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
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
use std::sync::atomic::{self, AtomicBool, AtomicU16, AtomicU32, Ordering};
use super::{AccessTime, KeyHash};
use crate::common::time::{AtomicInstant, Instant};
use portable_atomic::AtomicU64;
#[derive(Debug)]
pub(crate) struct EntryInfo<K> {
key_hash: KeyHash<K>,
/// `is_admitted` indicates that the entry has been admitted to the cache. When
/// `false`, it means the entry is _temporary_ admitted to the cache or evicted
/// from the cache (so it should not have LRU nodes).
///
/// This is distinct from [`is_retired`](Self::is_retired): `is_admitted`
/// tracks deque membership, while `is_retired` tracks CHT membership. A
/// freshly inserted entry is **not retired** and **not admitted** until its
/// `WriteOp::Upsert` is drained; a retired entry may still be `is_admitted`
/// until `handle_remove*` unlinks it.
is_admitted: AtomicBool,
/// CHT-membership flag: `false` while the entry is alive (still backed by
/// the concurrent hash table), `true` once the entry has been removed.
///
/// The flag is set by `retire()` at the moment the entry is unlinked from
/// the CHT and is never cleared (`false → true`, monotonic). The
/// invariant it enforces is:
///
/// > If `is_retired == true`, no policy-tier mutation may run against
/// > this `EntryInfo`.
///
/// Examples of policy-tier mutations:
///
/// - admitting the entry to an LRU deque,
/// - moving its deque node to the back,
/// - scheduling a timer-wheel node,
/// - incrementing `weighted_size`.
///
/// ### Why we need it
///
/// moka stores entries in a lock-free CHT (source of truth) and mirrors
/// their ordering into LRU deques updated asynchronously via a bounded
/// `WriteOp` channel. Without this flag, a `WriteOp::Upsert` can sit in the
/// channel long enough that the backing CHT entry is evicted before the op
/// is drained; the drain then pushes an "orphan" deque node that the
/// eviction loop cannot reach, blocking progress and stalling eviction
/// indefinitely (see <https://github.com/moka-rs/moka/issues/590>). The
/// retire store is performed inside the CHT's post-success
/// `with_previous_entry` callback so the transition linearises with the
/// bucket-pointer CAS that physically unlinked the entry; any consumer
/// draining a stale op thereafter observes `is_retired == true` and
/// declines to touch policy state.
///
/// ### Why a single bit suffices (no `Dead` state)
///
/// Caffeine encodes a three-state lifecycle (`Alive → Retired → Dead`)
/// because its `makeDead()` can be called twice for the same node: once
/// inline from `evictEntry()` during admission, and again later from
/// `RemovalTask.run()` draining the write buffer. The `isDead()` guard at
/// the top of `makeDead` prevents double-subtraction of `weightedSize`.
///
/// moka has no such double-enter path: every CHT removal site gates its
/// `handle_remove_*` call on `if let Some(entry) = …`; i.e. on winning
/// the CHT bucket CAS. So `handle_remove_*` runs exactly once per
/// retirement, structurally.
///
/// ### When a third state would become necessary
///
/// A third state (`Dead`) earns its keep when there are **multiple
/// finalization paths** that may both target the same entry. Currently,
/// moka has one at the CAS-gated inline `handle_remove_*`.
///
/// If moka ever grows a second finalization path that runs alongside the
/// `WriteOp::Remove` drain (e.g., an inline emergency-evict, an
/// async load-completion that touches policy state, or any code that
/// can re-enter `handle_remove_*` for an already-retired entry), the
/// idempotency requirement returns and a third state (or an equivalent
/// "weight already subtracted" marker) becomes necessary. Until then, a
/// single bit is sufficient.
is_retired: AtomicBool,
/// `entry_gen` (entry generation) is incremented every time the entry is updated
/// in the concurrent hash table.
entry_gen: AtomicU16,
/// `policy_gen` (policy generation) is incremented every time entry's `WriteOp`
/// is applied to the cache policies including the access-order queue (the LRU
/// deque).
policy_gen: AtomicU16,
/// Packed expiration state: contains both `expiration_time` (upper 52 bits) and
/// `expiry_gen` (lower 12 bits) in a single atomic u64 for consistent reads.
///
/// Encoding:
/// - Bits 0-11: 12-bit expiry_gen (wraps from 4095 to 0)
/// - Bits 12-63: 52-bit expiration timestamp (nanoseconds from a monotonic clock/Instant)
/// - Sentinel: When time field (bits 12-63) is all 1s (0xFFFF_FFFF_FFFF_F000),
/// represents "None" (no expiration). Gen bits are always preserved.
/// - Valid timestamps are clamped to 52-bit range to avoid collisions with sentinel.
///
/// NOTE: The time value is relative to the runtime's monotonic clock (not Unix epoch).
/// It is obtained from `Instant` and should not be interpreted as an absolute time
/// suitable for serialization or cross-process comparison.
expiration_state: AtomicU64,
last_accessed: AtomicInstant,
last_modified: AtomicInstant,
policy_weight: AtomicU32,
}
impl<K> EntryInfo<K> {
#[inline]
pub(crate) fn new(key_hash: KeyHash<K>, timestamp: Instant, policy_weight: u32) -> Self {
#[cfg(feature = "unstable-debug-counters")]
super::debug_counters::InternalGlobalDebugCounters::entry_info_created();
Self {
key_hash,
is_admitted: AtomicBool::default(),
is_retired: AtomicBool::default(),
// `entry_gen` starts at 1 and `policy_gen` start at 0.
entry_gen: AtomicU16::new(1),
policy_gen: AtomicU16::new(0),
// Initial state: None (time field all 1s), gen=0
expiration_state: AtomicU64::new(0xFFFF_FFFF_FFFF_F000),
last_accessed: AtomicInstant::new(timestamp),
last_modified: AtomicInstant::new(timestamp),
policy_weight: AtomicU32::new(policy_weight),
}
}
#[inline]
pub(crate) fn key_hash(&self) -> &KeyHash<K> {
&self.key_hash
}
#[inline]
pub(crate) fn is_admitted(&self) -> bool {
self.is_admitted.load(Ordering::Acquire)
}
#[inline]
pub(crate) fn set_admitted(&self, value: bool) {
self.is_admitted.store(value, Ordering::Release);
}
/// Returns `true` iff the entry has been retired (removed from the CHT).
#[inline]
pub(crate) fn is_retired(&self) -> bool {
self.is_retired.load(Ordering::Acquire)
}
/// Returns `true` iff the entry is still alive (not yet retired).
/// Consumers of buffered operations (`WriteOp::Upsert`, `ReadOp::Hit`,
/// TinyLFU inline victim scans) MUST check this before mutating policy
/// state; see the `is_retired` field docs.
#[inline]
pub(crate) fn is_alive(&self) -> bool {
!self.is_retired()
}
/// Transitions the entry to the retired state. Must be called inside the
/// post-CAS `with_previous_entry` callback of `cht::remove_entry_if_and`
/// (or an equivalent linearisation point) so the transition is atomic
/// with the CHT unlink.
///
/// The callsite is exclusive: the `with_previous_entry` closure fires at
/// most once per `EntryInfo` (it only runs for the thread that won the
/// bucket-pointer CAS), so no other thread can be transitioning the same
/// field in parallel. A plain `Release` store is therefore sufficient
/// (no RMW needed) and matches the `is_admitted` pattern. Double-retire
/// would indicate a broken invariant (caught by a debug-mode assertion).
#[inline]
pub(crate) fn retire(&self) {
debug_assert!(
!self.is_retired.load(Ordering::Acquire),
"retire() called on already-retired entry; lifecycle invariant broken",
);
self.is_retired.store(true, Ordering::Release);
}
/// Returns `true` if the `ValueEntry` having this `EntryInfo` is dirty.
///
/// Dirty means that the entry has been updated in the concurrent hash table but
/// not yet in the cache policies such as access-order queue.
#[inline]
pub(crate) fn is_dirty(&self) -> bool {
let result =
self.entry_gen.load(Ordering::Relaxed) != self.policy_gen.load(Ordering::Relaxed);
atomic::fence(Ordering::Acquire);
result
}
#[inline]
pub(crate) fn entry_gen(&self) -> u16 {
self.entry_gen.load(Ordering::Acquire)
}
/// Increments the entry generation and returns the new value.
#[inline]
pub(crate) fn incr_entry_gen(&self) -> u16 {
// NOTE: This operation wraps around on overflow.
let prev = self.entry_gen.fetch_add(1, Ordering::AcqRel);
// Need to add `1` to the previous value to get the current value.
prev.wrapping_add(1)
}
/// Sets the policy generation to the given value.
#[inline]
pub(crate) fn set_policy_gen(&self, value: u16) {
let g = &self.policy_gen;
loop {
let current = g.load(Ordering::Acquire);
// Do not set the given value if it is smaller than the current value of
// `policy_gen`. Note that the current value may have been wrapped
// around. If the value is much larger than the current value, it is
// likely that the value of `policy_gen` has been wrapped around.
if current >= value || value.wrapping_sub(current) > u16::MAX / 2 {
break;
}
// Try to set the value.
if g.compare_exchange_weak(current, value, Ordering::AcqRel, Ordering::Acquire)
.is_ok()
{
break;
}
}
}
#[inline]
pub(crate) fn policy_weight(&self) -> u32 {
self.policy_weight.load(Ordering::Acquire)
}
pub(crate) fn set_policy_weight(&self, size: u32) {
self.policy_weight.store(size, Ordering::Release);
}
/// Atomically reads both `expiration_time` and `expiry_gen` as a single unit.
/// Returns `(expiration_time, expiry_gen)` where expiration_time is None if unset.
///
/// This provides a consistent snapshot of the expiration state, avoiding TOCTOU
/// issues where the time and generation could be read separately while being
/// modified by another thread.
#[inline]
pub(crate) fn expiration_state(&self) -> (Option<Instant>, u32) {
const GEN_MASK: u64 = 0xFFF;
const TIME_MASK: u64 = 0xFFFF_FFFF_FFFF_F000;
let packed = self.expiration_state.load(Ordering::Acquire);
// Extract time field and gen bits
let time_nanos = packed & TIME_MASK;
let gen = (packed & GEN_MASK) as u32;
// Check if time field (upper 52 bits) is all 1s (sentinel for None)
if time_nanos == TIME_MASK {
(None, gen)
} else {
(Some(Instant::from_nanos(time_nanos)), gen)
}
}
/// Sets the expiration time and returns the new expiry generation.
pub(crate) fn set_expiration_time(&self, time: Option<Instant>) -> u32 {
const GEN_MASK: u64 = 0xFFF;
const TIME_MASK: u64 = 0xFFFF_FFFF_FFFF_F000;
// Use compare_exchange to atomically update the expiration state.
// This prevents race conditions where multiple threads try to update
// the expiration time simultaneously.
loop {
let prev_packed = self.expiration_state.load(Ordering::Acquire);
// Extract previous generation (always preserved, even for None state)
let prev_gen = (prev_packed & GEN_MASK) as u32;
let new_gen = prev_gen.wrapping_add(1) & GEN_MASK as u32;
// Pack the new state
let new_packed = if let Some(t) = time {
// Clamp timestamp to 52-bit range. Ensure it's strictly less than TIME_MASK
// to avoid collision with the sentinel (None) value.
let mut nanos = t.as_nanos() & TIME_MASK;
// If nanos equals TIME_MASK, adjust it down by one time unit (4096 nanos)
// to avoid corrupting the generation counter bits (should never happen in practice).
if nanos == TIME_MASK {
nanos = TIME_MASK - 0x1000; // Subtract one 12-bit unit to keep lower bits clear
}
debug_assert!(nanos < TIME_MASK, "Timestamp value collides with sentinel");
// Pack: store nanos in upper 52 bits, gen in lower 12 bits
nanos | (new_gen as u64)
} else {
// Sentinel: time field all 1s, gen preserved in lower bits
TIME_MASK | (new_gen as u64)
};
// Try to atomically update if the state hasn't changed
match self.expiration_state.compare_exchange_weak(
prev_packed,
new_packed,
Ordering::AcqRel,
Ordering::Acquire,
) {
Ok(_) => return new_gen, // Successfully updated
Err(_) => continue, // State changed, retry
}
}
}
}
#[cfg(feature = "unstable-debug-counters")]
impl<K> Drop for EntryInfo<K> {
fn drop(&mut self) {
super::debug_counters::InternalGlobalDebugCounters::entry_info_dropped();
}
}
impl<K> AccessTime for EntryInfo<K> {
#[inline]
fn last_accessed(&self) -> Option<Instant> {
self.last_accessed.instant()
}
#[inline]
fn set_last_accessed(&self, timestamp: Instant) {
self.last_accessed.set_instant(timestamp);
}
#[inline]
fn last_modified(&self) -> Option<Instant> {
self.last_modified.instant()
}
#[inline]
fn set_last_modified(&self, timestamp: Instant) {
self.last_modified.set_instant(timestamp);
}
}
#[cfg(test)]
mod tests {
use super::{EntryInfo, KeyHash};
use crate::common::time::Instant;
use std::sync::Arc;
fn make_entry_info() -> EntryInfo<u64> {
let kh = KeyHash::new(Arc::new(42u64), 0xdead_beef);
EntryInfo::new(kh, Instant::from_nanos(0), 1)
}
#[test]
fn lifecycle_initial_is_alive() {
let info = make_entry_info();
assert!(info.is_alive());
assert!(!info.is_retired());
}
#[test]
fn lifecycle_retire_transitions_to_retired() {
let info = make_entry_info();
info.retire();
assert!(!info.is_alive());
assert!(info.is_retired());
}
#[test]
#[cfg(debug_assertions)]
#[should_panic(expected = "retire() called on already-retired entry")]
fn lifecycle_double_retire_debug() {
let info = make_entry_info();
info.retire();
// Second retire violates the exclusive-call invariant; debug-asserted.
info.retire();
}
}