Line data Source code
1 : use std::os::fd::AsRawFd;
2 : use std::{
3 : borrow::Cow,
4 : fs::{self, File},
5 : io::{self, Write},
6 : };
7 :
8 : use camino::{Utf8Path, Utf8PathBuf};
9 :
10 : /// Similar to [`std::fs::create_dir`], except we fsync the
11 : /// created directory and its parent.
12 415 : pub fn create_dir(path: impl AsRef<Utf8Path>) -> io::Result<()> {
13 415 : let path = path.as_ref();
14 415 :
15 415 : fs::create_dir(path)?;
16 413 : fsync_file_and_parent(path)?;
17 413 : Ok(())
18 415 : }
19 :
20 : /// Similar to [`std::fs::create_dir_all`], except we fsync all
21 : /// newly created directories and the pre-existing parent.
22 5 : pub fn create_dir_all(path: impl AsRef<Utf8Path>) -> io::Result<()> {
23 5 : let mut path = path.as_ref();
24 5 :
25 5 : let mut dirs_to_create = Vec::new();
26 :
27 : // Figure out which directories we need to create.
28 : loop {
29 8 : match path.metadata() {
30 4 : Ok(metadata) if metadata.is_dir() => break,
31 : Ok(_) => {
32 1 : return Err(io::Error::new(
33 1 : io::ErrorKind::AlreadyExists,
34 1 : format!("non-directory found in path: {path}"),
35 1 : ));
36 : }
37 4 : Err(ref e) if e.kind() == io::ErrorKind::NotFound => {}
38 1 : Err(e) => return Err(e),
39 : }
40 :
41 3 : dirs_to_create.push(path);
42 3 :
43 3 : match path.parent() {
44 3 : Some(parent) => path = parent,
45 : None => {
46 0 : return Err(io::Error::new(
47 0 : io::ErrorKind::InvalidInput,
48 0 : format!("can't find parent of path '{path}'"),
49 0 : ));
50 : }
51 : }
52 : }
53 :
54 : // Create directories from parent to child.
55 3 : for &path in dirs_to_create.iter().rev() {
56 3 : fs::create_dir(path)?;
57 : }
58 :
59 : // Fsync the created directories from child to parent.
60 3 : for &path in dirs_to_create.iter() {
61 3 : fsync(path)?;
62 : }
63 :
64 : // If we created any new directories, fsync the parent.
65 3 : if !dirs_to_create.is_empty() {
66 2 : fsync(path)?;
67 1 : }
68 :
69 3 : Ok(())
70 5 : }
71 :
72 : /// Adds a suffix to the file(directory) name, either appending the suffix to the end of its extension,
73 : /// or if there's no extension, creates one and puts a suffix there.
74 2918 : pub fn path_with_suffix_extension(
75 2918 : original_path: impl AsRef<Utf8Path>,
76 2918 : suffix: &str,
77 2918 : ) -> Utf8PathBuf {
78 2918 : let new_extension = match original_path.as_ref().extension() {
79 1478 : Some(extension) => Cow::Owned(format!("{extension}.{suffix}")),
80 1440 : None => Cow::Borrowed(suffix),
81 : };
82 2918 : original_path.as_ref().with_extension(new_extension)
83 2918 : }
84 :
85 413 : pub fn fsync_file_and_parent(file_path: &Utf8Path) -> io::Result<()> {
86 413 : let parent = file_path.parent().ok_or_else(|| {
87 0 : io::Error::new(
88 0 : io::ErrorKind::Other,
89 0 : format!("File {file_path:?} has no parent"),
90 0 : )
91 413 : })?;
92 :
93 413 : fsync(file_path)?;
94 413 : fsync(parent)?;
95 413 : Ok(())
96 413 : }
97 :
98 831 : pub fn fsync(path: &Utf8Path) -> io::Result<()> {
99 831 : File::open(path)
100 831 : .map_err(|e| io::Error::new(e.kind(), format!("Failed to open the file {path:?}: {e}")))
101 831 : .and_then(|file| {
102 831 : file.sync_all().map_err(|e| {
103 0 : io::Error::new(
104 0 : e.kind(),
105 0 : format!("Failed to sync file {path:?} data and metadata: {e}"),
106 0 : )
107 831 : })
108 831 : })
109 831 : .map_err(|e| io::Error::new(e.kind(), format!("Failed to fsync file {path:?}: {e}")))
110 831 : }
111 :
112 2 : pub async fn fsync_async(path: impl AsRef<Utf8Path>) -> Result<(), std::io::Error> {
113 2 : tokio::fs::File::open(path.as_ref()).await?.sync_all().await
114 2 : }
115 :
116 8 : pub async fn fsync_async_opt(
117 8 : path: impl AsRef<Utf8Path>,
118 8 : do_fsync: bool,
119 8 : ) -> Result<(), std::io::Error> {
120 8 : if do_fsync {
121 0 : fsync_async(path.as_ref()).await?;
122 8 : }
123 8 : Ok(())
124 8 : }
125 :
126 : /// Like postgres' durable_rename, renames a file and issues fsyncs to make it durable. After
127 : /// returning, both the file and rename are guaranteed to be persisted. Both paths must be on the
128 : /// same file system.
129 : ///
130 : /// Unlike postgres, it only fsyncs 1) the file to make contents durable, and 2) the directory to
131 : /// make the rename durable. This sequence ensures the target file will never be incomplete.
132 : ///
133 : /// Postgres also:
134 : ///
135 : /// * Fsyncs the target file, if it exists, before the rename, to ensure either the new or existing
136 : /// file survives a crash. Current callers don't need this as it should already be fsynced if
137 : /// durability is needed.
138 : ///
139 : /// * Fsyncs the file after the rename. This can be required with certain OSes or file systems (e.g.
140 : /// NFS), but not on Linux with most common file systems like ext4 (which we currently use).
141 : ///
142 : /// An audit of 8 other databases found that none fsynced the file after a rename:
143 : /// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837180535>
144 : ///
145 : /// eBPF probes confirmed that this is sufficient with ext4, XFS, and ZFS, but possibly not Btrfs:
146 : /// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837926218>
147 : ///
148 : /// virtual_file.rs has similar code, but it doesn't use vfs.
149 : ///
150 : /// Useful links: <https://lwn.net/Articles/457667/>
151 : /// <https://www.postgresql.org/message-id/flat/56583BDD.9060302%402ndquadrant.com>
152 : /// <https://thunk.org/tytso/blog/2009/03/15/dont-fear-the-fsync/>
153 4 : pub async fn durable_rename(
154 4 : old_path: impl AsRef<Utf8Path>,
155 4 : new_path: impl AsRef<Utf8Path>,
156 4 : do_fsync: bool,
157 4 : ) -> io::Result<()> {
158 4 : // first fsync the file
159 4 : fsync_async_opt(old_path.as_ref(), do_fsync).await?;
160 :
161 : // Time to do the real deal.
162 4 : tokio::fs::rename(old_path.as_ref(), new_path.as_ref()).await?;
163 :
164 : // Now fsync the parent
165 4 : let parent = match new_path.as_ref().parent() {
166 4 : Some(p) => p,
167 0 : None => Utf8Path::new("./"), // assume current dir if there is no parent
168 : };
169 4 : fsync_async_opt(parent, do_fsync).await?;
170 :
171 4 : Ok(())
172 4 : }
173 :
174 : /// Writes a file to the specified `final_path` in a crash safe fasion, using [`std::fs`].
175 : ///
176 : /// The file is first written to the specified `tmp_path`, and in a second
177 : /// step, the `tmp_path` is renamed to the `final_path`. Intermediary fsync
178 : /// and atomic rename guarantee that, if we crash at any point, there will never
179 : /// be a partially written file at `final_path` (but maybe at `tmp_path`).
180 : ///
181 : /// Callers are responsible for serializing calls of this function for a given `final_path`.
182 : /// If they don't, there may be an error due to conflicting `tmp_path`, or there will
183 : /// be no error and the content of `final_path` will be the "winner" caller's `content`.
184 : /// I.e., the atomticity guarantees still hold.
185 28 : pub fn overwrite(
186 28 : final_path: &Utf8Path,
187 28 : tmp_path: &Utf8Path,
188 28 : content: &[u8],
189 28 : ) -> std::io::Result<()> {
190 28 : let Some(final_path_parent) = final_path.parent() else {
191 0 : return Err(std::io::Error::from_raw_os_error(
192 0 : nix::errno::Errno::EINVAL as i32,
193 0 : ));
194 : };
195 28 : std::fs::remove_file(tmp_path).or_else(crate::fs_ext::ignore_not_found)?;
196 28 : let mut file = std::fs::OpenOptions::new()
197 28 : .write(true)
198 28 : // Use `create_new` so that, if we race with ourselves or something else,
199 28 : // we bail out instead of causing damage.
200 28 : .create_new(true)
201 28 : .open(tmp_path)?;
202 28 : file.write_all(content)?;
203 28 : file.sync_all()?;
204 28 : drop(file); // don't keep the fd open for longer than we have to
205 28 :
206 28 : std::fs::rename(tmp_path, final_path)?;
207 :
208 28 : let final_parent_dirfd = std::fs::OpenOptions::new()
209 28 : .read(true)
210 28 : .open(final_path_parent)?;
211 :
212 28 : final_parent_dirfd.sync_all()?;
213 28 : Ok(())
214 28 : }
215 :
216 : /// Syncs the filesystem for the given file descriptor.
217 : #[cfg_attr(target_os = "macos", allow(unused_variables))]
218 0 : pub fn syncfs(fd: impl AsRawFd) -> anyhow::Result<()> {
219 : // Linux guarantees durability for syncfs.
220 : // POSIX doesn't have syncfs, and further does not actually guarantee durability of sync().
221 : #[cfg(target_os = "linux")]
222 0 : {
223 0 : use anyhow::Context;
224 0 : nix::unistd::syncfs(fd.as_raw_fd()).context("syncfs")?;
225 : }
226 : #[cfg(target_os = "macos")]
227 : {
228 : // macOS is not a production platform for Neon, don't even bother.
229 : }
230 : #[cfg(not(any(target_os = "linux", target_os = "macos")))]
231 : {
232 : compile_error!("Unsupported OS");
233 : }
234 0 : Ok(())
235 0 : }
236 :
237 : #[cfg(test)]
238 : mod tests {
239 :
240 : use super::*;
241 :
242 : #[test]
243 1 : fn test_create_dir_fsyncd() {
244 1 : let dir = camino_tempfile::tempdir().unwrap();
245 1 :
246 1 : let existing_dir_path = dir.path();
247 1 : let err = create_dir(existing_dir_path).unwrap_err();
248 1 : assert_eq!(err.kind(), io::ErrorKind::AlreadyExists);
249 :
250 1 : let child_dir = existing_dir_path.join("child");
251 1 : create_dir(child_dir).unwrap();
252 1 :
253 1 : let nested_child_dir = existing_dir_path.join("child1").join("child2");
254 1 : let err = create_dir(nested_child_dir).unwrap_err();
255 1 : assert_eq!(err.kind(), io::ErrorKind::NotFound);
256 1 : }
257 :
258 : #[test]
259 1 : fn test_create_dir_all_fsyncd() {
260 1 : let dir = camino_tempfile::tempdir().unwrap();
261 1 :
262 1 : let existing_dir_path = dir.path();
263 1 : create_dir_all(existing_dir_path).unwrap();
264 1 :
265 1 : let child_dir = existing_dir_path.join("child");
266 1 : assert!(!child_dir.exists());
267 1 : create_dir_all(&child_dir).unwrap();
268 1 : assert!(child_dir.exists());
269 :
270 1 : let nested_child_dir = existing_dir_path.join("child1").join("child2");
271 1 : assert!(!nested_child_dir.exists());
272 1 : create_dir_all(&nested_child_dir).unwrap();
273 1 : assert!(nested_child_dir.exists());
274 :
275 1 : let file_path = existing_dir_path.join("file");
276 1 : std::fs::write(&file_path, b"").unwrap();
277 1 :
278 1 : let err = create_dir_all(&file_path).unwrap_err();
279 1 : assert_eq!(err.kind(), io::ErrorKind::AlreadyExists);
280 :
281 1 : let invalid_dir_path = file_path.join("folder");
282 1 : create_dir_all(invalid_dir_path).unwrap_err();
283 1 : }
284 :
285 : #[test]
286 1 : fn test_path_with_suffix_extension() {
287 1 : let p = Utf8PathBuf::from("/foo/bar");
288 1 : assert_eq!(
289 1 : &path_with_suffix_extension(p, "temp").to_string(),
290 1 : "/foo/bar.temp"
291 1 : );
292 1 : let p = Utf8PathBuf::from("/foo/bar");
293 1 : assert_eq!(
294 1 : &path_with_suffix_extension(p, "temp.temp").to_string(),
295 1 : "/foo/bar.temp.temp"
296 1 : );
297 1 : let p = Utf8PathBuf::from("/foo/bar.baz");
298 1 : assert_eq!(
299 1 : &path_with_suffix_extension(p, "temp.temp").to_string(),
300 1 : "/foo/bar.baz.temp.temp"
301 1 : );
302 1 : let p = Utf8PathBuf::from("/foo/bar.baz");
303 1 : assert_eq!(
304 1 : &path_with_suffix_extension(p, ".temp").to_string(),
305 1 : "/foo/bar.baz..temp"
306 1 : );
307 1 : let p = Utf8PathBuf::from("/foo/bar/dir/");
308 1 : assert_eq!(
309 1 : &path_with_suffix_extension(p, ".temp").to_string(),
310 1 : "/foo/bar/dir..temp"
311 1 : );
312 1 : }
313 : }
|
