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
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
/// Per-connection byte accumulator for contiguous recv data.
///
/// Handlers always see a contiguous `&[u8]` and return the number of bytes consumed.
/// Unconsumed bytes are retained via O(1) `advance()` instead of shifting.
use bytes::{Bytes, BytesMut};
pub struct RecvAccumulator {
buf: BytesMut,
/// Upper bound on `buf.len()` after an `append`. `append` reports
/// overflow rather than growing past this.
max_size: usize,
}
impl RecvAccumulator {
/// Create a new accumulator with the given initial capacity and an
/// unlimited size cap. For runtime use, prefer
/// [`new_with_max`](Self::new_with_max).
#[allow(dead_code)]
pub fn new(capacity: usize) -> Self {
Self::new_with_max(capacity, usize::MAX)
}
/// Create a new accumulator with an initial capacity and an upper-bound
/// `max_size`. `append` rejects data that would push `buf.len()` past
/// `max_size`, leaving the existing contents intact so the caller can
/// fail the connection rather than OOM.
pub fn new_with_max(capacity: usize, max_size: usize) -> Self {
RecvAccumulator {
buf: BytesMut::with_capacity(capacity),
max_size,
}
}
/// Append received bytes. Returns `false` if the append would push the
/// accumulator past its `max_size` — in that case the existing contents
/// are preserved and the caller should close the connection (or accept
/// that intermediate-flush data is dropped, depending on context).
///
/// Not marked `#[must_use]`: not every caller is in a position to fail
/// the connection (e.g. intermediate buffer-shuffling paths inside
/// `WithDataFuture::poll`). The authoritative cap-enforcement sites are
/// the kernel-recv handlers in `backend/uring/event_loop.rs`.
pub fn append(&mut self, data: &[u8]) -> bool {
if self.buf.len().saturating_add(data.len()) > self.max_size {
return false;
}
self.buf.extend_from_slice(data);
true
}
/// Get a reference to the accumulated data.
pub fn data(&self) -> &[u8] {
&self.buf[..]
}
/// Consume `n` bytes from the front — O(1) via `BytesMut::advance`.
pub fn consume(&mut self, n: usize) {
if n == 0 {
return;
}
debug_assert!(
n <= self.buf.len(),
"consume({n}) exceeds buffer length {}",
self.buf.len()
);
let n = n.min(self.buf.len());
self.buf.advance(n);
}
/// Reset the accumulator (discard all data).
pub fn reset(&mut self) {
self.buf.clear();
}
/// Return the current backing-buffer capacity. Test-only.
#[cfg(test)]
pub(crate) fn capacity(&self) -> usize {
self.buf.capacity()
}
}
use bytes::Buf;
/// Parallel `Vec<RecvAccumulator>` indexed by connection index.
/// Stored as a separate field in EventLoop for borrow splitting.
pub struct AccumulatorTable {
accumulators: Vec<RecvAccumulator>,
}
impl AccumulatorTable {
/// Create a table with `count` accumulators, each with the given initial
/// capacity and no upper-bound size.
#[allow(dead_code)]
pub fn new(count: u32, capacity: usize) -> Self {
Self::new_with_max(count, capacity, usize::MAX)
}
/// Create a table with `count` accumulators, each with the given initial
/// capacity and upper-bound `max_size`.
pub fn new_with_max(count: u32, capacity: usize, max_size: usize) -> Self {
let mut accumulators = Vec::with_capacity(count as usize);
for _ in 0..count {
accumulators.push(RecvAccumulator::new_with_max(capacity, max_size));
}
AccumulatorTable { accumulators }
}
/// Append data to the accumulator at the given index. Returns `false`
/// if the append would exceed the accumulator's `max_size`; in that
/// case the existing buffer is unchanged. The authoritative
/// cap-enforcement sites are the kernel-recv handlers; intermediate
/// flush callers may ignore the return value.
pub fn append(&mut self, index: u32, data: &[u8]) -> bool {
self.accumulators[index as usize].append(data)
}
/// Get accumulated data at the given index.
pub fn data(&self, index: u32) -> &[u8] {
self.accumulators[index as usize].data()
}
/// Consume `n` bytes from the accumulator at the given index.
pub fn consume(&mut self, index: u32, n: usize) {
self.accumulators[index as usize].consume(n);
}
/// Reset the accumulator at the given index.
pub fn reset(&mut self, index: u32) {
self.accumulators[index as usize].reset();
}
/// Return the current backing-buffer capacity for the accumulator at
/// the given index. Test-only.
#[cfg(test)]
pub(crate) fn capacity(&self, index: u32) -> usize {
self.accumulators[index as usize].capacity()
}
/// Detach the accumulator's buffer as a frozen `Bytes` (O(1)).
///
/// The accumulator is left empty but retains the tail capacity of the
/// same allocation via `split_to`. Any subsequent `prepend()` of an
/// unconsumed remainder (the hot pipelined-parse path) reuses that
/// capacity instead of heap-allocating a fresh `BytesMut`.
///
/// Note: when the returned `Bytes` (or sub-slices the parser keeps, as in
/// `with_bytes`) outlive the next `append()`, the shared allocation cannot
/// reclaim its front, so the tail capacity shrinks across cycles and a
/// later `append()` may still reallocate. The win is the avoided per-parse
/// remainder allocation, not elimination of all reallocation.
pub fn take_frozen(&mut self, index: u32) -> Bytes {
let acc = &mut self.accumulators[index as usize];
let len = acc.buf.len();
// Fast path: empty accumulator — `split_to(0)` would unnecessarily
// upgrade `BytesMut` to shared mode, blocking later front-reclaim.
if len == 0 {
return Bytes::new();
}
// `split_to(len)` hands back the filled front as `other` and leaves
// `acc.buf` empty but still owning the tail capacity of the same
// allocation — O(1), no copy. The `freeze()` on the front is also
// O(1). Both operations are non-allocating.
acc.buf.split_to(len).freeze()
}
/// Put unconsumed data back into the accumulator.
///
/// Called after `take_frozen()` when the parser didn't consume
/// everything (e.g. pipelined remainder or incomplete next message).
pub fn prepend(&mut self, index: u32, data: &[u8]) {
if data.is_empty() {
return;
}
let acc = &mut self.accumulators[index as usize];
// Within the single-threaded per-worker event loop, `prepend` is
// always called immediately after `take_frozen` (within the same
// `poll`). No other task can append to this accumulator between
// `take_frozen` and `prepend`, so the buffer must be empty here.
debug_assert!(
acc.buf.is_empty(),
"prepend: accumulator unexpectedly non-empty (single-threaded invariant violated); \
len={}",
acc.buf.len()
);
if acc.buf.is_empty() {
// Fast path (invariant): buffer is empty, tail capacity is
// retained from `take_frozen`'s `split_to` — no allocation.
acc.buf.extend_from_slice(data);
} else {
// Slow path: new data already present (cannot happen in the
// single-threaded runtime). Prepend by building a new buffer
// with remainder first.
let mut new_buf = BytesMut::with_capacity(data.len() + acc.buf.len());
new_buf.extend_from_slice(data);
new_buf.extend_from_slice(&acc.buf);
acc.buf = new_buf;
}
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn append_and_consume() {
let mut acc = RecvAccumulator::new(64);
assert!(acc.append(b"hello "));
assert!(acc.append(b"world"));
assert_eq!(acc.data(), b"hello world");
acc.consume(6);
assert_eq!(acc.data(), b"world");
acc.consume(5);
assert_eq!(acc.data(), b"");
}
#[test]
fn grow_on_overflow() {
let mut acc = RecvAccumulator::new(4);
assert!(acc.append(b"abcdef")); // exceeds initial capacity but not max
assert_eq!(acc.data(), b"abcdef");
}
#[test]
fn reset_clears() {
let mut acc = RecvAccumulator::new(16);
assert!(acc.append(b"data"));
acc.reset();
assert_eq!(acc.data(), b"");
}
#[test]
fn append_past_max_returns_false_and_preserves_contents() {
let mut acc = RecvAccumulator::new_with_max(8, 8);
assert!(acc.append(b"abcdef"));
// Would push to 9 bytes, exceeding max=8.
assert!(!acc.append(b"xyz"));
// Existing contents intact.
assert_eq!(acc.data(), b"abcdef");
}
#[test]
fn table_operations() {
let mut table = AccumulatorTable::new(4, 64);
assert!(table.append(2, b"hello"));
assert_eq!(table.data(2), b"hello");
table.consume(2, 3);
assert_eq!(table.data(2), b"lo");
table.reset(2);
assert_eq!(table.data(2), b"");
}
#[test]
fn table_append_past_max_returns_false() {
let mut table = AccumulatorTable::new_with_max(1, 4, 4);
assert!(table.append(0, b"abcd"));
assert!(!table.append(0, b"e"));
assert_eq!(table.data(0), b"abcd");
}
#[test]
fn take_frozen_and_prepend() {
let mut table = AccumulatorTable::new(2, 64);
assert!(table.append(0, b"$5\r\nhello\r\n$3\r\nbar\r\n"));
let frozen = table.take_frozen(0);
assert_eq!(&frozen[..], b"$5\r\nhello\r\n$3\r\nbar\r\n");
// Accumulator is now empty.
assert_eq!(table.data(0), b"");
// Put back the unconsumed remainder.
table.prepend(0, &frozen[11..]);
assert_eq!(table.data(0), b"$3\r\nbar\r\n");
}
/// After `take_frozen` + `prepend(remainder)`, the accumulator must NOT
/// have allocated a new backing buffer — the tail capacity from the
/// split must still be in place (`capacity() > 0` immediately after
/// `take_frozen`, and the `prepend` does not reallocate).
#[test]
fn take_frozen_preserves_tail_capacity() {
let mut table = AccumulatorTable::new(1, 256);
// Fill with data that will have a remainder after the first parse.
assert!(table.append(0, b"$5\r\nhello\r\n$3\r\nbar\r\n"));
let frozen = table.take_frozen(0);
// Capacity must be non-zero — the tail allocation is retained.
assert!(
table.capacity(0) > 0,
"take_frozen must retain tail capacity, got 0"
);
// Prepend the unconsumed remainder — must not reallocate.
let cap_after_take = table.capacity(0);
let remainder = &frozen[11..];
table.prepend(0, remainder);
assert_eq!(
table.capacity(0),
cap_after_take,
"prepend(remainder) must reuse the retained capacity, not reallocate"
);
assert_eq!(table.data(0), b"$3\r\nbar\r\n");
// Multiple cycles must each preserve capacity (no per-cycle realloc).
for _ in 0..3 {
assert!(table.append(0, b" extra"));
let f = table.take_frozen(0);
assert!(
table.capacity(0) > 0,
"cycle: take_frozen must retain capacity"
);
let cap_before = table.capacity(0);
table.prepend(0, &f[..3]);
assert_eq!(
table.capacity(0),
cap_before,
"cycle: prepend must reuse retained capacity"
);
drop(f);
}
}
#[test]
fn take_frozen_empty() {
let mut table = AccumulatorTable::new(1, 16);
let frozen = table.take_frozen(0);
assert!(frozen.is_empty());
}
#[test]
fn prepend_to_empty() {
let mut table = AccumulatorTable::new(1, 16);
table.prepend(0, b"leftover");
assert_eq!(table.data(0), b"leftover");
}
#[test]
fn prepend_empty_is_noop() {
let mut table = AccumulatorTable::new(1, 16);
assert!(table.append(0, b"existing"));
table.prepend(0, b"");
assert_eq!(table.data(0), b"existing");
}
}