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::future::Future;
11 : use std::io::Read;
12 : use std::ops::Deref;
13 : use std::path::Path;
14 : use std::time::Instant;
15 :
16 : use crate::control_file_upgrade::downgrade_v9_to_v8;
17 : use crate::metrics::PERSIST_CONTROL_FILE_SECONDS;
18 : use crate::state::{EvictionState, TimelinePersistentState};
19 : use crate::{control_file_upgrade::upgrade_control_file, timeline::get_timeline_dir};
20 : use utils::{bin_ser::LeSer, id::TenantTimelineId};
21 :
22 : use crate::SafeKeeperConf;
23 :
24 : pub const SK_MAGIC: u32 = 0xcafeceefu32;
25 : pub const SK_FORMAT_VERSION: u32 = 9;
26 :
27 : // contains persistent metadata for safekeeper
28 : pub const CONTROL_FILE_NAME: &str = "safekeeper.control";
29 : // needed to atomically update the state using `rename`
30 : const CONTROL_FILE_NAME_PARTIAL: &str = "safekeeper.control.partial";
31 : pub const CHECKSUM_SIZE: usize = size_of::<u32>();
32 :
33 : /// Storage should keep actual state inside of it. It should implement Deref
34 : /// trait to access state fields and have persist method for updating that state.
35 : pub trait Storage: Deref<Target = TimelinePersistentState> {
36 : /// Persist safekeeper state on disk and update internal state.
37 : fn persist(&mut self, s: &TimelinePersistentState) -> impl Future<Output = Result<()>> + Send;
38 :
39 : /// Timestamp of last persist.
40 : fn last_persist_at(&self) -> Instant;
41 : }
42 :
43 : #[derive(Debug)]
44 : pub struct FileStorage {
45 : // save timeline dir to avoid reconstructing it every time
46 : timeline_dir: Utf8PathBuf,
47 : no_sync: bool,
48 :
49 : /// Last state persisted to disk.
50 : state: TimelinePersistentState,
51 : /// Not preserved across restarts.
52 : last_persist_at: Instant,
53 : }
54 :
55 : impl FileStorage {
56 : /// Initialize storage by loading state from disk.
57 2 : pub fn restore_new(ttid: &TenantTimelineId, conf: &SafeKeeperConf) -> Result<FileStorage> {
58 2 : let timeline_dir = get_timeline_dir(conf, ttid);
59 2 : let state = Self::load_control_file_from_dir(&timeline_dir)?;
60 :
61 1 : Ok(FileStorage {
62 1 : timeline_dir,
63 1 : no_sync: conf.no_sync,
64 1 : state,
65 1 : last_persist_at: Instant::now(),
66 1 : })
67 2 : }
68 :
69 : /// Create file storage for a new timeline, but don't persist it yet.
70 2 : pub fn create_new(
71 2 : timeline_dir: Utf8PathBuf,
72 2 : conf: &SafeKeeperConf,
73 2 : state: TimelinePersistentState,
74 2 : ) -> Result<FileStorage> {
75 2 : // we don't support creating new timelines in offloaded state
76 2 : assert!(matches!(state.eviction_state, EvictionState::Present));
77 :
78 2 : let store = FileStorage {
79 2 : timeline_dir,
80 2 : no_sync: conf.no_sync,
81 2 : state,
82 2 : last_persist_at: Instant::now(),
83 2 : };
84 2 :
85 2 : Ok(store)
86 2 : }
87 :
88 : /// Check the magic/version in the on-disk data and deserialize it, if possible.
89 2 : fn deser_sk_state(buf: &mut &[u8]) -> Result<TimelinePersistentState> {
90 : // Read the version independent part
91 2 : let magic = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
92 2 : if magic != SK_MAGIC {
93 0 : bail!(
94 0 : "bad control file magic: {:X}, expected {:X}",
95 0 : magic,
96 0 : SK_MAGIC
97 0 : );
98 2 : }
99 2 : let version = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
100 2 : if version == SK_FORMAT_VERSION {
101 0 : let res = TimelinePersistentState::des(buf)?;
102 0 : return Ok(res);
103 2 : }
104 2 : // try to upgrade
105 2 : upgrade_control_file(buf, version)
106 2 : }
107 :
108 : /// Load control file from given directory.
109 3 : fn load_control_file_from_dir(timeline_dir: &Utf8Path) -> Result<TimelinePersistentState> {
110 3 : let path = timeline_dir.join(CONTROL_FILE_NAME);
111 3 : Self::load_control_file(path)
112 3 : }
113 :
114 : /// Read in the control file.
115 3 : pub fn load_control_file<P: AsRef<Path>>(
116 3 : control_file_path: P,
117 3 : ) -> Result<TimelinePersistentState> {
118 3 : let mut control_file = std::fs::OpenOptions::new()
119 3 : .read(true)
120 3 : .write(true)
121 3 : .open(&control_file_path)
122 3 : .with_context(|| {
123 0 : format!(
124 0 : "failed to open control file at {}",
125 0 : control_file_path.as_ref().display(),
126 0 : )
127 3 : })?;
128 :
129 3 : let mut buf = Vec::new();
130 3 : control_file
131 3 : .read_to_end(&mut buf)
132 3 : .context("failed to read control file")?;
133 :
134 3 : let calculated_checksum = crc32c::crc32c(&buf[..buf.len() - CHECKSUM_SIZE]);
135 :
136 3 : let expected_checksum_bytes: &[u8; CHECKSUM_SIZE] =
137 3 : buf[buf.len() - CHECKSUM_SIZE..].try_into()?;
138 3 : let expected_checksum = u32::from_le_bytes(*expected_checksum_bytes);
139 3 :
140 3 : ensure!(
141 3 : calculated_checksum == expected_checksum,
142 1 : format!(
143 1 : "safekeeper control file checksum mismatch: expected {} got {}",
144 1 : expected_checksum, calculated_checksum
145 1 : )
146 : );
147 :
148 2 : let state = FileStorage::deser_sk_state(&mut &buf[..buf.len() - CHECKSUM_SIZE])
149 2 : .with_context(|| {
150 0 : format!(
151 0 : "while reading control file {}",
152 0 : control_file_path.as_ref().display(),
153 0 : )
154 2 : })?;
155 2 : Ok(state)
156 3 : }
157 : }
158 :
159 : impl Deref for FileStorage {
160 : type Target = TimelinePersistentState;
161 :
162 0 : fn deref(&self) -> &Self::Target {
163 0 : &self.state
164 0 : }
165 : }
166 :
167 : impl TimelinePersistentState {
168 2 : pub(crate) fn write_to_buf(&self) -> Result<Vec<u8>> {
169 2 : let mut buf: Vec<u8> = Vec::new();
170 2 : WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_MAGIC)?;
171 :
172 2 : if self.eviction_state == EvictionState::Present {
173 2 : // temp hack for forward compatibility
174 2 : const PREV_FORMAT_VERSION: u32 = 8;
175 2 : let prev = downgrade_v9_to_v8(self);
176 2 : WriteBytesExt::write_u32::<LittleEndian>(&mut buf, PREV_FORMAT_VERSION)?;
177 2 : prev.ser_into(&mut buf)?;
178 : } else {
179 : // otherwise, we write the current format version
180 0 : WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_FORMAT_VERSION)?;
181 0 : self.ser_into(&mut buf)?;
182 : }
183 :
184 : // calculate checksum before resize
185 2 : let checksum = crc32c::crc32c(&buf);
186 2 : buf.extend_from_slice(&checksum.to_le_bytes());
187 2 : Ok(buf)
188 2 : }
189 : }
190 :
191 : impl Storage for FileStorage {
192 : /// Persists state durably to the underlying storage.
193 : ///
194 : /// For a description, see <https://lwn.net/Articles/457667/>.
195 2 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()> {
196 2 : let _timer = PERSIST_CONTROL_FILE_SECONDS.start_timer();
197 2 :
198 2 : // write data to safekeeper.control.partial
199 2 : let control_partial_path = self.timeline_dir.join(CONTROL_FILE_NAME_PARTIAL);
200 2 : let mut control_partial = File::create(&control_partial_path).await.with_context(|| {
201 0 : format!(
202 0 : "failed to create partial control file at: {}",
203 0 : &control_partial_path
204 0 : )
205 2 : })?;
206 :
207 2 : let buf: Vec<u8> = s.write_to_buf()?;
208 :
209 2 : control_partial.write_all(&buf).await.with_context(|| {
210 0 : format!(
211 0 : "failed to write safekeeper state into control file at: {}",
212 0 : control_partial_path
213 0 : )
214 2 : })?;
215 2 : control_partial.flush().await.with_context(|| {
216 0 : format!(
217 0 : "failed to flush safekeeper state into control file at: {}",
218 0 : control_partial_path
219 0 : )
220 2 : })?;
221 :
222 2 : let control_path = self.timeline_dir.join(CONTROL_FILE_NAME);
223 14 : durable_rename(&control_partial_path, &control_path, !self.no_sync).await?;
224 :
225 : // update internal state
226 2 : self.state = s.clone();
227 2 : Ok(())
228 2 : }
229 :
230 0 : fn last_persist_at(&self) -> Instant {
231 0 : self.last_persist_at
232 0 : }
233 : }
234 :
235 : #[cfg(test)]
236 : mod test {
237 : use super::*;
238 : use tokio::fs;
239 : use utils::lsn::Lsn;
240 :
241 2 : fn stub_conf() -> SafeKeeperConf {
242 2 : let workdir = camino_tempfile::tempdir().unwrap().into_path();
243 2 : SafeKeeperConf {
244 2 : workdir,
245 2 : ..SafeKeeperConf::dummy()
246 2 : }
247 2 : }
248 :
249 2 : async fn load_from_control_file(
250 2 : conf: &SafeKeeperConf,
251 2 : ttid: &TenantTimelineId,
252 2 : ) -> Result<(FileStorage, TimelinePersistentState)> {
253 2 : let timeline_dir = get_timeline_dir(conf, ttid);
254 2 : fs::create_dir_all(&timeline_dir)
255 2 : .await
256 2 : .expect("failed to create timeline dir");
257 2 : Ok((
258 2 : FileStorage::restore_new(ttid, conf)?,
259 1 : FileStorage::load_control_file_from_dir(&timeline_dir)?,
260 : ))
261 2 : }
262 :
263 2 : async fn create(
264 2 : conf: &SafeKeeperConf,
265 2 : ttid: &TenantTimelineId,
266 2 : ) -> Result<(FileStorage, TimelinePersistentState)> {
267 2 : let timeline_dir = get_timeline_dir(conf, ttid);
268 2 : fs::create_dir_all(&timeline_dir)
269 2 : .await
270 2 : .expect("failed to create timeline dir");
271 2 : let state = TimelinePersistentState::empty();
272 2 : let storage = FileStorage::create_new(timeline_dir, conf, state.clone())?;
273 2 : Ok((storage, state))
274 2 : }
275 :
276 : #[tokio::test]
277 1 : async fn test_read_write_safekeeper_state() {
278 1 : let conf = stub_conf();
279 1 : let ttid = TenantTimelineId::generate();
280 1 : {
281 1 : let (mut storage, mut state) =
282 1 : create(&conf, &ttid).await.expect("failed to create state");
283 1 : // change something
284 1 : state.commit_lsn = Lsn(42);
285 1 : storage
286 1 : .persist(&state)
287 9 : .await
288 1 : .expect("failed to persist state");
289 1 : }
290 1 :
291 1 : let (_, state) = load_from_control_file(&conf, &ttid)
292 1 : .await
293 1 : .expect("failed to read state");
294 1 : assert_eq!(state.commit_lsn, Lsn(42));
295 1 : }
296 :
297 : #[tokio::test]
298 1 : async fn test_safekeeper_state_checksum_mismatch() {
299 1 : let conf = stub_conf();
300 1 : let ttid = TenantTimelineId::generate();
301 1 : {
302 1 : let (mut storage, mut state) =
303 1 : create(&conf, &ttid).await.expect("failed to read state");
304 1 :
305 1 : // change something
306 1 : state.commit_lsn = Lsn(42);
307 1 : storage
308 1 : .persist(&state)
309 9 : .await
310 1 : .expect("failed to persist state");
311 1 : }
312 1 : let control_path = get_timeline_dir(&conf, &ttid).join(CONTROL_FILE_NAME);
313 1 : let mut data = fs::read(&control_path).await.unwrap();
314 1 : data[0] += 1; // change the first byte of the file to fail checksum validation
315 1 : fs::write(&control_path, &data)
316 1 : .await
317 1 : .expect("failed to write control file");
318 1 :
319 1 : match load_from_control_file(&conf, &ttid).await {
320 1 : Err(err) => assert!(err
321 1 : .to_string()
322 1 : .contains("safekeeper control file checksum mismatch")),
323 1 : Ok(_) => panic!("expected error"),
324 1 : }
325 1 : }
326 : }
|
