summaryrefslogtreecommitdiff
path: root/absl/container/internal
diff options
context:
space:
mode:
authorGravatar Abseil Team <absl-team@google.com>2018-12-20 12:29:59 -0800
committerGravatar Xiaoyi Zhang <zhangxy988@gmail.com>2018-12-21 14:43:20 -0500
commit968a34ffdaadd7db062a9621dfbdf8b2d16e05af (patch)
tree6db3f5237087d2c51b264ecb33cd57f1e17b69b9 /absl/container/internal
parent3e2e9b5557e76d098de4b8a2a659125b98ca519b (diff)
Export of internal Abseil changes.
-- 7fa1107161a03dac53fb84c2b06d8092616c7b13 by Abseil Team <absl-team@google.com>: Harden the generic stacktrace implementation for use during early program execution PiperOrigin-RevId: 226375950 -- 079f9969329f5eb66f647dd3c44b17541b1bf217 by Matt Kulukundis <kfm@google.com>: Workaround platforms that have over-aggressive warnings on -Wexit-time-destructors PiperOrigin-RevId: 226362948 -- 1447943f509be681ca5495add0162c750ef237f1 by Matt Kulukundis <kfm@google.com>: Switch from 64 to size_t atomics so they work on embedded platforms that do not have 64 bit atomics. PiperOrigin-RevId: 226210704 -- d14d49837ae2bcde74051e0c79c18ee0f43866b9 by Tom Manshreck <shreck@google.com>: Develop initial documentation for API breaking changes process: PiperOrigin-RevId: 226210021 -- 7ea3d7fe0e86979dab83a5fc9cc3bf1d6cb3bd53 by Abseil Team <absl-team@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 226195522 -- 7de873e880d7f016a4fa1e08d626f0535cc470af by Abseil Team <absl-team@google.com>: Make Abseil LICENSE files newline terminated, with a single trailing blank line. Also remove line-ending whitespace. PiperOrigin-RevId: 226182949 -- 7d00643fadfad7f0d992c68bd9d2ed2e5bc960b0 by Matt Kulukundis <kfm@google.com>: Internal cleanup PiperOrigin-RevId: 226045282 -- c4a0a11c0ce2875271191e477f3d36eaaeca4613 by Matt Kulukundis <kfm@google.com>: Internal cleanup PiperOrigin-RevId: 226038273 -- 8ee4ebbb1ae5cda119e436e5ff7e3aa966608c10 by Matt Kulukundis <kfm@google.com>: Adds a global sampler which tracks a fraction of live tables for collecting telemetry data. PiperOrigin-RevId: 226032080 -- d576446f050518cd1b0ae447d682d8552f0e7e30 by Mark Barolak <mbar@google.com>: Replace an internal CaseEqual function with calls to the identical absl::EqualsIgnoreCase. This closes out a rather old TODO. PiperOrigin-RevId: 226024779 -- 6b23f1ee028a5ffa608c920424f1220a117a8f3d by Abseil Team <absl-team@google.com>: Add December 2018 LTS branch to list of LTS branches. PiperOrigin-RevId: 226011333 -- bb0833a43bdaef4c8c059b17bcd27ba9a085a114 by Mark Barolak <mbar@google.com>: Explicitly state that when the SimpleAtoi family of functions encounter an error, the value of their output parameter is unspecified. Also standardize the name of the output parameter to be `out`. PiperOrigin-RevId: 225997035 -- 46c1876b1a248eabda7545daa61a74a4cdfe9077 by Abseil Team <absl-team@google.com>: Remove deprecated CMake function absl_test, absl_library and absl_header_library PiperOrigin-RevId: 225950041 GitOrigin-RevId: 7fa1107161a03dac53fb84c2b06d8092616c7b13 Change-Id: I2ca9d3aada9292614527d1339a7557494139b806
Diffstat (limited to 'absl/container/internal')
-rw-r--r--absl/container/internal/hashtablez_sampler.cc289
-rw-r--r--absl/container/internal/hashtablez_sampler.h236
-rw-r--r--absl/container/internal/hashtablez_sampler_test.cc307
-rw-r--r--absl/container/internal/raw_hash_set.h53
-rw-r--r--absl/container/internal/raw_hash_set_test.cc22
5 files changed, 892 insertions, 15 deletions
diff --git a/absl/container/internal/hashtablez_sampler.cc b/absl/container/internal/hashtablez_sampler.cc
new file mode 100644
index 00000000..6cc10c20
--- /dev/null
+++ b/absl/container/internal/hashtablez_sampler.cc
@@ -0,0 +1,289 @@
+// Copyright 2018 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/container/internal/hashtablez_sampler.h"
+
+#include <atomic>
+#include <cassert>
+#include <functional>
+#include <limits>
+
+#include "absl/base/attributes.h"
+#include "absl/container/internal/have_sse.h"
+#include "absl/debugging/stacktrace.h"
+#include "absl/memory/memory.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+namespace container_internal {
+constexpr int HashtablezInfo::kMaxStackDepth;
+
+namespace {
+ABSL_CONST_INIT std::atomic<bool> g_hashtablez_enabled{
+ false
+};
+ABSL_CONST_INIT std::atomic<int32_t> g_hashtablez_sample_parameter{1 << 10};
+ABSL_CONST_INIT std::atomic<int32_t> g_hashtablez_max_samples{1 << 20};
+
+// Returns the next pseudo-random value.
+// pRNG is: aX+b mod c with a = 0x5DEECE66D, b = 0xB, c = 1<<48
+// This is the lrand64 generator.
+uint64_t NextRandom(uint64_t rnd) {
+ const uint64_t prng_mult = uint64_t{0x5DEECE66D};
+ const uint64_t prng_add = 0xB;
+ const uint64_t prng_mod_power = 48;
+ const uint64_t prng_mod_mask = ~(~uint64_t{0} << prng_mod_power);
+ return (prng_mult * rnd + prng_add) & prng_mod_mask;
+}
+
+// Generates a geometric variable with the specified mean.
+// This is done by generating a random number between 0 and 1 and applying
+// the inverse cumulative distribution function for an exponential.
+// Specifically: Let m be the inverse of the sample period, then
+// the probability distribution function is m*exp(-mx) so the CDF is
+// p = 1 - exp(-mx), so
+// q = 1 - p = exp(-mx)
+// log_e(q) = -mx
+// -log_e(q)/m = x
+// log_2(q) * (-log_e(2) * 1/m) = x
+// In the code, q is actually in the range 1 to 2**26, hence the -26 below
+//
+int64_t GetGeometricVariable(int64_t mean) {
+#if ABSL_HAVE_THREAD_LOCAL
+ thread_local
+#else // ABSL_HAVE_THREAD_LOCAL
+ // SampleSlow and hence GetGeometricVariable is guarded by a single mutex when
+ // there are not thread locals. Thus, a single global rng is acceptable for
+ // that case.
+ static
+#endif // ABSL_HAVE_THREAD_LOCAL
+ uint64_t rng = []() {
+ // We don't get well distributed numbers from this so we call
+ // NextRandom() a bunch to mush the bits around. We use a global_rand
+ // to handle the case where the same thread (by memory address) gets
+ // created and destroyed repeatedly.
+ ABSL_CONST_INIT static std::atomic<uint32_t> global_rand(0);
+ uint64_t r = reinterpret_cast<uint64_t>(&rng) +
+ global_rand.fetch_add(1, std::memory_order_relaxed);
+ for (int i = 0; i < 20; ++i) {
+ r = NextRandom(r);
+ }
+ return r;
+ }();
+
+ rng = NextRandom(rng);
+
+ // Take the top 26 bits as the random number
+ // (This plus the 1<<58 sampling bound give a max possible step of
+ // 5194297183973780480 bytes.)
+ const uint64_t prng_mod_power = 48; // Number of bits in prng
+ // The uint32_t cast is to prevent a (hard-to-reproduce) NAN
+ // under piii debug for some binaries.
+ double q = static_cast<uint32_t>(rng >> (prng_mod_power - 26)) + 1.0;
+ // Put the computed p-value through the CDF of a geometric.
+ double interval = (std::log2(q) - 26) * (-std::log(2.0) * mean);
+
+ // Very large values of interval overflow int64_t. If we happen to
+ // hit such improbable condition, we simply cheat and clamp interval
+ // to largest supported value.
+ if (interval > static_cast<double>(std::numeric_limits<int64_t>::max() / 2)) {
+ return std::numeric_limits<int64_t>::max() / 2;
+ }
+
+ // Small values of interval are equivalent to just sampling next time.
+ if (interval < 1) {
+ return 1;
+ }
+ return static_cast<int64_t>(interval);
+}
+
+} // namespace
+
+HashtablezSampler& HashtablezSampler::Global() {
+ static auto* sampler = new HashtablezSampler();
+ return *sampler;
+}
+
+HashtablezInfo::HashtablezInfo() { PrepareForSampling(); }
+HashtablezInfo::~HashtablezInfo() = default;
+
+void HashtablezInfo::PrepareForSampling() {
+ capacity.store(0, std::memory_order_relaxed);
+ size.store(0, std::memory_order_relaxed);
+ num_erases.store(0, std::memory_order_relaxed);
+ max_probe_length.store(0, std::memory_order_relaxed);
+ total_probe_length.store(0, std::memory_order_relaxed);
+ hashes_bitwise_or.store(0, std::memory_order_relaxed);
+ hashes_bitwise_and.store(~size_t{}, std::memory_order_relaxed);
+
+ create_time = absl::Now();
+ // The inliner makes hardcoded skip_count difficult (especially when combined
+ // with LTO). We use the ability to exclude stacks by regex when encoding
+ // instead.
+ depth = absl::GetStackTrace(stack, HashtablezInfo::kMaxStackDepth,
+ /* skip_count= */ 0);
+ dead = nullptr;
+}
+
+HashtablezSampler::HashtablezSampler()
+ : dropped_samples_(0), size_estimate_(0), all_(nullptr) {
+ absl::MutexLock l(&graveyard_.init_mu);
+ graveyard_.dead = &graveyard_;
+}
+
+HashtablezSampler::~HashtablezSampler() {
+ HashtablezInfo* s = all_.load(std::memory_order_acquire);
+ while (s != nullptr) {
+ HashtablezInfo* next = s->next;
+ delete s;
+ s = next;
+ }
+}
+
+void HashtablezSampler::PushNew(HashtablezInfo* sample) {
+ sample->next = all_.load(std::memory_order_relaxed);
+ while (!all_.compare_exchange_weak(sample->next, sample,
+ std::memory_order_release,
+ std::memory_order_relaxed)) {
+ }
+}
+
+void HashtablezSampler::PushDead(HashtablezInfo* sample) {
+ absl::MutexLock graveyard_lock(&graveyard_.init_mu);
+ absl::MutexLock sample_lock(&sample->init_mu);
+ sample->dead = graveyard_.dead;
+ graveyard_.dead = sample;
+}
+
+HashtablezInfo* HashtablezSampler::PopDead() {
+ absl::MutexLock graveyard_lock(&graveyard_.init_mu);
+
+ // The list is circular, so eventually it collapses down to
+ // graveyard_.dead == &graveyard_
+ // when it is empty.
+ HashtablezInfo* sample = graveyard_.dead;
+ if (sample == &graveyard_) return nullptr;
+
+ absl::MutexLock sample_lock(&sample->init_mu);
+ graveyard_.dead = sample->dead;
+ sample->PrepareForSampling();
+ return sample;
+}
+
+HashtablezInfo* HashtablezSampler::Register() {
+ int64_t size = size_estimate_.fetch_add(1, std::memory_order_relaxed);
+ if (size > g_hashtablez_max_samples.load(std::memory_order_relaxed)) {
+ size_estimate_.fetch_sub(1, std::memory_order_relaxed);
+ dropped_samples_.fetch_add(1, std::memory_order_relaxed);
+ return nullptr;
+ }
+
+ HashtablezInfo* sample = PopDead();
+ if (sample == nullptr) {
+ // Resurrection failed. Hire a new warlock.
+ sample = new HashtablezInfo();
+ PushNew(sample);
+ }
+
+ return sample;
+}
+
+void HashtablezSampler::Unregister(HashtablezInfo* sample) {
+ PushDead(sample);
+ size_estimate_.fetch_sub(1, std::memory_order_relaxed);
+}
+
+int64_t HashtablezSampler::Iterate(
+ const std::function<void(const HashtablezInfo& stack)>& f) {
+ HashtablezInfo* s = all_.load(std::memory_order_acquire);
+ while (s != nullptr) {
+ absl::MutexLock l(&s->init_mu);
+ if (s->dead == nullptr) {
+ f(*s);
+ }
+ s = s->next;
+ }
+
+ return dropped_samples_.load(std::memory_order_relaxed);
+}
+
+HashtablezInfo* SampleSlow(int64_t* next_sample) {
+ bool first = *next_sample < 0;
+ *next_sample = GetGeometricVariable(
+ g_hashtablez_sample_parameter.load(std::memory_order_relaxed));
+
+ // g_hashtablez_enabled can be dynamically flipped, we need to set a threshold
+ // low enough that we will start sampling in a reasonable time, so we just use
+ // the default sampling rate.
+ if (!g_hashtablez_enabled.load(std::memory_order_relaxed)) return nullptr;
+
+ // We will only be negative on our first count, so we should just retry in
+ // that case.
+ if (first) {
+ if (ABSL_PREDICT_TRUE(--*next_sample > 0)) return nullptr;
+ return SampleSlow(next_sample);
+ }
+
+ return HashtablezSampler::Global().Register();
+}
+
+void UnsampleSlow(HashtablezInfo* info) {
+ HashtablezSampler::Global().Unregister(info);
+}
+
+void RecordInsertSlow(HashtablezInfo* info, size_t hash,
+ size_t distance_from_desired) {
+ // SwissTables probe in groups of 16, so scale this to count items probes and
+ // not offset from desired.
+ size_t probe_length = distance_from_desired;
+#if SWISSTABLE_HAVE_SSE2
+ probe_length /= 16;
+#else
+ probe_length /= 8;
+#endif
+
+ info->hashes_bitwise_and.fetch_and(hash, std::memory_order_relaxed);
+ info->hashes_bitwise_or.fetch_or(hash, std::memory_order_relaxed);
+ info->max_probe_length.store(
+ std::max(info->max_probe_length.load(std::memory_order_relaxed),
+ probe_length),
+ std::memory_order_relaxed);
+ info->total_probe_length.fetch_add(probe_length, std::memory_order_relaxed);
+ info->size.fetch_add(1, std::memory_order_relaxed);
+}
+
+void SetHashtablezEnabled(bool enabled) {
+ g_hashtablez_enabled.store(enabled, std::memory_order_release);
+}
+
+void SetHashtablezSampleParameter(int32_t rate) {
+ if (rate > 0) {
+ g_hashtablez_sample_parameter.store(rate, std::memory_order_release);
+ } else {
+ ABSL_RAW_LOG(ERROR, "Invalid hashtablez sample rate: %lld",
+ static_cast<long long>(rate)); // NOLINT(runtime/int)
+ }
+}
+
+void SetHashtablezMaxSamples(int32_t max) {
+ if (max > 0) {
+ g_hashtablez_max_samples.store(max, std::memory_order_release);
+ } else {
+ ABSL_RAW_LOG(ERROR, "Invalid hashtablez max samples: %lld",
+ static_cast<long long>(max)); // NOLINT(runtime/int)
+ }
+}
+
+} // namespace container_internal
+} // namespace absl
diff --git a/absl/container/internal/hashtablez_sampler.h b/absl/container/internal/hashtablez_sampler.h
new file mode 100644
index 00000000..4aea3ffa
--- /dev/null
+++ b/absl/container/internal/hashtablez_sampler.h
@@ -0,0 +1,236 @@
+// Copyright 2018 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// This is a low level library to sample hashtables and collect runtime
+// statistics about them.
+//
+// `HashtablezSampler` controls the lifecycle of `HashtablezInfo` objects which
+// store information about a single sample.
+//
+// `Record*` methods store information into samples.
+// `Sample()` and `Unsample()` make use of a single global sampler with
+// properties controlled by the flags hashtablez_enabled,
+// hashtablez_sample_rate, and hashtablez_max_samples.
+
+#ifndef ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
+#define ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
+
+#include <atomic>
+#include <functional>
+#include <memory>
+#include <vector>
+
+#include "absl/base/optimization.h"
+#include "absl/synchronization/mutex.h"
+#include "absl/utility/utility.h"
+
+namespace absl {
+namespace container_internal {
+
+// Stores information about a sampled hashtable. All mutations to this *must*
+// be made through `Record*` functions below. All reads from this *must* only
+// occur in the callback to `HashtablezSampler::Iterate`.
+struct HashtablezInfo {
+ // Constructs the object but does not fill in any fields.
+ HashtablezInfo();
+ ~HashtablezInfo();
+ HashtablezInfo(const HashtablezInfo&) = delete;
+ HashtablezInfo& operator=(const HashtablezInfo&) = delete;
+
+ // Puts the object into a clean state, fills in the logically `const` members,
+ // blocking for any readers that are currently sampling the object.
+ void PrepareForSampling() EXCLUSIVE_LOCKS_REQUIRED(init_mu);
+
+ // These fields are mutated by the various Record* APIs and need to be
+ // thread-safe.
+ std::atomic<size_t> capacity;
+ std::atomic<size_t> size;
+ std::atomic<size_t> num_erases;
+ std::atomic<size_t> max_probe_length;
+ std::atomic<size_t> total_probe_length;
+ std::atomic<size_t> hashes_bitwise_or;
+ std::atomic<size_t> hashes_bitwise_and;
+
+ // `HashtablezSampler` maintains intrusive linked lists for all samples. See
+ // comments on `HashtablezSampler::all_` for details on these. `init_mu`
+ // guards the ability to restore the sample to a pristine state. This
+ // prevents races with sampling and resurrecting an object.
+ absl::Mutex init_mu;
+ HashtablezInfo* next;
+ HashtablezInfo* dead GUARDED_BY(init_mu);
+
+ // All of the fields below are set by `PrepareForSampling`, they must not be
+ // mutated in `Record*` functions. They are logically `const` in that sense.
+ // These are guarded by init_mu, but that is not externalized to clients, who
+ // can only read them during `HashtablezSampler::Iterate` which will hold the
+ // lock.
+ static constexpr int kMaxStackDepth = 64;
+ absl::Time create_time;
+ int32_t depth;
+ void* stack[kMaxStackDepth];
+};
+
+inline void RecordStorageChangedSlow(HashtablezInfo* info, size_t size,
+ size_t capacity) {
+ info->size.store(size, std::memory_order_relaxed);
+ info->capacity.store(capacity, std::memory_order_relaxed);
+}
+
+void RecordInsertSlow(HashtablezInfo* info, size_t hash,
+ size_t distance_from_desired);
+
+inline void RecordEraseSlow(HashtablezInfo* info) {
+ info->size.fetch_sub(1, std::memory_order_relaxed);
+ info->num_erases.fetch_add(1, std::memory_order_relaxed);
+}
+
+HashtablezInfo* SampleSlow(int64_t* next_sample);
+void UnsampleSlow(HashtablezInfo* info);
+
+class HashtablezInfoHandle {
+ public:
+ explicit HashtablezInfoHandle() : info_(nullptr) {}
+ explicit HashtablezInfoHandle(HashtablezInfo* info) : info_(info) {}
+ ~HashtablezInfoHandle() {
+ if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
+ UnsampleSlow(info_);
+ }
+
+ HashtablezInfoHandle(const HashtablezInfoHandle&) = delete;
+ HashtablezInfoHandle& operator=(const HashtablezInfoHandle&) = delete;
+
+ HashtablezInfoHandle(HashtablezInfoHandle&& o) noexcept
+ : info_(absl::exchange(o.info_, nullptr)) {}
+ HashtablezInfoHandle& operator=(HashtablezInfoHandle&& o) noexcept {
+ if (ABSL_PREDICT_FALSE(info_ != nullptr)) {
+ UnsampleSlow(info_);
+ }
+ info_ = absl::exchange(o.info_, nullptr);
+ return *this;
+ }
+
+ inline void RecordStorageChanged(size_t size, size_t capacity) {
+ if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
+ RecordStorageChangedSlow(info_, size, capacity);
+ }
+
+ inline void RecordInsert(size_t hash, size_t distance_from_desired) {
+ if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
+ RecordInsertSlow(info_, hash, distance_from_desired);
+ }
+
+ inline void RecordErase() {
+ if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
+ RecordEraseSlow(info_);
+ }
+
+ friend inline void swap(HashtablezInfoHandle& lhs,
+ HashtablezInfoHandle& rhs) {
+ std::swap(lhs.info_, rhs.info_);
+ }
+
+ private:
+ friend class HashtablezInfoHandlePeer;
+ HashtablezInfo* info_;
+};
+
+// Returns an RAII sampling handle that manages registration and unregistation
+// with the global sampler.
+inline HashtablezInfoHandle Sample() {
+#if ABSL_HAVE_THREAD_LOCAL
+ thread_local int64_t next_sample = 0;
+#else // ABSL_HAVE_THREAD_LOCAL
+ static auto* mu = new absl::Mutex;
+ static int64_t next_sample = 0;
+ absl::MutexLock l(mu);
+#endif // ABSL_HAVE_THREAD_LOCAL
+
+ if (ABSL_PREDICT_TRUE(--next_sample > 0)) {
+ return HashtablezInfoHandle(nullptr);
+ }
+ return HashtablezInfoHandle(SampleSlow(&next_sample));
+}
+
+// Holds samples and their associated stack traces with a soft limit of
+// `SetHashtablezMaxSamples()`.
+//
+// Thread safe.
+class HashtablezSampler {
+ public:
+ // Returns a global Sampler.
+ static HashtablezSampler& Global();
+
+ HashtablezSampler();
+ ~HashtablezSampler();
+
+ // Registers for sampling. Returns an opaque registration info.
+ HashtablezInfo* Register();
+
+ // Unregisters the sample.
+ void Unregister(HashtablezInfo* sample);
+
+ // Iterates over all the registered `StackInfo`s. Returning the number of
+ // samples that have been dropped.
+ int64_t Iterate(const std::function<void(const HashtablezInfo& stack)>& f);
+
+ private:
+ void PushNew(HashtablezInfo* sample);
+ void PushDead(HashtablezInfo* sample);
+ HashtablezInfo* PopDead();
+
+ std::atomic<size_t> dropped_samples_;
+ std::atomic<size_t> size_estimate_;
+
+ // Intrusive lock free linked lists for tracking samples.
+ //
+ // `all_` records all samples (they are never removed from this list) and is
+ // terminated with a `nullptr`.
+ //
+ // `graveyard_.dead` is a circular linked list. When it is empty,
+ // `graveyard_.dead == &graveyard`. The list is circular so that
+ // every item on it (even the last) has a non-null dead pointer. This allows
+ // `Iterate` to determine if a given sample is live or dead using only
+ // information on the sample itself.
+ //
+ // For example, nodes [A, B, C, D, E] with [A, C, E] alive and [B, D] dead
+ // looks like this (G is the Graveyard):
+ //
+ // +---+ +---+ +---+ +---+ +---+
+ // all -->| A |--->| B |--->| C |--->| D |--->| E |
+ // | | | | | | | | | |
+ // +---+ | | +->| |-+ | | +->| |-+ | |
+ // | G | +---+ | +---+ | +---+ | +---+ | +---+
+ // | | | | | |
+ // | | --------+ +--------+ |
+ // +---+ |
+ // ^ |
+ // +--------------------------------------+
+ //
+ std::atomic<HashtablezInfo*> all_;
+ HashtablezInfo graveyard_;
+};
+
+// Enables or disables sampling for Swiss tables.
+void SetHashtablezEnabled(bool enabled);
+
+// Sets the rate at which Swiss tables will be sampled.
+void SetHashtablezSampleParameter(int32_t rate);
+
+// Sets a soft max for the number of samples that will be kept.
+void SetHashtablezMaxSamples(int32_t max);
+
+} // namespace container_internal
+} // namespace absl
+
+#endif // ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
diff --git a/absl/container/internal/hashtablez_sampler_test.cc b/absl/container/internal/hashtablez_sampler_test.cc
new file mode 100644
index 00000000..31e7641a
--- /dev/null
+++ b/absl/container/internal/hashtablez_sampler_test.cc
@@ -0,0 +1,307 @@
+// Copyright 2018 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/container/internal/hashtablez_sampler.h"
+
+#include <atomic>
+#include <limits>
+#include <random>
+
+#include "gmock/gmock.h"
+#include "gtest/gtest.h"
+#include "absl/base/attributes.h"
+#include "absl/container/internal/have_sse.h"
+#include "absl/synchronization/blocking_counter.h"
+#include "absl/synchronization/internal/thread_pool.h"
+#include "absl/synchronization/mutex.h"
+#include "absl/synchronization/notification.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+
+#if SWISSTABLE_HAVE_SSE2
+constexpr int kProbeLength = 16;
+#else
+constexpr int kProbeLength = 8;
+#endif
+
+namespace absl {
+namespace container_internal {
+class HashtablezInfoHandlePeer {
+ public:
+ static bool IsSampled(const HashtablezInfoHandle& h) {
+ return h.info_ != nullptr;
+ }
+
+ static HashtablezInfo* GetInfo(HashtablezInfoHandle* h) { return h->info_; }
+};
+
+namespace {
+using ::absl::synchronization_internal::ThreadPool;
+using ::testing::IsEmpty;
+using ::testing::UnorderedElementsAre;
+
+std::vector<size_t> GetSizes(HashtablezSampler* s) {
+ std::vector<size_t> res;
+ s->Iterate([&](const HashtablezInfo& info) {
+ res.push_back(info.size.load(std::memory_order_acquire));
+ });
+ return res;
+}
+
+HashtablezInfo* Register(HashtablezSampler* s, size_t size) {
+ auto* info = s->Register();
+ assert(info != nullptr);
+ info->size.store(size);
+ return info;
+}
+
+TEST(HashtablezInfoTest, PrepareForSampling) {
+ absl::Time test_start = absl::Now();
+ HashtablezInfo info;
+ absl::MutexLock l(&info.init_mu);
+ info.PrepareForSampling();
+
+ EXPECT_EQ(info.capacity.load(), 0);
+ EXPECT_EQ(info.size.load(), 0);
+ EXPECT_EQ(info.num_erases.load(), 0);
+ EXPECT_EQ(info.max_probe_length.load(), 0);
+ EXPECT_EQ(info.total_probe_length.load(), 0);
+ EXPECT_EQ(info.hashes_bitwise_or.load(), 0);
+ EXPECT_EQ(info.hashes_bitwise_and.load(), ~size_t{});
+ EXPECT_GE(info.create_time, test_start);
+
+ info.capacity.store(1, std::memory_order_relaxed);
+ info.size.store(1, std::memory_order_relaxed);
+ info.num_erases.store(1, std::memory_order_relaxed);
+ info.max_probe_length.store(1, std::memory_order_relaxed);
+ info.total_probe_length.store(1, std::memory_order_relaxed);
+ info.hashes_bitwise_or.store(1, std::memory_order_relaxed);
+ info.hashes_bitwise_and.store(1, std::memory_order_relaxed);
+ info.create_time = test_start - absl::Hours(20);
+
+ info.PrepareForSampling();
+ EXPECT_EQ(info.capacity.load(), 0);
+ EXPECT_EQ(info.size.load(), 0);
+ EXPECT_EQ(info.num_erases.load(), 0);
+ EXPECT_EQ(info.max_probe_length.load(), 0);
+ EXPECT_EQ(info.total_probe_length.load(), 0);
+ EXPECT_EQ(info.hashes_bitwise_or.load(), 0);
+ EXPECT_EQ(info.hashes_bitwise_and.load(), ~size_t{});
+ EXPECT_GE(info.create_time, test_start);
+}
+
+TEST(HashtablezInfoTest, RecordStorageChanged) {
+ HashtablezInfo info;
+ absl::MutexLock l(&info.init_mu);
+ info.PrepareForSampling();
+ RecordStorageChangedSlow(&info, 17, 47);
+ EXPECT_EQ(info.size.load(), 17);
+ EXPECT_EQ(info.capacity.load(), 47);
+ RecordStorageChangedSlow(&info, 20, 20);
+ EXPECT_EQ(info.size.load(), 20);
+ EXPECT_EQ(info.capacity.load(), 20);
+}
+
+TEST(HashtablezInfoTest, RecordInsert) {
+ HashtablezInfo info;
+ absl::MutexLock l(&info.init_mu);
+ info.PrepareForSampling();
+ EXPECT_EQ(info.max_probe_length.load(), 0);
+ RecordInsertSlow(&info, 0x0000FF00, 6 * kProbeLength);
+ EXPECT_EQ(info.max_probe_length.load(), 6);
+ EXPECT_EQ(info.hashes_bitwise_and.load(), 0x0000FF00);
+ EXPECT_EQ(info.hashes_bitwise_or.load(), 0x0000FF00);
+ RecordInsertSlow(&info, 0x000FF000, 4 * kProbeLength);
+ EXPECT_EQ(info.max_probe_length.load(), 6);
+ EXPECT_EQ(info.hashes_bitwise_and.load(), 0x0000F000);
+ EXPECT_EQ(info.hashes_bitwise_or.load(), 0x000FFF00);
+ RecordInsertSlow(&info, 0x00FF0000, 12 * kProbeLength);
+ EXPECT_EQ(info.max_probe_length.load(), 12);
+ EXPECT_EQ(info.hashes_bitwise_and.load(), 0x00000000);
+ EXPECT_EQ(info.hashes_bitwise_or.load(), 0x00FFFF00);
+}
+
+TEST(HashtablezInfoTest, RecordErase) {
+ HashtablezInfo info;
+ absl::MutexLock l(&info.init_mu);
+ info.PrepareForSampling();
+ EXPECT_EQ(info.num_erases.load(), 0);
+ EXPECT_EQ(info.size.load(), 0);
+ RecordInsertSlow(&info, 0x0000FF00, 6 * kProbeLength);
+ EXPECT_EQ(info.size.load(), 1);
+ RecordEraseSlow(&info);
+ EXPECT_EQ(info.size.load(), 0);
+ EXPECT_EQ(info.num_erases.load(), 1);
+}
+
+TEST(HashtablezSamplerTest, SmallSampleParameter) {
+ SetHashtablezEnabled(true);
+ SetHashtablezSampleParameter(100);
+
+ for (int i = 0; i < 1000; ++i) {
+ int64_t next_sample = 0;
+ HashtablezInfo* sample = SampleSlow(&next_sample);
+ EXPECT_GT(next_sample, 0);
+ EXPECT_NE(sample, nullptr);
+ UnsampleSlow(sample);
+ }
+}
+
+TEST(HashtablezSamplerTest, LargeSampleParameter) {
+ SetHashtablezEnabled(true);
+ SetHashtablezSampleParameter(std::numeric_limits<int32_t>::max());
+
+ for (int i = 0; i < 1000; ++i) {
+ int64_t next_sample = 0;
+ HashtablezInfo* sample = SampleSlow(&next_sample);
+ EXPECT_GT(next_sample, 0);
+ EXPECT_NE(sample, nullptr);
+ UnsampleSlow(sample);
+ }
+}
+
+TEST(HashtablezSamplerTest, Sample) {
+ SetHashtablezEnabled(true);
+ SetHashtablezSampleParameter(100);
+ int64_t num_sampled = 0;
+ int64_t total = 0;
+ double sample_rate;
+ for (int i = 0; i < 1000000; ++i) {
+ HashtablezInfoHandle h = Sample();
+ ++total;
+ if (HashtablezInfoHandlePeer::IsSampled(h)) {
+ ++num_sampled;
+ }
+ sample_rate = static_cast<double>(num_sampled) / total;
+ if (0.005 < sample_rate && sample_rate < 0.015) break;
+ }
+ EXPECT_NEAR(sample_rate, 0.01, 0.005);
+}
+
+TEST(HashtablezSamplerTest, Handle) {
+ auto& sampler = HashtablezSampler::Global();
+ HashtablezInfoHandle h(sampler.Register());
+ auto* info = HashtablezInfoHandlePeer::GetInfo(&h);
+ info->hashes_bitwise_and.store(0x12345678, std::memory_order_relaxed);
+
+ bool found = false;
+ sampler.Iterate([&](const HashtablezInfo& h) {
+ if (&h == info) {
+ EXPECT_EQ(h.hashes_bitwise_and.load(), 0x12345678);
+ found = true;
+ }
+ });
+ EXPECT_TRUE(found);
+
+ h = HashtablezInfoHandle();
+ found = false;
+ sampler.Iterate([&](const HashtablezInfo& h) {
+ if (&h == info) {
+ // this will only happen if some other thread has resurrected the info
+ // the old handle was using.
+ if (h.hashes_bitwise_and.load() == 0x12345678) {
+ found = true;
+ }
+ }
+ });
+ EXPECT_FALSE(found);
+}
+
+TEST(HashtablezSamplerTest, Registration) {
+ HashtablezSampler sampler;
+ auto* info1 = Register(&sampler, 1);
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(1));
+
+ auto* info2 = Register(&sampler, 2);
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(1, 2));
+ info1->size.store(3);
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(3, 2));
+
+ sampler.Unregister(info1);
+ sampler.Unregister(info2);
+}
+
+TEST(HashtablezSamplerTest, Unregistration) {
+ HashtablezSampler sampler;
+ std::vector<HashtablezInfo*> infos;
+ for (size_t i = 0; i < 3; ++i) {
+ infos.push_back(Register(&sampler, i));
+ }
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(0, 1, 2));
+
+ sampler.Unregister(infos[1]);
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(0, 2));
+
+ infos.push_back(Register(&sampler, 3));
+ infos.push_back(Register(&sampler, 4));
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(0, 2, 3, 4));
+ sampler.Unregister(infos[3]);
+ EXPECT_THAT(GetSizes(&sampler), UnorderedElementsAre(0, 2, 4));
+
+ sampler.Unregister(infos[0]);
+ sampler.Unregister(infos[2]);
+ sampler.Unregister(infos[4]);
+ EXPECT_THAT(GetSizes(&sampler), IsEmpty());
+}
+
+TEST(HashtablezSamplerTest, MultiThreaded) {
+ HashtablezSampler sampler;
+ Notification stop;
+ ThreadPool pool(10);
+
+ for (int i = 0; i < 10; ++i) {
+ pool.Schedule([&sampler, &stop]() {
+ std::random_device rd;
+ std::mt19937 gen(rd());
+
+ std::vector<HashtablezInfo*> infoz;
+ while (!stop.HasBeenNotified()) {
+ if (infoz.empty()) {
+ infoz.push_back(sampler.Register());
+ }
+ switch (std::uniform_int_distribution<>(0, 2)(gen)) {
+ case 0: {
+ infoz.push_back(sampler.Register());
+ break;
+ }
+ case 1: {
+ size_t p =
+ std::uniform_int_distribution<>(0, infoz.size() - 1)(gen);
+ HashtablezInfo* info = infoz[p];
+ infoz[p] = infoz.back();
+ infoz.pop_back();
+ sampler.Unregister(info);
+ break;
+ }
+ case 2: {
+ absl::Duration oldest = absl::ZeroDuration();
+ sampler.Iterate([&](const HashtablezInfo& info) {
+ oldest = std::max(oldest, absl::Now() - info.create_time);
+ });
+ ASSERT_GE(oldest, absl::ZeroDuration());
+ break;
+ }
+ }
+ }
+ });
+ }
+ // The threads will hammer away. Give it a little bit of time for tsan to
+ // spot errors.
+ absl::SleepFor(absl::Seconds(3));
+ stop.Notify();
+}
+
+} // namespace
+} // namespace container_internal
+} // namespace absl
diff --git a/absl/container/internal/raw_hash_set.h b/absl/container/internal/raw_hash_set.h
index b7b5ef8c..34d69d7a 100644
--- a/absl/container/internal/raw_hash_set.h
+++ b/absl/container/internal/raw_hash_set.h
@@ -109,6 +109,7 @@
#include "absl/container/internal/container_memory.h"
#include "absl/container/internal/hash_policy_traits.h"
#include "absl/container/internal/hashtable_debug_hooks.h"
+#include "absl/container/internal/hashtablez_sampler.h"
#include "absl/container/internal/have_sse.h"
#include "absl/container/internal/layout.h"
#include "absl/memory/memory.h"
@@ -943,9 +944,10 @@ class raw_hash_set {
// than a full `insert`.
for (const auto& v : that) {
const size_t hash = PolicyTraits::apply(HashElement{hash_ref()}, v);
- const size_t i = find_first_non_full(hash);
- set_ctrl(i, H2(hash));
- emplace_at(i, v);
+ auto target = find_first_non_full(hash);
+ set_ctrl(target.offset, H2(hash));
+ emplace_at(target.offset, v);
+ infoz_.RecordInsert(hash, target.probe_length);
}
size_ = that.size();
growth_left() -= that.size();
@@ -959,6 +961,7 @@ class raw_hash_set {
slots_(absl::exchange(that.slots_, nullptr)),
size_(absl::exchange(that.size_, 0)),
capacity_(absl::exchange(that.capacity_, 0)),
+ infoz_(absl::exchange(that.infoz_, HashtablezInfoHandle())),
// Hash, equality and allocator are copied instead of moved because
// `that` must be left valid. If Hash is std::function<Key>, moving it
// would create a nullptr functor that cannot be called.
@@ -979,6 +982,7 @@ class raw_hash_set {
std::swap(size_, that.size_);
std::swap(capacity_, that.capacity_);
std::swap(growth_left(), that.growth_left());
+ std::swap(infoz_, that.infoz_);
} else {
reserve(that.size());
// Note: this will copy elements of dense_set and unordered_set instead of
@@ -1049,6 +1053,7 @@ class raw_hash_set {
growth_left() = static_cast<size_t>(capacity_ * kMaxLoadFactor);
}
assert(empty());
+ infoz_.RecordStorageChanged(size_, capacity_);
}
// This overload kicks in when the argument is an rvalue of insertable and
@@ -1323,6 +1328,7 @@ class raw_hash_set {
swap(growth_left(), that.growth_left());
swap(hash_ref(), that.hash_ref());
swap(eq_ref(), that.eq_ref());
+ swap(infoz_, that.infoz_);
if (AllocTraits::propagate_on_container_swap::value) {
swap(alloc_ref(), that.alloc_ref());
} else {
@@ -1333,7 +1339,11 @@ class raw_hash_set {
void rehash(size_t n) {
if (n == 0 && capacity_ == 0) return;
- if (n == 0 && size_ == 0) return destroy_slots();
+ if (n == 0 && size_ == 0) {
+ destroy_slots();
+ infoz_.RecordStorageChanged(size_, capacity_);
+ return;
+ }
auto m = NormalizeCapacity((std::max)(n, NumSlotsFast(size())));
// n == 0 unconditionally rehashes as per the standard.
if (n == 0 || m > capacity_) {
@@ -1550,10 +1560,15 @@ class raw_hash_set {
set_ctrl(index, was_never_full ? kEmpty : kDeleted);
growth_left() += was_never_full;
+ infoz_.RecordErase();
}
void initialize_slots() {
assert(capacity_);
+ if (slots_ == nullptr) {
+ infoz_ = Sample();
+ }
+
auto layout = MakeLayout(capacity_);
char* mem = static_cast<char*>(
Allocate<Layout::Alignment()>(&alloc_ref(), layout.AllocSize()));
@@ -1561,6 +1576,7 @@ class raw_hash_set {
slots_ = layout.template Pointer<1>(mem);
reset_ctrl();
growth_left() = static_cast<size_t>(capacity_ * kMaxLoadFactor) - size_;
+ infoz_.RecordStorageChanged(size_, capacity_);
}
void destroy_slots() {
@@ -1593,7 +1609,7 @@ class raw_hash_set {
if (IsFull(old_ctrl[i])) {
size_t hash = PolicyTraits::apply(HashElement{hash_ref()},
PolicyTraits::element(old_slots + i));
- size_t new_i = find_first_non_full(hash);
+ size_t new_i = find_first_non_full(hash).offset;
set_ctrl(new_i, H2(hash));
PolicyTraits::transfer(&alloc_ref(), slots_ + new_i, old_slots + i);
}
@@ -1633,7 +1649,7 @@ class raw_hash_set {
if (!IsDeleted(ctrl_[i])) continue;
size_t hash = PolicyTraits::apply(HashElement{hash_ref()},
PolicyTraits::element(slots_ + i));
- size_t new_i = find_first_non_full(hash);
+ size_t new_i = find_first_non_full(hash).offset;
// Verify if the old and new i fall within the same group wrt the hash.
// If they do, we don't need to move the object as it falls already in the
@@ -1706,7 +1722,11 @@ class raw_hash_set {
// - the input is already a set
// - there are enough slots
// - the element with the hash is not in the table
- size_t find_first_non_full(size_t hash) {
+ struct FindInfo {
+ size_t offset;
+ size_t probe_length;
+ };
+ FindInfo find_first_non_full(size_t hash) {
auto seq = probe(hash);
while (true) {
Group g{ctrl_ + seq.offset()};
@@ -1718,11 +1738,11 @@ class raw_hash_set {
// the group.
// TODO(kfm,sbenza): revisit after we do unconditional mixing
if (ShouldInsertBackwards(hash, ctrl_))
- return seq.offset(mask.HighestBitSet());
+ return {seq.offset(mask.HighestBitSet()), seq.index()};
else
- return seq.offset(mask.LowestBitSet());
+ return {seq.offset(mask.LowestBitSet()), seq.index()};
#else
- return seq.offset(mask.LowestBitSet());
+ return {seq.offset(mask.LowestBitSet()), seq.index()};
#endif
}
assert(seq.index() < capacity_ && "full table!");
@@ -1762,15 +1782,17 @@ class raw_hash_set {
}
size_t prepare_insert(size_t hash) ABSL_ATTRIBUTE_NOINLINE {
- size_t target = find_first_non_full(hash);
- if (ABSL_PREDICT_FALSE(growth_left() == 0 && !IsDeleted(ctrl_[target]))) {
+ auto target = find_first_non_full(hash);
+ if (ABSL_PREDICT_FALSE(growth_left() == 0 &&
+ !IsDeleted(ctrl_[target.offset]))) {
rehash_and_grow_if_necessary();
target = find_first_non_full(hash);
}
++size_;
- growth_left() -= IsEmpty(ctrl_[target]);
- set_ctrl(target, H2(hash));
- return target;
+ growth_left() -= IsEmpty(ctrl_[target.offset]);
+ set_ctrl(target.offset, H2(hash));
+ infoz_.RecordInsert(hash, target.probe_length);
+ return target.offset;
}
// Constructs the value in the space pointed by the iterator. This only works
@@ -1847,6 +1869,7 @@ class raw_hash_set {
slot_type* slots_ = nullptr; // [capacity * slot_type]
size_t size_ = 0; // number of full slots
size_t capacity_ = 0; // total number of slots
+ HashtablezInfoHandle infoz_;
absl::container_internal::CompressedTuple<size_t /* growth_left */, hasher,
key_equal, allocator_type>
settings_{0, hasher{}, key_equal{}, allocator_type{}};
diff --git a/absl/container/internal/raw_hash_set_test.cc b/absl/container/internal/raw_hash_set_test.cc
index 5ad4904f..78b62755 100644
--- a/absl/container/internal/raw_hash_set_test.cc
+++ b/absl/container/internal/raw_hash_set_test.cc
@@ -342,6 +342,7 @@ TEST(Table, EmptyFunctorOptimization) {
size_t size;
size_t capacity;
size_t growth_left;
+ void* infoz;
};
struct StatelessHash {
size_t operator()(absl::string_view) const { return 0; }
@@ -1798,6 +1799,27 @@ TEST(TableDeathTest, EraseOfEndAsserts) {
EXPECT_DEATH_IF_SUPPORTED(t.erase(t.end()), kDeathMsg);
}
+TEST(RawHashSamplerTest, Sample) {
+ // Enable the feature even if the prod default is off.
+ SetHashtablezEnabled(true);
+ SetHashtablezSampleParameter(100);
+
+ auto& sampler = HashtablezSampler::Global();
+ size_t start_size = 0;
+ start_size += sampler.Iterate([&](const HashtablezInfo&) { ++start_size; });
+
+ std::vector<IntTable> tables;
+ for (int i = 0; i < 1000000; ++i) {
+ tables.emplace_back();
+ tables.back().insert(1);
+ }
+ size_t end_size = 0;
+ end_size += sampler.Iterate([&](const HashtablezInfo&) { ++end_size; });
+
+ EXPECT_NEAR((end_size - start_size) / static_cast<double>(tables.size()),
+ 0.01, 0.005);
+}
+
#ifdef ADDRESS_SANITIZER
TEST(Sanitizer, PoisoningUnused) {
IntTable t;