| 1 | // Package historywork bounds reconstructable history work independently of |
| 2 | // persistence and runtime ownership. Cancelling a reader never cancels a save. |
| 3 | package historywork |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "errors" |
| 8 | "io" |
| 9 | "sync" |
| 10 | "sync/atomic" |
| 11 | "time" |
| 12 | ) |
| 13 | |
| 14 | const ( |
| 15 | BatchEntries = 128 |
| 16 | BatchBytes = 4 << 20 |
| 17 | ReadChunk = 64 << 10 |
| 18 | BytesPerSecond = 8 << 20 |
| 19 | SliceDuration = 50 * time.Millisecond |
| 20 | PauseDuration = 100 * time.Millisecond |
| 21 | ) |
| 22 | |
| 23 | // Coordinator is shared by the catalog and historical source discovery. |
| 24 | // Background permits are held for one bounded slice, never a directory scan. |
| 25 | type Coordinator struct { |
| 26 | once sync.Once |
| 27 | background chan struct{} |
| 28 | foreground chan struct{} |
| 29 | mu sync.Mutex |
| 30 | readers int |
| 31 | changed chan struct{} |
| 32 | backgroundReady time.Time |
| 33 | readBytes atomic.Int64 |
| 34 | readCalls atomic.Uint64 |
| 35 | canceledReads atomic.Uint64 |
| 36 | slices atomic.Uint64 |
| 37 | lastSliceNanos atomic.Int64 |
| 38 | } |
| 39 | |
| 40 | func (c *Coordinator) init() { |
| 41 | c.once.Do(func() { |
| 42 | c.background = make(chan struct{}, 1) |
| 43 | c.foreground = make(chan struct{}, 1) |
| 44 | c.changed = make(chan struct{}) |
| 45 | }) |
| 46 | } |
| 47 | |
| 48 | func (c *Coordinator) Foreground(ctx context.Context) (func(), error) { |
| 49 | c.init() |
| 50 | if err := ctx.Err(); err != nil { |
| 51 | return nil, err |
| 52 | } |
| 53 | select { |
| 54 | case c.foreground <- struct{}{}: |
| 55 | case <-ctx.Done(): |
| 56 | return nil, ctx.Err() |
| 57 | } |
| 58 | if err := ctx.Err(); err != nil { |
| 59 | <-c.foreground |
| 60 | return nil, err |
| 61 | } |
| 62 | c.mu.Lock() |
| 63 | c.readers++ |
| 64 | c.mu.Unlock() |
| 65 | var once sync.Once |
| 66 | return func() { |
| 67 | once.Do(func() { |
| 68 | c.mu.Lock() |
| 69 | c.readers-- |
| 70 | close(c.changed) |
| 71 | c.changed = make(chan struct{}) |
| 72 | c.mu.Unlock() |
| 73 | <-c.foreground |
| 74 | }) |
| 75 | }, nil |
| 76 | } |
| 77 | |
| 78 | func (c *Coordinator) Background(ctx context.Context) (func(), error) { |
| 79 | release, err := c.BackgroundSlice(ctx, false) |
| 80 | if err != nil { |
| 81 | return nil, err |
| 82 | } |
| 83 | return func() { release(0) }, nil |
| 84 | } |
| 85 | |
| 86 | func (c *Coordinator) ForegroundActive() bool { |
| 87 | c.init() |
| 88 | c.mu.Lock() |
| 89 | defer c.mu.Unlock() |
| 90 | return c.readers > 0 |
| 91 | } |
| 92 | |
| 93 | // BackgroundSlice applies one shared rate budget across all discovery owners. |
| 94 | // P1 metadata may proceed during a foreground preparation; P2 waits for it. |
| 95 | func (c *Coordinator) BackgroundSlice(ctx context.Context, foregroundAllowed bool) (func(int64), error) { |
| 96 | return c.backgroundSlice(ctx, foregroundAllowed, false) |
| 97 | } |
| 98 | |
| 99 | var ErrForegroundActive = errors.New("foreground history preparation active") |
| 100 | |
| 101 | // A single scheduling loop must be able to reconsider P1 while P2 is paused. |
| 102 | // Yielding preserves the caller's iterator; this is not a failed slice. |
| 103 | func (c *Coordinator) BackgroundSliceYielding(ctx context.Context, foregroundAllowed bool) (func(int64), error) { |
| 104 | return c.backgroundSlice(ctx, foregroundAllowed, true) |
| 105 | } |
| 106 | |
| 107 | func (c *Coordinator) backgroundSlice(ctx context.Context, foregroundAllowed, yield bool) (func(int64), error) { |
| 108 | c.init() |
| 109 | for { |
| 110 | if err := ctx.Err(); err != nil { |
| 111 | return nil, err |
| 112 | } |
| 113 | c.mu.Lock() |
| 114 | busy, changed, ready := c.readers > 0 && !foregroundAllowed, c.changed, c.backgroundReady |
| 115 | c.mu.Unlock() |
| 116 | if busy { |
| 117 | if yield { |
| 118 | return nil, ErrForegroundActive |
| 119 | } |
| 120 | select { |
| 121 | case <-changed: |
| 122 | continue |
| 123 | case <-ctx.Done(): |
| 124 | return nil, ctx.Err() |
| 125 | } |
| 126 | } |
| 127 | if delay := time.Until(ready); delay > 0 { |
| 128 | timer := time.NewTimer(delay) |
| 129 | select { |
| 130 | case <-ctx.Done(): |
| 131 | timer.Stop() |
| 132 | return nil, ctx.Err() |
| 133 | case <-timer.C: |
| 134 | } |
| 135 | } |
| 136 | select { |
| 137 | case c.background <- struct{}{}: |
| 138 | case <-ctx.Done(): |
| 139 | return nil, ctx.Err() |
| 140 | } |
| 141 | c.mu.Lock() |
| 142 | foregroundBusy := c.readers > 0 && !foregroundAllowed |
| 143 | busy = foregroundBusy || time.Now().Before(c.backgroundReady) |
| 144 | c.mu.Unlock() |
| 145 | if busy { |
| 146 | <-c.background |
| 147 | if foregroundBusy && yield { |
| 148 | return nil, ErrForegroundActive |
| 149 | } |
| 150 | continue |
| 151 | } |
| 152 | var once sync.Once |
| 153 | started := time.Now() |
| 154 | return func(bytes int64) { |
| 155 | once.Do(func() { |
| 156 | c.slices.Add(1) |
| 157 | c.lastSliceNanos.Store(time.Since(started).Nanoseconds()) |
| 158 | c.mu.Lock() |
| 159 | c.backgroundReady = time.Now().Add(max(PauseDuration, time.Duration(max(0, bytes))*time.Second/BytesPerSecond)) |
| 160 | c.mu.Unlock() |
| 161 | <-c.background |
| 162 | }) |
| 163 | }, nil |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | // Pause includes both the per-slice yield and the byte-rate allowance. Waiting |
| 168 | // is cancellable; tests may inject a clock through a caller's slice runner. |
| 169 | func Pause(ctx context.Context, bytes int64) error { |
| 170 | delay := max(PauseDuration, time.Duration(bytes)*time.Second/BytesPerSecond) |
| 171 | timer := time.NewTimer(delay) |
| 172 | defer timer.Stop() |
| 173 | select { |
| 174 | case <-timer.C: |
| 175 | return ctx.Err() |
| 176 | case <-ctx.Done(): |
| 177 | return ctx.Err() |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | // Reader checks cancellation before each bounded source read, including reads |
| 182 | // performed inside a JSON decoder rather than only between decoded records. |
| 183 | type Reader struct { |
| 184 | Context context.Context |
| 185 | Source io.Reader |
| 186 | Bytes int64 |
| 187 | } |
| 188 | |
| 189 | func (r *Reader) Read(p []byte) (int, error) { |
| 190 | meter, _ := r.Context.Value(meterKey{}).(*Coordinator) |
| 191 | if err := r.Context.Err(); err != nil { |
| 192 | if meter != nil { |
| 193 | meter.canceledReads.Add(1) |
| 194 | } |
| 195 | return 0, err |
| 196 | } |
| 197 | if len(p) > ReadChunk { |
| 198 | p = p[:ReadChunk] |
| 199 | } |
| 200 | n, err := r.Source.Read(p) |
| 201 | r.Bytes += int64(n) |
| 202 | if meter != nil { |
| 203 | meter.readBytes.Add(int64(n)) |
| 204 | meter.readCalls.Add(1) |
| 205 | } |
| 206 | return n, err |
| 207 | } |
| 208 | |
| 209 | type meterKey struct{} |
| 210 | |
| 211 | func (c *Coordinator) Context(ctx context.Context) context.Context { |
| 212 | return context.WithValue(ctx, meterKey{}, c) |
| 213 | } |
| 214 | |
| 215 | // Counters contain no source identities or transcript text. Bytes counts only |
| 216 | // reads routed through Reader, rather than pretending a stat size was read. |
| 217 | type Diagnostics struct { |
| 218 | ForegroundActive int `json:"foregroundActive"` |
| 219 | BackgroundActive int `json:"backgroundActive"` |
| 220 | BackgroundSlices uint64 `json:"backgroundSlices"` |
| 221 | InstrumentedReadBytes int64 `json:"instrumentedReadBytes"` |
| 222 | InstrumentedReadCalls uint64 `json:"instrumentedReadCalls"` |
| 223 | CanceledReadCheckpoints uint64 `json:"canceledReadCheckpoints"` |
| 224 | LastBackgroundSliceMS float64 `json:"lastBackgroundSliceMs"` |
| 225 | } |
| 226 | |
| 227 | func (c *Coordinator) Diagnostics() Diagnostics { |
| 228 | c.init() |
| 229 | c.mu.Lock() |
| 230 | defer c.mu.Unlock() |
| 231 | return Diagnostics{ForegroundActive: c.readers, BackgroundActive: len(c.background), BackgroundSlices: c.slices.Load(), InstrumentedReadBytes: c.readBytes.Load(), InstrumentedReadCalls: c.readCalls.Load(), CanceledReadCheckpoints: c.canceledReads.Load(), LastBackgroundSliceMS: float64(c.lastSliceNanos.Load()) / float64(time.Millisecond)} |
| 232 | } |
| 233 |