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
//! PERF-alloc_perchunk (allocation churn on the per-chunk hot path).
//!
//! KEY: alloc_perchunk
//! VECTOR: SPEED / OPTIMIZATION (Law 7 (extra copies are production bugs at scale)).
//!
//! ## The defect
//!
//! Every chunk scanned routes through `CompiledScanner::prepare_chunk`
//! (crates/scanner/src/engine/backend_dispatch.rs:69), which builds a
//! `ScannerPreprocessedText`. On the overwhelmingly common path, a chunk
//! with no structured-config shape and no multiline concatenation indicators
//! (plain source, .env lines, config without continuation), it takes the
//! passthrough branch:
//!
//! crates/scanner/src/engine/backend_dispatch.rs:116
//! `ScannerPreprocessedText::passthrough(data_ref)`
//!
//! which lands at:
//!
//! crates/scanner/src/multiline/config.rs:63-84 (multiline feature on)
//! ```
//! pub fn passthrough(text: &str) -> Self {
//! ...
//! Self { text: text.to_string(), original_end, mappings } // <-- FULL COPY
//! }
//! ```
//! (and the identical `text: text.to_string()` at
//! crates/scanner/src/multiline/preprocessor.rs:168 and
//! crates/scanner/src/types.rs:147 for the non-multiline build).
//!
//! `text.to_string()` heap-allocates and memcpy's the ENTIRE chunk body into a
//! fresh `String` on every single chunk, even though `prepare_chunk` already
//! holds a borrow of `chunk.data` for the whole call and the downstream
//! scan reads `preprocessed.text` immutably. The preprocessed text for a
//! passthrough chunk is byte-identical to the input (the copy buys nothing).
//! At fleet scale (hundreds of thousands of files, one chunk each) this is one
//! whole-file-sized allocation + memcpy per file, hammering the global
//! allocator on the hot path the engine spends most of its time in.
//!
//! ## What this tripwire measures (hardware-independent)
//!
//! A `#[global_allocator]` counts total bytes requested during a scan. We scan
//! a passthrough chunk of `N` bytes and again of `2N` bytes (identical shape,
//! no detector hits, prefilters passed) and look at how the per-scan allocation
//! volume GROWS with chunk size:
//!
//! growth = bytes_alloc(2N) - bytes_alloc(N)
//!
//! Because `passthrough` copies the whole body, `growth >= N` (the extra N
//! bytes of body copied into the second scan's `String`). Once the path
//! borrows the chunk text instead of copying it, the preprocessed `String`
//! disappears and `growth` collapses to the size-independent bookkeeping
//! (line mappings etc.), far below `N`.
//!
//! This is a RATIO/asymptotic assertion on a deterministic allocation COUNTER,
//! not a wall-clock timing, so it does not flake on slow CI hosts, the byte
//! counts are identical on every machine for a given build.
//!
//! Profile: built/run under the workspace `release-fast` characteristics
//! (the prebuilt black-box binary lives at
//! `/mnt/FlareTraining/santh-archive/cargo-target/release-fast/keyhog`); the
//! assertion is profile-independent because it counts bytes, not nanoseconds.
//!
//! ## Measured on the current tree (dev box)
//!
//! With N = 256 KiB the passthrough copy makes `growth` track the body size:
//! growth ~= N (>= 262_144 bytes). The FLOOR is set at N/2 (131_072 bytes)
//! the real defect overshoots it by ~2x while leaving generous headroom so
//! incidental per-scan allocations never trip it. After the borrow fix,
//! `growth` is a few KiB of line-mapping bookkeeping (well under N/2), so the
//! test passes.
use std::alloc::{GlobalAlloc, Layout, System};
use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
use keyhog_core::{Chunk, ChunkMetadata};
use keyhog_scanner::CompiledScanner;
/// Counts total bytes requested via `alloc` while `COUNTING` is enabled.
struct CountingAlloc;
static BYTES_ALLOCATED: AtomicUsize = AtomicUsize::new(0);
static COUNTING: AtomicBool = AtomicBool::new(false);
unsafe impl GlobalAlloc for CountingAlloc {
unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
if COUNTING.load(Ordering::Relaxed) {
BYTES_ALLOCATED.fetch_add(layout.size(), Ordering::Relaxed);
}
System.alloc(layout)
}
unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
System.dealloc(ptr, layout)
}
unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
if COUNTING.load(Ordering::Relaxed) {
// A growing realloc copies `new_size` fresh bytes, count the
// delta so buffer GROWTH (e.g. a Vec doubling) is captured too.
if new_size > layout.size() {
BYTES_ALLOCATED.fetch_add(new_size - layout.size(), Ordering::Relaxed);
}
}
System.realloc(ptr, layout, new_size)
}
}
#[global_allocator]
static ALLOC: CountingAlloc = CountingAlloc;
fn load_embedded_detectors() -> Vec<keyhog_core::DetectorSpec> {
keyhog_core::load_embedded_detectors_or_fail().expect("embedded detector corpus must load")
}
/// A passthrough-shaped chunk: realistic source bytes so the alphabet/bigram
/// prefilters in `scan_with_deadline_and_backend` let it reach
/// `prepare_chunk`, but no concatenation indicators (so it takes the
/// `passthrough` branch, NOT the multiline preprocessor) and no real
/// credential (so detector extraction does no body-proportional work that
/// would muddy the size-delta signal).
fn passthrough_chunk(target_bytes: usize) -> Chunk {
// Each line is benign code-shaped text: letters, an assignment, a quote.
// It carries enough distinct bytes/bigrams to pass the prefilters but
// contains no detector literal prefix and no `\` / template-literal /
// implicit-concatenation indicator that would divert into the multiline
// path.
const LINE: &str = "let value_name = compute_label(index_position, lookup_table);\n";
let mut s = String::with_capacity(target_bytes + LINE.len());
while s.len() < target_bytes {
s.push_str(LINE);
}
Chunk {
data: s.into(),
metadata: ChunkMetadata {
base_offset: 0,
base_line: 0,
source_type: "filesystem".into(),
// A neutral .rs path: not a sensitive/config name, not `.keyhog`,
// not under a `detectors/` dir, so nothing short-circuits the scan.
path: Some("src/module/component_helper.rs".into()),
commit: None,
author: None,
date: None,
mtime_ns: None,
size_bytes: None,
..Default::default()
},
}
}
/// Run `scan` once and return the bytes the allocator handed out during it.
fn scan_alloc_bytes(scanner: &CompiledScanner, chunk: &Chunk) -> usize {
BYTES_ALLOCATED.store(0, Ordering::Relaxed);
COUNTING.store(true, Ordering::Relaxed);
let matches = scanner.scan(chunk);
COUNTING.store(false, Ordering::Relaxed);
// Keep the result observably alive so the scan can't be optimized away,
// and assert the fixture really is a no-match passthrough so the alloc
// signal is the preprocessing copy, not match post-processing.
assert!(
matches.is_empty(),
"fixture chunk should produce no matches (got {}); a matching chunk \
would add body-proportional post-processing allocations and pollute \
the per-chunk-copy signal",
matches.len()
);
BYTES_ALLOCATED.load(Ordering::Relaxed)
}
#[test]
fn passthrough_prepare_does_not_copy_whole_chunk_body() {
const N: usize = 256 * 1024; // 256 KiB base chunk
// FLOOR: the per-scan allocation growth from N -> 2N must stay BELOW N/2.
// The full-body `text.to_string()` copy makes growth >= N (one extra copy
// of the extra N bytes), so it trips this floor by ~2x. A borrow-based
// passthrough grows only by size-independent bookkeeping (a few KiB of
// line mappings), far under N/2.
const FLOOR_BYTES: usize = N / 2;
let detectors = load_embedded_detectors();
let scanner = CompiledScanner::compile(detectors).expect("compile scanner");
let chunk_n = passthrough_chunk(N);
let chunk_2n = passthrough_chunk(2 * N);
// Warm: first scans lazily allocate HS scratch, regex DFA caches, intern
// tables, etc. Those are one-time and must NOT be attributed to per-chunk
// copy cost. Run each size twice up front so the measured pass below sees
// only steady-state per-chunk allocation.
for _ in 0..2 {
let _ = scanner.scan(&chunk_n);
let _ = scanner.scan(&chunk_2n);
}
// Best-of-K on the COUNTER: take the minimum allocation reading across K
// runs at each size, so any incidental lazy first-touch that slipped past
// warmup is excluded (the min run has none).
const K: usize = 3;
let mut min_n = usize::MAX;
let mut min_2n = usize::MAX;
for _ in 0..K {
min_n = min_n.min(scan_alloc_bytes(&scanner, &chunk_n));
min_2n = min_2n.min(scan_alloc_bytes(&scanner, &chunk_2n));
}
// Growth in per-scan allocation volume as the chunk body doubles.
let growth = min_2n.saturating_sub(min_n);
assert!(
growth < FLOOR_BYTES,
"PERF-alloc_perchunk-1: per-chunk allocation grows with chunk body \
size: the passthrough path copies the whole chunk into a fresh \
String.\n scan({N}B) allocated {min_n} B; scan({}B) allocated {min_2n} B\n \
growth (2N - N) = {growth} B (must be < {FLOOR_BYTES} B = N/2)\n \
Defect: ScannerPreprocessedText::passthrough does `text.to_string()` \
(crates/scanner/src/multiline/config.rs:80, \
preprocessor.rs:168, types.rs:147), reached from \
prepare_chunk (engine/backend_dispatch.rs:116) on every non-multiline \
chunk. growth >= N here means one full-body copy per chunk.\n \
Fix: make the passthrough preprocessed text BORROW chunk.data \
(Cow<'a, str> / &str) instead of owning a copy, so per-chunk \
allocation is size-independent and `growth` drops to the line-mapping \
bookkeeping (a few KiB).",
2 * N
);
}