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