1 // Copyright 2018 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 // https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 //
15 // -----------------------------------------------------------------------------
16 // File: hashtablez_sampler.h
17 // -----------------------------------------------------------------------------
18 //
19 // This header file defines the API for a low level library to sample hashtables
20 // and collect runtime statistics about them.
21 //
22 // `HashtablezSampler` controls the lifecycle of `HashtablezInfo` objects which
23 // store information about a single sample.
24 //
25 // `Record*` methods store information into samples.
26 // `Sample()` and `Unsample()` make use of a single global sampler with
27 // properties controlled by the flags hashtablez_enabled,
28 // hashtablez_sample_rate, and hashtablez_max_samples.
29 //
30 // WARNING
31 //
32 // Using this sampling API may cause sampled Swiss tables to use the global
33 // allocator (operator `new`) in addition to any custom allocator. If you
34 // are using a table in an unusual circumstance where allocation or calling a
35 // linux syscall is unacceptable, this could interfere.
36 //
37 // This utility is internal-only. Use at your own risk.
38
39 #ifndef ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
40 #define ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
41
42 #include <atomic>
43 #include <functional>
44 #include <memory>
45 #include <vector>
46
47 #include "absl/base/internal/per_thread_tls.h"
48 #include "absl/base/optimization.h"
49 #include "absl/container/internal/have_sse.h"
50 #include "absl/synchronization/mutex.h"
51 #include "absl/utility/utility.h"
52
53 namespace absl {
54 ABSL_NAMESPACE_BEGIN
55 namespace container_internal {
56
57 // Stores information about a sampled hashtable. All mutations to this *must*
58 // be made through `Record*` functions below. All reads from this *must* only
59 // occur in the callback to `HashtablezSampler::Iterate`.
60 struct HashtablezInfo {
61 // Constructs the object but does not fill in any fields.
62 HashtablezInfo();
63 ~HashtablezInfo();
64 HashtablezInfo(const HashtablezInfo&) = delete;
65 HashtablezInfo& operator=(const HashtablezInfo&) = delete;
66
67 // Puts the object into a clean state, fills in the logically `const` members,
68 // blocking for any readers that are currently sampling the object.
69 void PrepareForSampling() ABSL_EXCLUSIVE_LOCKS_REQUIRED(init_mu);
70
71 // These fields are mutated by the various Record* APIs and need to be
72 // thread-safe.
73 std::atomic<size_t> capacity;
74 std::atomic<size_t> size;
75 std::atomic<size_t> num_erases;
76 std::atomic<size_t> max_probe_length;
77 std::atomic<size_t> total_probe_length;
78 std::atomic<size_t> hashes_bitwise_or;
79 std::atomic<size_t> hashes_bitwise_and;
80
81 // `HashtablezSampler` maintains intrusive linked lists for all samples. See
82 // comments on `HashtablezSampler::all_` for details on these. `init_mu`
83 // guards the ability to restore the sample to a pristine state. This
84 // prevents races with sampling and resurrecting an object.
85 absl::Mutex init_mu;
86 HashtablezInfo* next;
87 HashtablezInfo* dead ABSL_GUARDED_BY(init_mu);
88
89 // All of the fields below are set by `PrepareForSampling`, they must not be
90 // mutated in `Record*` functions. They are logically `const` in that sense.
91 // These are guarded by init_mu, but that is not externalized to clients, who
92 // can only read them during `HashtablezSampler::Iterate` which will hold the
93 // lock.
94 static constexpr int kMaxStackDepth = 64;
95 absl::Time create_time;
96 int32_t depth;
97 void* stack[kMaxStackDepth];
98 };
99
RecordRehashSlow(HashtablezInfo * info,size_t total_probe_length)100 inline void RecordRehashSlow(HashtablezInfo* info, size_t total_probe_length) {
101 #if SWISSTABLE_HAVE_SSE2
102 total_probe_length /= 16;
103 #else
104 total_probe_length /= 8;
105 #endif
106 info->total_probe_length.store(total_probe_length, std::memory_order_relaxed);
107 info->num_erases.store(0, std::memory_order_relaxed);
108 }
109
RecordStorageChangedSlow(HashtablezInfo * info,size_t size,size_t capacity)110 inline void RecordStorageChangedSlow(HashtablezInfo* info, size_t size,
111 size_t capacity) {
112 info->size.store(size, std::memory_order_relaxed);
113 info->capacity.store(capacity, std::memory_order_relaxed);
114 if (size == 0) {
115 // This is a clear, reset the total/num_erases too.
116 RecordRehashSlow(info, 0);
117 }
118 }
119
120 void RecordInsertSlow(HashtablezInfo* info, size_t hash,
121 size_t distance_from_desired);
122
RecordEraseSlow(HashtablezInfo * info)123 inline void RecordEraseSlow(HashtablezInfo* info) {
124 info->size.fetch_sub(1, std::memory_order_relaxed);
125 info->num_erases.fetch_add(1, std::memory_order_relaxed);
126 }
127
128 HashtablezInfo* SampleSlow(int64_t* next_sample);
129 void UnsampleSlow(HashtablezInfo* info);
130
131 class HashtablezInfoHandle {
132 public:
HashtablezInfoHandle()133 explicit HashtablezInfoHandle() : info_(nullptr) {}
HashtablezInfoHandle(HashtablezInfo * info)134 explicit HashtablezInfoHandle(HashtablezInfo* info) : info_(info) {}
~HashtablezInfoHandle()135 ~HashtablezInfoHandle() {
136 if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
137 UnsampleSlow(info_);
138 }
139
140 HashtablezInfoHandle(const HashtablezInfoHandle&) = delete;
141 HashtablezInfoHandle& operator=(const HashtablezInfoHandle&) = delete;
142
HashtablezInfoHandle(HashtablezInfoHandle && o)143 HashtablezInfoHandle(HashtablezInfoHandle&& o) noexcept
144 : info_(absl::exchange(o.info_, nullptr)) {}
145 HashtablezInfoHandle& operator=(HashtablezInfoHandle&& o) noexcept {
146 if (ABSL_PREDICT_FALSE(info_ != nullptr)) {
147 UnsampleSlow(info_);
148 }
149 info_ = absl::exchange(o.info_, nullptr);
150 return *this;
151 }
152
RecordStorageChanged(size_t size,size_t capacity)153 inline void RecordStorageChanged(size_t size, size_t capacity) {
154 if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
155 RecordStorageChangedSlow(info_, size, capacity);
156 }
157
RecordRehash(size_t total_probe_length)158 inline void RecordRehash(size_t total_probe_length) {
159 if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
160 RecordRehashSlow(info_, total_probe_length);
161 }
162
RecordInsert(size_t hash,size_t distance_from_desired)163 inline void RecordInsert(size_t hash, size_t distance_from_desired) {
164 if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
165 RecordInsertSlow(info_, hash, distance_from_desired);
166 }
167
RecordErase()168 inline void RecordErase() {
169 if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
170 RecordEraseSlow(info_);
171 }
172
swap(HashtablezInfoHandle & lhs,HashtablezInfoHandle & rhs)173 friend inline void swap(HashtablezInfoHandle& lhs,
174 HashtablezInfoHandle& rhs) {
175 std::swap(lhs.info_, rhs.info_);
176 }
177
178 private:
179 friend class HashtablezInfoHandlePeer;
180 HashtablezInfo* info_;
181 };
182
183 #if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
184 #error ABSL_INTERNAL_HASHTABLEZ_SAMPLE cannot be directly set
185 #endif // defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
186
187 #if (ABSL_PER_THREAD_TLS == 1) && !defined(ABSL_BUILD_DLL) && \
188 !defined(ABSL_CONSUME_DLL)
189 #define ABSL_INTERNAL_HASHTABLEZ_SAMPLE
190 #endif
191
192 #if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
193 extern ABSL_PER_THREAD_TLS_KEYWORD int64_t global_next_sample;
194 #endif // ABSL_PER_THREAD_TLS
195
196 // Returns an RAII sampling handle that manages registration and unregistation
197 // with the global sampler.
Sample()198 inline HashtablezInfoHandle Sample() {
199 #if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
200 if (ABSL_PREDICT_TRUE(--global_next_sample > 0)) {
201 return HashtablezInfoHandle(nullptr);
202 }
203 return HashtablezInfoHandle(SampleSlow(&global_next_sample));
204 #else
205 return HashtablezInfoHandle(nullptr);
206 #endif // !ABSL_PER_THREAD_TLS
207 }
208
209 // Holds samples and their associated stack traces with a soft limit of
210 // `SetHashtablezMaxSamples()`.
211 //
212 // Thread safe.
213 class HashtablezSampler {
214 public:
215 // Returns a global Sampler.
216 static HashtablezSampler& Global();
217
218 HashtablezSampler();
219 ~HashtablezSampler();
220
221 // Registers for sampling. Returns an opaque registration info.
222 HashtablezInfo* Register();
223
224 // Unregisters the sample.
225 void Unregister(HashtablezInfo* sample);
226
227 // The dispose callback will be called on all samples the moment they are
228 // being unregistered. Only affects samples that are unregistered after the
229 // callback has been set.
230 // Returns the previous callback.
231 using DisposeCallback = void (*)(const HashtablezInfo&);
232 DisposeCallback SetDisposeCallback(DisposeCallback f);
233
234 // Iterates over all the registered `StackInfo`s. Returning the number of
235 // samples that have been dropped.
236 int64_t Iterate(const std::function<void(const HashtablezInfo& stack)>& f);
237
238 private:
239 void PushNew(HashtablezInfo* sample);
240 void PushDead(HashtablezInfo* sample);
241 HashtablezInfo* PopDead();
242
243 std::atomic<size_t> dropped_samples_;
244 std::atomic<size_t> size_estimate_;
245
246 // Intrusive lock free linked lists for tracking samples.
247 //
248 // `all_` records all samples (they are never removed from this list) and is
249 // terminated with a `nullptr`.
250 //
251 // `graveyard_.dead` is a circular linked list. When it is empty,
252 // `graveyard_.dead == &graveyard`. The list is circular so that
253 // every item on it (even the last) has a non-null dead pointer. This allows
254 // `Iterate` to determine if a given sample is live or dead using only
255 // information on the sample itself.
256 //
257 // For example, nodes [A, B, C, D, E] with [A, C, E] alive and [B, D] dead
258 // looks like this (G is the Graveyard):
259 //
260 // +---+ +---+ +---+ +---+ +---+
261 // all -->| A |--->| B |--->| C |--->| D |--->| E |
262 // | | | | | | | | | |
263 // +---+ | | +->| |-+ | | +->| |-+ | |
264 // | G | +---+ | +---+ | +---+ | +---+ | +---+
265 // | | | | | |
266 // | | --------+ +--------+ |
267 // +---+ |
268 // ^ |
269 // +--------------------------------------+
270 //
271 std::atomic<HashtablezInfo*> all_;
272 HashtablezInfo graveyard_;
273
274 std::atomic<DisposeCallback> dispose_;
275 };
276
277 // Enables or disables sampling for Swiss tables.
278 void SetHashtablezEnabled(bool enabled);
279
280 // Sets the rate at which Swiss tables will be sampled.
281 void SetHashtablezSampleParameter(int32_t rate);
282
283 // Sets a soft max for the number of samples that will be kept.
284 void SetHashtablezMaxSamples(int32_t max);
285
286 // Configuration override.
287 // This allows process-wide sampling without depending on order of
288 // initialization of static storage duration objects.
289 // The definition of this constant is weak, which allows us to inject a
290 // different value for it at link time.
291 extern "C" bool AbslContainerInternalSampleEverything();
292
293 } // namespace container_internal
294 ABSL_NAMESPACE_END
295 } // namespace absl
296
297 #endif // ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
298