| 1 | package dbreads |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "sync" |
| 8 | "time" |
| 9 | ) |
| 10 | |
| 11 | var ( |
| 12 | errBusy = errors.New("dbreads: this box is running its two reads") |
| 13 | errOrphans = errors.New("dbreads: a carlos-dbread@ unit this agent is not running is still active") |
| 14 | errNoRoom = errors.New("dbreads: not enough disk for the read") |
| 15 | ) |
| 16 | |
| 17 | // admission is the agent's one table (spec §5.7, J5): how many worker |
| 18 | // connections it holds, which reads it is running, and every disk |
| 19 | // reservation. The agent is the only process that starts workers, so one |
| 20 | // mutex covers it all. |
| 21 | type admission struct { |
| 22 | units func(context.Context) (int, error) |
| 23 | free func() (avail, total uint64, err error) |
| 24 | |
| 25 | mu sync.Mutex |
| 26 | conns int |
| 27 | running map[string]bool |
| 28 | held map[string]int64 |
| 29 | probing bool |
| 30 | } |
| 31 | |
| 32 | func newAdmission(units func(context.Context) (int, error), free func() (uint64, uint64, error)) *admission { |
| 33 | return &admission{units: units, free: free, running: map[string]bool{}, held: map[string]int64{}} |
| 34 | } |
| 35 | |
| 36 | // reconcileLocked is the refusal-based reconciliation (spec §5.7): while any |
| 37 | // carlos-dbread@ unit is active, activating or deactivating beyond the |
| 38 | // connections this agent holds, nothing is admitted. A finished read's |
| 39 | // reservation is released only when no unit is left at all: then every |
| 40 | // stopped unit's scratch has been removed by systemd. |
| 41 | func (a *admission) reconcileLocked(ctx context.Context) error { |
| 42 | n, err := a.units(ctx) |
| 43 | if err != nil { |
| 44 | return fmt.Errorf("dbreads: list carlos-dbread@ units: %w", err) |
| 45 | } |
| 46 | if n > a.conns { |
| 47 | return errOrphans |
| 48 | } |
| 49 | if n == 0 { |
| 50 | for id := range a.held { |
| 51 | if !a.running[id] { |
| 52 | delete(a.held, id) |
| 53 | } |
| 54 | } |
| 55 | } |
| 56 | return nil |
| 57 | } |
| 58 | |
| 59 | // fitsLocked: after subtracting every other reservation and need, the |
| 60 | // filesystem holding /var/tmp keeps DiskFloor and DiskFreePercent free. |
| 61 | func (a *admission) fitsLocked(id string, need int64) bool { |
| 62 | avail, total, err := a.free() |
| 63 | if err != nil { |
| 64 | return false |
| 65 | } |
| 66 | left := int64(avail) - need |
| 67 | for k, v := range a.held { |
| 68 | if k != id { |
| 69 | left -= v |
| 70 | } |
| 71 | } |
| 72 | return left >= DiskFloor && left >= int64(total)*DiskFreePercent/100 |
| 73 | } |
| 74 | |
| 75 | // admitRead admits a read before it enters the election: a free read slot, |
| 76 | // no orphan, and room for the floor reservation. |
| 77 | func (a *admission) admitRead(ctx context.Context, id string) error { |
| 78 | a.mu.Lock() |
| 79 | defer a.mu.Unlock() |
| 80 | if err := a.reconcileLocked(ctx); err != nil { |
| 81 | return err |
| 82 | } |
| 83 | if len(a.running) >= ReadsPerBox { |
| 84 | return errBusy |
| 85 | } |
| 86 | if !a.fitsLocked(id, DiskFloor) { |
| 87 | return errNoRoom |
| 88 | } |
| 89 | a.running[id] = true |
| 90 | a.held[id] = DiskFloor |
| 91 | return nil |
| 92 | } |
| 93 | |
| 94 | // grow raises a running read's reservation to need (never below the floor) |
| 95 | // once its plan is known. |
| 96 | func (a *admission) grow(id string, need int64) error { |
| 97 | a.mu.Lock() |
| 98 | defer a.mu.Unlock() |
| 99 | need = max(need, DiskFloor) |
| 100 | if !a.fitsLocked(id, need) { |
| 101 | return errNoRoom |
| 102 | } |
| 103 | a.held[id] = need |
| 104 | return nil |
| 105 | } |
| 106 | |
| 107 | // finish ends a read's slot. Its reservation stays until reconcileLocked |
| 108 | // sees no unit left. |
| 109 | func (a *admission) finish(id string) { |
| 110 | a.mu.Lock() |
| 111 | defer a.mu.Unlock() |
| 112 | delete(a.running, id) |
| 113 | } |
| 114 | |
| 115 | // abandon gives back a read this agent admitted and then did not run (it |
| 116 | // lost the election, or stopped before it): no worker was started for it, |
| 117 | // so its reservation goes at once rather than waiting for the units to drain. |
| 118 | func (a *admission) abandon(id string) { |
| 119 | a.mu.Lock() |
| 120 | defer a.mu.Unlock() |
| 121 | delete(a.running, id) |
| 122 | delete(a.held, id) |
| 123 | } |
| 124 | |
| 125 | // retry is a read starting a fresh plan after a planned file vanished (spec |
| 126 | // §8.2). Its previous worker may still hold scratch until its unit stops, so |
| 127 | // that reservation moves to a retired key, held until no unit is left (J5), |
| 128 | // and the read starts again from the floor (plan review finding 10). |
| 129 | func (a *admission) retry(id string, attempt int) { |
| 130 | a.mu.Lock() |
| 131 | defer a.mu.Unlock() |
| 132 | if v, ok := a.held[id]; ok { |
| 133 | a.held[fmt.Sprintf("%s#%d", id, attempt)] = v |
| 134 | } |
| 135 | a.held[id] = DiskFloor |
| 136 | } |
| 137 | |
| 138 | // settle waits until no carlos-dbread@ unit is running beyond this agent's |
| 139 | // open connections, so a retry opens its next worker only once the last one |
| 140 | // is gone, and never into the probe's connection. It releases retired |
| 141 | // reservations once no unit is left at all. |
| 142 | func (a *admission) settle(ctx context.Context, poll time.Duration) error { |
| 143 | for { |
| 144 | a.mu.Lock() |
| 145 | err := a.reconcileLocked(ctx) |
| 146 | a.mu.Unlock() |
| 147 | if !errors.Is(err, errOrphans) { |
| 148 | return err |
| 149 | } |
| 150 | select { |
| 151 | case <-ctx.Done(): |
| 152 | return ctx.Err() |
| 153 | case <-time.After(poll): |
| 154 | } |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | // admitProbe admits the readiness probe: no orphan and no other probe. The |
| 159 | // probe's connection is the third (opened checks the count). |
| 160 | func (a *admission) admitProbe(ctx context.Context) bool { |
| 161 | a.mu.Lock() |
| 162 | defer a.mu.Unlock() |
| 163 | if a.probing || a.reconcileLocked(ctx) != nil { |
| 164 | return false |
| 165 | } |
| 166 | a.probing = true |
| 167 | return true |
| 168 | } |
| 169 | |
| 170 | func (a *admission) probeDone() { |
| 171 | a.mu.Lock() |
| 172 | defer a.mu.Unlock() |
| 173 | a.probing = false |
| 174 | } |
| 175 | |
| 176 | // opened takes a connection slot; the agent never opens a fourth. |
| 177 | func (a *admission) opened() bool { |
| 178 | a.mu.Lock() |
| 179 | defer a.mu.Unlock() |
| 180 | if a.conns >= MaxConnections { |
| 181 | return false |
| 182 | } |
| 183 | a.conns++ |
| 184 | return true |
| 185 | } |
| 186 | |
| 187 | func (a *admission) closed() { |
| 188 | a.mu.Lock() |
| 189 | defer a.mu.Unlock() |
| 190 | a.conns-- |
| 191 | } |
| 192 | |
amadan