/",
113 auto cacheType = mapping(dir + "type");
114 auto equivStr = mapping(dir + "shared_cpu_list");
115 if (cacheType.size() == 0 || equivStr.size() == 0) {
119 if (cacheType[0] == 'I') {
120 // cacheType in { "Data", "Instruction", "Unified" }. skip icache
123 auto equiv = parseLeadingNumber(equivStr);
124 auto level = levels.size();
125 levels.push_back(equiv);
128 // we only want to count the equiv classes once, so we do it when
129 // we first encounter them
130 while (numCachesByLevel.size() <= level) {
131 numCachesByLevel.push_back(0);
133 numCachesByLevel[level]++;
137 if (levels.size() == 0) {
138 // no levels at all for this cpu, we must be done
141 equivClassesByCpu.emplace_back(std::move(levels));
145 if (cpus.size() == 0) {
146 throw std::runtime_error("unable to load cache sharing info");
149 std::sort(cpus.begin(), cpus.end(), [&](size_t lhs, size_t rhs) -> bool {
150 // sort first by equiv class of cache with highest index, direction
151 // doesn't matter. If different cpus have different numbers of
152 // caches then this code might produce a sub-optimal ordering, but
154 auto& lhsEquiv = equivClassesByCpu[lhs];
155 auto& rhsEquiv = equivClassesByCpu[rhs];
156 for (int i = std::min(lhsEquiv.size(), rhsEquiv.size()) - 1; i >= 0; --i) {
157 if (lhsEquiv[i] != rhsEquiv[i]) {
158 return lhsEquiv[i] < rhsEquiv[i];
162 // break ties deterministically by cpu
166 // the cpus are now sorted by locality, with neighboring entries closer
167 // to each other than entries that are far away. For striping we want
168 // the inverse map, since we are starting with the cpu
169 std::vector<size_t> indexes(cpus.size());
170 for (size_t i = 0; i < cpus.size(); ++i) {
171 indexes[cpus[i]] = i;
174 return CacheLocality{
175 cpus.size(), std::move(numCachesByLevel), std::move(indexes) };
178 CacheLocality CacheLocality::readFromSysfs() {
179 return readFromSysfsTree([](std::string name) {
180 std::ifstream xi(name.c_str());
182 std::getline(xi, rv);
188 CacheLocality CacheLocality::uniform(size_t numCpus) {
191 rv.numCpus = numCpus;
193 // one cache shared by all cpus
194 rv.numCachesByLevel.push_back(numCpus);
196 // no permutations in locality index mapping
197 for (size_t cpu = 0; cpu < numCpus; ++cpu) {
198 rv.localityIndexByCpu.push_back(cpu);
204 ////////////// Getcpu
206 /// Resolves the dynamically loaded symbol __vdso_getcpu, returning null
208 static Getcpu::Func loadVdsoGetcpu() {
212 void* h = dlopen("linux-vdso.so.1", RTLD_LAZY | RTLD_LOCAL | RTLD_NOLOAD);
217 auto func = Getcpu::Func(dlsym(h, "__vdso_getcpu"));
218 if (func == nullptr) {
219 // technically a null result could either be a failure or a successful
220 // lookup of a symbol with the null value, but the second can't actually
221 // happen for this symbol. No point holding the handle forever if
222 // we don't need the code
230 Getcpu::Func Getcpu::vdsoFunc() {
231 static Func func = loadVdsoGetcpu();
235 /////////////// SequentialThreadId
238 std::atomic<size_t> SequentialThreadId<std::atomic>::prevId(0);
241 FOLLY_TLS size_t SequentialThreadId<std::atomic>::currentId(0);
243 /////////////// AccessSpreader
246 const AccessSpreader<std::atomic>
247 AccessSpreader<std::atomic>::stripeByCore(
248 CacheLocality::system<>().numCachesByLevel.front());
251 const AccessSpreader<std::atomic>
252 AccessSpreader<std::atomic>::stripeByChip(
253 CacheLocality::system<>().numCachesByLevel.back());
256 AccessSpreaderArray<std::atomic,128>
257 AccessSpreaderArray<std::atomic,128>::sharedInstance = {};
259 /// Always claims to be on CPU zero, node zero
260 static int degenerateGetcpu(unsigned* cpu, unsigned* node, void* unused) {
261 if (cpu != nullptr) {
264 if (node != nullptr) {
271 Getcpu::Func AccessSpreader<std::atomic>::pickGetcpuFunc(size_t numStripes) {
272 if (numStripes == 1) {
273 // there's no need to call getcpu if there is only one stripe.
274 // This should not be common, so we don't want to waste a test and
275 // branch in the main code path, but we might as well use a faster
277 return °enerateGetcpu;
279 auto best = Getcpu::vdsoFunc();
280 return best ? best : &SequentialThreadId<std::atomic>::getcpu;
284 } } // namespace folly::detail