yo_kv/keyspace.rs
1//! One database, and the parts of it that are not about any particular type.
2//!
3//! This is the `dict` a Redis `SELECT` picks between, and one of these is what a
4//! shard owns. It was called `Strings` while strings were the only thing in it,
5//! which was accurate for M2 and stopped being accurate the moment a set needed
6//! somewhere to live.
7//!
8//! The commands hang off this as separate `impl` blocks, one file per type, so
9//! that `SET` lives in [`strings`](crate::strings) next to the other twenty five
10//! string commands rather than in a file that is the whole of Redis. They are
11//! methods on the keyspace and not on some per type object because a key belongs
12//! to the database and not to a type: `DEL` does not care what it is deleting,
13//! and `SADD` against a string has to be able to see that it is a string.
14//!
15//! # Not Sync
16//!
17//! Like everything a stripe holds. One of these is worked on by one thread at a
18//! time and the lock around it is what says which, so nothing in here takes a
19//! lock or an atomic of its own.
20//!
21//! It is `Send`, because which thread that is changes from one command to the
22//! next: a stripe goes to whoever takes it. Nothing in here is tied to a
23//! particular thread, and the arena underneath says the same thing for the same
24//! reason.
25
26use std::sync::atomic::{AtomicU64, Ordering};
27
28use yo_common::{Addr, Code, Error, Result, Rng, bytes_eq};
29use yo_index::RawMap;
30
31use crate::Clock;
32use crate::access::{Access, Lfu, Policy};
33use crate::array::Array;
34use crate::cold::Store;
35use crate::evict;
36use crate::foreign::Foreign;
37use crate::hash::{self, Hash};
38use crate::list::{self, List};
39use crate::lookups;
40use crate::news;
41use crate::set::{self, Set};
42use crate::slab::{Bytes, Slab};
43use crate::stream::{self, Stream};
44use crate::tier::{self, Faulted, Relief, Tier};
45use crate::ttl::{self, Applied, Ask, Cond};
46use crate::value::{self, Kind, Str};
47use crate::zset::{self, Zset};
48
49/// Every collection already answered this question, and this is the answer said
50/// once more in a shape the slab can ask for without knowing what it is holding.
51///
52/// Here rather than in the five type files because it is one fact about the
53/// keyspace and not five facts about five types, and because a reader looking
54/// for how the memory total is kept should find it next to the slabs it counts.
55macro_rules! bytes {
56 ($($t:ty),*) => { $(impl Bytes for $t {
57 #[inline]
58 fn memory_bytes(&self) -> usize {
59 <$t>::memory_bytes(self)
60 }
61 })* };
62}
63bytes!(Set, Hash, List, Zset, Array, Stream);
64
65/// A foreign body counts what it says it counts, plus the box around it.
66///
67/// Not through the macro, because the macro calls an inherent method of the
68/// same name and this one is a trait method reached through a vtable. The
69/// pointer itself is two words on top of whatever the engine reports, which is
70/// the price of the escape and is worth naming rather than losing.
71impl Bytes for Box<dyn Foreign> {
72 #[inline]
73 fn memory_bytes(&self) -> usize {
74 self.as_ref().memory_bytes() + std::mem::size_of::<Box<dyn Foreign>>()
75 }
76}
77
78/// One database: every key, whatever type it holds.
79pub struct Keyspace {
80 pub(crate) map: RawMap,
81 pub(crate) clock: Clock,
82 /// Keys that were found dead on the way to answering something else.
83 pub(crate) expired: u64,
84 /// Keys thrown away to make room, which is a different number entirely.
85 ///
86 /// Redis keeps `expired_keys` and `evicted_keys` apart in `INFO` and the
87 /// distinction is the one people watch: expiry is the client getting what it
88 /// asked for, and eviction is the server deciding it cannot keep a promise
89 /// nobody asked it to break.
90 pub(crate) evicted: u64,
91 /// Lookups a client's read made that found the key, which Redis reports as
92 /// `keyspace_hits`.
93 pub(crate) hits: u64,
94 /// Lookups a client's read made that did not, which Redis reports as
95 /// `keyspace_misses`.
96 pub(crate) misses: u64,
97 /// Hash fields that reached their own deadline, whoever found them.
98 ///
99 /// Redis's `expired_subkeys`, and kept apart from
100 /// [`Keyspace::expired`] there and here, because a field going is not a key
101 /// going and a hash of a thousand fields can lose all thousand without the
102 /// key count moving at all.
103 pub(crate) expired_fields: u64,
104 /// How many of those the cycle found rather than a command tripping over.
105 ///
106 /// Redis's `expired_subkeys_active`, and the only reason it is a second
107 /// number is that it is the one that says whether the cycle is keeping up.
108 pub(crate) expired_fields_active: u64,
109 /// Keys of hashes that have been given a field deadline at some point.
110 ///
111 /// The list [`Keyspace::field_expire_cycle`] sweeps, and the reason it can
112 /// be a sweep at all. A field deadline is not in the record, so the marked
113 /// index the key cycle draws from cannot see one, and the alternative to a
114 /// list is walking every hash in the database asking each whether it has a
115 /// deadline that has passed.
116 ///
117 /// Keys and not slab slots, because the sweep has to name the key: it
118 /// publishes `hexpired` against it and deletes it when the last field goes,
119 /// and a body does not know what it is called.
120 ///
121 /// A name goes on here the first time a hash takes a deadline, which each
122 /// hash does once, and comes off when the key is no longer a hash. It is
123 /// empty and costs nothing on a server that has never used the `HEXPIRE`
124 /// family, which is nearly all of them.
125 pub(crate) field_deadlines: Vec<Box<[u8]>>,
126 /// Where the next sweep of that list starts.
127 pub(crate) field_at: usize,
128 /// Every set in this database, addressed by the number in its record.
129 pub(crate) sets: Slab<Set>,
130 /// Every hash in this database, addressed the same way.
131 ///
132 /// A slab per type rather than one slab of an enum, so that a record's four
133 /// bytes index a `Hash` directly and reaching one is a load and not a load
134 /// followed by a discriminant check. The type tag in the record already
135 /// says which slab to look in, so the discriminant would be a second copy
136 /// of a fact the record has.
137 pub(crate) hashes: Slab<Hash>,
138 /// Every list in this database, addressed the same way.
139 pub(crate) lists: Slab<List>,
140 /// Every sorted set in this database, addressed the same way.
141 pub(crate) zsets: Slab<Zset>,
142 /// Every sparse array in this database, addressed the same way.
143 pub(crate) arrays: Slab<Array>,
144 /// Every stream in this database, addressed the same way.
145 pub(crate) streams: Slab<Stream>,
146 /// Every foreign body in this database, addressed the same way.
147 ///
148 /// A box per slot rather than a value, because the thing in it is not sized
149 /// here and could not be. That is one indirection more than the other
150 /// slabs pay, and it buys the graph, document and vector engines a place in
151 /// the keyspace without this crate depending on any of them. See
152 /// [`crate::foreign`].
153 ///
154 /// Empty on a server that has never held one, which is every server today,
155 /// and an empty slab is three words.
156 pub(crate) foreign: Slab<Box<dyn Foreign>>,
157 /// How many keys hold a body that is in a slab right now.
158 ///
159 /// Not how many hold something that is not a string, which is what it used to
160 /// be and what it still is on a database with no file behind it. A collection
161 /// that has been demoted has no slab slot, so it does not count here, and
162 /// that is what every reader of this number wants: [`Keyspace::free_body`]
163 /// has nothing to free for one, and a sweep has nothing to move.
164 ///
165
166 /// This exists so that a database of nothing but strings, which is every
167 /// benchmark today and most of what `SET` sees, can skip the body check in
168 /// [`Keyspace::free_body`] on one predictable branch against a field that is
169 /// already hot, rather than paying a second lookup per write forever.
170 pub(crate) bodies: usize,
171 /// Where a set changes representation.
172 pub(crate) limits: set::Limits,
173 /// Where a hash changes representation.
174 pub(crate) hash_limits: hash::Limits,
175 /// Where a list changes representation.
176 pub(crate) list_limits: list::Limits,
177 /// Where a sorted set changes representation.
178 pub(crate) zset_limits: zset::Limits,
179 /// Where a stream starts a new node, which is two `CONFIG` values.
180 pub(crate) stream_limits: stream::Limits,
181 /// What this database would evict, and therefore what a read writes back.
182 ///
183 /// One server wide setting in Redis, carried per database here for the same
184 /// reason the size ladder is: a `Keyspace` is reached without a server and
185 /// has to be able to answer on its own. `CONFIG SET maxmemory-policy` writes
186 /// it to all of them.
187 pub(crate) policy: Policy,
188 /// The two numbers the LFU counter moves by, which are `CONFIG` values.
189 pub(crate) lfu: Lfu,
190 /// How many keys a round of eviction sampling looks at.
191 ///
192 /// `maxmemory-samples`, carried per database for the same reason the policy
193 /// is. See [`evict::SAMPLES`] for why the default is five.
194 pub(crate) samples: usize,
195 /// The good candidates from earlier rounds of eviction sampling.
196 ///
197 /// See [`evict::Pool`]. Empty and costing nothing until the first eviction,
198 /// which on most databases is never.
199 pub(crate) pool: evict::Pool,
200 /// Where `SPOP` and `SRANDMEMBER` draw from.
201 pub(crate) rng: Rng,
202 /// Where a demoted value goes and comes back from, if this database has one.
203 ///
204 /// `None` on a database with no file behind it, which is every embedded
205 /// caller that never opened one and every test that does not care, and on
206 /// such a database no record is ever cold and every check against this is a
207 /// null test on a field in the same cache line as the map.
208 ///
209 /// Boxed rather than a type parameter on `Keyspace`. The parameter would
210 /// have to be named by `yo-resp`, by the typed API and by every caller of
211 /// either, all to spell a type that only the code opening the file knows,
212 /// and it would be a parameter on the hot path to describe the cold one. See
213 /// [`Blocks`].
214 pub(crate) tier: Option<Tier<Store>>,
215 /// The last value read off the file, for the read that is being answered.
216 ///
217 /// A demoted value that is served rather than promoted has to live
218 /// somewhere for the length of one command, because the record it came from
219 /// holds an address and the caller was promised bytes. One buffer, cleared
220 /// and refilled, on the same argument as [`Keyspace::scratch`]: a fault is
221 /// already a device read and a malloc on top of it is free by comparison,
222 /// but it is also unnecessary.
223 ///
224 /// It holds the value of the last key that was faulted and nothing says
225 /// which key that was, which is why nothing reads it without having faulted
226 /// in the same call. Everything that does goes through
227 /// [`Keyspace::warm`].
228 pub(crate) cold: Vec<u8>,
229 /// Whose value is in [`Keyspace::cold`].
230 ///
231 /// Only ever read by a debug assertion, and it is there because the failure
232 /// it catches is silent: a read that forgets to warm and then finds a cold
233 /// record would hand back whatever the last fault put in the buffer, which
234 /// is a real value belonging to a different key. A test would see plausible
235 /// bytes and pass. The copy costs a key's worth of memcpy on a path that
236 /// has just read a device, which is nothing next to what it is guarding.
237 pub(crate) cold_key: Vec<u8>,
238 /// A collection body on its way to the file or on its way back.
239 ///
240 /// Not [`Keyspace::cold`], which holds the value of the last string that was
241 /// faulted and is read after the fact by [`Keyspace::value_of`]. This buffer
242 /// is written and consumed inside one call and nothing looks at it
243 /// afterwards, so sharing the other one would mean a `SADD` on a demoted set
244 /// quietly replacing the bytes a `GET` in the same pipeline was about to hand
245 /// back. One `Vec` each, cleared and refilled, is three words of a struct
246 /// that already has a hundred.
247 pub(crate) frozen: Vec<u8>,
248 /// The last collection key that was resolved, for the command behind it.
249 memo: Memo,
250 /// One buffer for the commands that have to hold an element while the
251 /// structure it came out of is being written.
252 ///
253 /// [`Keyspace::lmove`] is the reason this is here: it takes an element out
254 /// of one list and puts it into another, so there is a moment where the
255 /// bytes belong to nothing, and the borrow it would need to avoid that is a
256 /// borrow of two lists at once when the two lists may be the same one. A
257 /// `Vec` per call is the obvious way to cover that moment and it is a malloc
258 /// and a free on a command that a queue sends millions of. This is the same
259 /// `Vec` every time, cleared rather than freed, so the steady state is no
260 /// allocator call at all.
261 ///
262 /// [`Keyspace::append`], [`Keyspace::setrange`] and the string arm of
263 /// [`Keyspace::set_expiry`] use it for the same shape of problem: each of
264 /// them has to hold the old value while it writes the new record, and each
265 /// of them was doing that with a fresh `Vec` of the whole value. They cannot
266 /// overlap, because each one puts the buffer back before it returns and one
267 /// command runs at a time.
268 ///
269 /// It lives on the database and not on the caller because the callers are
270 /// wire handlers that are handed a `&mut Keyspace` and nothing else.
271 ///
272 /// It starts at [`SCRATCH`] bytes rather than empty. An empty one grows on
273 /// the first command that uses it, and that growth is a real allocation on a
274 /// command path even though it happens once. Buying it here, where nobody is
275 /// waiting, makes the rule Y7 enforces true without an exception written for
276 /// it. A value larger than that still grows it, and that one is allocation
277 /// proportional to what the caller sent rather than overhead per command.
278 pub(crate) scratch: Vec<u8>,
279
280 /// The same idea for indices rather than bytes.
281 ///
282 /// `ZRANDMEMBER` with a positive count under the size of the set does a
283 /// partial Fisher-Yates, and that needs the permutation somewhere while it
284 /// draws from it. One buffer, cleared and refilled, rather than one `Vec`
285 /// per call, because sampling is a thing callers do in a loop.
286 ///
287 /// It does not start at a capacity, unlike [`Keyspace::scratch`]. There is
288 /// no size to guess: the buffer has to be as long as the set, so the first
289 /// call on a set larger than anything seen before grows it whatever it was
290 /// given to start with. That growth is proportional to the data rather than
291 /// per command.
292 pub(crate) rows: Vec<usize>,
293
294 /// The tables set algebra fills in, kept rather than built per call.
295 ///
296 /// Same idea again, one level up: a union walks everything into a hash
297 /// table and lets the table be the duplicate check, and building that table
298 /// was the largest single allocation left on any command path. See
299 /// [`setops::Scratch`], which is where the two tables and the argument for
300 /// them live.
301 ///
302 /// It is one table per database and not one per command, so a database that
303 /// has answered a union over a million members holds a million member table
304 /// until it answers a smaller one. That is the trade and it is the right way
305 /// round: the same database had to build that table anyway, and the version
306 /// that threw it away afterwards built it again on the next call.
307 pub(crate) setops: crate::setops::Scratch,
308
309 /// What the last geo search found, kept for the same reason.
310 ///
311 /// A search cannot answer in the order it walks: the nine hash boxes come
312 /// out in hash order and the reply is in distance order, so every candidate
313 /// has to be in hand before the first one can be written. See
314 /// [`crate::geos::Scratch`], which is the two buffers and the argument for
315 /// keeping them here.
316 pub(crate) geo: crate::geos::Scratch,
317}
318
319/// How big [`Keyspace::scratch`] starts.
320///
321/// A kibibyte, which covers a value of any ordinary size and costs one
322/// allocation per database. The number is not tuned and does not need to be: too
323/// small only means the buffer grows once more on some later command, and too
324/// large only means a kibibyte nobody used.
325const SCRATCH: usize = 1024;
326
327/// How many segments one call to [`Keyspace::victim`] will draw from.
328///
329/// A round is a whole segment, and a segment is sixty four buckets of seven
330/// entries each before its overflow chains are counted, so one round almost
331/// always answers. The retries are for the case where a round came back with
332/// nothing usable, which happens when the segment it drew was empty or when a
333/// `volatile` policy filtered out everything in it.
334///
335/// Four rather than more, because a round that comes back with nothing is
336/// telling you something a fifth round will not change. That used to be the
337/// wrong shape for the `volatile` policies, which could draw four rounds of
338/// keys with no deadline on a database that had almost none, on a path where a
339/// client is waiting. Those policies now draw from the second index of just the
340/// keys that carry a deadline, the way Redis draws from `db->expires`, so every
341/// key a round looks at is a key it is allowed to take.
342const ROUNDS: usize = 4;
343
344/// Where the last collection key resolved to, if it still resolves there.
345///
346/// Y13 says a batch of `SADD` on one key should be one table growth check, and
347/// the same argument applies a step earlier: it should be one resolve. A
348/// resolve is a hash, a bucket walk and a record read, and on a hot key every
349/// command in the batch was paying for all three to be told the same answer the
350/// command in front of it got.
351///
352/// One entry and not a cache, because one entry is the shape of the problem.
353/// Single key `SADD` is the case with no spread to exploit, so the only reuse
354/// there is to find is the command immediately before, and a bigger structure
355/// would cost a lookup to avoid a lookup.
356///
357/// It holds a slot rather than reaching the body through an address, because a
358/// slot is an index into the slab for its type and stays right for as long as
359/// the key is there. That is also why nothing here memoizes a string: a string
360/// lives in the record itself and moves when the record does.
361///
362/// It does carry the record's address alongside, and that is safe for a narrower
363/// reason than the slot is. An address is only good until the next write, and
364/// this whole memo is thrown away by the next write, so inside the window where
365/// the memo answers at all the address is exactly as valid as the slot. What it
366/// buys is the eviction stamp: a memo hit skips the probe, and without an address
367/// the stamp would have to put the probe back and there would be no memo left.
368///
369/// What it is worth is measured rather than argued about, by the pair of rows
370/// `engine/sadd` and `engine/sadd-alternating` in `yo-resp`'s `engine` bench.
371/// The second one alternates between two keys, which defeats this on every
372/// command and leaves both keys as warm in the cache as the one key was, so the
373/// difference between the rows is close to this and nothing else. On an Apple M4
374/// it is about nineteen nanoseconds a command, which is 1.25x at pipeline 64.
375struct Memo {
376 /// What the map's write counter said when this was taken.
377 writes: u64,
378 /// Whether there is anything here. Separate from the length because the
379 /// empty key is a key, and `SADD "" m` is a command Redis accepts.
380 live: bool,
381 /// The type the key held, so a hit can still answer `WRONGTYPE`.
382 kind: Kind,
383 /// Where the body is in the slab for `kind`.
384 slot: u32,
385 /// Where the record is, for the stamp a hit still owes.
386 addr: Addr,
387 /// How much of `key` is the key.
388 len: u8,
389 key: [u8; Memo::MAX],
390}
391
392impl Memo {
393 /// The longest key worth remembering.
394 ///
395 /// Thirty two bytes is half a cache line and covers every hot key anyone
396 /// writes down, including the `myset:{tag}` the generators send. A longer
397 /// key is not memoized rather than heap allocated, because the whole point
398 /// of this is to not touch memory it does not have to.
399 const MAX: usize = 32;
400
401 const fn empty() -> Memo {
402 Memo {
403 writes: 0,
404 live: false,
405 kind: Kind::String,
406 slot: 0,
407 addr: Addr::NONE,
408 len: 0,
409 key: [0; Memo::MAX],
410 }
411 }
412
413 /// What `key` resolved to last time, if that answer still stands.
414 ///
415 /// `writes` is the map's counter now. Any write at all since this was taken
416 /// and the answer is thrown away, which is stricter than it has to be and is
417 /// the version that cannot be wrong.
418 #[inline]
419 fn get(&self, writes: u64, key: &[u8]) -> Option<(Kind, u32, Addr)> {
420 if !self.live || self.writes != writes || key.len() != self.len as usize {
421 return None;
422 }
423 // `bytes_eq` and not `==`, which is a call into the platform's `memcmp`
424 // for a key of a length the compiler cannot see. This is the one
425 // comparison the hot key path always does, and on a profile of `SADD`
426 // it was most of what the lookup cost.
427 bytes_eq(&self.key[..key.len()], key).then_some((self.kind, self.slot, self.addr))
428 }
429
430 /// Remember that `key` is at `slot`, in the record at `addr`.
431 #[inline]
432 fn put(&mut self, writes: u64, key: &[u8], kind: Kind, slot: u32, addr: Addr) {
433 if key.len() > Memo::MAX {
434 self.live = false;
435 return;
436 }
437 self.writes = writes;
438 self.live = true;
439 self.kind = kind;
440 self.slot = slot;
441 self.addr = addr;
442 self.len = key.len() as u8;
443 self.key[..key.len()].copy_from_slice(key);
444 }
445}
446
447/// How many databases this process has made.
448///
449/// Mixed into a new database's seed so that the eight shards a server starts in
450/// the same millisecond do not all draw the same members in the same order. It
451/// is the only atomic in this file and it is touched once per database rather
452/// than once per command, so it is not on any path Y1 cares about.
453static MADE: AtomicU64 = AtomicU64::new(0);
454
455impl Keyspace {
456 /// An empty database on the system clock.
457 #[must_use]
458 pub fn new() -> Keyspace {
459 Keyspace::with_clock(Clock::system())
460 }
461
462 /// An empty database on a clock of the caller's choosing.
463 #[must_use]
464 pub fn with_clock(clock: Clock) -> Keyspace {
465 let made = MADE.fetch_add(1, Ordering::Relaxed);
466 let seed = clock.now_ms() ^ made.wrapping_mul(0x9e37_79b9_7f4a_7c15);
467 Keyspace {
468 map: RawMap::new(),
469 clock,
470 expired: 0,
471 evicted: 0,
472 hits: 0,
473 misses: 0,
474 expired_fields: 0,
475 expired_fields_active: 0,
476 field_deadlines: Vec::new(),
477 field_at: 0,
478 sets: Slab::new(),
479 hashes: Slab::new(),
480 lists: Slab::new(),
481 zsets: Slab::new(),
482 arrays: Slab::new(),
483 streams: Slab::new(),
484 foreign: Slab::new(),
485 bodies: 0,
486 limits: set::Limits::DEFAULT,
487 hash_limits: hash::Limits::DEFAULT,
488 list_limits: list::Limits::default(),
489 zset_limits: zset::Limits::DEFAULT,
490 stream_limits: stream::Limits::default(),
491 policy: Policy::default(),
492 lfu: Lfu::DEFAULT,
493 samples: evict::SAMPLES,
494 pool: evict::Pool::new(),
495 rng: Rng::new(seed),
496 tier: None,
497 cold: Vec::new(),
498 cold_key: Vec::new(),
499 frozen: Vec::new(),
500 memo: Memo::empty(),
501 scratch: Vec::with_capacity(SCRATCH),
502 rows: Vec::new(),
503 setops: crate::setops::Scratch::new(),
504 geo: crate::geos::Scratch::default(),
505 }
506 }
507
508 /// Pin what `SPOP` and `SRANDMEMBER` draw.
509 ///
510 /// A database seeds itself from the clock and a counter, which is what a
511 /// server wants and what a test cannot assert against. Every test in this
512 /// crate that cares which member comes back calls this first, the same way
513 /// every expiry test drives a fixed clock, and for the same reason: the one
514 /// input that makes a result unrepeatable is better handed in than reached
515 /// for.
516 ///
517 /// It is public because reproducing a bug report is the same problem. A
518 /// seed printed in a crash report is worth having somewhere to put.
519 #[inline]
520 pub const fn seed(&mut self, seed: u64) {
521 self.rng = Rng::new(seed);
522 }
523
524 /// The next number from that same stream.
525 ///
526 /// For a command whose value lives in a [`Foreign`] body and so cannot
527 /// reach the draw any other way. `VRANDMEMBER` is the one, and it should be
528 /// as repeatable under [`Keyspace::seed`] as `SRANDMEMBER` is, which it
529 /// would not be if it carried a generator of its own.
530 #[inline]
531 pub const fn random(&mut self) -> u64 {
532 self.rng.next_u64()
533 }
534
535 /// What this database would evict, which is `CONFIG GET maxmemory-policy`.
536 #[inline]
537 #[must_use]
538 pub const fn policy(&self) -> Policy {
539 self.policy
540 }
541
542 /// Change what this database would evict.
543 ///
544 /// Every key already stored keeps whatever is in its access field, which is
545 /// why Redis warns on `OBJECT FREQ` that switching at runtime takes time to
546 /// adjust. Under the new policy those bits mean something else, and the only
547 /// honest thing to do about it is to let them be corrected by use. A key
548 /// nobody has touched since the switch reads as freshly used rather than as
549 /// stale, which is the safe direction: the other one evicts the working set
550 /// on the first pass after an operator changes a setting.
551 ///
552 /// The candidate pool does go, because a score only means anything against
553 /// another score under the same rule and every number in there was worked
554 /// out under the old one.
555 #[inline]
556 pub fn set_policy(&mut self, policy: Policy) {
557 if policy != self.policy {
558 self.pool.clear();
559 }
560 self.policy = policy;
561 }
562
563 /// The two numbers the LFU counter moves by, which are two `CONFIG` values.
564 #[inline]
565 #[must_use]
566 pub const fn lfu(&self) -> Lfu {
567 self.lfu
568 }
569
570 /// Change how fast the LFU counter climbs and decays.
571 #[inline]
572 pub const fn set_lfu(&mut self, lfu: Lfu) {
573 self.lfu = lfu;
574 }
575
576 /// Seconds since `key` was last used, which is `OBJECT IDLETIME`.
577 ///
578 /// `None` for a key that is not there. A key that has never been stamped
579 /// reads as zero rather than as ancient, which is what
580 /// [`Access::is_unset`] is for.
581 ///
582 /// This does not count as a use. Redis looks the key up with its no touch
583 /// flag here, and it has to: a diagnostic that resets the number it reports
584 /// would answer zero every time it was asked.
585 pub fn idle_secs(&mut self, key: &[u8]) -> Option<u64> {
586 let addr = self.live_rec_untouched(key)?;
587 let now = self.clock.now_ms();
588 Some(self.access_at(addr).idle_secs(now))
589 }
590
591 /// How often `key` is used, which is `OBJECT FREQ`.
592 ///
593 /// The eight bit counter, decayed to now, on the same terms as
594 /// [`Keyspace::idle_secs`]: `None` for a key that is not there, and asking
595 /// is not using.
596 ///
597 /// The caller is the one that has to check the policy first. This reports
598 /// what the bits say, and under a policy that is not LFU they say something
599 /// else, which is a refusal on the wire rather than a number.
600 pub fn freq(&mut self, key: &[u8]) -> Option<u8> {
601 let addr = self.live_rec_untouched(key)?;
602 let (now, lfu) = (self.clock.now_ms(), self.lfu);
603 Some(self.access_at(addr).freq(now, lfu))
604 }
605
606 /// Say that whatever is under `key` is about to be thrown away for a
607 /// `becoming`, if there is anything under it and anybody is listening.
608 ///
609 /// The distinction the two events draw is between a write that replaces the
610 /// whole value and a write that changes part of one. `SET`, `MSET`,
611 /// `GETSET` and every store form replace it. `APPEND`, `SETRANGE`, `INCR`,
612 /// `RPUSH`, `SADD` and the rest do not, however much of the value they end
613 /// up moving, and they say nothing here. That is the line Redis draws too,
614 /// and it draws it in the same place: at the call that swaps one value
615 /// object for another rather than at the ones that reach into the object
616 /// that is already there.
617 ///
618 /// Said in front of the write, so a listener that goes looking still finds
619 /// the old value under the name, and so that the news comes out ahead of
620 /// whatever the command itself says it did.
621 ///
622 /// The keys that arrive rather than change, which is `RENAME`, `COPY`,
623 /// `MOVE` and `RESTORE`, are not here. They take the whole record away
624 /// first, so there is nothing left for this to find, and the layer that
625 /// runs them says it afterwards instead. That is not a detail of this
626 /// implementation: a real server says it after the rename too, and before
627 /// the write everywhere else.
628 pub(crate) fn replacing(&mut self, key: &[u8], becoming: Kind) {
629 if !news::listening() {
630 return;
631 }
632 // Through the reaping lookup, because a key whose deadline has passed is
633 // a key that is not there, and a write landing on one is a key arriving
634 // rather than a value being replaced.
635 let Some(was) = self.kind_of(key) else {
636 return;
637 };
638 news::say(key, news::What::Overwritten);
639 if was != becoming {
640 news::say(key, news::What::TypeChanged);
641 }
642 }
643
644 /// Write a record under `key`, with the access field the policy wants on it.
645 ///
646 /// Every record this crate writes goes through here, which is the point of
647 /// it. A record is written fresh whenever a key is created and whenever a
648 /// string's value changes, and a fresh record starts with the blank field
649 /// [`value::write_record`] leaves behind. Blank reads as freshly used, which
650 /// is right at the moment of writing and wrong a minute later, so something
651 /// has to stamp it and this is the only place that knows the clock.
652 ///
653 /// Redis stamps at the same moment, in `createObject`, and for the same
654 /// reason.
655 ///
656 /// It is also where a key is noticed to be new. The map hands the old value
657 /// to the peek closure and only calls it when there was one, so a call that
658 /// went past without it is a name that was free a moment ago, which is the
659 /// one piece of news [`crate::news`] carries that a command provoked rather
660 /// than nobody. Said after the write and not before it, because a listener
661 /// that goes looking should find the key there.
662 pub(crate) fn write_rec(
663 &mut self,
664 key: &[u8],
665 len: usize,
666 fill: impl FnOnce(&mut [u8]),
667 ) -> Option<usize> {
668 let a = self.access_for_write(key);
669 let mut had = false;
670 // The one bit in the record that says the key has a deadline, handed
671 // straight back to the map. What the map does with it is keep a second
672 // index of just those records, so the expire cycle and the volatile
673 // eviction policies have somewhere to sample from that is not the whole
674 // keyspace. Nothing here counts anything: the map's own count of marked
675 // records is the number, and a number kept in two places is a number
676 // that eventually disagrees with itself.
677 let wrote = self.map.set_with(
678 key,
679 len,
680 |_| had = true,
681 |out| {
682 fill(out);
683 value::set_access(out, a);
684 value::has_expiry(out)
685 },
686 );
687 if !had {
688 news::say(key, news::What::Born);
689 }
690 wrote
691 }
692
693 /// Take `key` out of the map, keeping the deadline count right.
694 ///
695 /// The other half of [`Keyspace::write_rec`], and every path that removes a
696 /// record goes through one of the two. That is `DEL`, lazy expiry, eviction
697 /// and the source key of a `RENAME`, and the last one is why this is not
698 /// simply folded into [`Keyspace::drop_key`]: a rename hands the body to the
699 /// destination and must not free it, so it deletes the source record without
700 /// dropping the key, and it still has to be counted.
701 #[inline]
702 pub(crate) fn del_rec(&mut self, key: &[u8]) -> bool {
703 self.map.del(key)
704 }
705
706 /// What the access field of a record about to be written should say.
707 ///
708 /// Under the eight policies that read the field as a clock this is the time,
709 /// with no probe and no thought: writing a key is using it, and under the LRM
710 /// pair writing it is the only thing that counts as using it.
711 ///
712 /// Under LFU it is the counter that is already there, carried across the
713 /// rewrite unchanged. Unchanged rather than incremented, because the lookup
714 /// that resolved the key for this write already counted the access, and
715 /// counting it twice would rank a key that is written more highly than a key
716 /// that is read the same number of times. A key that is not there yet starts
717 /// at [`crate::access::LFU_INIT`], which is where Redis starts a new object.
718 ///
719 /// The probe is the reason this is written as two cases rather than one. It
720 /// is paid only under an LFU policy, so the default policy and every other
721 /// one write a record for exactly what it cost before.
722 fn access_for_write(&mut self, key: &[u8]) -> Access {
723 let now = self.clock.now_ms();
724 if !self.policy.is_lfu() {
725 return Access::lru(now);
726 }
727 match self.map.get(key).and_then(value::access) {
728 Some(a) if !a.is_unset() => a,
729 _ => Access::lfu(now),
730 }
731 }
732
733 /// The access field of the record at `addr`, or the unset one for a record
734 /// written before the field existed.
735 #[inline]
736 fn access_at(&self, addr: Addr) -> Access {
737 value::access(self.map.value_at(addr)).unwrap_or_default()
738 }
739
740 /// Write the access field back to the record at `addr`.
741 ///
742 /// The whole reason the field exists, and it runs on nearly every command,
743 /// so what it does is a load, an arithmetic step and a three byte store into
744 /// a cache line the caller has just read. It does not count as a write to the
745 /// map, because nothing moves and counting it would throw the [`Memo`] away
746 /// once per command. See [`RawMap::value_at_mut`].
747 ///
748 /// The LFU arm reads before it writes, because the counter it produces is a
749 /// function of the counter that is there. The clock arm does not, because the
750 /// time is the time whatever the record used to say.
751 #[inline]
752 fn stamp(&mut self, addr: Addr) {
753 let now = self.clock.now_ms();
754 if self.policy.is_lfu() {
755 let (lfu, current) = (self.lfu, self.access_at(addr));
756 let next = current.touched(now, lfu, &mut self.rng);
757 value::set_access(self.map.value_at_mut(addr), next);
758 } else {
759 value::set_access(self.map.value_at_mut(addr), Access::lru(now));
760 }
761 }
762
763 /// Where a set changes representation, which is three `CONFIG` values.
764 #[inline]
765 pub const fn limits(&self) -> &set::Limits {
766 &self.limits
767 }
768
769 /// Change where a set changes representation.
770 ///
771 /// Moving these does not rewrite the sets that already exist, which is what
772 /// Redis does too: `CONFIG SET set-max-listpack-entries 0` leaves every
773 /// listpack alone and only decides what the next `SADD` builds.
774 #[inline]
775 pub const fn set_limits(&mut self, limits: set::Limits) {
776 self.limits = limits;
777 }
778
779 /// Where a hash changes representation, which is two `CONFIG` values.
780 #[inline]
781 pub const fn hash_limits(&self) -> &hash::Limits {
782 &self.hash_limits
783 }
784
785 /// Change where a hash changes representation.
786 ///
787 /// Same rule as the set: moving these leaves every hash that already exists
788 /// exactly as it is, and only decides what the next `HSET` builds.
789 #[inline]
790 pub const fn set_hash_limits(&mut self, limits: hash::Limits) {
791 self.hash_limits = limits;
792 }
793
794 /// Where a list changes representation, which is one `CONFIG` value.
795 #[inline]
796 pub const fn list_limits(&self) -> &list::Limits {
797 &self.list_limits
798 }
799
800 /// Change where a list changes representation.
801 ///
802 /// Same rule again: this decides what the next `LPUSH` builds and leaves
803 /// every list that already exists alone. `list-max-listpack-size` is one
804 /// number rather than two, and [`list::Limits::of`] is what turns it into
805 /// the pair this holds.
806 #[inline]
807 pub const fn set_list_limits(&mut self, limits: list::Limits) {
808 self.list_limits = limits;
809 }
810
811 /// Where a stream starts a new node, which is two `CONFIG` values.
812 #[inline]
813 pub const fn stream_limits(&self) -> &stream::Limits {
814 &self.stream_limits
815 }
816
817 /// Change where a stream starts a new node.
818 ///
819 /// Same rule as the other four: this decides what the next `XADD` builds
820 /// and leaves every node that is already full exactly as it is, which is
821 /// also what Redis does, since a node is never resized after it is written.
822 #[inline]
823 pub const fn set_stream_limits(&mut self, limits: stream::Limits) {
824 self.stream_limits = limits;
825 }
826
827 /// Where a sorted set changes representation, which is two `CONFIG` values.
828 #[inline]
829 pub const fn zset_limits(&self) -> &zset::Limits {
830 &self.zset_limits
831 }
832
833 /// Change where a sorted set changes representation.
834 ///
835 /// Same rule as the other three: this decides what the next `ZADD` builds
836 /// and leaves every sorted set that already exists exactly as it is.
837 #[inline]
838 pub const fn set_zset_limits(&mut self, limits: zset::Limits) {
839 self.zset_limits = limits;
840 }
841
842 /// The clock expiry compares against.
843 #[inline]
844 pub const fn clock(&self) -> &Clock {
845 &self.clock
846 }
847
848 /// The map underneath, for statistics and for compaction.
849 #[inline]
850 pub const fn map(&self) -> &RawMap {
851 &self.map
852 }
853
854 /// How many keys are stored, including any that are dead and not yet
855 /// noticed. This is Redis's `DBSIZE`, which counts the same way.
856 #[inline]
857 pub fn len(&self) -> usize {
858 self.map.len()
859 }
860
861 /// Whether anything is stored.
862 #[inline]
863 pub fn is_empty(&self) -> bool {
864 self.map.is_empty()
865 }
866
867 /// What `key` holds, or `None` if there is nothing under it.
868 ///
869 /// This is `TYPE`. A key past its deadline is reaped first, so a dead key
870 /// answers `None` and not the type it used to be.
871 ///
872 /// One lookup, because the tag and the deadline are both in the record the
873 /// lookup returned. Reading the kind out before the reap rather than after
874 /// is what keeps it to one.
875 ///
876 /// It does not go through the lookup that stamps, and so it leaves the
877 /// eviction clock where it was, which is right and is worth saying rather than
878 /// leaving to be inferred from the shape of the code. `TYPE` is one of the
879 /// commands Redis looks up with its no touch flag, along with the `OBJECT`
880 /// subcommands underneath this, which read through `reap` for the same
881 /// reason.
882 pub fn kind_of(&mut self, key: &[u8]) -> Option<Kind> {
883 let now = self.clock.now_ms();
884 let found = self
885 .map
886 .get(key)
887 .map(|rec| (value::kind(rec), value::is_expired(rec, now)));
888 let Some((kind, dead)) = found else {
889 self.looked(false);
890 return None;
891 };
892 if dead {
893 self.reaped(key);
894 self.looked(false);
895 return None;
896 }
897 self.looked(true);
898 Some(kind)
899 }
900
901 /// How a set is represented, or `None` if `key` is not a set.
902 ///
903 /// This follows the slot and asks the body rather than reading the record,
904 /// because the record only holds a number. Putting a copy of the
905 /// representation in the record's two spare encoding bits would mean
906 /// rewriting the record every time a set was promoted, for the sake of a
907 /// command nobody calls in a loop, and would leave two places able to
908 /// disagree about the same fact.
909 /// A demoted set is brought back to answer, which is the one thing this
910 /// costs that the others do not. The word is a property of the body and the
911 /// body is on the device, so there is nothing else to read it off. It is one
912 /// device read for a command nobody sends in a loop, and the alternative is a
913 /// copy of the word in the record's two spare encoding bits with two places
914 /// then able to disagree about it.
915 pub fn set_encoding(&mut self, key: &[u8]) -> Option<set::Encoding> {
916 self.reap(key);
917 let rec = self.map.get(key)?;
918 if value::kind(rec) != Kind::Set {
919 return None;
920 }
921 let cold = value::Meta::from_byte(rec[0]).is_cold();
922 let at = if cold {
923 // The record is given up here, because promotion writes a new one.
924 self.promote_body(key).ok()??
925 } else {
926 value::slot(rec)
927 };
928 Some(self.sets.get(at)?.encoding())
929 }
930
931 /// How a hash is represented, or `None` if `key` is not a hash.
932 ///
933 /// The same shape as [`Keyspace::set_encoding`] and for the same reason: the
934 /// record holds a slot number and the body is the thing that knows which of
935 /// the two it currently is. A demoted hash is brought back to answer, which
936 /// is the same trade the set makes.
937 pub fn hash_encoding(&mut self, key: &[u8]) -> Option<hash::Encoding> {
938 self.reap(key);
939 let rec = self.map.get(key)?;
940 if value::kind(rec) != Kind::Hash {
941 return None;
942 }
943 let cold = value::Meta::from_byte(rec[0]).is_cold();
944 let at = if cold {
945 self.promote_body(key).ok()??
946 } else {
947 value::slot(rec)
948 };
949 Some(self.hashes.get(at)?.encoding())
950 }
951
952 /// How a list is represented, or `None` if `key` is not a list.
953 ///
954 /// The same shape as [`Keyspace::set_encoding`], and the same argument for
955 /// asking the body rather than reading a copy out of the record. A demoted
956 /// list is brought back to answer, as the set and the hash are.
957 pub fn list_encoding(&mut self, key: &[u8]) -> Option<list::Encoding> {
958 self.reap(key);
959 let rec = self.map.get(key)?;
960 if value::kind(rec) != Kind::List {
961 return None;
962 }
963 let cold = value::Meta::from_byte(rec[0]).is_cold();
964 let at = if cold {
965 self.promote_body(key).ok()??
966 } else {
967 value::slot(rec)
968 };
969 Some(self.lists.get(at)?.encoding())
970 }
971
972 /// How a sorted set is represented, or `None` if `key` is not one.
973 pub fn zset_encoding(&mut self, key: &[u8]) -> Option<zset::Encoding> {
974 self.reap(key);
975 let rec = self.map.get(key)?;
976 if value::kind(rec) != Kind::Zset {
977 return None;
978 }
979 let cold = value::Meta::from_byte(rec[0]).is_cold();
980 let at = if cold {
981 self.promote_body(key).ok()??
982 } else {
983 value::slot(rec)
984 };
985 Some(self.zsets.get(at)?.encoding())
986 }
987
988 /// Put a foreign body under `key`, over whatever was there.
989 ///
990 /// The keyspace takes the box and frees it when the key goes, which is the
991 /// whole reason a graph lives in here rather than in a table beside it. See
992 /// [`crate::foreign`] for why that mattered enough to spend the last tag
993 /// pattern on.
994 ///
995 /// Overwriting is allowed and is what a caller that has just decided to
996 /// replace a key wants. A caller that did not mean to overwrite asks
997 /// [`Keyspace::kind_of`] first, which is what the commands above do so they
998 /// can answer WRONGTYPE rather than quietly throw a hash away.
999 pub fn put_foreign(&mut self, key: &[u8], body: Box<dyn Foreign>) -> u32 {
1000 self.free_body(key);
1001 let at = self.foreign.insert(body);
1002 let len = value::slot_record_len(false);
1003 self.write_rec(key, len, |out| {
1004 value::write_slot_record(out, Kind::Foreign, at, None);
1005 });
1006 self.bodies += 1;
1007 at
1008 }
1009
1010 /// The foreign body under `key`.
1011 ///
1012 /// `None` for a key that is not there or has expired, an error for a key
1013 /// holding something this crate does understand, which is the same three
1014 /// way answer every other type's entry point gives.
1015 ///
1016 /// The caller turns the `&dyn Foreign` back into its own type with
1017 /// [`downcast_ref`], and a `None` from that is a key holding a different
1018 /// foreign body, which is also WRONGTYPE and is the caller's to report
1019 /// because only it knows which one it wanted.
1020 ///
1021 /// [`downcast_ref`]: crate::Foreign
1022 pub fn foreign(&mut self, key: &[u8]) -> Result<Option<&dyn Foreign>> {
1023 let Some(at) = self.live_slot(key, Kind::Foreign)? else {
1024 return Ok(None);
1025 };
1026 Ok(Some(
1027 self.foreign
1028 .get(at)
1029 .expect("the record points at its body")
1030 .as_ref(),
1031 ))
1032 }
1033
1034 /// The same, with a mutable borrow.
1035 pub fn foreign_mut(&mut self, key: &[u8]) -> Result<Option<&mut dyn Foreign>> {
1036 let Some(at) = self.live_slot(key, Kind::Foreign)? else {
1037 return Ok(None);
1038 };
1039 Ok(Some(
1040 self.foreign
1041 .get_mut(at)
1042 .expect("the record points at its body")
1043 .as_mut(),
1044 ))
1045 }
1046
1047 /// Drop `key` if the foreign body under it has gone empty.
1048 ///
1049 /// Redis deletes a key when its collection empties, and a client can see
1050 /// the difference, so every command that removes something calls this
1051 /// afterwards rather than each of them deciding what empty means.
1052 pub fn reap_foreign(&mut self, key: &[u8]) {
1053 let gone = matches!(self.foreign(key), Ok(Some(b)) if b.is_empty());
1054 if gone {
1055 self.drop_key(key);
1056 }
1057 }
1058
1059 /// The foreign body under `key`, without the reap or the type check.
1060 ///
1061 /// For the arms that already have a live record in hand and only want the
1062 /// body, where going through [`Keyspace::foreign`] would mean reaping a key
1063 /// that was read a line ago.
1064 fn foreign_at(&mut self, key: &[u8]) -> Option<&dyn Foreign> {
1065 let rec = self.map.get(key)?;
1066 let at = value::slot(rec);
1067 Some(self.foreign.get(at)?.as_ref())
1068 }
1069
1070 /// What `TYPE` should say about `key`.
1071 ///
1072 /// [`Kind::name`] for everything this crate knows, and the body's own word
1073 /// for a foreign one, because a client asking about a graph is told `graph`
1074 /// and not `foreign`. `None` for a key that is not there, which is the
1075 /// `none` Redis answers with.
1076 pub fn type_name(&mut self, key: &[u8]) -> Option<&'static str> {
1077 match self.kind_of(key)? {
1078 Kind::Foreign => self.foreign_at(key).map(Foreign::type_name),
1079 kind => Some(kind.name()),
1080 }
1081 }
1082
1083 /// `OBJECT ENCODING key`, as the word Redis puts on the wire.
1084 ///
1085 /// One place that knows every type's answer, so that adding the hash means
1086 /// adding an arm here and not finding the four callers that each worked it
1087 /// out for themselves.
1088 pub fn encoding_name(&mut self, key: &[u8]) -> Option<&'static str> {
1089 match self.kind_of(key)? {
1090 Kind::String => self.encoding(key).map(value::Encoding::name),
1091 Kind::Set => self.set_encoding(key).map(set::Encoding::name),
1092 Kind::Hash => self.hash_encoding(key).map(hash::Encoding::name),
1093 Kind::List => self.list_encoding(key).map(list::Encoding::name),
1094 Kind::Zset => self.zset_encoding(key).map(zset::Encoding::name),
1095 // The one type with one encoding, so there is nothing to ask.
1096 Kind::Array => Some("sliced-array"),
1097 // The same, and Redis's word for it rather than a description of
1098 // the node layout.
1099 Kind::Stream => Some("stream"),
1100 // The body knows and this does not, which is the whole point of it.
1101 Kind::Foreign => self.foreign_at(key).map(Foreign::encoding),
1102 }
1103 }
1104
1105 /// Put a deadline on `key`, or take one off. Answers whether it was there.
1106 ///
1107 /// Any type. A deadline lives in the record and changes its length, so this
1108 /// writes the record again rather than patching it, and for a set that is
1109 /// five bytes or thirteen and never the members. The body is left exactly
1110 /// where it is, which is why this writes through the map instead of taking
1111 /// the free the body path an overwrite takes.
1112 ///
1113 /// This is the raw write. [`Keyspace::expire`] and [`Keyspace::persist`] are
1114 /// what `EXPIRE` and its family call, and they come through here once they
1115 /// have worked out whether the deadline is allowed to move.
1116 pub fn set_expiry(&mut self, key: &[u8], at: Option<u64>) -> bool {
1117 self.reap(key);
1118 let Some(rec) = self.map.get(key) else {
1119 return false;
1120 };
1121 if value::expire_at(rec) == at {
1122 return true;
1123 }
1124 // A deadline does not move a value that is on the file, it changes the
1125 // eight bytes in front of the address. So this writes a new pointer
1126 // rather than reading the value back to write it out again, and `EXPIRE`
1127 // on a demoted key costs no device read at all. That is the same promise
1128 // `TTL` and `STRLEN` keep and it is why the header stayed in memory.
1129 //
1130 // Ahead of the type split rather than inside the string arm, because a
1131 // demoted set is a pointer in exactly the same way and writing a slot
1132 // record over one would leave a record pointing at a slab slot that
1133 // belongs to something else.
1134 if let Some(c) = value::cold(rec) {
1135 let m = value::Meta::from_byte(rec[0]);
1136 let (kind, enc) = (m.kind(), m.encoding());
1137 let was = value::access(rec).unwrap_or_default();
1138 self.map.set_with(
1139 key,
1140 value::cold_record_len(at.is_some()),
1141 |_| {},
1142 |out| {
1143 value::write_cold_record(out, kind, enc, c.at, c.len, at);
1144 value::set_access(out, was);
1145 value::has_expiry(out)
1146 },
1147 );
1148 return true;
1149 }
1150 // Read what has to survive out of the record before writing over it.
1151 match value::kind(rec) {
1152 Kind::String => {
1153 // Through the scratch buffer rather than a fresh `Vec`, since
1154 // `EXPIRE` on a string is a command a cache sends as often as
1155 // the `SET` before it.
1156 let mut bytes = std::mem::take(&mut self.scratch);
1157 bytes.clear();
1158 value::read(rec).write_to(&mut bytes);
1159 self.store(key, &bytes, at);
1160 self.scratch = bytes;
1161 }
1162 // Every body type writes the same record: a tag and a slot number.
1163 // The body is not touched and does not need to be, which is the
1164 // whole point of keeping it out of the record.
1165 kind @ (Kind::Set
1166 | Kind::Hash
1167 | Kind::List
1168 | Kind::Zset
1169 | Kind::Array
1170 | Kind::Stream
1171 | Kind::Foreign) => {
1172 let slot = value::slot(rec);
1173 let len = value::slot_record_len(at.is_some());
1174 self.write_rec(key, len, |out| {
1175 value::write_slot_record(out, kind, slot, at);
1176 });
1177 }
1178 }
1179 true
1180 }
1181
1182 /// The key's deadline, as the three way answer `TTL` and `PTTL` are built on.
1183 ///
1184 /// [`Ask::Missing`] for a key that is not there, [`Ask::NoDeadline`] for one
1185 /// that is and has no deadline, and the absolute millisecond otherwise. A key
1186 /// past its deadline is reaped on the way through, so it answers `Missing`
1187 /// and not the moment that has gone.
1188 ///
1189 /// Asking when a key dies is not using it, so this does not stamp the
1190 /// eviction clock. Redis reads the key with its no touch flag here for the
1191 /// same reason, and it matters more than it looks: a client polling `TTL` on
1192 /// a key would otherwise keep that key at the top of the working set for as
1193 /// long as it kept asking whether it was about to go.
1194 pub fn deadline_of(&mut self, key: &[u8]) -> Ask {
1195 let Some(addr) = self.live_rec_untouched(key) else {
1196 return Ask::Missing;
1197 };
1198 match value::expire_at(self.map.value_at(addr)) {
1199 Some(at) => Ask::At(at),
1200 None => Ask::NoDeadline,
1201 }
1202 }
1203
1204 /// Move `key`'s deadline to `at`, if `cond` lets it.
1205 ///
1206 /// This is `EXPIRE`, `PEXPIRE`, `EXPIREAT` and `PEXPIREAT`, which differ only
1207 /// in the unit and the origin of the number. All four turn it into one
1208 /// absolute millisecond before they get here, so the condition rules live in
1209 /// one place and the four commands cannot drift apart.
1210 ///
1211 /// A deadline that has already passed deletes the key rather than being
1212 /// stored, and the answer says so. `EXPIRE` cannot report the difference
1213 /// because it replies 1 either way, but the caller is not always `EXPIRE`,
1214 /// and a delete is a different thing from a deadline.
1215 ///
1216 /// The condition is checked before the past check, which is the order Redis
1217 /// uses and is the one that matters: `EXPIRE key 0 XX` on a key with no
1218 /// deadline answers 0 and leaves the key alone, rather than deleting it.
1219 pub fn expire(&mut self, key: &[u8], at: u64, cond: Cond) -> Applied {
1220 let prev = match self.deadline_of(key) {
1221 Ask::Missing => return Applied::Missing,
1222 Ask::NoDeadline => None,
1223 Ask::At(at) => Some(at),
1224 };
1225 let done = ttl::decide(prev, at, cond, self.clock.now_ms());
1226 match done {
1227 Applied::Ok => {
1228 self.set_expiry(key, Some(at));
1229 }
1230 // The structure that answered `Deleted` for a field only holds
1231 // deadlines, so its caller has to remove the field. Here the caller
1232 // is us and the key is ours, so it goes now.
1233 Applied::Deleted => {
1234 self.drop_key(key);
1235 }
1236 Applied::Missing | Applied::NotMet => {}
1237 }
1238 done
1239 }
1240
1241 /// Take `key`'s deadline off. Answers whether there was one to take.
1242 ///
1243 /// This is `PERSIST`, and the reply is the same 0 for a key that is not there
1244 /// and a key that was never going to expire, which is Redis's answer and not
1245 /// a shortcut here.
1246 pub fn persist(&mut self, key: &[u8]) -> bool {
1247 if !matches!(self.deadline_of(key), Ask::At(_)) {
1248 return false;
1249 }
1250 self.set_expiry(key, None);
1251 true
1252 }
1253
1254 /// Give back whatever `key` holds outside its record, if it holds anything.
1255 ///
1256 /// Every path that deletes a key or writes over one has to come through
1257 /// here, because a set that loses its record without losing its slab slot is
1258 /// a leak that nothing ever notices: the memory is reachable, the slot is
1259 /// never reused, and `DBSIZE` looks right. Six delete sites and four string
1260 /// writers each remembering to do it themselves is five chances to forget,
1261 /// and one of them would be forgotten. So this is the funnel, and when the
1262 /// hash type lands the only place that changes is the match below.
1263 ///
1264 /// The record is left alone. This frees the body and the caller either
1265 /// deletes the record or writes a new one over it.
1266 pub(crate) fn free_body(&mut self, key: &[u8]) {
1267 if self.bodies == 0 {
1268 return;
1269 }
1270 let Some(rec) = self.map.get(key) else {
1271 return;
1272 };
1273 // A body that has been moved to the file has no slab slot to give back,
1274 // and the four bytes where the slot number would be are the front of an
1275 // address. Reading them as a slot and freeing it would hand back a slot
1276 // belonging to a different key. The chunks on the file are left where
1277 // they are, which is what the log's compaction collects. See
1278 // [`crate::tier`].
1279 if value::Meta::from_byte(rec[0]).is_cold() {
1280 return;
1281 }
1282 match value::kind(rec) {
1283 Kind::String => {}
1284 Kind::Set => {
1285 let at = value::slot(rec);
1286 self.sets.remove(at);
1287 self.bodies -= 1;
1288 }
1289 Kind::Hash => {
1290 let at = value::slot(rec);
1291 self.hashes.remove(at);
1292 self.bodies -= 1;
1293 }
1294 Kind::List => {
1295 let at = value::slot(rec);
1296 self.lists.remove(at);
1297 self.bodies -= 1;
1298 }
1299 Kind::Zset => {
1300 let at = value::slot(rec);
1301 self.zsets.remove(at);
1302 self.bodies -= 1;
1303 }
1304 Kind::Array => {
1305 let at = value::slot(rec);
1306 self.arrays.remove(at);
1307 self.bodies -= 1;
1308 }
1309 Kind::Stream => {
1310 let at = value::slot(rec);
1311 self.streams.remove(at);
1312 self.bodies -= 1;
1313 }
1314 Kind::Foreign => {
1315 let at = value::slot(rec);
1316 self.foreign.remove(at);
1317 self.bodies -= 1;
1318 }
1319 }
1320 }
1321
1322 /// Delete `key` and whatever it held. Answers whether it was there.
1323 #[inline]
1324 pub(crate) fn drop_key(&mut self, key: &[u8]) -> bool {
1325 self.free_body(key);
1326 self.del_rec(key)
1327 }
1328
1329 /// Drop `key` if its deadline has passed.
1330 ///
1331 /// This is lazy expiry and it is half of the story. The other half is
1332 /// [`Keyspace::expire_cycle`], which the maintenance slice runs and which is
1333 /// what stops a key nobody ever reads again from holding its memory forever
1334 /// (`14` section 1).
1335 ///
1336 /// Every public read calls this first, whatever type it is reading, which
1337 /// is why it is here and not in the file for any one type.
1338 #[inline]
1339 pub(crate) fn reap(&mut self, key: &[u8]) {
1340 let now = self.clock.now_ms();
1341 let found = self.map.get(key).map(|r| value::is_expired(r, now));
1342 if found == Some(true) {
1343 self.reaped(key);
1344 }
1345 // The other half of what [`Keyspace::looked`] is for. The groups whose
1346 // values are read straight out of the record come through here and not
1347 // through the funnels below, because a string is its own record and
1348 // there is no slot to resolve, so this is where `BITCOUNT` and `PFCOUNT`
1349 // and `GETDEL` are counted.
1350 self.looked(found == Some(false));
1351 }
1352
1353 /// Take a key whose deadline has passed, count it, and say that it went.
1354 ///
1355 /// Every lazy reap in the crate ends here, and so does the cycle, because
1356 /// all of them owe the same three things and `INFO stats` reports one number
1357 /// over the lot. Answers whether there was anything there to take, which the
1358 /// cycle reads and the readers do not, since a reader has already looked.
1359 ///
1360 /// The caller decides whether the deadline has passed. This does not check
1361 /// again, because every caller has just read the record to find out and a
1362 /// second look would be a second probe for an answer already in hand.
1363 #[inline]
1364 pub(crate) fn reaped(&mut self, key: &[u8]) -> bool {
1365 let gone = self.drop_key(key);
1366 if gone {
1367 self.expired += 1;
1368 news::say(key, news::What::Expired);
1369 }
1370 gone
1371 }
1372
1373 /// Give this database somewhere to keep values that are not in memory.
1374 ///
1375 /// Until this is called nothing is ever demoted, no record is ever cold and
1376 /// every command takes exactly the path it took before, which is why a
1377 /// database that never opens a file pays nothing for this existing.
1378 ///
1379 /// The store is whatever the caller wants it to be. In a server it is the
1380 /// shard's log. In a test it is a vector. This crate does not depend on
1381 /// either and does not want to: [`Blocks`](crate::cold::Blocks) is an append
1382 /// that hands back an address and a read that takes one, and that is the whole
1383 /// of the contract between the memory engine and whatever is under it.
1384 pub fn attach(&mut self, blocks: Store) {
1385 self.tier = Some(Tier::new(blocks));
1386 }
1387
1388 /// The tier, if one was attached, for its counters.
1389 #[must_use]
1390 pub const fn tier(&self) -> Option<&Tier<Store>> {
1391 self.tier.as_ref()
1392 }
1393
1394 /// The tier, mutably, for a caller driving a sweep.
1395 pub const fn tier_mut(&mut self) -> Option<&mut Tier<Store>> {
1396 self.tier.as_mut()
1397 }
1398
1399 /// Move one key's value out to the file.
1400 ///
1401 /// Answers whether it went. A key that is not there, one that is int
1402 /// encoded, one holding a type that does not move yet and one whose value is
1403 /// shorter than the pointer that would replace it all answer false, and so
1404 /// does every key on a database with nothing attached.
1405 ///
1406 /// This is the single key form, which is what a test and a `DEBUG`
1407 /// subcommand want. What a server under memory pressure wants is
1408 /// [`Keyspace::relieve`].
1409 ///
1410 /// One entry point for both kinds of value, because a caller naming a key
1411 /// should not have to know whether its body is in the record or in a slab.
1412 /// The two paths underneath are different all the way down: a string goes
1413 /// through the tier, and a collection goes through `demote_body` beside
1414 /// this, which frees a slab slot and grows a record.
1415 ///
1416 /// # Errors
1417 ///
1418 /// Whatever the store says when it will not take the bytes.
1419 pub fn demote(&mut self, key: &[u8]) -> Result<bool> {
1420 if self.tier.is_none() {
1421 return Ok(false);
1422 }
1423 let Some(addr) = self.map.find(key) else {
1424 return Ok(false);
1425 };
1426 if value::kind(self.map.value_at(addr)).is_body() {
1427 return self.demote_body(key);
1428 }
1429 let tier = self.tier.as_mut().expect("checked at the top");
1430 tier.demote(&mut self.map, key)
1431 }
1432
1433 /// How many bytes the attached store is holding, or `None` if there is not
1434 /// one.
1435 ///
1436 /// `None` and `Some(0)` are different answers and the difference is the one
1437 /// `maxstore` turns on. A database with nothing attached cannot migrate and
1438 /// has to evict, and a database with an empty file attached can migrate the
1439 /// moment it needs to.
1440 #[must_use]
1441 pub fn store_bytes(&self) -> Option<u64> {
1442 self.tier.as_ref().map(Tier::store_bytes)
1443 }
1444
1445 /// Move values out to the file until at least `shed` bytes of memory have
1446 /// gone.
1447 ///
1448 /// Answers with a [`Relief`], which is how many keys went and how much
1449 /// memory that gave back. This is `maxmemory` under the inversion `14`
1450 /// describes: the limit that used to throw keys away now moves them, and
1451 /// what a client stored is still there afterwards.
1452 ///
1453 /// Bytes to shed rather than a target to reach, because the caller with the
1454 /// limit is a server holding sixteen databases against one number and what
1455 /// it knows is how far over it is, not what any one database should be
1456 /// holding. `usize::MAX` means everything that can go, which is what a sweep
1457 /// wants.
1458 ///
1459 /// Victims are chosen by the same policy `maxmemory-policy` names, so a
1460 /// database set to `allkeys-lru` demotes the coldest keys and one set to
1461 /// `volatile-ttl` demotes the ones closest to expiring. See
1462 /// [`Tier::relieve`] for what a sweep does and where it stops.
1463 ///
1464 /// # Why `noeviction` still moves values
1465 ///
1466 /// Because it says do not lose data, and moving a value to the file does not
1467 /// lose any. The policy is two things at once in Redis, whether to give
1468 /// memory back at all and which keys to take it from, and only the second of
1469 /// those means anything here. So the default policy picks victims the way
1470 /// `allkeys-lru` does and the promise it was set for is kept: every key a
1471 /// client stored is still readable afterwards.
1472 ///
1473 /// The alternative is a server that was given a file, was given a limit, and
1474 /// answers writes with OOM until somebody finds the third setting that turns
1475 /// the file on. That is a trap and not a default.
1476 ///
1477 /// # Two passes, because the two kinds of value are counted in different
1478 /// places
1479 ///
1480 /// Strings go first, through [`Tier::relieve`], which measures itself against
1481 /// the arena because a string is its record and moving one makes the arena
1482 /// smaller. Collections cannot be swept that way. A collection's body is in a
1483 /// slab the arena knows nothing about, and moving one makes the arena
1484 /// **bigger**, because a twenty byte pointer replaces an eight byte slot
1485 /// number. A loop that watched the arena would demote every collection in the
1486 /// database, watch its number go up the whole time, and never stop.
1487 ///
1488 /// So the second pass is here rather than in the tier, and it measures itself
1489 /// against [`Keyspace::memory_bytes`], which is the arena and the slabs
1490 /// together. That is the only number that goes down when a body moves, and it
1491 /// is the number the server's limit is compared against anyway.
1492 ///
1493 /// # Errors
1494 ///
1495 /// Whatever the store says when it will not take the bytes.
1496 pub fn relieve(&mut self, shed: usize) -> Result<Relief> {
1497 if self.tier.is_none() {
1498 return Ok(Relief::default());
1499 }
1500 let now = self.clock.now_ms();
1501 let policy = match self.policy {
1502 Policy::NoEviction => Policy::AllKeysLru,
1503 chosen => chosen,
1504 };
1505 let lfu = self.lfu;
1506 let start = self.memory_bytes();
1507 let target = start.saturating_sub(shed);
1508 let budget = self.map.memory_bytes().saturating_sub(shed);
1509 let tier = self.tier.as_mut().expect("checked just above");
1510 let mut relief = tier.relieve(&mut self.map, budget, policy, now, lfu)?;
1511 // A sweep compacts the arena, which moves records, and the memo holds a
1512 // record's address. It is normally thrown away by the map's write
1513 // counter moving, and a sweep that demoted nothing and compacted anyway
1514 // is the one case where that counter does not move.
1515 self.memo = Memo::empty();
1516 if self.bodies > 0 && self.memory_bytes() > target {
1517 relief.moved += self.shed_bodies(target, policy, now, lfu)?;
1518 }
1519 relief.freed = start.saturating_sub(self.memory_bytes());
1520 Ok(relief)
1521 }
1522
1523 /// Sample the keyspace for collection bodies and move them out until
1524 /// [`Keyspace::memory_bytes`] is under `target`.
1525 ///
1526 /// The same shape as [`Tier::relieve`]'s loop, with the same stop rule and
1527 /// for the same reason: a round that finds nothing is a collision rather than
1528 /// a conclusion, so it takes [`tier::BARREN`] of them in a row to give up.
1529 /// What is different is the number being watched, which is explained on
1530 /// [`Keyspace::relieve`], and that there is no compaction step in here. The
1531 /// arena grows on this path rather than shrinking, so there is nothing for a
1532 /// compaction to hand back that the string pass has not already taken.
1533 fn shed_bodies(
1534 &mut self,
1535 target: usize,
1536 policy: Policy,
1537 now_ms: u64,
1538 lfu: Lfu,
1539 ) -> Result<usize> {
1540 let mut moved = 0;
1541 let mut barren = 0;
1542 // The pool comes out of the keyspace for the length of the sweep, because
1543 // it hands back a borrow of itself and demoting one victim needs the
1544 // whole keyspace. `kb` is the same borrow copied somewhere it can live
1545 // across that call, and it is one allocation for the sweep rather than
1546 // one per victim.
1547 let mut pool = core::mem::take(&mut self.pool);
1548 let mut kb: Vec<u8> = Vec::new();
1549 while self.memory_bytes() > target {
1550 pool.clear();
1551 let r = self.rng.next_u64();
1552 let pool = &mut pool;
1553 let mut seen = 0usize;
1554 let mut found = 0usize;
1555 // The body check is on the record and not on the slab, so the sample
1556 // closure does not need a second borrow of the keyspace. A record
1557 // that turns out to hold a body too small to be worth moving is
1558 // refused by `demote_body`, which has the slab in hand by then.
1559 self.map.sample(r, |k, v, _| {
1560 seen += 1;
1561 let m = value::Meta::from_byte(v[0]);
1562 if !m.is_cold() && moves(m.kind()) {
1563 pool.offer(k, evict::score(v, policy, now_ms, lfu));
1564 found += 1;
1565 }
1566 found < evict::CANDIDATES && seen < tier::WALK
1567 });
1568
1569 let mut round = 0;
1570 while let Some(k) = pool.take() {
1571 kb.clear();
1572 kb.extend_from_slice(k);
1573 if self.demote_body(&kb)? {
1574 round += 1;
1575 }
1576 }
1577 if round == 0 {
1578 barren += 1;
1579 if barren == tier::BARREN {
1580 break;
1581 }
1582 continue;
1583 }
1584 barren = 0;
1585 moved += round;
1586 }
1587 self.pool = pool;
1588 // Every demotion rewrote a record, so nothing the memo holds is worth
1589 // keeping and one of the slot numbers in it is now a freed slab slot.
1590 self.memo = Memo::empty();
1591 Ok(moved)
1592 }
1593
1594 /// Move `key`'s collection body out to the file.
1595 ///
1596 /// `Ok(false)` for a key that is not there, that holds a type this does not
1597 /// move yet, that is already on the file, or whose body is smaller than the
1598 /// pointer that would replace it. None of those is an error, for the same
1599 /// reason none of them is in [`crate::tier::demote`](Tier::demote): a sweep
1600 /// asks about a lot of keys and most of the answers are no.
1601 ///
1602 /// # What is different from a string
1603 ///
1604 /// A string's value is its record, so the tier can read it, write it out and
1605 /// rewrite the record without anyone else being involved. A collection's body
1606 /// is in a slab and its record holds a four byte number, so three things have
1607 /// to happen here and only here: the body is turned into bytes that mean
1608 /// something on a device, the slab slot is freed, and the record grows from
1609 /// eight bytes to twenty because an address and a length are longer than a
1610 /// slot number.
1611 ///
1612 /// That last part is why the memory a sweep frees does not show up in the
1613 /// arena. Demoting a collection makes the arena bigger and the slab smaller,
1614 /// and it is the sum that goes down. See [`Keyspace::relieve`].
1615 ///
1616 /// # Errors
1617 ///
1618 /// Whatever the store says when it cannot take the bytes.
1619 pub(crate) fn demote_body(&mut self, key: &[u8]) -> Result<bool> {
1620 if self.tier.is_none() {
1621 return Ok(false);
1622 }
1623 let Some(addr) = self.map.find(key) else {
1624 return Ok(false);
1625 };
1626 let rec = self.map.value_at(addr);
1627 let m = value::Meta::from_byte(rec[0]);
1628 if m.is_cold() || !moves(m.kind()) {
1629 return Ok(false);
1630 }
1631 let kind = m.kind();
1632 let expire_at = value::expire_at(rec);
1633 // Carried across and not restamped, as in `Tier::demote`. A key that was
1634 // moved out was not used, and a demotion that looked like a use would
1635 // make the next sweep pick the wrong victim.
1636 let was = value::access(rec).unwrap_or_default();
1637 let slot = value::slot(rec);
1638 // The same arithmetic the string side uses and not a tunable: a body
1639 // that costs less to keep than the pointer to it would is left where it
1640 // is. It is even more favourable here, because what a table costs in
1641 // memory is well above what its members weigh on a device.
1642 let grows_by = value::cold_record_len(expire_at.is_some())
1643 - value::slot_record_len(expire_at.is_some());
1644
1645 let mut buf = core::mem::take(&mut self.frozen);
1646 buf.clear();
1647 let worth = self.freeze_body(kind, slot, grows_by, &mut buf);
1648 let stashed = if worth {
1649 let tier = self.tier.as_mut().expect("checked at the top");
1650 Some(tier.stash(&buf))
1651 } else {
1652 None
1653 };
1654 self.frozen = buf;
1655 let Some(wrote) = stashed else {
1656 return Ok(false);
1657 };
1658 let chain = wrote?;
1659
1660 self.free_slot(kind, slot);
1661 self.bodies -= 1;
1662 let len = chain.len as u32;
1663 let wrote = self.map.set_with(
1664 key,
1665 value::cold_record_len(expire_at.is_some()),
1666 |_| {},
1667 |out| {
1668 // The encoding bits mean nothing on a collection record, which
1669 // is what `Meta::slot` already writes and says. `OBJECT
1670 // ENCODING` on a demoted body brings it back and asks the body,
1671 // which is one device read for a command nobody sends in a
1672 // loop, and is one place holding the fact rather than two.
1673 value::write_cold_record(out, kind, value::Encoding::Int, chain.at, len, expire_at);
1674 value::set_access(out, was);
1675 value::has_expiry(out)
1676 },
1677 );
1678 debug_assert!(wrote.is_some(), "the key was found a moment ago");
1679 // Not cleared by hand. Writing the record moves the map's write counter
1680 // and that is what the memo checks, so the slot number it is holding for
1681 // this key is already unreachable.
1682 Ok(true)
1683 }
1684
1685 /// Turn the body in `slot` into bytes, or say it is not worth moving.
1686 ///
1687 /// The one place that knows which slab a kind indexes on the way out, and
1688 /// the size check is in here because it needs the body in hand and the body
1689 /// is the thing this is holding. `false` leaves `buf` in whatever state it
1690 /// was, which the caller does not read.
1691 fn freeze_body(&self, kind: Kind, slot: u32, grows_by: usize, buf: &mut Vec<u8>) -> bool {
1692 match kind {
1693 Kind::Set => match self.sets.get(slot) {
1694 Some(body) if body.memory_bytes() > grows_by => {
1695 body.freeze(buf);
1696 true
1697 }
1698 Some(_) => false,
1699 None => {
1700 debug_assert!(false, "a set record with no set behind it");
1701 false
1702 }
1703 },
1704 Kind::Hash => match self.hashes.get(slot) {
1705 Some(body) if body.memory_bytes() > grows_by => {
1706 body.freeze(buf);
1707 true
1708 }
1709 Some(_) => false,
1710 None => {
1711 debug_assert!(false, "a hash record with no hash behind it");
1712 false
1713 }
1714 },
1715 Kind::List => match self.lists.get(slot) {
1716 Some(body) if body.memory_bytes() > grows_by => {
1717 body.freeze(buf);
1718 true
1719 }
1720 Some(_) => false,
1721 None => {
1722 debug_assert!(false, "a list record with no list behind it");
1723 false
1724 }
1725 },
1726 Kind::Zset => match self.zsets.get(slot) {
1727 Some(body) if body.memory_bytes() > grows_by => {
1728 body.freeze(buf);
1729 true
1730 }
1731 Some(_) => false,
1732 None => {
1733 debug_assert!(false, "a sorted set record with no sorted set behind it");
1734 false
1735 }
1736 },
1737 Kind::Array => match self.arrays.get(slot) {
1738 Some(body) if body.memory_bytes() > grows_by => {
1739 body.freeze(buf);
1740 true
1741 }
1742 Some(_) => false,
1743 None => {
1744 debug_assert!(false, "an array record with no array behind it");
1745 false
1746 }
1747 },
1748 Kind::Stream => match self.streams.get(slot) {
1749 Some(body) if body.memory_bytes() > grows_by => {
1750 body.freeze(buf);
1751 true
1752 }
1753 Some(_) => false,
1754 None => {
1755 debug_assert!(false, "a stream record with no stream behind it");
1756 false
1757 }
1758 },
1759 _ => {
1760 debug_assert!(false, "a kind `moves` said yes to and this does not know");
1761 false
1762 }
1763 }
1764 }
1765
1766 /// Give a slab slot back once its body has been written out.
1767 ///
1768 /// The twin of [`Keyspace::freeze_body`], kept beside it so that adding a
1769 /// type means touching two arms next to each other rather than hunting for
1770 /// the second one.
1771 fn free_slot(&mut self, kind: Kind, slot: u32) {
1772 match kind {
1773 Kind::Set => {
1774 self.sets.remove(slot);
1775 }
1776 Kind::Hash => {
1777 self.hashes.remove(slot);
1778 }
1779 Kind::List => {
1780 self.lists.remove(slot);
1781 }
1782 Kind::Zset => {
1783 self.zsets.remove(slot);
1784 }
1785 Kind::Array => {
1786 self.arrays.remove(slot);
1787 }
1788 Kind::Stream => {
1789 self.streams.remove(slot);
1790 }
1791 _ => debug_assert!(false, "freeing a slot in a slab that was never found"),
1792 }
1793 }
1794
1795 /// Send a cold record that holds a body back to [`Keyspace::promote_body`],
1796 /// and say whether that is what happened.
1797 ///
1798 /// The guard in front of both fault entry points, and it is here rather than
1799 /// in the tier because the tier cannot do this job. [`Tier::fault`] puts a
1800 /// value back by writing a string record, so handing it a demoted set would
1801 /// turn the set into a string holding the bytes a set freezes to. The kind is
1802 /// in the record and only this side has a slab to put a body in.
1803 ///
1804 /// One probe of the map on a path that is about to read a device, on a
1805 /// database that has a file at all. Every other database is refused by the
1806 /// caller before it gets here.
1807 ///
1808 /// # Errors
1809 ///
1810 /// Whatever the store says when the chain will not read back.
1811 fn body_came_back(&mut self, key: &[u8]) -> Result<bool> {
1812 let Some(addr) = self.map.find(key) else {
1813 return Ok(false);
1814 };
1815 let m = value::Meta::from_byte(self.map.value_at(addr)[0]);
1816 if !m.is_cold() || !m.kind().is_body() {
1817 return Ok(false);
1818 }
1819 self.promote_body(key)?;
1820 Ok(true)
1821 }
1822
1823 /// Bring `key`'s collection body back into a slab and answer its new slot.
1824 ///
1825 /// The record has to be rewritten either way, because a resident collection
1826 /// is a slot number and there is no way to hold one without being in the
1827 /// slab, so the doorkeeper does not get a vote. See [`Tier::fetch`].
1828 ///
1829 /// # Errors
1830 ///
1831 /// Whatever the store says when the chain will not read back, and
1832 /// [`Code::Corrupt`] when it reads back as something that is not a body.
1833 fn promote_body(&mut self, key: &[u8]) -> Result<Option<u32>> {
1834 let Some(addr) = self.map.find(key) else {
1835 return Ok(None);
1836 };
1837 let rec = self.map.value_at(addr);
1838 let Some(c) = value::cold(rec) else {
1839 return Ok(None);
1840 };
1841 let kind = value::kind(rec);
1842 let expire_at = value::expire_at(rec);
1843 let was = value::access(rec).unwrap_or_default();
1844 let Some(tier) = self.tier.as_mut() else {
1845 debug_assert!(false, "a cold record on a database with no file");
1846 return Ok(None);
1847 };
1848
1849 let mut buf = core::mem::take(&mut self.frozen);
1850 let read = tier.fetch(
1851 crate::cold::Chain {
1852 at: c.at,
1853 len: u64::from(c.len),
1854 },
1855 &mut buf,
1856 );
1857 self.frozen = buf;
1858 read?;
1859
1860 let slot = match kind {
1861 Kind::Set => {
1862 let set = Set::thaw(&self.frozen).map_err(|e| {
1863 Error::new(Code::Corrupt, "a demoted set did not read back")
1864 .with_detail(e.to_string())
1865 })?;
1866 self.sets.insert(set)
1867 }
1868 Kind::Hash => {
1869 let hash = Hash::thaw(&self.frozen).map_err(|e| {
1870 Error::new(Code::Corrupt, "a demoted hash did not read back")
1871 .with_detail(e.to_string())
1872 })?;
1873 self.hashes.insert(hash)
1874 }
1875 Kind::List => {
1876 let list = List::thaw(&self.frozen).map_err(|e| {
1877 Error::new(Code::Corrupt, "a demoted list did not read back")
1878 .with_detail(e.to_string())
1879 })?;
1880 self.lists.insert(list)
1881 }
1882 Kind::Zset => {
1883 let zset = Zset::thaw(&self.frozen).map_err(|e| {
1884 Error::new(Code::Corrupt, "a demoted sorted set did not read back")
1885 .with_detail(e.to_string())
1886 })?;
1887 self.zsets.insert(zset)
1888 }
1889 Kind::Array => {
1890 let array = Array::thaw(&self.frozen).map_err(|e| {
1891 Error::new(Code::Corrupt, "a demoted array did not read back")
1892 .with_detail(e.to_string())
1893 })?;
1894 self.arrays.insert(array)
1895 }
1896 Kind::Stream => {
1897 let stream = Stream::thaw(&self.frozen).map_err(|e| {
1898 Error::new(Code::Corrupt, "a demoted stream did not read back")
1899 .with_detail(e.to_string())
1900 })?;
1901 self.streams.insert(stream)
1902 }
1903 // Nothing else is written cold with a body yet, so arriving here
1904 // means a record that says one thing and a demoter that did another.
1905 _ => {
1906 return Err(Error::new(
1907 Code::Corrupt,
1908 "a demoted body of a type that is not moved out",
1909 )
1910 .with_detail(kind.name().to_string()));
1911 }
1912 };
1913 self.bodies += 1;
1914 let wrote = self.map.set_with(
1915 key,
1916 value::slot_record_len(expire_at.is_some()),
1917 |_| {},
1918 |out| {
1919 value::write_slot_record(out, kind, slot, expire_at);
1920 value::set_access(out, was);
1921 value::has_expiry(out)
1922 },
1923 );
1924 debug_assert!(wrote.is_some(), "the key was found a moment ago");
1925 Ok(Some(slot))
1926 }
1927
1928 /// Read `key`'s value off the file into [`Keyspace::cold`], if that is where
1929 /// it is.
1930 ///
1931 /// The doorkeeper decides whether the value also goes back into memory, so
1932 /// after this the record under `key` is either resident or still cold with
1933 /// its bytes in the buffer, and [`Keyspace::value_of`] is what tells the two
1934 /// apart. The address the caller was holding is not valid afterwards on the
1935 /// promoting path, because promotion rewrites the record.
1936 ///
1937 /// # Errors
1938 ///
1939 /// Whatever the store says when the chain will not read back.
1940 pub(crate) fn warm(&mut self, key: &[u8]) -> Result<Faulted> {
1941 if self.tier.is_none() {
1942 return Ok(Faulted::Warm);
1943 }
1944 if self.body_came_back(key)? {
1945 return Ok(Faulted::Promoted);
1946 }
1947 let tier = self.tier.as_mut().expect("checked at the top");
1948 let mut buf = core::mem::take(&mut self.cold);
1949 let r = tier.fault(&mut self.map, key, &mut buf);
1950 self.cold = buf;
1951 if r == Ok(Faulted::Served) {
1952 self.cold_key.clear();
1953 self.cold_key.extend_from_slice(key);
1954 }
1955 r
1956 }
1957
1958 /// Put `key`'s value back in memory if it was not there, for a command that
1959 /// is about to write it.
1960 ///
1961 /// See [`Tier::thaw`] for why the doorkeeper does not get a vote here. The
1962 /// short of it is that a read modify write leaves a resident record either
1963 /// way, so there is nothing for an answer to change.
1964 ///
1965 /// A caller can carry on exactly as it did before after this returns, with
1966 /// no cold case left to handle, which is why it is one line at the top of
1967 /// `APPEND` and `INCR` rather than a second path through them.
1968 ///
1969 /// # Errors
1970 ///
1971 /// Whatever the store says when the chain will not read back.
1972 pub(crate) fn thaw(&mut self, key: &[u8]) -> Result<()> {
1973 if self.tier.is_none() {
1974 return Ok(());
1975 }
1976 if self.body_came_back(key)? {
1977 return Ok(());
1978 }
1979 let tier = self.tier.as_mut().expect("checked at the top");
1980 let mut buf = core::mem::take(&mut self.cold);
1981 let r = tier.thaw(&mut self.map, key, &mut buf);
1982 self.cold = buf;
1983 r.map(|_| ())
1984 }
1985
1986 /// The value in a record that may have been left on the file.
1987 ///
1988 /// Only correct straight after a [`Keyspace::warm`] of the same key, since
1989 /// the buffer holds one value at a time. A debug build says so rather than
1990 /// handing back a value that belongs to somebody else.
1991 pub(crate) fn value_of<'a>(&'a self, key: &[u8], rec: &'a [u8]) -> Str<'a> {
1992 if value::cold(rec).is_some() {
1993 debug_assert!(
1994 bytes_eq(&self.cold_key, key),
1995 "a read found a value on the file without faulting it in first"
1996 );
1997 Str::Bytes(&self.cold)
1998 } else {
1999 value::read(rec)
2000 }
2001 }
2002
2003 /// [`Keyspace::warm`] and then the value, for the readers that do nothing
2004 /// else with the record.
2005 ///
2006 /// `None` when the key went away, which it cannot do here but which the
2007 /// second lookup has to allow for anyway: the first lookup's address does
2008 /// not survive a promotion, so this has to find the key again rather than
2009 /// trust a number from before.
2010 ///
2011 /// # Errors
2012 ///
2013 /// Whatever the store says when the chain will not read back.
2014 pub(crate) fn warmed(&mut self, key: &[u8]) -> Result<Option<Str<'_>>> {
2015 self.warm(key)?;
2016 let Some(addr) = self.map.find(key) else {
2017 return Ok(None);
2018 };
2019 Ok(Some(self.value_of(key, self.map.value_at(addr))))
2020 }
2021
2022 /// Where `key`'s record is, having thrown the key away first if it is dead.
2023 ///
2024 /// The same fold as [`Keyspace::live_slot`] for a caller that wants the
2025 /// record itself rather than a slot number, which is every string command.
2026 /// `GET` used to be a reap, then a type check, then a read, and each of the
2027 /// three hashed the key and walked a bucket for the same record. It is one
2028 /// walk now and two arena reads, and an arena read at a known address is a
2029 /// load.
2030 ///
2031 /// The address dies at the next write, which is why this is `pub(crate)`
2032 /// and why every caller reads it and drops it inside one command.
2033 ///
2034 /// Finding a key counts as using it, so this stamps the access field on the
2035 /// way past under every policy that wants it stamped, which is eight of the
2036 /// ten. A command that has to look at a key without using it calls
2037 /// [`Keyspace::live_rec_untouched`] instead, and the list of those is short
2038 /// and is Redis's list rather than ours.
2039 pub(crate) fn live_rec(&mut self, key: &[u8]) -> Option<Addr> {
2040 let addr = self.live_rec_untouched(key)?;
2041 if self.policy.stamps_on_read() {
2042 self.stamp(addr);
2043 }
2044 Some(addr)
2045 }
2046
2047 /// [`Keyspace::live_rec`] for a command that is asking about a key rather
2048 /// than using it.
2049 ///
2050 /// `TYPE`, `EXISTS`, the `TTL` family and every `OBJECT` subcommand look
2051 /// without touching, which is Redis's `LOOKUP_NOTOUCH` and is not an
2052 /// optimisation. `OBJECT IDLETIME` that counted as a use would report zero
2053 /// every time, and `EXISTS` in a health check loop would keep a dead key at
2054 /// the top of the working set forever.
2055 ///
2056 /// Untouched means the access field only. A key past its deadline is still
2057 /// reaped here, because a command asking whether a key exists has to be told
2058 /// that it does not.
2059 pub(crate) fn live_rec_untouched(&mut self, key: &[u8]) -> Option<Addr> {
2060 let now = self.clock.now_ms();
2061 let Some(addr) = self.map.find(key) else {
2062 self.looked(false);
2063 return None;
2064 };
2065 if value::is_expired(self.map.value_at(addr), now) {
2066 self.reaped(key);
2067 self.looked(false);
2068 return None;
2069 }
2070 self.looked(true);
2071 Some(addr)
2072 }
2073
2074 /// Count a lookup a client's read made, whichever way it went.
2075 ///
2076 /// Only what a read did. Every command comes through the funnels above this
2077 /// and the ones that write are not what the hit rate is about, so the
2078 /// dispatcher says which kind is running and [`lookups`] carries the answer
2079 /// down. See that module for why it is a thread local.
2080 ///
2081 /// A key found and then refused for holding the wrong type is a hit, since
2082 /// the lookup did find what was under the name. That is Redis's rule as
2083 /// well, and it falls out of where the count sits rather than being decided
2084 /// here: the type check is above this and this has already counted.
2085 #[inline]
2086 fn looked(&mut self, found: bool) {
2087 if lookups::is_reading() {
2088 if found {
2089 self.hits += 1;
2090 } else {
2091 self.misses += 1;
2092 }
2093 }
2094 }
2095
2096 /// The slot under `key`, having thrown the key away first if it is dead.
2097 ///
2098 /// `None` for a key that is not there or that was and is now reaped, and
2099 /// `WRONGTYPE` for a key holding something other than `want`.
2100 ///
2101 /// One probe of the map, where a [`Keyspace::reap`] followed by a `get`
2102 /// costs two. That pair is how every collection command used to start, so a
2103 /// pipeline of sixty four `SADD` on one key hashed and probed for that key a
2104 /// hundred and twenty eight times to do sixty four inserts. The reap has to
2105 /// read the record and the command has to read the same record, and there
2106 /// was never a reason for those to be two visits.
2107 ///
2108 /// It answers a number rather than the record it just read because of the
2109 /// borrow checker and not because a number is nicer. A method that hands
2110 /// back a borrow of the map on one path and takes a mutable borrow to reap
2111 /// on the other is the case the borrow checker still refuses without
2112 /// Polonius. A slot is four bytes and copies out, so the borrow ends here
2113 /// and the caller reaches its body through the slab.
2114 ///
2115 /// And no probe at all when the command in front of it asked for the same
2116 /// key and nothing has been written since, which is the [`Memo`] and is what
2117 /// Y13 asks for on single key `SADD`.
2118 pub(crate) fn live_slot(&mut self, key: &[u8], want: Kind) -> Result<Option<u32>> {
2119 // A memo hit skips the record, so the stamp has to happen on the way out
2120 // of it as well. It is the same address every time, which is what makes
2121 // this cheap: no probe, just the store. Getting this wrong is the trap
2122 // worth naming, because the key that hits the memo most often is the
2123 // hottest key in the database, and it is the one that would have looked
2124 // steadily more idle the harder it was used.
2125 if let Some((kind, slot, addr)) = self.memo.get(self.map.writes(), key) {
2126 // Remembering where a key was is still finding it, so the count
2127 // happens on this path too. A pipeline of `SISMEMBER` on one key
2128 // would otherwise report a hit rate that fell as the key got hotter.
2129 self.looked(true);
2130 if kind != want {
2131 return Err(wrong_type());
2132 }
2133 if self.policy.stamps_on_read() {
2134 self.stamp(addr);
2135 }
2136 return Ok(Some(slot));
2137 }
2138 // `find` and then `value_at` rather than `get`, which is the same two
2139 // steps, so that the address is still in hand for the stamp below. `get`
2140 // would mean probing a second time for a record already read.
2141 let now = self.clock.now_ms();
2142 let Some(addr) = self.map.find(key) else {
2143 self.looked(false);
2144 return Ok(None);
2145 };
2146 let rec = self.map.value_at(addr);
2147 if value::is_expired(rec, now) {
2148 self.reaped(key);
2149 self.looked(false);
2150 return Ok(None);
2151 }
2152 if value::kind(rec) != want {
2153 // Counted on the way out of each of the three exits below rather
2154 // than once above them, because the record is borrowed here and
2155 // counting needs the keyspace mutably. On the path that takes an
2156 // exit the borrow is over, which is why this compiles and a single
2157 // call in front of them would not.
2158 self.looked(true);
2159 return Err(wrong_type());
2160 }
2161 // One test of a bit in a byte that is already in a register, on the
2162 // funnel every collection command comes through. A database with no file
2163 // behind it never sets it and pays that test and nothing else.
2164 if value::Meta::from_byte(rec[0]).is_cold() {
2165 self.looked(true);
2166 return self.promote_body(key);
2167 }
2168 let slot = value::slot(rec);
2169 // A key with a deadline is not memoized. The memo is invalidated by
2170 // writes and a deadline passes without one, so remembering a dated key
2171 // would be remembering it past the moment it should have been reaped.
2172 // Both of these are read off the record before the stamp, which needs it
2173 // mutably and is the end of this borrow.
2174 let dated = value::expire_at(rec).is_some();
2175 self.looked(true);
2176 if self.policy.stamps_on_read() {
2177 self.stamp(addr);
2178 }
2179 if !dated {
2180 self.memo.put(self.map.writes(), key, want, slot, addr);
2181 }
2182 Ok(Some(slot))
2183 }
2184
2185 /// Where `key` is, when either of two types will do.
2186 ///
2187 /// Every input to a sorted set operation may be a sorted set or a plain set,
2188 /// which is Redis's rule and means the type check there is a membership test
2189 /// rather than an equality. The kind comes back with the slot because the
2190 /// caller has to know which slab the number indexes.
2191 pub(crate) fn live_slot_either(
2192 &mut self,
2193 key: &[u8],
2194 a: Kind,
2195 b: Kind,
2196 ) -> Result<Option<(Kind, u32)>> {
2197 if let Some((kind, slot, addr)) = self.memo.get(self.map.writes(), key) {
2198 self.looked(true);
2199 if kind != a && kind != b {
2200 return Err(wrong_type());
2201 }
2202 if self.policy.stamps_on_read() {
2203 self.stamp(addr);
2204 }
2205 return Ok(Some((kind, slot)));
2206 }
2207 let now = self.clock.now_ms();
2208 let Some(addr) = self.map.find(key) else {
2209 self.looked(false);
2210 return Ok(None);
2211 };
2212 let rec = self.map.value_at(addr);
2213 if value::is_expired(rec, now) {
2214 self.reaped(key);
2215 self.looked(false);
2216 return Ok(None);
2217 }
2218 let kind = value::kind(rec);
2219 if kind != a && kind != b {
2220 self.looked(true);
2221 return Err(wrong_type());
2222 }
2223 // As in `live_slot`, and the kind was read before the record was given
2224 // up because promotion rewrites it.
2225 if value::Meta::from_byte(rec[0]).is_cold() {
2226 self.looked(true);
2227 return Ok(self.promote_body(key)?.map(|slot| (kind, slot)));
2228 }
2229 let slot = value::slot(rec);
2230 let dated = value::expire_at(rec).is_some();
2231 self.looked(true);
2232 if self.policy.stamps_on_read() {
2233 self.stamp(addr);
2234 }
2235 if !dated {
2236 self.memo.put(self.map.writes(), key, kind, slot, addr);
2237 }
2238 Ok(Some((kind, slot)))
2239 }
2240
2241 /// Throw every key away. This is `FLUSHDB` on one database.
2242 ///
2243 /// The expiry counter is not reset, because Redis does not reset it either:
2244 /// `expired_keys` in `INFO stats` counts what this process has expired since
2245 /// it started, and emptying a database is not expiring anything. The count of
2246 /// keys that carry a deadline is a different number and it does go to zero,
2247 /// because it is a fact about what is in the database right now and there is
2248 /// nothing in it.
2249 pub fn clear(&mut self) {
2250 self.map.clear();
2251 self.sets.clear();
2252 self.hashes.clear();
2253 self.lists.clear();
2254 self.zsets.clear();
2255 self.arrays.clear();
2256 self.streams.clear();
2257 self.foreign.clear();
2258 self.pool.clear();
2259 self.bodies = 0;
2260 // The same argument as the deadline count above. It is a list of keys
2261 // that are in the database, and there are none.
2262 self.field_deadlines.clear();
2263 self.field_at = 0;
2264 }
2265
2266 /// Keys reclaimed by running into them after their deadline.
2267 ///
2268 /// Redis calls this `expired_keys` in `INFO stats` and counts both lazy and
2269 /// active expiry into it, and so does this. [`Keyspace::expire_cycle`] is
2270 /// the active half and it counts into the same number, which is what makes
2271 /// this the total a dashboard can compare against a write rate rather than
2272 /// the share of it that happened to be reclaimed by a read.
2273 #[inline]
2274 pub const fn expired_keys(&self) -> u64 {
2275 self.expired
2276 }
2277
2278 /// Hash fields reclaimed after their own deadline passed.
2279 ///
2280 /// Redis calls this `expired_subkeys` and counts both halves into it, the
2281 /// fields a command tripped over on its way past and the fields
2282 /// [`Keyspace::field_expire_cycle`] went looking for.
2283 #[inline]
2284 pub const fn expired_fields(&self) -> u64 {
2285 self.expired_fields
2286 }
2287
2288 /// The share of those the cycle found.
2289 ///
2290 /// Redis calls this `expired_subkeys_active`. It is the number that says
2291 /// whether the cycle is keeping up: a server where it stays far behind the
2292 /// total is one where the fields are only going because clients keep asking
2293 /// for the hashes they are in.
2294 #[inline]
2295 pub const fn expired_fields_active(&self) -> u64 {
2296 self.expired_fields_active
2297 }
2298
2299 /// Keys thrown away to make room.
2300 ///
2301 /// Redis calls this `evicted_keys` in `INFO stats`. It stays at zero under
2302 /// `noeviction`, which is the whole point of that policy, and a monitoring
2303 /// dashboard that sees it move on a server configured that way is looking at
2304 /// a bug rather than at load.
2305 #[inline]
2306 pub const fn evicted_keys(&self) -> u64 {
2307 self.evicted
2308 }
2309
2310 /// Lookups a client's read made here that found the key.
2311 ///
2312 /// Redis calls this `keyspace_hits` in `INFO stats`, and with the number
2313 /// below it is the hit rate every dashboard watching a cache is drawn from.
2314 /// Only reads are in it: the lookup a `SET` does on its way to writing a key
2315 /// is not a hit and its failing to find one is not a miss. See [`lookups`]
2316 /// for how the two are told apart.
2317 #[inline]
2318 pub const fn hits(&self) -> u64 {
2319 self.hits
2320 }
2321
2322 /// Lookups a client's read made here that did not find the key.
2323 ///
2324 /// Redis calls this `keyspace_misses`. A key found and refused for holding
2325 /// the wrong type is not one of these, and neither is a field a hash does not
2326 /// have: both of those are lookups that found the key.
2327 #[inline]
2328 pub const fn misses(&self) -> u64 {
2329 self.misses
2330 }
2331
2332 /// Put every running total this database keeps back to zero, which is
2333 /// `CONFIG RESETSTAT`.
2334 ///
2335 /// All six and not just the two lookup counters, because Redis's
2336 /// `resetServerStats` clears the expiry and eviction totals in the same
2337 /// breath. They are totals since the server started and this is the command
2338 /// that says start again from here, so a dashboard that resets and then
2339 /// watches has to see all of them move together or none of them.
2340 ///
2341 /// Not the same thing as `FLUSHDB`, which throws away the keys and leaves
2342 /// every one of these alone. See [`Keyspace::clear`].
2343 pub const fn zero_stats(&mut self) {
2344 self.hits = 0;
2345 self.misses = 0;
2346 self.expired = 0;
2347 self.evicted = 0;
2348 self.expired_fields = 0;
2349 self.expired_fields_active = 0;
2350 }
2351
2352 /// How many live keys carry a deadline.
2353 ///
2354 /// This is what `INFO keyspace` reports as `expires=`, and it is the live
2355 /// count rather than a running total: a key that gets a `TTL` and then has it
2356 /// taken away with `PERSIST` is in it and then is not.
2357 ///
2358 /// The map keeps it, because the map keeps the second index these keys are
2359 /// in. Nothing in this file counts it, which is deliberate: a count kept
2360 /// alongside the thing it counts is a count that eventually disagrees with
2361 /// it, and the one place that can be wrong should be the one place that owns
2362 /// the entries.
2363 #[inline]
2364 pub fn expires(&self) -> usize {
2365 self.map.tagged_len()
2366 }
2367
2368 /// How many keys a round of eviction sampling looks at.
2369 #[inline]
2370 pub const fn samples(&self) -> usize {
2371 self.samples
2372 }
2373
2374 /// Set how many keys a round of eviction sampling looks at.
2375 ///
2376 /// Zero is not refused here, because the caller doing the refusing is
2377 /// `CONFIG SET` and it has a message to produce. A zero that reaches here
2378 /// samples one bucket and takes the best of it, because the loop runs its
2379 /// body before it checks, which is a better answer than dividing by nothing.
2380 #[inline]
2381 pub const fn set_samples(&mut self, samples: usize) {
2382 self.samples = samples;
2383 }
2384
2385 /// Throw away one key, chosen by the policy. Answers whether one went.
2386 ///
2387 /// This is one step and not a loop on purpose. The caller is the thing that
2388 /// knows how much room it needs back, and a loop in here would either take
2389 /// too much or have to be told the same number twice. It also means the
2390 /// caller can put a bound on how long it spends evicting before it answers
2391 /// the client, which matters because the client is waiting on a write that
2392 /// this is making room for.
2393 ///
2394 /// It answers false without doing anything under `noeviction`, and also when
2395 /// a `volatile` policy is set on a database where nothing has a deadline.
2396 /// Those are the same answer to the caller and they mean the same thing: this
2397 /// server cannot give memory back and is about to have to refuse a write.
2398 pub fn evict_one(&mut self) -> bool {
2399 let Some(addr) = self.victim() else {
2400 return false;
2401 };
2402 // The key has to outlive the borrow that found it, because deleting is a
2403 // write and the address came out of a read. One copy into the scratch
2404 // buffer rather than a `Vec` per eviction, for the reason written on
2405 // [`Keyspace::scratch`]: this runs in a loop when it runs at all.
2406 let mut buf = core::mem::take(&mut self.scratch);
2407 buf.clear();
2408 buf.extend_from_slice(self.map.entry_at(addr).0);
2409 let gone = self.drop_key(&buf);
2410 if gone {
2411 self.evicted += 1;
2412 // Before the buffer goes back, since that is what is holding the
2413 // key, and a listener wants the name of what it lost.
2414 news::say(&buf, news::What::Evicted);
2415 }
2416 self.scratch = buf;
2417 gone
2418 }
2419
2420 /// Where the key this policy would throw away lives, if there is one.
2421 ///
2422 /// The sampling loop. It draws buckets until it has looked at `samples` keys
2423 /// the policy would consider, scores each one, and hands back the best. See
2424 /// [`evict`] for what the score means and [`yo_index::RawMap::sample`] for
2425 /// why a bucket is the unit.
2426 ///
2427 /// The round cap is the part that is not obvious. A database with a hundred
2428 /// keys in a directory sized for a million is mostly empty buckets, and a
2429 /// `volatile` policy on a database where nothing has a deadline has no
2430 /// eligible keys at all however many buckets it looks in. Without the cap the
2431 /// second case is an infinite loop, and it is not a rare configuration, it is
2432 /// the classic eviction surprise. With it, the worst case is a fixed number
2433 /// of cache misses and a false, which is exactly what the caller needs to
2434 /// hear.
2435 ///
2436 /// A key past its deadline is skipped rather than taken. It is dead memory
2437 /// and evicting it would look like a win, but it would be counted as an
2438 /// eviction when it is an expiry, and those two numbers are watched
2439 /// separately for a reason. Lazy expiry takes it the next time anything asks
2440 /// for it, and the active cycle takes it before that.
2441 ///
2442 /// What comes back is not only the worst of this round. Everything sampled
2443 /// goes into [`evict::Pool`], which holds the sixteen best across rounds, so
2444 /// the answer is the worst key seen since the pool was last emptied. The
2445 /// price is that a candidate is a key rather than an address and so has to
2446 /// be looked up and rechecked here, because it can have been deleted or have
2447 /// expired or have lost its deadline since the round that spotted it.
2448 fn victim(&mut self) -> Option<Addr> {
2449 if matches!(self.policy, Policy::NoEviction) || self.map.is_empty() {
2450 self.pool.clear();
2451 return None;
2452 }
2453 // The classic eviction surprise, answered before it costs anything. A
2454 // `volatile` policy on a database where no key has a deadline has no
2455 // eligible key anywhere, and the loop below can only find that out by
2456 // drawing four rounds of buckets and being told so by every key in them,
2457 // on a path where a client is waiting for the write this is making room
2458 // for. The count knows.
2459 if self.policy.volatile_only() && self.expires() == 0 {
2460 self.pool.clear();
2461 return None;
2462 }
2463 let now = self.clock.now_ms();
2464 let (policy, lfu, want) = (self.policy, self.lfu, self.samples);
2465 if policy.is_random() {
2466 return self.draw(now, want);
2467 }
2468 // Which index to draw from. The `volatile` policies can only take a key
2469 // that has a deadline, and the map keeps a second index of exactly
2470 // those, so drawing from the whole keyspace and then throwing most of it
2471 // away is work with a cheaper alternative sitting right there. The
2472 // `allkeys` policies draw from everything, because everything is
2473 // eligible.
2474 let volatile = policy.volatile_only();
2475 let mut seen = 0usize;
2476 for _ in 0..ROUNDS {
2477 let r = self.rng.next_u64();
2478 let pool = &mut self.pool;
2479 // By reference, so the same closure can go to either sampler. A
2480 // `&mut F` is an `FnMut` when `F` is, which is what makes the two
2481 // calls below one closure rather than two copies of it.
2482 let mut offer = |key: &[u8], rec: &[u8], _addr: Addr| {
2483 if !value::is_expired(rec, now) && evict::eligible(rec, policy) {
2484 seen += 1;
2485 pool.offer(key, evict::score(rec, policy, now, lfu));
2486 }
2487 seen < want
2488 };
2489 if volatile {
2490 self.map.sample_tagged(r, &mut offer);
2491 } else {
2492 self.map.sample(r, &mut offer);
2493 }
2494 if seen >= want {
2495 break;
2496 }
2497 }
2498 while let Some(key) = self.pool.take() {
2499 let Some(addr) = self.map.find(key) else {
2500 continue;
2501 };
2502 let rec = self.map.value_at(addr);
2503 if value::is_expired(rec, now) || !evict::eligible(rec, policy) {
2504 continue;
2505 }
2506 return Some(addr);
2507 }
2508 None
2509 }
2510
2511 /// A fair draw among the eligible keys, which is what the random pair want.
2512 ///
2513 /// No pool, because there is no ordering for one to approximate: under
2514 /// `allkeys-random` and `volatile-random` every eligible key is as good a
2515 /// victim as every other, and remembering sixteen of them across rounds
2516 /// would only mean the same sixteen going first. The sampling is what does
2517 /// the choosing, so the address it lands on is used straight away and the
2518 /// key never has to be copied at all.
2519 fn draw(&mut self, now: u64, want: usize) -> Option<Addr> {
2520 let policy = self.policy;
2521 let volatile = policy.volatile_only();
2522 let mut best = evict::Best::EMPTY;
2523 let mut seen = 0usize;
2524 for _ in 0..ROUNDS {
2525 let r = self.rng.next_u64();
2526 let mut offer = |_key: &[u8], rec: &[u8], addr: Addr| {
2527 if !value::is_expired(rec, now) && evict::eligible(rec, policy) {
2528 seen += 1;
2529 best.offer(addr, evict::ANY);
2530 }
2531 seen < want
2532 };
2533 if volatile {
2534 self.map.sample_tagged(r, &mut offer);
2535 } else {
2536 self.map.sample(r, &mut offer);
2537 }
2538 if seen >= want {
2539 break;
2540 }
2541 }
2542 (!best.is_empty()).then_some(best.addr)
2543 }
2544
2545 /// Bytes held by the index, the arena and every body hanging off them.
2546 ///
2547 /// Asks every collection, so this is O(the number of collections) and is for
2548 /// the places that want the number exactly and are asked for it rarely:
2549 /// `INFO memory`, `MEMORY USAGE` and the tests.
2550 /// [`Keyspace::settled_memory_bytes`] is the one a memory limit uses.
2551 #[inline]
2552 pub fn memory_bytes(&self) -> usize {
2553 self.slab_bytes()
2554 + self.sets.value_bytes()
2555 + self.hashes.value_bytes()
2556 + self.lists.value_bytes()
2557 + self.zsets.value_bytes()
2558 + self.arrays.value_bytes()
2559 + self.streams.value_bytes()
2560 + self.foreign.value_bytes()
2561 }
2562
2563 /// The same number, asked only of the collections that could have moved.
2564 ///
2565 /// See [`Slab::track_bytes`] for how that is known. With tracking on this
2566 /// costs what the batch touched instead of what the database holds, which is
2567 /// what lets a server with a `maxmemory` ask once a batch. With tracking off
2568 /// it is [`Keyspace::memory_bytes`] and the two cannot disagree, because
2569 /// they are the same sum over the same values either way.
2570 #[inline]
2571 pub fn settled_memory_bytes(&mut self) -> usize {
2572 self.slab_bytes()
2573 + self.sets.settled_bytes()
2574 + self.hashes.settled_bytes()
2575 + self.lists.settled_bytes()
2576 + self.zsets.settled_bytes()
2577 + self.arrays.settled_bytes()
2578 + self.streams.settled_bytes()
2579 + self.foreign.settled_bytes()
2580 }
2581
2582 /// Start or stop keeping the running total in every slab.
2583 ///
2584 /// One call for all seven, because a limit is a property of the server and
2585 /// not of a type, and a database tracking its sets but not its hashes would
2586 /// answer a number that is neither of the two things it could mean.
2587 pub fn track_memory(&mut self, on: bool) {
2588 self.sets.track_bytes(on);
2589 self.hashes.track_bytes(on);
2590 self.lists.track_bytes(on);
2591 self.zsets.track_bytes(on);
2592 self.arrays.track_bytes(on);
2593 self.streams.track_bytes(on);
2594 self.foreign.track_bytes(on);
2595 }
2596
2597 /// The index, the arena and the slot arrays, none of which need asking
2598 /// twice, plus the tier when there is one.
2599 ///
2600 /// The tier is in here because a doorkeeper and a directory buffer are real
2601 /// memory and a limit that did not count them would be a limit on part of
2602 /// the server. It is also the honest way round: the thing that gives memory
2603 /// back costs some to keep, and both numbers belong in the same total.
2604 #[inline]
2605 fn slab_bytes(&self) -> usize {
2606 self.tier.as_ref().map_or(0, Tier::memory_bytes)
2607 + self.map.memory_bytes()
2608 + self.sets.slot_bytes()
2609 + self.hashes.slot_bytes()
2610 + self.lists.slot_bytes()
2611 + self.zsets.slot_bytes()
2612 + self.arrays.slot_bytes()
2613 + self.streams.slot_bytes()
2614 + self.foreign.slot_bytes()
2615 }
2616
2617 /// Give back one segment's worth of space if one has gone mostly dead.
2618 ///
2619 /// Overwriting a key does not reuse its bytes, it writes the new record at
2620 /// the bump pointer and counts the old one as dead, so a workload that sets
2621 /// the same keys over and over holds far more than it is storing until
2622 /// something compacts. This is that something, and it does at most one
2623 /// segment per call so that the loop can afford to ask every turn.
2624 #[inline]
2625 pub fn compact_step(&mut self) -> Option<usize> {
2626 self.map.compact_step()
2627 }
2628
2629 /// The same, for a store that is over a memory limit and has to give pages
2630 /// back rather than wait for a segment to be worth collecting.
2631 ///
2632 /// See [`RawMap::compact_hard`] for why the choice of segment changes and
2633 /// why it only changes under pressure.
2634 #[inline]
2635 pub fn compact_hard(&mut self) -> Option<usize> {
2636 self.map.compact_hard()
2637 }
2638
2639 /// Ask the cache for the bucket this key will land in.
2640 ///
2641 /// The first of the loop's two walks (`04` section 3) calls this.
2642 #[inline]
2643 pub fn prefetch(&self, hash: u64) {
2644 self.map.prefetch(hash);
2645 }
2646
2647 /// The hash this database files `key` under.
2648 #[inline]
2649 #[must_use]
2650 pub fn hash_of(key: &[u8]) -> u64 {
2651 RawMap::hash_of(key)
2652 }
2653}
2654
2655/// Whether a body of this kind can leave memory yet.
2656///
2657/// One list rather than a check in each of the three places that need it, so a
2658/// type that gains a `freeze` cannot be demotable in the sweep and unreadable on
2659/// the way back. A kind that is not in here stays in memory, which costs a
2660/// demotion that did not happen and is never a wrong answer.
2661const fn moves(kind: Kind) -> bool {
2662 matches!(
2663 kind,
2664 Kind::Set | Kind::Hash | Kind::List | Kind::Zset | Kind::Array | Kind::Stream
2665 )
2666}
2667
2668/// What Redis says when a command is sent at a key holding another type.
2669///
2670/// The text is Redis's, word for word, because it goes on the wire verbatim and
2671/// clients match on it. The `WRONGTYPE` at the front is not part of the message:
2672/// the protocol layer puts it there from the [`Code`], which is what lets an
2673/// embedded caller match on a value instead of on a string (P5).
2674pub fn wrong_type() -> Error {
2675 Error::new(
2676 Code::WrongType,
2677 "Operation against a key holding the wrong kind of value",
2678 )
2679}
2680
2681impl Default for Keyspace {
2682 fn default() -> Keyspace {
2683 Keyspace::new()
2684 }
2685}
2686
2687#[cfg(test)]
2688mod tests {
2689 use super::*;
2690
2691 fn db() -> Keyspace {
2692 Keyspace::with_clock(Clock::fixed(1_000))
2693 }
2694
2695 #[test]
2696 fn the_two_counters_move_only_while_a_read_is_armed() {
2697 let mut d = db();
2698 d.set_plain(b"k", b"v").expect("room");
2699 // The write above and the reads below it are the same lookups as far as
2700 // anything in here can tell, and the only thing that separates them is
2701 // the setting the wire layer arms.
2702 assert_eq!((d.hits(), d.misses()), (0, 0));
2703 assert!(d.get(b"k").expect("a string").is_some());
2704 assert_eq!((d.hits(), d.misses()), (0, 0), "nothing was armed");
2705
2706 let armed = lookups::reading(true);
2707 assert!(d.get(b"k").expect("a string").is_some());
2708 assert!(d.get(b"nope").expect("nothing there").is_none());
2709 assert_eq!((d.hits(), d.misses()), (1, 1));
2710
2711 // And a lookup on the way to a write is not one of them, which is what
2712 // every command that reads a key and then writes it turns on.
2713 {
2714 let _quiet = lookups::quiet();
2715 assert!(d.get(b"k").expect("a string").is_some());
2716 assert!(d.get(b"nope").expect("nothing there").is_none());
2717 }
2718 assert_eq!(
2719 (d.hits(), d.misses()),
2720 (1, 1),
2721 "the quiet ones are not in it"
2722 );
2723
2724 drop(armed);
2725 assert!(d.get(b"k").expect("a string").is_some());
2726 assert_eq!(
2727 (d.hits(), d.misses()),
2728 (1, 1),
2729 "and it stopped when it was dropped"
2730 );
2731
2732 d.zero_stats();
2733 assert_eq!((d.hits(), d.misses()), (0, 0));
2734 }
2735
2736 /// A key that is there but holds the wrong thing was found, so it is a hit,
2737 /// and a key that has passed its deadline was not, so it is a miss. Both are
2738 /// Redis's answers and neither falls out of what the command returned.
2739 #[test]
2740 fn a_wrong_type_is_a_hit_and_a_key_past_its_deadline_is_a_miss() {
2741 let mut d = db();
2742 d.set_plain(b"s", b"v").expect("room");
2743 d.psetex(b"dead", 100, b"v").expect("room");
2744 let _armed = lookups::reading(true);
2745 assert!(d.lrange(b"s", 0, -1).is_err(), "a list read of a string");
2746 assert_eq!((d.hits(), d.misses()), (1, 0));
2747
2748 d.clock().advance(100);
2749 assert!(d.get(b"dead").expect("nothing there").is_none());
2750 assert_eq!((d.hits(), d.misses()), (1, 1));
2751 }
2752
2753 #[test]
2754 fn type_answers_string_for_a_string_and_nothing_for_a_missing_key() {
2755 let mut d = db();
2756 d.set_plain(b"k", b"v").expect("room");
2757 assert_eq!(d.kind_of(b"k"), Some(Kind::String));
2758 assert_eq!(d.kind_of(b"nope"), None);
2759 }
2760
2761 #[test]
2762 fn type_does_not_report_a_key_whose_deadline_has_gone() {
2763 let mut d = db();
2764 d.psetex(b"k", 100, b"v").expect("room");
2765 assert_eq!(d.kind_of(b"k"), Some(Kind::String));
2766
2767 d.clock().advance(100);
2768 assert_eq!(
2769 d.kind_of(b"k"),
2770 None,
2771 "the deadline was 1100 and it is 1100"
2772 );
2773 assert_eq!(d.len(), 0, "and asking reaped it rather than leaving it");
2774 assert_eq!(d.expired_keys(), 1);
2775 }
2776
2777 /// The default policy evicts nothing and still keeps the clock, which is
2778 /// Redis's behaviour and is the configuration nearly every server runs.
2779 #[test]
2780 fn the_clock_runs_under_the_default_policy() {
2781 let mut d = db();
2782 assert_eq!(d.policy(), Policy::NoEviction);
2783 d.set_plain(b"k", b"v").expect("room");
2784 assert_eq!(d.idle_secs(b"k"), Some(0));
2785
2786 d.clock().advance(60_000);
2787 assert_eq!(d.idle_secs(b"k"), Some(60), "a minute of nobody asking");
2788
2789 d.get(b"k").expect("a string").expect("still there");
2790 assert_eq!(d.idle_secs(b"k"), Some(0), "and reading it is using it");
2791 }
2792
2793 /// The commands that ask about a key rather than use it. Getting this wrong
2794 /// makes `OBJECT IDLETIME` answer zero every time it is called, because
2795 /// calling it would be the most recent use.
2796 #[test]
2797 fn asking_about_a_key_is_not_using_it() {
2798 let mut d = db();
2799 d.set_plain(b"k", b"v").expect("room");
2800 d.clock().advance(30_000);
2801
2802 assert!(d.exists(b"k"));
2803 assert_eq!(d.kind_of(b"k"), Some(Kind::String));
2804 assert_eq!(d.encoding_name(b"k"), Some("embstr"));
2805 assert_eq!(d.deadline_of(b"k"), Ask::NoDeadline);
2806 assert_eq!(d.expire_at(b"k"), None);
2807 assert_eq!(d.idle_secs(b"k"), Some(30));
2808
2809 assert_eq!(
2810 d.idle_secs(b"k"),
2811 Some(30),
2812 "and asking twice is still not using it"
2813 );
2814 }
2815
2816 /// Least recently modified is the one policy where a read must leave the
2817 /// clock where it is, because the clock is the only thing it measures.
2818 #[test]
2819 fn a_read_moves_the_clock_under_lru_and_leaves_it_under_lrm() {
2820 for (policy, idle_after_read) in [(Policy::AllKeysLru, 0), (Policy::AllKeysLrm, 45)] {
2821 let mut d = db();
2822 d.set_policy(policy);
2823 d.set_plain(b"k", b"v").expect("room");
2824 d.clock().advance(45_000);
2825
2826 d.get(b"k").expect("a string").expect("still there");
2827 assert_eq!(
2828 d.idle_secs(b"k"),
2829 Some(idle_after_read),
2830 "{}",
2831 policy.name()
2832 );
2833
2834 // Both of them move it on a write, which is the whole of what LRM
2835 // is measuring and is a side effect of the resolve under LRU.
2836 d.set_plain(b"k", b"w").expect("room");
2837 assert_eq!(
2838 d.idle_secs(b"k"),
2839 Some(0),
2840 "{} after a write",
2841 policy.name()
2842 );
2843 }
2844 }
2845
2846 /// The trap the memo sets. A hit skips the record entirely, so a stamp that
2847 /// only happened on a miss would leave the hottest key in the database
2848 /// looking steadily more idle the harder it was used.
2849 #[test]
2850 fn the_hot_key_path_still_stamps() {
2851 let mut d = db();
2852 d.set_policy(Policy::AllKeysLru);
2853 d.sadd(b"s", [&b"a"[..]].into_iter()).expect("room");
2854
2855 // Warm the memo, then run the key hard with nothing written in between,
2856 // which is the case the memo exists for.
2857 d.scard(b"s").expect("a set");
2858 d.clock().advance(120_000);
2859 for _ in 0..64 {
2860 d.scard(b"s").expect("a set");
2861 }
2862 assert_eq!(d.idle_secs(b"s"), Some(0), "the memo swallowed the stamp");
2863 }
2864
2865 /// Under LFU the same bits are a counter, and it climbs with use rather than
2866 /// resetting to now.
2867 #[test]
2868 fn the_counter_climbs_under_an_lfu_policy() {
2869 let mut d = db();
2870 d.set_policy(Policy::AllKeysLfu);
2871 d.seed(7);
2872 d.set_plain(b"k", b"v").expect("room");
2873 let start = d.freq(b"k").expect("there");
2874
2875 for _ in 0..200 {
2876 d.get(b"k").expect("a string").expect("still there");
2877 }
2878 let hot = d.freq(b"k").expect("there");
2879 assert!(hot > start, "{hot} did not climb from {start}");
2880
2881 // And a key nobody reads decays rather than holding its place forever.
2882 d.set_plain(b"cold", b"v").expect("room");
2883 d.clock().advance(60_000 * 10);
2884 assert!(d.freq(b"cold").expect("there") < start);
2885 }
2886
2887 /// Nothing goes under `noeviction`, which is the only promise that policy
2888 /// makes and the reason it is the default.
2889 #[test]
2890 fn noeviction_evicts_nothing() {
2891 let mut d = db();
2892 for i in 0..200u32 {
2893 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
2894 }
2895 assert!(!d.evict_one());
2896 assert_eq!(d.len(), 200);
2897 assert_eq!(d.evicted_keys(), 0);
2898 }
2899
2900 /// A volatile policy on a database where nothing has a deadline is the
2901 /// classic surprise: it looks configured and it cannot free a byte.
2902 #[test]
2903 fn a_volatile_policy_with_no_deadlines_anywhere_cannot_evict() {
2904 let mut d = db();
2905 d.set_policy(Policy::VolatileLru);
2906 for i in 0..200u32 {
2907 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
2908 }
2909 assert!(!d.evict_one(), "it found a key it had no business taking");
2910 assert_eq!(d.len(), 200);
2911
2912 // Give one key a deadline and it becomes the only thing that can go,
2913 // however many rounds of sampling that takes.
2914 let deadline = d.clock().now_ms() + 100_000;
2915 d.set_expiry(b"k7", Some(deadline));
2916 assert!(d.evict_one());
2917 assert!(!d.exists(b"k7"));
2918 assert_eq!(d.evicted_keys(), 1);
2919 }
2920
2921 /// The count against a walk, over everything that can move it. If these two
2922 /// ever disagree the count is worse than useless, because `INFO` would be
2923 /// reporting a number that looks like a measurement.
2924 #[test]
2925 fn the_deadline_count_says_what_a_walk_of_the_keyspace_says() {
2926 let mut d = db();
2927 let now = d.clock().now_ms();
2928 let check = |d: &mut Keyspace, note: &str| {
2929 let mut names = Vec::new();
2930 d.keys(|k| names.push(k.to_vec()));
2931 let walked = names
2932 .iter()
2933 .filter(|k| matches!(d.deadline_of(k), Ask::At(_)))
2934 .count();
2935 assert_eq!(d.expires(), walked, "{note}");
2936 };
2937
2938 for i in 0..40u32 {
2939 d.set_plain(format!("s{i}").as_bytes(), b"v").expect("room");
2940 d.sadd(format!("c{i}").as_bytes(), [b"m".as_slice()].into_iter())
2941 .expect("room");
2942 }
2943 check(&mut d, "nothing has a deadline yet");
2944 assert_eq!(d.expires(), 0);
2945
2946 // On, on again with a different deadline, and off.
2947 for i in (0..40u32).step_by(2) {
2948 d.set_expiry(format!("s{i}").as_bytes(), Some(now + 500_000));
2949 d.set_expiry(format!("c{i}").as_bytes(), Some(now + 500_000));
2950 }
2951 check(&mut d, "half of each type has one");
2952 assert_eq!(d.expires(), 40);
2953 for i in (0..40u32).step_by(4) {
2954 d.set_expiry(format!("s{i}").as_bytes(), Some(now + 900_000));
2955 }
2956 check(&mut d, "moving a deadline is not gaining one");
2957 assert_eq!(d.expires(), 40);
2958 for i in (0..40u32).step_by(4) {
2959 d.set_expiry(format!("c{i}").as_bytes(), None);
2960 }
2961 check(&mut d, "and PERSIST gives them back");
2962 assert_eq!(d.expires(), 30);
2963
2964 // Written over, which is the path where the record loses its deadline
2965 // without anybody saying so.
2966 d.set_plain(b"s2", b"fresh").expect("room");
2967 check(&mut d, "a plain SET drops the deadline it wrote over");
2968
2969 // Renamed, deleted, expired and evicted.
2970 d.rename(b"s6", b"s6new", false);
2971 check(&mut d, "a rename moved one rather than losing it");
2972 d.drop_key(b"s6new");
2973 d.drop_key(b"c2");
2974 check(&mut d, "two deleted");
2975 d.psetex(b"gone", 50, b"v").expect("room");
2976 check(&mut d, "and one more with a short deadline");
2977 d.clock().advance(60);
2978 assert_eq!(d.kind_of(b"gone"), None, "which the read reaped");
2979 check(&mut d, "so the count lost it too");
2980 d.set_policy(Policy::VolatileRandom);
2981 assert!(d.evict_one());
2982 check(&mut d, "eviction under a volatile policy takes one of them");
2983
2984 d.clear();
2985 assert_eq!(d.expires(), 0, "and FLUSHDB takes the lot");
2986 }
2987
2988 /// The point of the count on the eviction path. A volatile policy with
2989 /// nothing to evict answers on the comparison rather than on four rounds of
2990 /// buckets, and it has to still answer `false`.
2991 #[test]
2992 fn a_volatile_policy_asks_the_count_before_it_samples() {
2993 let mut d = db();
2994 d.set_policy(Policy::VolatileLfu);
2995 for i in 0..500u32 {
2996 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
2997 }
2998 assert_eq!(d.expires(), 0);
2999 assert!(!d.evict_one(), "nothing is eligible and nothing went");
3000 assert_eq!(d.len(), 500);
3001
3002 // And the fast path gets out of the way the moment one key qualifies.
3003 d.set_expiry(b"k123", Some(d.clock().now_ms() + 100_000));
3004 assert_eq!(d.expires(), 1);
3005 assert!(d.evict_one());
3006 assert!(!d.exists(b"k123"));
3007 assert_eq!(d.expires(), 0, "and the count went with it");
3008 }
3009
3010 /// A needle in a haystack, which is the case sampling the whole keyspace
3011 /// could not do.
3012 ///
3013 /// Fifty thousand keys and three of them with a deadline. Under a volatile
3014 /// policy those three are the only ones that may go, and four rounds of
3015 /// sixty four buckets drawn from the whole index would land on one of them
3016 /// about once in a hundred tries. Drawn from the index of just the keys that
3017 /// carry a deadline it is the only thing there is to land on.
3018 ///
3019 /// Three of them and not one, so that the test is about finding an eligible
3020 /// key rather than about a table with a single entry in it.
3021 #[test]
3022 fn a_volatile_policy_finds_the_one_key_in_a_database_that_is_not_volatile() {
3023 let mut d = db();
3024 d.set_policy(Policy::VolatileLru);
3025 for i in 0..50_000u32 {
3026 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3027 }
3028 let deadline = d.clock().now_ms() + 100_000;
3029 for k in [b"k7".as_slice(), b"k30000", b"k49999"] {
3030 assert!(d.set_expiry(k, Some(deadline)));
3031 }
3032 assert_eq!(d.expires(), 3);
3033
3034 for round in 0..3 {
3035 assert!(d.evict_one(), "round {round} found nothing to take");
3036 }
3037 assert_eq!(d.expires(), 0, "all three went");
3038 assert_eq!(d.len(), 50_000 - 3, "and nothing else did");
3039 assert!(!d.evict_one(), "and now there is nothing eligible left");
3040 assert_eq!(d.len(), 50_000 - 3);
3041 }
3042
3043 /// The direction of the score, which is the thing worth pinning. A test that
3044 /// only checked something was evicted would pass just as happily on a cache
3045 /// that keeps the cold keys and throws away the hot ones.
3046 #[test]
3047 fn the_stale_key_goes_before_the_fresh_one() {
3048 let mut d = db();
3049 d.set_policy(Policy::AllKeysLru);
3050 // Two keys is a small enough database that a bucket holds both of them
3051 // and the pick is between them rather than between whatever turned up.
3052 d.set_plain(b"cold", b"v").expect("room");
3053 d.clock().advance(600_000);
3054 d.set_plain(b"hot", b"v").expect("room");
3055
3056 assert!(d.evict_one());
3057 assert!(!d.exists(b"cold"), "it kept the stale one");
3058 assert!(d.exists(b"hot"), "it took the fresh one");
3059 }
3060
3061 /// Under `volatile-ttl` the ordering is by deadline and not by use, so the
3062 /// key about to expire anyway is the one that goes.
3063 #[test]
3064 fn the_soonest_deadline_goes_first() {
3065 let mut d = db();
3066 d.set_policy(Policy::VolatileTtl);
3067 let now = d.clock().now_ms();
3068 d.set_plain(b"soon", b"v").expect("room");
3069 d.set_plain(b"later", b"v").expect("room");
3070 d.set_expiry(b"soon", Some(now + 10_000));
3071 d.set_expiry(b"later", Some(now + 900_000));
3072
3073 assert!(d.evict_one());
3074 assert!(!d.exists(b"soon"));
3075 assert!(d.exists(b"later"));
3076 }
3077
3078 /// Under LFU the key nobody reads goes, even though it was written more
3079 /// recently than the one that survives. That is the difference between the
3080 /// two families and it is invisible to a test written against the clock.
3081 #[test]
3082 fn the_least_used_key_goes_under_lfu() {
3083 let mut d = db();
3084 d.set_policy(Policy::AllKeysLfu);
3085 d.seed(11);
3086 d.set_plain(b"popular", b"v").expect("room");
3087 for _ in 0..300 {
3088 d.get(b"popular").expect("a string").expect("still there");
3089 }
3090 // Written after the reads above, so under any clock policy this would be
3091 // the freshest key in the database and the last thing to go.
3092 d.set_plain(b"ignored", b"v").expect("room");
3093
3094 assert!(d.evict_one());
3095 assert!(!d.exists(b"ignored"));
3096 assert!(d.exists(b"popular"));
3097 }
3098
3099 /// Sampling has to keep working when almost every bucket it looks in is
3100 /// empty, which is what a database looks like after most of it is deleted.
3101 #[test]
3102 fn a_nearly_empty_database_still_gives_up_a_key() {
3103 let mut d = db();
3104 d.set_policy(Policy::AllKeysRandom);
3105 for i in 0..4000u32 {
3106 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3107 }
3108 for i in 0..3999u32 {
3109 d.drop_key(format!("k{i}").as_bytes());
3110 }
3111 assert_eq!(d.len(), 1);
3112
3113 // One key in a directory sized for four thousand. It may take more than
3114 // one round to land on it, and it may take more than one call, but the
3115 // rounds are bounded and so is this loop.
3116 let mut went = false;
3117 for _ in 0..500 {
3118 if d.evict_one() {
3119 went = true;
3120 break;
3121 }
3122 }
3123 assert!(went, "sampling never found the one key that was left");
3124 assert_eq!(d.len(), 0);
3125 assert!(!d.evict_one(), "and an empty database has nothing to give");
3126 }
3127
3128 /// Eviction and expiry are counted apart, so a key that was already dead
3129 /// when sampling found it is not billed as an eviction.
3130 #[test]
3131 fn a_dead_key_is_not_evicted() {
3132 let mut d = db();
3133 d.set_policy(Policy::AllKeysLru);
3134 let now = d.clock().now_ms();
3135 d.set_plain(b"k", b"v").expect("room");
3136 d.set_expiry(b"k", Some(now + 1000));
3137 d.clock().advance(5000);
3138
3139 assert!(!d.evict_one(), "it evicted a key that was already dead");
3140 assert_eq!(d.evicted_keys(), 0);
3141 }
3142
3143 /// The point of the pool. A round looks at five keys, takes one, and used to
3144 /// throw the other four away, so the second worst key in the database had to
3145 /// be found again from scratch every time.
3146 #[test]
3147 fn a_candidate_that_was_not_taken_is_still_in_the_running() {
3148 let mut d = db();
3149 d.set_policy(Policy::AllKeysLru);
3150 for i in 0..40u32 {
3151 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3152 d.clock().advance(1000);
3153 }
3154 assert!(d.pool.is_empty(), "nothing has sampled anything yet");
3155
3156 assert!(d.evict_one());
3157 assert!(
3158 !d.pool.is_empty(),
3159 "every key it looked at and did not take was thrown away"
3160 );
3161 }
3162
3163 /// A candidate is a key and not an address, so it can stop being a key
3164 /// between the round that spotted it and the round that wants it. Every one
3165 /// of them going at once is the worst case, and the answer has to be the
3166 /// live key rather than a shrug.
3167 #[test]
3168 fn a_candidate_that_went_away_is_stepped_over() {
3169 let mut d = db();
3170 d.set_policy(Policy::AllKeysLru);
3171 for i in 0..40u32 {
3172 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3173 d.clock().advance(1000);
3174 }
3175 assert!(d.evict_one());
3176 assert!(!d.pool.is_empty());
3177
3178 // By hand, so every candidate still held names a key that is not there.
3179 for i in 0..40u32 {
3180 d.drop_key(format!("k{i}").as_bytes());
3181 }
3182 assert_eq!(d.len(), 0);
3183 d.set_plain(b"fresh", b"v").expect("room");
3184
3185 assert!(d.evict_one(), "it gave up on a database with a key in it");
3186 assert!(!d.exists(b"fresh"));
3187 assert!(d.pool.is_empty(), "and the stale ones went with it");
3188 }
3189
3190 /// A score only means something against another score under the same rule,
3191 /// so a pool full of them is worth nothing the moment the rule changes.
3192 #[test]
3193 fn changing_the_policy_throws_the_candidates_away() {
3194 let mut d = db();
3195 d.set_policy(Policy::AllKeysLru);
3196 let now = d.clock().now_ms();
3197 for i in 0..40u32 {
3198 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3199 d.set_expiry(
3200 format!("k{i}").as_bytes(),
3201 Some(now + 100_000 + u64::from(i)),
3202 );
3203 d.clock().advance(1000);
3204 }
3205 assert!(d.evict_one());
3206 assert!(!d.pool.is_empty());
3207
3208 d.set_policy(Policy::VolatileTtl);
3209 assert!(d.pool.is_empty(), "idle seconds against a countdown");
3210
3211 // And the same policy set again is not a change and costs nothing.
3212 d.set_policy(Policy::VolatileTtl);
3213 assert!(d.evict_one());
3214 assert!(!d.pool.is_empty());
3215 d.set_policy(Policy::VolatileTtl);
3216 assert!(!d.pool.is_empty());
3217 }
3218
3219 /// A fair draw has no ordering for a pool to get closer to, so the random
3220 /// pair never copy a key at all.
3221 #[test]
3222 fn a_random_policy_keeps_no_candidates() {
3223 let mut d = db();
3224 d.set_policy(Policy::AllKeysRandom);
3225 for i in 0..40u32 {
3226 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3227 }
3228
3229 assert!(d.evict_one());
3230 assert_eq!(d.len(), 39);
3231 assert!(d.pool.is_empty());
3232 assert_eq!(
3233 d.pool.memory_bytes(),
3234 0,
3235 "and it allocated nothing to do it"
3236 );
3237 }
3238
3239 /// A flush leaves the pool naming keys that are all gone, which the recheck
3240 /// would survive and would pay sixteen lookups for.
3241 #[test]
3242 fn a_flush_takes_the_candidates_with_it() {
3243 let mut d = db();
3244 d.set_policy(Policy::AllKeysLru);
3245 for i in 0..40u32 {
3246 d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
3247 d.clock().advance(1000);
3248 }
3249 assert!(d.evict_one());
3250 assert!(!d.pool.is_empty());
3251
3252 d.clear();
3253 assert!(d.pool.is_empty());
3254 }
3255}