1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
//! Mapped appends: the log file is extended ahead of the writes in chunks,
//! each chunk is mapped, and an append is a copy into the mapping — no
//! `write()` and no user-space buffer. A store into a shared mapping is in
//! the page cache the moment it is made, so a killed process keeps every
//! append that returned. The file ends in zeros past the last record, the
//! part of the last chunk no record reached; replay stops there cleanly
//! and the open cuts it off.
//!
//! Only worth it where a fresh mapped page is cheap: on Apple platforms
//! a 4 KiB append this way costs about a tenth of going through `write()`,
//! while on Linux the page fault makes it slower than `write()`.
use kevy_sys::{FileMap, MapSync};
use std::fs::File;
use std::io::{self, Write};
use crate::aof::Aof;
const FIRST_CHUNK: u64 = 4 << 20;
const LAST_CHUNK: u64 = 64 << 20;
/// Mappings start on a multiple of this, which is a whole number of pages
/// on every platform (4 KiB and 16 KiB pages alike).
const ALIGN: u64 = 64 * 1024;
/// A log's mapped chunks and where its records end.
#[derive(Debug)]
pub(crate) struct Mapped {
file: File,
/// Each chunk and the file offset it starts at, in file order.
chunks: Vec<(u64, FileMap)>,
/// The logical end: where the next record goes.
end: u64,
/// The file's length: the end of the last chunk.
phys: u64,
next_chunk: u64,
}
impl Mapped {
/// Map `file` for appends after its first `end` bytes.
fn open(file: File, end: u64) -> io::Result<Mapped> {
let mut m = Mapped { file, chunks: Vec::new(), end, phys: end, next_chunk: FIRST_CHUNK };
let start = end - end % ALIGN;
m.extend_from(start, FIRST_CHUNK)?;
Ok(m)
}
fn extend_from(&mut self, start: u64, len: u64) -> io::Result<()> {
kevy_sys::preallocate(&self.file, len)?;
self.file.set_len(start + len)?;
let map = FileMap::map_at(&self.file, start, len as usize)?;
self.chunks.push((start, map));
self.phys = start + len;
self.next_chunk = (len * 2).min(LAST_CHUNK);
Ok(())
}
/// Sync handles for the chunks, for a sync that runs after the lock is
/// released: each chunk stays mapped until its last holder lets go.
pub(crate) fn sync_handles(&self) -> Vec<MapSync> {
self.chunks.iter().map(|(_, m)| m.sync_handle()).collect()
}
/// Put the mapped pages on the medium: `msync` each chunk, then the
/// file's own sync (`F_FULLFSYNC` on Apple platforms).
pub(crate) fn sync(&self) -> io::Result<()> {
for (_, m) in &self.chunks {
m.sync()?;
}
self.file.sync_data()
}
/// Put the mapped pages on their way to the file, unmap, and cut the
/// file back to its last record.
fn close(self) -> io::Result<()> {
let synced = self.chunks.iter().try_for_each(|(_, m)| m.sync());
drop(self);
synced
}
}
impl Drop for Mapped {
fn drop(&mut self) {
self.chunks.clear();
// a failed cut leaves the zero tail, which the next open removes
#[expect(clippy::let_underscore_must_use, reason = "replay handles the tail it leaves")]
let _ = self.file.set_len(self.end);
}
}
/// A chunk's sync handle, as an off-lock sync carries it.
pub(crate) type MapHandle = MapSync;
/// `msync` each handle's chunk.
pub(crate) fn sync_handles(maps: &[MapHandle]) -> io::Result<()> {
maps.iter().try_for_each(MapSync::sync)
}
impl Write for Mapped {
fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
if self.end == self.phys {
let len = self.next_chunk;
self.extend_from(self.phys, len)?;
}
let (start, map) = self.chunks.last_mut().expect("a mapped log has a chunk");
let at = (self.end - *start) as usize;
let n = buf.len().min(map.len() - at);
map.bytes_mut(at, n).copy_from_slice(&buf[..n]);
self.end += n as u64;
Ok(n)
}
fn flush(&mut self) -> io::Result<()> {
Ok(())
}
}
impl Aof {
/// Append by mapping from here on: the rest of this log's life, until a
/// switch to `Always` turns it back. Only a v2 log that is not in queued
/// mode maps; call it after the log was replayed and any staging ring it
/// had was settled, so no record is still on its way through `write()`.
///
/// ```
/// use kevy_persist::{Aof, Argv, Fsync, replay_aof};
///
/// let path = std::env::temp_dir().join(format!("mapped-doc-{}.aof", std::process::id()));
/// let mut log = Aof::open(&path, Fsync::EverySec)?;
/// assert!(log.map_appends()?);
/// assert!(log.maps_appends());
/// log.append(&Argv::from(vec![b"SET".to_vec(), b"k".to_vec(), b"v".to_vec()]))?;
/// drop(log); // a clean close cuts the unused preallocation off
/// let mut n = 0;
/// replay_aof(&path, |_| n += 1)?;
/// assert_eq!(n, 1);
/// # std::fs::remove_file(&path)?;
/// # Ok::<(), std::io::Error>(())
/// ```
pub fn map_appends(&mut self) -> io::Result<bool> {
let can = self.format == crate::AofFormat::V2
&& self.queue.is_none()
&& self.stage.is_none()
&& !matches!(self.fsync, crate::Fsync::Always);
self.maps = can;
if can && self.mapped.is_none() {
self.file.flush()?;
let end = self.file.get_ref().metadata()?.len();
// a shared writable mapping needs a descriptor open for reading
// too; the log's own is append-only
let file = File::options().read(true).write(true).open(&self.path)?;
self.mapped = Some(Mapped::open(file, end)?);
}
Ok(can)
}
/// Whether appends go into a mapping.
///
/// ```
/// let path = std::env::temp_dir().join(format!("maps-doc-{}.aof", std::process::id()));
/// let log = kevy_persist::Aof::open(&path, kevy_persist::Fsync::No)?;
/// assert!(!log.maps_appends(), "a log maps only once asked to");
/// # drop(log);
/// # std::fs::remove_file(&path)?;
/// # Ok::<(), std::io::Error>(())
/// ```
pub fn maps_appends(&self) -> bool {
self.mapped.is_some()
}
/// Sync the log's bytes to the medium: the mapped chunks first, when
/// there are any, then the file.
pub(crate) fn sync_file(&self) -> io::Result<()> {
match &self.mapped {
Some(m) => m.sync(),
None => self.file.get_ref().sync_data(),
}
}
/// Handles for an off-lock sync of the mapped chunks (none unmapped).
pub(crate) fn map_handles(&self) -> Vec<MapHandle> {
self.mapped.as_ref().map(Mapped::sync_handles).unwrap_or_default()
}
/// Stop mapping for good (a switch to `Always`): later appends go
/// through `write()` again.
pub(crate) fn stop_mapping(&mut self) -> io::Result<()> {
self.maps = false;
self.unmap()
}
/// Unmap and cut the file to its records before the file is renamed
/// over, truncated or synced through the write path.
pub(crate) fn unmap(&mut self) -> io::Result<()> {
match self.mapped.take() {
Some(m) => m.close(),
None => Ok(()),
}
}
/// The file was replaced or emptied: point the staging ring at it, or
/// map its tail.
pub(crate) fn after_file_change(&mut self) -> io::Result<()> {
self.rebase_stage()?;
self.remap()
}
/// Map again after the file was replaced or emptied.
pub(crate) fn remap(&mut self) -> io::Result<()> {
if !self.maps {
return Ok(());
}
self.unmap()?;
self.map_appends().map(|_| ())
}
}