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.
27 #include <type_traits>
28 #include <unordered_map>
31 #include <folly/Indestructible.h>
32 #include <folly/Likely.h>
33 #include <folly/Memory.h>
34 #include <folly/Portability.h>
35 #include <folly/ThreadId.h>
36 #include <folly/hash/Hash.h>
37 #include <folly/portability/BitsFunctexcept.h>
38 #include <folly/portability/Memory.h>
42 // This file contains several classes that might be useful if you are
43 // trying to dynamically optimize cache locality: CacheLocality reads
44 // cache sharing information from sysfs to determine how CPUs should be
45 // grouped to minimize contention, Getcpu provides fast access to the
46 // current CPU via __vdso_getcpu, and AccessSpreader uses these two to
47 // optimally spread accesses among a predetermined number of stripes.
49 // AccessSpreader<>::current(n) microbenchmarks at 22 nanos, which is
50 // substantially less than the cost of a cache miss. This means that we
51 // can effectively use it to reduce cache line ping-pong on striped data
52 // structures such as IndexedMemPool or statistics counters.
54 // Because CacheLocality looks at all of the cache levels, it can be
55 // used for different levels of optimization. AccessSpreader(2) does
56 // per-chip spreading on a dual socket system. AccessSpreader(numCpus)
57 // does perfect per-cpu spreading. AccessSpreader(numCpus / 2) does
58 // perfect L1 spreading in a system with hyperthreading enabled.
60 struct CacheLocality {
61 /// 1 more than the maximum value that can be returned from sched_getcpu
62 /// or getcpu. This is the number of hardware thread contexts provided
66 /// Holds the number of caches present at each cache level (0 is
67 /// the closest to the cpu). This is the number of AccessSpreader
68 /// stripes needed to avoid cross-cache communication at the specified
69 /// layer. numCachesByLevel.front() is the number of L1 caches and
70 /// numCachesByLevel.back() is the number of last-level caches.
71 std::vector<size_t> numCachesByLevel;
73 /// A map from cpu (from sched_getcpu or getcpu) to an index in the
74 /// range 0..numCpus-1, where neighboring locality indices are more
75 /// likely to share caches then indices far away. All of the members
76 /// of a particular cache level be contiguous in their locality index.
77 /// For example, if numCpus is 32 and numCachesByLevel.back() is 2,
78 /// then cpus with a locality index < 16 will share one last-level
79 /// cache and cpus with a locality index >= 16 will share the other.
80 std::vector<size_t> localityIndexByCpu;
82 /// Returns the best CacheLocality information available for the current
83 /// system, cached for fast access. This will be loaded from sysfs if
84 /// possible, otherwise it will be correct in the number of CPUs but
85 /// not in their sharing structure.
87 /// If you are into yo dawgs, this is a shared cache of the local
88 /// locality of the shared caches.
90 /// The template parameter here is used to allow injection of a
91 /// repeatable CacheLocality structure during testing. Rather than
92 /// inject the type of the CacheLocality provider into every data type
93 /// that transitively uses it, all components select between the default
94 /// sysfs implementation and a deterministic implementation by keying
95 /// off the type of the underlying atomic. See DeterministicScheduler.
96 template <template <typename> class Atom = std::atomic>
97 static const CacheLocality& system();
99 /// Reads CacheLocality information from a tree structured like
100 /// the sysfs filesystem. The provided function will be evaluated
101 /// for each sysfs file that needs to be queried. The function
102 /// should return a string containing the first line of the file
103 /// (not including the newline), or an empty string if the file does
104 /// not exist. The function will be called with paths of the form
105 /// /sys/devices/system/cpu/cpu*/cache/index*/{type,shared_cpu_list} .
106 /// Throws an exception if no caches can be parsed at all.
107 static CacheLocality readFromSysfsTree(
108 const std::function<std::string(std::string)>& mapping);
110 /// Reads CacheLocality information from the real sysfs filesystem.
111 /// Throws an exception if no cache information can be loaded.
112 static CacheLocality readFromSysfs();
114 /// Returns a usable (but probably not reflective of reality)
115 /// CacheLocality structure with the specified number of cpus and a
116 /// single cache level that associates one cpu per cache.
117 static CacheLocality uniform(size_t numCpus);
120 /// Memory locations on the same cache line are subject to false
121 /// sharing, which is very bad for performance. Microbenchmarks
122 /// indicate that pairs of cache lines also see interference under
123 /// heavy use of atomic operations (observed for atomic increment on
124 /// Sandy Bridge). See FOLLY_ALIGN_TO_AVOID_FALSE_SHARING
125 kFalseSharingRange = 128
129 kFalseSharingRange == 128,
130 "FOLLY_ALIGN_TO_AVOID_FALSE_SHARING should track kFalseSharingRange");
133 // TODO replace __attribute__ with alignas and 128 with kFalseSharingRange
135 /// An attribute that will cause a variable or field to be aligned so that
136 /// it doesn't have false sharing with anything at a smaller memory address.
137 #define FOLLY_ALIGN_TO_AVOID_FALSE_SHARING FOLLY_ALIGNED(128)
139 /// Knows how to derive a function pointer to the VDSO implementation of
140 /// getcpu(2), if available
142 /// Function pointer to a function with the same signature as getcpu(2).
143 typedef int (*Func)(unsigned* cpu, unsigned* node, void* unused);
145 /// Returns a pointer to the VDSO implementation of getcpu(2), if
146 /// available, or nullptr otherwise. This function may be quite
147 /// expensive, be sure to cache the result.
148 static Func resolveVdsoFunc();
152 template <template <typename> class Atom>
153 struct SequentialThreadId {
154 /// Returns the thread id assigned to the current thread
155 static unsigned get() {
157 if (UNLIKELY(rv == 0)) {
158 rv = currentId = ++prevId;
164 static Atom<unsigned> prevId;
166 static FOLLY_TLS unsigned currentId;
169 template <template <typename> class Atom>
170 Atom<unsigned> SequentialThreadId<Atom>::prevId(0);
172 template <template <typename> class Atom>
173 FOLLY_TLS unsigned SequentialThreadId<Atom>::currentId(0);
175 // Suppress this instantiation in other translation units. It is
176 // instantiated in CacheLocality.cpp
177 extern template struct SequentialThreadId<std::atomic>;
180 struct HashingThreadId {
181 static unsigned get() {
182 return hash::twang_32from64(getCurrentThreadID());
186 /// A class that lazily binds a unique (for each implementation of Atom)
187 /// identifier to a thread. This is a fallback mechanism for the access
188 /// spreader if __vdso_getcpu can't be loaded
189 template <typename ThreadId>
190 struct FallbackGetcpu {
191 /// Fills the thread id into the cpu and node out params (if they
192 /// are non-null). This method is intended to act like getcpu when a
193 /// fast-enough form of getcpu isn't available or isn't desired
194 static int getcpu(unsigned* cpu, unsigned* node, void* /* unused */) {
195 auto id = ThreadId::get();
207 typedef FallbackGetcpu<SequentialThreadId<std::atomic>> FallbackGetcpuType;
209 typedef FallbackGetcpu<HashingThreadId> FallbackGetcpuType;
212 /// AccessSpreader arranges access to a striped data structure in such a
213 /// way that concurrently executing threads are likely to be accessing
214 /// different stripes. It does NOT guarantee uncontended access.
215 /// Your underlying algorithm must be thread-safe without spreading, this
216 /// is merely an optimization. AccessSpreader::current(n) is typically
217 /// much faster than a cache miss (12 nanos on my dev box, tested fast
218 /// in both 2.6 and 3.2 kernels).
220 /// If available (and not using the deterministic testing implementation)
221 /// AccessSpreader uses the getcpu system call via VDSO and the
222 /// precise locality information retrieved from sysfs by CacheLocality.
223 /// This provides optimal anti-sharing at a fraction of the cost of a
226 /// When there are not as many stripes as processors, we try to optimally
227 /// place the cache sharing boundaries. This means that if you have 2
228 /// stripes and run on a dual-socket system, your 2 stripes will each get
229 /// all of the cores from a single socket. If you have 16 stripes on a
230 /// 16 core system plus hyperthreading (32 cpus), each core will get its
231 /// own stripe and there will be no cache sharing at all.
233 /// AccessSpreader has a fallback mechanism for when __vdso_getcpu can't be
234 /// loaded, or for use during deterministic testing. Using sched_getcpu
235 /// or the getcpu syscall would negate the performance advantages of
236 /// access spreading, so we use a thread-local value and a shared atomic
237 /// counter to spread access out. On systems lacking both a fast getcpu()
238 /// and TLS, we hash the thread id to spread accesses.
240 /// AccessSpreader is templated on the template type that is used
241 /// to implement atomics, as a way to instantiate the underlying
242 /// heuristics differently for production use and deterministic unit
243 /// testing. See DeterministicScheduler for more. If you aren't using
244 /// DeterministicScheduler, you can just use the default template parameter
246 template <template <typename> class Atom = std::atomic>
247 struct AccessSpreader {
248 /// Returns the stripe associated with the current CPU. The returned
249 /// value will be < numStripes.
250 static size_t current(size_t numStripes) {
251 // widthAndCpuToStripe[0] will actually work okay (all zeros), but
252 // something's wrong with the caller
253 assert(numStripes > 0);
256 getcpuFunc(&cpu, nullptr, nullptr);
257 return widthAndCpuToStripe[std::min(size_t(kMaxCpus), numStripes)]
262 /// If there are more cpus than this nothing will crash, but there
263 /// might be unnecessary sharing
264 enum { kMaxCpus = 128 };
266 typedef uint8_t CompactStripe;
269 (kMaxCpus & (kMaxCpus - 1)) == 0,
270 "kMaxCpus should be a power of two so modulo is fast");
272 kMaxCpus - 1 <= std::numeric_limits<CompactStripe>::max(),
273 "stripeByCpu element type isn't wide enough");
275 /// Points to the getcpu-like function we are using to obtain the
276 /// current cpu. It should not be assumed that the returned cpu value
277 /// is in range. We use a static for this so that we can prearrange a
278 /// valid value in the pre-constructed state and avoid the need for a
279 /// conditional on every subsequent invocation (not normally a big win,
280 /// but 20% on some inner loops here).
281 static Getcpu::Func getcpuFunc;
283 /// For each level of splitting up to kMaxCpus, maps the cpu (mod
284 /// kMaxCpus) to the stripe. Rather than performing any inequalities
285 /// or modulo on the actual number of cpus, we just fill in the entire
287 static CompactStripe widthAndCpuToStripe[kMaxCpus + 1][kMaxCpus];
289 static bool initialized;
291 /// Returns the best getcpu implementation for Atom
292 static Getcpu::Func pickGetcpuFunc() {
293 auto best = Getcpu::resolveVdsoFunc();
294 return best ? best : &FallbackGetcpuType::getcpu;
297 /// Always claims to be on CPU zero, node zero
298 static int degenerateGetcpu(unsigned* cpu, unsigned* node, void*) {
299 if (cpu != nullptr) {
302 if (node != nullptr) {
308 // The function to call for fast lookup of getcpu is a singleton, as
309 // is the precomputed table of locality information. AccessSpreader
310 // is used in very tight loops, however (we're trying to race an L1
311 // cache miss!), so the normal singleton mechanisms are noticeably
312 // expensive. Even a not-taken branch guarding access to getcpuFunc
313 // slows AccessSpreader::current from 12 nanos to 14. As a result, we
314 // populate the static members with simple (but valid) values that can
315 // be filled in by the linker, and then follow up with a normal static
316 // initializer call that puts in the proper version. This means that
317 // when there are initialization order issues we will just observe a
318 // zero stripe. Once a sanitizer gets smart enough to detect this as
319 // a race or undefined behavior, we can annotate it.
321 static bool initialize() {
322 getcpuFunc = pickGetcpuFunc();
324 auto& cacheLocality = CacheLocality::system<Atom>();
325 auto n = cacheLocality.numCpus;
326 for (size_t width = 0; width <= kMaxCpus; ++width) {
327 auto numStripes = std::max(size_t{1}, width);
328 for (size_t cpu = 0; cpu < kMaxCpus && cpu < n; ++cpu) {
329 auto index = cacheLocality.localityIndexByCpu[cpu];
331 // as index goes from 0..n, post-transform value goes from
333 widthAndCpuToStripe[width][cpu] =
334 CompactStripe((index * numStripes) / n);
335 assert(widthAndCpuToStripe[width][cpu] < numStripes);
337 for (size_t cpu = n; cpu < kMaxCpus; ++cpu) {
338 widthAndCpuToStripe[width][cpu] = widthAndCpuToStripe[width][cpu - n];
345 template <template <typename> class Atom>
346 Getcpu::Func AccessSpreader<Atom>::getcpuFunc =
347 AccessSpreader<Atom>::degenerateGetcpu;
349 template <template <typename> class Atom>
350 typename AccessSpreader<Atom>::CompactStripe
351 AccessSpreader<Atom>::widthAndCpuToStripe[kMaxCpus + 1][kMaxCpus] = {};
353 template <template <typename> class Atom>
354 bool AccessSpreader<Atom>::initialized = AccessSpreader<Atom>::initialize();
356 // Suppress this instantiation in other translation units. It is
357 // instantiated in CacheLocality.cpp
358 extern template struct AccessSpreader<std::atomic>;
361 * A simple freelist allocator. Allocates things of size sz, from
362 * slabs of size allocSize. Takes a lock on each
363 * allocation/deallocation.
365 class SimpleAllocator {
367 uint8_t* mem_{nullptr};
368 uint8_t* end_{nullptr};
369 void* freelist_{nullptr};
372 std::vector<void*> blocks_;
375 SimpleAllocator(size_t allocSize, size_t sz);
377 void* allocateHard();
379 // Inline fast-paths.
381 std::lock_guard<std::mutex> g(m_);
382 // Freelist allocation.
384 auto mem = freelist_;
385 freelist_ = *static_cast<void**>(freelist_);
389 // Bump-ptr allocation.
390 if (intptr_t(mem_) % 128 == 0) {
391 // Avoid allocating pointers that may look like malloc
393 mem_ += std::min(sz_, folly::max_align_v);
395 if (mem_ && (mem_ + sz_ <= end_)) {
399 assert(intptr_t(mem) % 128 != 0);
403 return allocateHard();
405 void deallocate(void* mem) {
406 std::lock_guard<std::mutex> g(m_);
407 *static_cast<void**>(mem) = freelist_;
413 * An allocator that can be used with CacheLocality to allocate
416 * There is actually nothing special about the memory itself (it is
417 * not bound to numa nodes or anything), but the allocator guarantees
418 * that memory allocatd from the same stripe will only come from cache
419 * lines also allocated to the same stripe. This means multiple
420 * things using CacheLocality can allocate memory in smaller-than
421 * cacheline increments, and be assured that it won't cause more false
422 * sharing than it otherwise would.
424 * Note that allocation and deallocation takes a per-sizeclass lock.
426 template <size_t Stripes>
427 class CoreAllocator {
430 static constexpr size_t AllocSize{4096};
432 uint8_t sizeClass(size_t size) {
435 } else if (size <= 16) {
437 } else if (size <= 32) {
439 } else if (size <= 64) {
441 } else { // punt to malloc.
446 std::array<SimpleAllocator, 4> allocators_{
447 {{AllocSize, 8}, {AllocSize, 16}, {AllocSize, 32}, {AllocSize, 64}}};
450 void* allocate(size_t size) {
451 auto cl = sizeClass(size);
454 CacheLocality::kFalseSharingRange == 128,
455 "kFalseSharingRange changed");
456 // Align to a cacheline
457 size = size + (CacheLocality::kFalseSharingRange - 1);
458 size &= ~size_t(CacheLocality::kFalseSharingRange - 1);
460 detail::aligned_malloc(size, CacheLocality::kFalseSharingRange);
462 std::__throw_bad_alloc();
466 return allocators_[cl].allocate();
468 void deallocate(void* mem) {
473 // See if it came from this allocator or malloc.
474 if (intptr_t(mem) % 128 != 0) {
476 reinterpret_cast<void*>(intptr_t(mem) & ~intptr_t(AllocSize - 1));
477 auto allocator = *static_cast<SimpleAllocator**>(addr);
478 allocator->deallocate(mem);
480 detail::aligned_free(mem);
485 Allocator* get(size_t stripe) {
486 assert(stripe < Stripes);
487 return &allocators_[stripe];
491 Allocator allocators_[Stripes];
494 template <size_t Stripes>
495 typename CoreAllocator<Stripes>::Allocator* getCoreAllocator(size_t stripe) {
496 // We cannot make sure that the allocator will be destroyed after
497 // all the objects allocated with it, so we leak it.
498 static Indestructible<CoreAllocator<Stripes>> allocator;
499 return allocator->get(stripe);
502 template <typename T, size_t Stripes>
503 StlAllocator<typename CoreAllocator<Stripes>::Allocator, T> getCoreAllocatorStl(
505 auto alloc = getCoreAllocator<Stripes>(stripe);
506 return StlAllocator<typename CoreAllocator<Stripes>::Allocator, T>(alloc);