Line data Source code
1 : //! Control file serialization, deserialization and persistence.
2 :
3 : use anyhow::{bail, ensure, Context, Result};
4 : use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
5 : use camino::{Utf8Path, Utf8PathBuf};
6 : use tokio::fs::File;
7 : use tokio::io::AsyncWriteExt;
8 : use utils::crashsafe::durable_rename;
9 :
10 : use std::io::Read;
11 : use std::ops::Deref;
12 : use std::path::Path;
13 : use std::time::Instant;
14 :
15 : use crate::metrics::PERSIST_CONTROL_FILE_SECONDS;
16 : use crate::state::TimelinePersistentState;
17 : use crate::{control_file_upgrade::upgrade_control_file, timeline::get_timeline_dir};
18 : use utils::{bin_ser::LeSer, id::TenantTimelineId};
19 :
20 : use crate::SafeKeeperConf;
21 :
22 : pub const SK_MAGIC: u32 = 0xcafeceefu32;
23 : pub const SK_FORMAT_VERSION: u32 = 8;
24 :
25 : // contains persistent metadata for safekeeper
26 : pub const CONTROL_FILE_NAME: &str = "safekeeper.control";
27 : // needed to atomically update the state using `rename`
28 : const CONTROL_FILE_NAME_PARTIAL: &str = "safekeeper.control.partial";
29 : pub const CHECKSUM_SIZE: usize = std::mem::size_of::<u32>();
30 :
31 : /// Storage should keep actual state inside of it. It should implement Deref
32 : /// trait to access state fields and have persist method for updating that state.
33 : #[async_trait::async_trait]
34 : pub trait Storage: Deref<Target = TimelinePersistentState> {
35 : /// Persist safekeeper state on disk and update internal state.
36 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()>;
37 :
38 : /// Timestamp of last persist.
39 : fn last_persist_at(&self) -> Instant;
40 : }
41 :
42 : #[derive(Debug)]
43 : pub struct FileStorage {
44 : // save timeline dir to avoid reconstructing it every time
45 : timeline_dir: Utf8PathBuf,
46 : no_sync: bool,
47 :
48 : /// Last state persisted to disk.
49 : state: TimelinePersistentState,
50 : /// Not preserved across restarts.
51 : last_persist_at: Instant,
52 : }
53 :
54 : impl FileStorage {
55 : /// Initialize storage by loading state from disk.
56 4 : pub fn restore_new(ttid: &TenantTimelineId, conf: &SafeKeeperConf) -> Result<FileStorage> {
57 4 : let timeline_dir = get_timeline_dir(conf, ttid);
58 4 : let state = Self::load_control_file_from_dir(&timeline_dir)?;
59 :
60 2 : Ok(FileStorage {
61 2 : timeline_dir,
62 2 : no_sync: conf.no_sync,
63 2 : state,
64 2 : last_persist_at: Instant::now(),
65 2 : })
66 4 : }
67 :
68 : /// Create file storage for a new timeline, but don't persist it yet.
69 4 : pub fn create_new(
70 4 : timeline_dir: Utf8PathBuf,
71 4 : conf: &SafeKeeperConf,
72 4 : state: TimelinePersistentState,
73 4 : ) -> Result<FileStorage> {
74 4 : let store = FileStorage {
75 4 : timeline_dir,
76 4 : no_sync: conf.no_sync,
77 4 : state,
78 4 : last_persist_at: Instant::now(),
79 4 : };
80 4 :
81 4 : Ok(store)
82 4 : }
83 :
84 : /// Check the magic/version in the on-disk data and deserialize it, if possible.
85 4 : fn deser_sk_state(buf: &mut &[u8]) -> Result<TimelinePersistentState> {
86 : // Read the version independent part
87 4 : let magic = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
88 4 : if magic != SK_MAGIC {
89 0 : bail!(
90 0 : "bad control file magic: {:X}, expected {:X}",
91 0 : magic,
92 0 : SK_MAGIC
93 0 : );
94 4 : }
95 4 : let version = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
96 4 : if version == SK_FORMAT_VERSION {
97 4 : let res = TimelinePersistentState::des(buf)?;
98 4 : return Ok(res);
99 0 : }
100 0 : // try to upgrade
101 0 : upgrade_control_file(buf, version)
102 4 : }
103 :
104 : /// Load control file from given directory.
105 6 : pub fn load_control_file_from_dir(timeline_dir: &Utf8Path) -> Result<TimelinePersistentState> {
106 6 : let path = timeline_dir.join(CONTROL_FILE_NAME);
107 6 : Self::load_control_file(path)
108 6 : }
109 :
110 : /// Read in the control file.
111 6 : pub fn load_control_file<P: AsRef<Path>>(
112 6 : control_file_path: P,
113 6 : ) -> Result<TimelinePersistentState> {
114 6 : let mut control_file = std::fs::OpenOptions::new()
115 6 : .read(true)
116 6 : .write(true)
117 6 : .open(&control_file_path)
118 6 : .with_context(|| {
119 0 : format!(
120 0 : "failed to open control file at {}",
121 0 : control_file_path.as_ref().display(),
122 0 : )
123 6 : })?;
124 :
125 6 : let mut buf = Vec::new();
126 6 : control_file
127 6 : .read_to_end(&mut buf)
128 6 : .context("failed to read control file")?;
129 :
130 6 : let calculated_checksum = crc32c::crc32c(&buf[..buf.len() - CHECKSUM_SIZE]);
131 :
132 6 : let expected_checksum_bytes: &[u8; CHECKSUM_SIZE] =
133 6 : buf[buf.len() - CHECKSUM_SIZE..].try_into()?;
134 6 : let expected_checksum = u32::from_le_bytes(*expected_checksum_bytes);
135 6 :
136 6 : ensure!(
137 6 : calculated_checksum == expected_checksum,
138 2 : format!(
139 2 : "safekeeper control file checksum mismatch: expected {} got {}",
140 2 : expected_checksum, calculated_checksum
141 2 : )
142 : );
143 :
144 4 : let state = FileStorage::deser_sk_state(&mut &buf[..buf.len() - CHECKSUM_SIZE])
145 4 : .with_context(|| {
146 0 : format!(
147 0 : "while reading control file {}",
148 0 : control_file_path.as_ref().display(),
149 0 : )
150 4 : })?;
151 4 : Ok(state)
152 6 : }
153 : }
154 :
155 : impl Deref for FileStorage {
156 : type Target = TimelinePersistentState;
157 :
158 0 : fn deref(&self) -> &Self::Target {
159 0 : &self.state
160 0 : }
161 : }
162 :
163 : #[async_trait::async_trait]
164 : impl Storage for FileStorage {
165 : /// Persists state durably to the underlying storage.
166 : ///
167 : /// For a description, see <https://lwn.net/Articles/457667/>.
168 4 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()> {
169 4 : let _timer = PERSIST_CONTROL_FILE_SECONDS.start_timer();
170 4 :
171 4 : // write data to safekeeper.control.partial
172 4 : let control_partial_path = self.timeline_dir.join(CONTROL_FILE_NAME_PARTIAL);
173 4 : let mut control_partial = File::create(&control_partial_path).await.with_context(|| {
174 0 : format!(
175 0 : "failed to create partial control file at: {}",
176 0 : &control_partial_path
177 0 : )
178 4 : })?;
179 4 : let mut buf: Vec<u8> = Vec::new();
180 4 : WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_MAGIC)?;
181 4 : WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_FORMAT_VERSION)?;
182 4 : s.ser_into(&mut buf)?;
183 4 :
184 4 : // calculate checksum before resize
185 4 : let checksum = crc32c::crc32c(&buf);
186 4 : buf.extend_from_slice(&checksum.to_le_bytes());
187 4 :
188 4 : control_partial.write_all(&buf).await.with_context(|| {
189 0 : format!(
190 0 : "failed to write safekeeper state into control file at: {}",
191 0 : control_partial_path
192 0 : )
193 4 : })?;
194 4 : control_partial.flush().await.with_context(|| {
195 0 : format!(
196 0 : "failed to flush safekeeper state into control file at: {}",
197 0 : control_partial_path
198 0 : )
199 4 : })?;
200 4 :
201 4 : let control_path = self.timeline_dir.join(CONTROL_FILE_NAME);
202 28 : durable_rename(&control_partial_path, &control_path, !self.no_sync).await?;
203 4 :
204 4 : // update internal state
205 4 : self.state = s.clone();
206 4 : Ok(())
207 4 : }
208 :
209 0 : fn last_persist_at(&self) -> Instant {
210 0 : self.last_persist_at
211 0 : }
212 : }
213 :
214 : #[cfg(test)]
215 : mod test {
216 : use super::*;
217 : use tokio::fs;
218 : use utils::lsn::Lsn;
219 :
220 4 : fn stub_conf() -> SafeKeeperConf {
221 4 : let workdir = camino_tempfile::tempdir().unwrap().into_path();
222 4 : SafeKeeperConf {
223 4 : workdir,
224 4 : ..SafeKeeperConf::dummy()
225 4 : }
226 4 : }
227 :
228 4 : async fn load_from_control_file(
229 4 : conf: &SafeKeeperConf,
230 4 : ttid: &TenantTimelineId,
231 4 : ) -> Result<(FileStorage, TimelinePersistentState)> {
232 4 : let timeline_dir = get_timeline_dir(conf, ttid);
233 4 : fs::create_dir_all(&timeline_dir)
234 4 : .await
235 4 : .expect("failed to create timeline dir");
236 4 : Ok((
237 4 : FileStorage::restore_new(ttid, conf)?,
238 2 : FileStorage::load_control_file_from_dir(&timeline_dir)?,
239 : ))
240 4 : }
241 :
242 4 : async fn create(
243 4 : conf: &SafeKeeperConf,
244 4 : ttid: &TenantTimelineId,
245 4 : ) -> Result<(FileStorage, TimelinePersistentState)> {
246 4 : let timeline_dir = get_timeline_dir(conf, ttid);
247 4 : fs::create_dir_all(&timeline_dir)
248 4 : .await
249 4 : .expect("failed to create timeline dir");
250 4 : let state = TimelinePersistentState::empty();
251 4 : let storage = FileStorage::create_new(timeline_dir, conf, state.clone())?;
252 4 : Ok((storage, state))
253 4 : }
254 :
255 : #[tokio::test]
256 2 : async fn test_read_write_safekeeper_state() {
257 2 : let conf = stub_conf();
258 2 : let ttid = TenantTimelineId::generate();
259 2 : {
260 2 : let (mut storage, mut state) =
261 2 : create(&conf, &ttid).await.expect("failed to create state");
262 2 : // change something
263 2 : state.commit_lsn = Lsn(42);
264 2 : storage
265 2 : .persist(&state)
266 18 : .await
267 2 : .expect("failed to persist state");
268 2 : }
269 2 :
270 2 : let (_, state) = load_from_control_file(&conf, &ttid)
271 2 : .await
272 2 : .expect("failed to read state");
273 2 : assert_eq!(state.commit_lsn, Lsn(42));
274 2 : }
275 :
276 : #[tokio::test]
277 2 : async fn test_safekeeper_state_checksum_mismatch() {
278 2 : let conf = stub_conf();
279 2 : let ttid = TenantTimelineId::generate();
280 2 : {
281 2 : let (mut storage, mut state) =
282 2 : create(&conf, &ttid).await.expect("failed to read state");
283 2 :
284 2 : // change something
285 2 : state.commit_lsn = Lsn(42);
286 2 : storage
287 2 : .persist(&state)
288 18 : .await
289 2 : .expect("failed to persist state");
290 2 : }
291 2 : let control_path = get_timeline_dir(&conf, &ttid).join(CONTROL_FILE_NAME);
292 2 : let mut data = fs::read(&control_path).await.unwrap();
293 2 : data[0] += 1; // change the first byte of the file to fail checksum validation
294 2 : fs::write(&control_path, &data)
295 2 : .await
296 2 : .expect("failed to write control file");
297 2 :
298 2 : match load_from_control_file(&conf, &ttid).await {
299 2 : Err(err) => assert!(err
300 2 : .to_string()
301 2 : .contains("safekeeper control file checksum mismatch")),
302 2 : Ok(_) => panic!("expected error"),
303 2 : }
304 2 : }
305 : }
|
