1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
//! Per-shard reactor glue for the replication subsystem: accept new
//! replica connections, drive their read/write handlers, reap closed
//! conns into the slot table, and periodically expire stale slots.
//!
//! Split from [`crate::replication`] (which holds the state machine
//! types + `close()` + handshake parser) so each file stays under the
//! 500-LOC house rule. All methods here are `impl<C: Commands> Shard<C>`.
use crate::Commands;
use crate::replication::{ReplicaConn, ReplicaState, advance_handshake};
use crate::shard::Shard;
use std::io;
/// Maximum bytes a replication conn may buffer before handshake
/// completes. The whole `REPLICATE FROM <offset> ID <id>` fits in
/// ~80 bytes for any realistic id; 4 KiB is generous + catches a
/// misbehaving / hostile peer trying to OOM the primary by holding a
/// handshake half-open.
const HANDSHAKE_MAX_INPUT: usize = 4 * 1024;
/// Maximum in-flight `accept(2)`s per `accept_ready_replication` tick.
/// We drain until `WouldBlock`, but cap to defend against an
/// accept-flood DoS. Real replica counts are < 16 so this is room
/// to spare.
const ACCEPT_BURST_CAP: usize = 64;
/// Cap on the input buffer a streaming replica may accumulate before
/// being dropped. Streaming input is the replica→primary ACK channel
/// (`REPLCONF ACK` lines parsed by `parse_replica_acks`); anything
/// unparseable is skipped, so a well-behaved peer never accumulates
/// much. The cap protects against a peer dumping arbitrary bytes
/// hoping to bloat memory.
const STREAMING_INPUT_DISCARD_CAP: usize = 64 * 1024;
impl<C: Commands> Shard<C> {
/// Drain the replication listener — accept until `WouldBlock` or
/// the burst cap. Each accepted socket goes into `self.replicas`
/// and gets registered with the poller for readability.
pub(crate) fn accept_ready_replication(&mut self) -> io::Result<()> {
let Some(listener) = self.replication_listener.as_ref() else {
return Ok(());
};
for _ in 0..ACCEPT_BURST_CAP {
match listener.accept() {
Ok(sock) => {
sock.set_nonblocking()?;
self.poller.add(sock.raw(), true, false)?;
// Capture the replica's peer addr at
// accept time so `INFO replication` / `ROLE` can
// report it. `peer_addr` errs on a peer that
// already vanished — fall back to 0.0.0.0:0,
// the connection will reap on the next read.
let peer = sock.peer_addr().unwrap_or((
std::net::Ipv4Addr::UNSPECIFIED,
0,
));
self.replicas.push(ReplicaConn::with_peer(sock, peer));
}
Err(e) if e.kind() == io::ErrorKind::WouldBlock => return Ok(()),
Err(e) => return Err(e),
}
}
Ok(())
}
/// Locate a replica conn by raw fd. Linear scan — replica counts
/// are < 16; a hashmap would cost more than the few comparisons.
pub(crate) fn replica_index_by_fd(&self, fd: i32) -> Option<usize> {
self.replicas.iter().position(|r| r.fd == fd)
}
/// Handle readability on a replica conn. In `HandshakePending`,
/// pull bytes, advance handshake state, queue `+ACK` on success,
/// `close()` on failure. In `Streaming`, parse `REPLCONF ACK`
/// lines (the replica→primary ACK channel). In `AckSent`/`Closed`,
/// ignore.
// LOC-WAIVER: replica-conn read state machine — one read loop over
// ReplicaState arms (handshake / ACK stream / drain); the loop and
// state arms are one indivisible protocol unit.
pub(crate) fn replica_readable(&mut self, idx: usize) -> io::Result<()> {
let mut scratch = [0u8; 256];
loop {
match self.replicas[idx].sock.read(&mut scratch) {
Ok(0) => {
self.replicas[idx].close();
return Ok(());
}
Ok(n) => match &self.replicas[idx].state {
ReplicaState::HandshakePending => {
let conn = &mut self.replicas[idx];
if conn.input.len() + n > HANDSHAKE_MAX_INPUT {
conn.close();
return Ok(());
}
conn.input.extend_from_slice(&scratch[..n]);
let feed_gen =
self.replicate.as_ref().map_or(0, |f| f.generation());
let conn = &mut self.replicas[idx];
if let Err(e) = advance_handshake(conn, feed_gen) {
eprintln!(
"kevy: replica handshake rejected on fd {}: {e}",
conn.fd,
);
conn.close();
return Ok(());
}
if !matches!(
self.replicas[idx].state,
ReplicaState::HandshakePending
) {
return Ok(());
}
}
ReplicaState::Streaming { .. } => {
// The replica→primary direction is
// the ACK channel — this readable handler is
// its SINGLE reader (an earlier design had a
// second reader in the pump that raced this
// one and lost most ACKs).
let conn = &mut self.replicas[idx];
if conn.input.len() + n > STREAMING_INPUT_DISCARD_CAP {
eprintln!(
"kevy: streaming replica {} sent > {} B \
of unparseable input; dropping link",
conn.fd, STREAMING_INPUT_DISCARD_CAP,
);
conn.close();
return Ok(());
}
conn.input.extend_from_slice(&scratch[..n]);
self.parse_replica_acks(idx);
}
ReplicaState::AckSent { .. }
| ReplicaState::SnapshotShipping { .. }
| ReplicaState::Closed { .. } => {
// No input expected; drain into the void.
}
},
Err(e) if e.kind() == io::ErrorKind::WouldBlock => return Ok(()),
Err(e) if e.kind() == io::ErrorKind::Interrupted => continue,
Err(e) => return Err(e),
}
}
}
/// Drain `output[write_off..]` non-blocking. When fully drained
/// in `AckSent`, transition to [`ReplicaState::Streaming`]
/// (carrying the handshake's replica id + from-offset).
pub(crate) fn replica_writable(&mut self, idx: usize) -> io::Result<()> {
loop {
let conn = &mut self.replicas[idx];
if conn.write_off >= conn.output.len() {
conn.output.clear();
conn.write_off = 0;
if let ReplicaState::AckSent { replica_id, from_offset, generation } =
&conn.state
{
let rid = replica_id.clone();
let (off, generation) = (*from_offset, *generation);
conn.state = ReplicaState::Streaming {
replica_id: rid,
sent_offset: off,
generation,
};
}
return Ok(());
}
match conn.sock.write(&conn.output[conn.write_off..]) {
Ok(0) => {
conn.close();
return Ok(());
}
Ok(n) => {
conn.write_off += n;
}
Err(e) if e.kind() == io::ErrorKind::WouldBlock => return Ok(()),
Err(e) if e.kind() == io::ErrorKind::Interrupted => continue,
Err(e) => return Err(e),
}
}
}
/// An I/O error on a replica link kills THAT LINK, never the shard.
///
/// A replica that goes away mid-stream makes the primary's next write
/// to it fail with `EPIPE` / `ECONNRESET`. Propagating that out of the
/// reactor ends the shard — and with it every client connection the
/// shard owns, none of which had anything to do with replication.
/// That is what `kevy: shard N exited with error: Broken pipe`
/// meant: killing a replica took down the primary's shards, and the
/// clients saw their connections close with no reply in flight.
///
/// Closing the link is also what the streaming-output cap already
/// does for the same situation reached a different way; a replica
/// reconnects and resumes from the backlog.
pub(crate) fn replica_io_failed(&mut self, idx: usize, what: &str, e: &io::Error) {
if let Some(conn) = self.replicas.get_mut(idx) {
eprintln!(
"kevy: shard {} replica fd {} {what} failed: {e} — dropping link \
(reconnect resumes from the backlog)",
self.id, conn.fd,
);
conn.close();
}
}
/// Remove every replica in [`ReplicaState::Closed`]. Conns whose
/// `Closed.replica_id` is `Some` get their slot's `last_seen_ns`
/// touched before dropping so a reconnect within the window stays
/// correlatable — WITHOUT advancing `acked_offset` (the conn
/// knows what it sent, not what the peer confirmed; a sent-as-
/// acked write here would let `WAIT` count unconfirmed bytes
/// through the reconnect window).
pub(crate) fn reap_closed_replicas(&mut self) {
// Fast path: no Closed conns. Avoids the Instant::now() cost
// on every reactor iteration.
if !self.replicas.iter().any(|r| matches!(r.state, ReplicaState::Closed { .. })) {
return;
}
let now_ns = std::time::Instant::now()
.duration_since(self.replication_epoch)
.as_nanos() as u64;
let mut i = self.replicas.len();
while i > 0 {
i -= 1;
if let ReplicaState::Closed { replica_id } = &self.replicas[i].state {
if let Some(id) = replica_id.as_ref() {
self.slots.touch_or_insert_unacked(id, now_ns);
}
let conn = self.replicas.swap_remove(i);
let _ = self.poller.delete(conn.fd);
// sock drops here → fd closed.
}
}
}
/// Periodic expiry of stale slots. Called from the shard tick.
pub(crate) fn tick_replication_slots(&mut self, now: std::time::Instant) {
if self.replicate.is_none() || self.slots.is_empty() {
return;
}
let now_ns = now
.duration_since(self.replication_epoch)
.as_nanos() as u64;
let window_ns = u64::from(self.replication_reconnect_window_ms) * 1_000_000;
let dropped = self.slots.expire(now_ns, window_ns);
if !dropped.is_empty() {
eprintln!(
"kevy: shard {} expired {} replication slot(s) past reconnect window",
self.id,
dropped.len(),
);
}
}
}