| 1 |
// Copyright The OpenTelemetry Authors |
| 2 |
// SPDX-License-Identifier: Apache-2.0 |
| 3 |
|
| 4 |
package aggregate // import "go.opentelemetry.io/otel/sdk/metric/internal/aggregate" |
| 5 |
|
| 6 |
import ( |
| 7 |
"math" |
| 8 |
"runtime" |
| 9 |
"sync" |
| 10 |
"sync/atomic" |
| 11 |
|
| 12 |
"go.opentelemetry.io/otel/attribute" |
| 13 |
) |
| 14 |
|
| 15 |
// atomicCounter is an efficient way of adding to a number which is either an |
| 16 |
// int64 or float64. It is designed to be efficient when adding whole |
| 17 |
// numbers, regardless of whether N is an int64 or float64. |
| 18 |
// |
| 19 |
// Inspired by the Prometheus counter implementation: |
| 20 |
// https://github.com/prometheus/client_golang/blob/14ccb93091c00f86b85af7753100aa372d63602b/prometheus/counter.go#L108 |
| 21 |
type atomicCounter[N int64 | float64] struct { |
| 22 |
// nFloatBits contains only the non-integer portion of the counter. |
| 23 |
nFloatBits atomic.Uint64 |
| 24 |
// nInt contains only the integer portion of the counter. |
| 25 |
nInt atomic.Int64 |
| 26 |
} |
| 27 |
|
| 28 |
// load returns the current value. The caller must ensure all calls to add have |
| 29 |
// returned prior to calling load. |
| 30 |
func (n *atomicCounter[N]) load() N { |
| 31 |
fval := math.Float64frombits(n.nFloatBits.Load()) |
| 32 |
ival := n.nInt.Load() |
| 33 |
return N(fval + float64(ival)) |
| 34 |
} |
| 35 |
|
| 36 |
func (n *atomicCounter[N]) add(value N) { |
| 37 |
ival := int64(value) |
| 38 |
// This case is where the value is an int, or if it is a whole-numbered float. |
| 39 |
if float64(ival) == float64(value) { |
| 40 |
n.nInt.Add(ival) |
| 41 |
return |
| 42 |
} |
| 43 |
|
| 44 |
// Value must be a float below. |
| 45 |
for { |
| 46 |
oldBits := n.nFloatBits.Load() |
| 47 |
newBits := math.Float64bits(math.Float64frombits(oldBits) + float64(value)) |
| 48 |
if n.nFloatBits.CompareAndSwap(oldBits, newBits) { |
| 49 |
return |
| 50 |
} |
| 51 |
} |
| 52 |
} |
| 53 |
|
| 54 |
// reset resets the internal state, and is not safe to call concurrently. |
| 55 |
func (n *atomicCounter[N]) reset() { |
| 56 |
n.nFloatBits.Store(0) |
| 57 |
n.nInt.Store(0) |
| 58 |
} |
| 59 |
|
| 60 |
// atomicN is a generic atomic number value. |
| 61 |
type atomicN[N int64 | float64] struct { |
| 62 |
val atomic.Uint64 |
| 63 |
} |
| 64 |
|
| 65 |
func (a *atomicN[N]) Load() (value N) { |
| 66 |
v := a.val.Load() |
| 67 |
switch any(value).(type) { |
| 68 |
case int64: |
| 69 |
value = N(v) |
| 70 |
case float64: |
| 71 |
value = N(math.Float64frombits(v)) |
| 72 |
default: |
| 73 |
panic("unsupported type") |
| 74 |
} |
| 75 |
return value |
| 76 |
} |
| 77 |
|
| 78 |
func (a *atomicN[N]) Store(v N) { |
| 79 |
var val uint64 |
| 80 |
switch any(v).(type) { |
| 81 |
case int64: |
| 82 |
val = uint64(v) |
| 83 |
case float64: |
| 84 |
val = math.Float64bits(float64(v)) |
| 85 |
default: |
| 86 |
panic("unsupported type") |
| 87 |
} |
| 88 |
a.val.Store(val) |
| 89 |
} |
| 90 |
|
| 91 |
func (a *atomicN[N]) CompareAndSwap(oldN, newN N) bool { |
| 92 |
var o, n uint64 |
| 93 |
switch any(oldN).(type) { |
| 94 |
case int64: |
| 95 |
o, n = uint64(oldN), uint64(newN) |
| 96 |
case float64: |
| 97 |
o, n = math.Float64bits(float64(oldN)), math.Float64bits(float64(newN)) |
| 98 |
default: |
| 99 |
panic("unsupported type") |
| 100 |
} |
| 101 |
return a.val.CompareAndSwap(o, n) |
| 102 |
} |
| 103 |
|
| 104 |
type atomicMinMax[N int64 | float64] struct { |
| 105 |
minimum, maximum atomicN[N] |
| 106 |
set atomic.Bool |
| 107 |
mu sync.Mutex |
| 108 |
} |
| 109 |
|
| 110 |
// init returns true if the value was used to initialize min and max. |
| 111 |
func (s *atomicMinMax[N]) init(val N) bool { |
| 112 |
s.mu.Lock() |
| 113 |
defer s.mu.Unlock() |
| 114 |
if !s.set.Load() { |
| 115 |
defer s.set.Store(true) |
| 116 |
s.minimum.Store(val) |
| 117 |
s.maximum.Store(val) |
| 118 |
return true |
| 119 |
} |
| 120 |
return false |
| 121 |
} |
| 122 |
|
| 123 |
func (s *atomicMinMax[N]) Update(val N) { |
| 124 |
if !s.set.Load() && s.init(val) { |
| 125 |
return |
| 126 |
} |
| 127 |
|
| 128 |
old := s.minimum.Load() |
| 129 |
for val < old { |
| 130 |
if s.minimum.CompareAndSwap(old, val) { |
| 131 |
return |
| 132 |
} |
| 133 |
old = s.minimum.Load() |
| 134 |
} |
| 135 |
|
| 136 |
old = s.maximum.Load() |
| 137 |
for old < val { |
| 138 |
if s.maximum.CompareAndSwap(old, val) { |
| 139 |
return |
| 140 |
} |
| 141 |
old = s.maximum.Load() |
| 142 |
} |
| 143 |
} |
| 144 |
|
| 145 |
// hotColdWaitGroup is a synchronization primitive which enables lockless |
| 146 |
// writes for concurrent writers and enables a reader to acquire exclusive |
| 147 |
// access to a snapshot of state including only completed operations. |
| 148 |
// Conceptually, it can be thought of as a "hot" wait group, |
| 149 |
// and a "cold" wait group, with the ability for the reader to atomically swap |
| 150 |
// the hot and cold wait groups, and wait for the now-cold wait group to |
| 151 |
// complete. |
| 152 |
// |
| 153 |
// Inspired by the prometheus/client_golang histogram implementation: |
| 154 |
// https://github.com/prometheus/client_golang/blob/a974e0d45e0aa54c65492559114894314d8a2447/prometheus/histogram.go#L725 |
| 155 |
// |
| 156 |
// Usage: |
| 157 |
// |
| 158 |
// var hcwg hotColdWaitGroup |
| 159 |
// var data [2]any |
| 160 |
// |
| 161 |
// func write() { |
| 162 |
// hotIdx := hcwg.start() |
| 163 |
// defer hcwg.done(hotIdx) |
| 164 |
// // modify data without locking |
| 165 |
// data[hotIdx].update() |
| 166 |
// } |
| 167 |
// |
| 168 |
// func read() { |
| 169 |
// coldIdx := hcwg.swapHotAndWait() |
| 170 |
// // read data now that all writes to the cold data have completed. |
| 171 |
// data[coldIdx].read() |
| 172 |
// } |
| 173 |
type hotColdWaitGroup struct { |
| 174 |
// startedCountAndHotIdx contains a 63-bit counter in the lower bits, |
| 175 |
// and a 1 bit hot index to denote which of the two data-points new |
| 176 |
// measurements to write to. These are contained together so that read() |
| 177 |
// can atomically swap the hot bit, reset the started writes to zero, and |
| 178 |
// read the number writes that were started prior to the hot bit being |
| 179 |
// swapped. |
| 180 |
startedCountAndHotIdx atomic.Uint64 |
| 181 |
// endedCounts is the number of writes that have completed to each |
| 182 |
// dataPoint. |
| 183 |
endedCounts [2]atomic.Uint64 |
| 184 |
} |
| 185 |
|
| 186 |
// start returns the hot index that the writer should write to. The returned |
| 187 |
// hot index is 0 or 1. The caller must call done(hot index) after it finishes |
| 188 |
// its operation. start() is safe to call concurrently with other methods. |
| 189 |
func (l *hotColdWaitGroup) start() uint64 { |
| 190 |
// We increment h.startedCountAndHotIdx so that the counter in the lower |
| 191 |
// 63 bits gets incremented. At the same time, we get the new value |
| 192 |
// back, which we can use to return the currently-hot index. |
| 193 |
return l.startedCountAndHotIdx.Add(1) >> 63 |
| 194 |
} |
| 195 |
|
| 196 |
// done signals to the reader that an operation has fully completed. |
| 197 |
// done is safe to call concurrently. |
| 198 |
func (l *hotColdWaitGroup) done(hotIdx uint64) { |
| 199 |
l.endedCounts[hotIdx].Add(1) |
| 200 |
} |
| 201 |
|
| 202 |
// swapHotAndWait swaps the hot bit, waits for all start() calls to be done(), |
| 203 |
// and then returns the now-cold index for the reader to read from. The |
| 204 |
// returned index is 0 or 1. swapHotAndWait must not be called concurrently. |
| 205 |
func (l *hotColdWaitGroup) swapHotAndWait() uint64 { |
| 206 |
n := l.startedCountAndHotIdx.Load() |
| 207 |
coldIdx := (^n) >> 63 |
| 208 |
// Swap the hot and cold index while resetting the started measurements |
| 209 |
// count to zero. |
| 210 |
n = l.startedCountAndHotIdx.Swap((coldIdx << 63)) |
| 211 |
hotIdx := n >> 63 |
| 212 |
startedCount := n & ((1 << 63) - 1) |
| 213 |
// Wait for all measurements to the previously-hot map to finish. |
| 214 |
for startedCount != l.endedCounts[hotIdx].Load() { |
| 215 |
runtime.Gosched() // Let measurements complete. |
| 216 |
} |
| 217 |
// reset the number of ended operations |
| 218 |
l.endedCounts[hotIdx].Store(0) |
| 219 |
return hotIdx |
| 220 |
} |
| 221 |
|
| 222 |
// limitedSyncMap is a sync.Map which enforces the aggregation limit on |
| 223 |
// attribute sets and provides a Len() function. |
| 224 |
type limitedSyncMap struct { |
| 225 |
sync.Map |
| 226 |
aggLimit int |
| 227 |
len int |
| 228 |
lenMux sync.Mutex |
| 229 |
} |
| 230 |
|
| 231 |
func (m *limitedSyncMap) LoadOrStoreAttr(fltrAttr attribute.Set, newValue func(attribute.Set) any) any { |
| 232 |
actual, loaded := m.Load(fltrAttr.Equivalent()) |
| 233 |
if loaded { |
| 234 |
return actual |
| 235 |
} |
| 236 |
// If the overflow set exists, assume we have already overflowed and don't |
| 237 |
// bother with the slow path below. |
| 238 |
actual, loaded = m.Load(overflowSet.Equivalent()) |
| 239 |
if loaded { |
| 240 |
return actual |
| 241 |
} |
| 242 |
// Slow path: add a new attribute set. |
| 243 |
m.lenMux.Lock() |
| 244 |
defer m.lenMux.Unlock() |
| 245 |
|
| 246 |
// re-fetch now that we hold the lock to ensure we don't use the overflow |
| 247 |
// set unless we are sure the attribute set isn't being written |
| 248 |
// concurrently. |
| 249 |
actual, loaded = m.Load(fltrAttr.Equivalent()) |
| 250 |
if loaded { |
| 251 |
return actual |
| 252 |
} |
| 253 |
|
| 254 |
if m.aggLimit > 0 && m.len >= m.aggLimit-1 { |
| 255 |
fltrAttr = overflowSet |
| 256 |
} |
| 257 |
actual, loaded = m.LoadOrStore(fltrAttr.Equivalent(), newValue(fltrAttr)) |
| 258 |
if !loaded { |
| 259 |
m.len++ |
| 260 |
} |
| 261 |
return actual |
| 262 |
} |
| 263 |
|
| 264 |
func (m *limitedSyncMap) Clear() { |
| 265 |
m.lenMux.Lock() |
| 266 |
defer m.lenMux.Unlock() |
| 267 |
m.len = 0 |
| 268 |
m.Map.Clear() |
| 269 |
} |
| 270 |
|
| 271 |
func (m *limitedSyncMap) Len() int { |
| 272 |
m.lenMux.Lock() |
| 273 |
defer m.lenMux.Unlock() |
| 274 |
return m.len |
| 275 |
} |