Line data Source code
1 : use std::borrow::Cow;
2 : use std::fs::{self, File};
3 : use std::io::{self, Write};
4 : use std::os::fd::AsRawFd;
5 :
6 : use camino::{Utf8Path, Utf8PathBuf};
7 :
8 : /// Similar to [`std::fs::create_dir`], except we fsync the
9 : /// created directory and its parent.
10 895 : pub fn create_dir(path: impl AsRef<Utf8Path>) -> io::Result<()> {
11 895 : let path = path.as_ref();
12 895 :
13 895 : fs::create_dir(path)?;
14 893 : fsync_file_and_parent(path)?;
15 893 : Ok(())
16 3 : }
17 :
18 : /// Similar to [`std::fs::create_dir_all`], except we fsync all
19 : /// newly created directories and the pre-existing parent.
20 5 : pub fn create_dir_all(path: impl AsRef<Utf8Path>) -> io::Result<()> {
21 5 : let mut path = path.as_ref();
22 5 :
23 5 : let mut dirs_to_create = Vec::new();
24 :
25 : // Figure out which directories we need to create.
26 : loop {
27 8 : match path.metadata() {
28 4 : Ok(metadata) if metadata.is_dir() => break,
29 : Ok(_) => {
30 1 : return Err(io::Error::new(
31 1 : io::ErrorKind::AlreadyExists,
32 1 : format!("non-directory found in path: {path}"),
33 1 : ));
34 : }
35 4 : Err(ref e) if e.kind() == io::ErrorKind::NotFound => {}
36 1 : Err(e) => return Err(e),
37 : }
38 :
39 3 : dirs_to_create.push(path);
40 3 :
41 3 : match path.parent() {
42 3 : Some(parent) => path = parent,
43 : None => {
44 0 : return Err(io::Error::new(
45 0 : io::ErrorKind::InvalidInput,
46 0 : format!("can't find parent of path '{path}'"),
47 0 : ));
48 : }
49 : }
50 : }
51 :
52 : // Create directories from parent to child.
53 3 : for &path in dirs_to_create.iter().rev() {
54 3 : fs::create_dir(path)?;
55 : }
56 :
57 : // Fsync the created directories from child to parent.
58 3 : for &path in dirs_to_create.iter() {
59 3 : fsync(path)?;
60 : }
61 :
62 : // If we created any new directories, fsync the parent.
63 3 : if !dirs_to_create.is_empty() {
64 2 : fsync(path)?;
65 1 : }
66 :
67 3 : Ok(())
68 5 : }
69 :
70 : /// Adds a suffix to the file(directory) name, either appending the suffix to the end of its extension,
71 : /// or if there's no extension, creates one and puts a suffix there.
72 6379 : pub fn path_with_suffix_extension(
73 6379 : original_path: impl AsRef<Utf8Path>,
74 6379 : suffix: &str,
75 6379 : ) -> Utf8PathBuf {
76 6379 : let new_extension = match original_path.as_ref().extension() {
77 3066 : Some(extension) => Cow::Owned(format!("{extension}.{suffix}")),
78 3313 : None => Cow::Borrowed(suffix),
79 : };
80 6379 : original_path.as_ref().with_extension(new_extension)
81 6379 : }
82 :
83 893 : pub fn fsync_file_and_parent(file_path: &Utf8Path) -> io::Result<()> {
84 893 : let parent = file_path
85 893 : .parent()
86 893 : .ok_or_else(|| io::Error::other(format!("File {file_path:?} has no parent")))?;
87 :
88 893 : fsync(file_path)?;
89 893 : fsync(parent)?;
90 893 : Ok(())
91 893 : }
92 :
93 1791 : pub fn fsync(path: &Utf8Path) -> io::Result<()> {
94 1791 : File::open(path)
95 1791 : .map_err(|e| io::Error::new(e.kind(), format!("Failed to open the file {path:?}: {e}")))
96 1791 : .and_then(|file| {
97 1791 : file.sync_all().map_err(|e| {
98 0 : io::Error::new(
99 0 : e.kind(),
100 0 : format!("Failed to sync file {path:?} data and metadata: {e}"),
101 0 : )
102 1791 : })
103 1791 : })
104 1791 : .map_err(|e| io::Error::new(e.kind(), format!("Failed to fsync file {path:?}: {e}")))
105 1791 : }
106 :
107 44 : pub async fn fsync_async(path: impl AsRef<Utf8Path>) -> Result<(), std::io::Error> {
108 44 : tokio::fs::File::open(path.as_ref()).await?.sync_all().await
109 0 : }
110 :
111 48 : pub async fn fsync_async_opt(
112 48 : path: impl AsRef<Utf8Path>,
113 48 : do_fsync: bool,
114 48 : ) -> Result<(), std::io::Error> {
115 48 : if do_fsync {
116 40 : fsync_async(path.as_ref()).await?;
117 0 : }
118 48 : Ok(())
119 0 : }
120 :
121 : /// Like postgres' durable_rename, renames a file and issues fsyncs to make it durable. After
122 : /// returning, both the file and rename are guaranteed to be persisted. Both paths must be on the
123 : /// same file system.
124 : ///
125 : /// Unlike postgres, it only fsyncs 1) the file to make contents durable, and 2) the directory to
126 : /// make the rename durable. This sequence ensures the target file will never be incomplete.
127 : ///
128 : /// Postgres also:
129 : ///
130 : /// * Fsyncs the target file, if it exists, before the rename, to ensure either the new or existing
131 : /// file survives a crash. Current callers don't need this as it should already be fsynced if
132 : /// durability is needed.
133 : ///
134 : /// * Fsyncs the file after the rename. This can be required with certain OSes or file systems (e.g.
135 : /// NFS), but not on Linux with most common file systems like ext4 (which we currently use).
136 : ///
137 : /// An audit of 8 other databases found that none fsynced the file after a rename:
138 : /// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837180535>
139 : ///
140 : /// eBPF probes confirmed that this is sufficient with ext4, XFS, and ZFS, but possibly not Btrfs:
141 : /// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837926218>
142 : ///
143 : /// virtual_file.rs has similar code, but it doesn't use vfs.
144 : ///
145 : /// Useful links: <https://lwn.net/Articles/457667/>
146 : /// <https://www.postgresql.org/message-id/flat/56583BDD.9060302%402ndquadrant.com>
147 : /// <https://thunk.org/tytso/blog/2009/03/15/dont-fear-the-fsync/>
148 24 : pub async fn durable_rename(
149 24 : old_path: impl AsRef<Utf8Path>,
150 24 : new_path: impl AsRef<Utf8Path>,
151 24 : do_fsync: bool,
152 24 : ) -> io::Result<()> {
153 24 : // first fsync the file
154 24 : fsync_async_opt(old_path.as_ref(), do_fsync).await?;
155 :
156 : // Time to do the real deal.
157 24 : tokio::fs::rename(old_path.as_ref(), new_path.as_ref()).await?;
158 :
159 : // Now fsync the parent
160 24 : let parent = match new_path.as_ref().parent() {
161 24 : Some(p) => p,
162 0 : None => Utf8Path::new("./"), // assume current dir if there is no parent
163 : };
164 24 : fsync_async_opt(parent, do_fsync).await?;
165 :
166 24 : Ok(())
167 0 : }
168 :
169 : /// Writes a file to the specified `final_path` in a crash safe fasion, using [`std::fs`].
170 : ///
171 : /// The file is first written to the specified `tmp_path`, and in a second
172 : /// step, the `tmp_path` is renamed to the `final_path`. Intermediary fsync
173 : /// and atomic rename guarantee that, if we crash at any point, there will never
174 : /// be a partially written file at `final_path` (but maybe at `tmp_path`).
175 : ///
176 : /// Callers are responsible for serializing calls of this function for a given `final_path`.
177 : /// If they don't, there may be an error due to conflicting `tmp_path`, or there will
178 : /// be no error and the content of `final_path` will be the "winner" caller's `content`.
179 : /// I.e., the atomticity guarantees still hold.
180 56 : pub fn overwrite(
181 56 : final_path: &Utf8Path,
182 56 : tmp_path: &Utf8Path,
183 56 : content: &[u8],
184 56 : ) -> std::io::Result<()> {
185 56 : let Some(final_path_parent) = final_path.parent() else {
186 0 : return Err(std::io::Error::from_raw_os_error(
187 0 : nix::errno::Errno::EINVAL as i32,
188 0 : ));
189 : };
190 56 : std::fs::remove_file(tmp_path).or_else(crate::fs_ext::ignore_not_found)?;
191 56 : let mut file = std::fs::OpenOptions::new()
192 56 : .write(true)
193 56 : // Use `create_new` so that, if we race with ourselves or something else,
194 56 : // we bail out instead of causing damage.
195 56 : .create_new(true)
196 56 : .open(tmp_path)?;
197 56 : file.write_all(content)?;
198 56 : file.sync_all()?;
199 56 : drop(file); // don't keep the fd open for longer than we have to
200 56 :
201 56 : std::fs::rename(tmp_path, final_path)?;
202 :
203 56 : let final_parent_dirfd = std::fs::OpenOptions::new()
204 56 : .read(true)
205 56 : .open(final_path_parent)?;
206 :
207 56 : final_parent_dirfd.sync_all()?;
208 56 : Ok(())
209 56 : }
210 :
211 : /// Syncs the filesystem for the given file descriptor.
212 : #[cfg_attr(target_os = "macos", allow(unused_variables))]
213 0 : pub fn syncfs(fd: impl AsRawFd) -> anyhow::Result<()> {
214 : // Linux guarantees durability for syncfs.
215 : // POSIX doesn't have syncfs, and further does not actually guarantee durability of sync().
216 : #[cfg(target_os = "linux")]
217 0 : {
218 0 : use anyhow::Context;
219 0 : nix::unistd::syncfs(fd.as_raw_fd()).context("syncfs")?;
220 : }
221 : #[cfg(target_os = "macos")]
222 : {
223 : // macOS is not a production platform for Neon, don't even bother.
224 : }
225 : #[cfg(not(any(target_os = "linux", target_os = "macos")))]
226 : {
227 : compile_error!("Unsupported OS");
228 : }
229 0 : Ok(())
230 0 : }
231 :
232 : #[cfg(test)]
233 : mod tests {
234 :
235 : use super::*;
236 :
237 : #[test]
238 1 : fn test_create_dir_fsyncd() {
239 1 : let dir = camino_tempfile::tempdir().unwrap();
240 1 :
241 1 : let existing_dir_path = dir.path();
242 1 : let err = create_dir(existing_dir_path).unwrap_err();
243 1 : assert_eq!(err.kind(), io::ErrorKind::AlreadyExists);
244 :
245 1 : let child_dir = existing_dir_path.join("child");
246 1 : create_dir(child_dir).unwrap();
247 1 :
248 1 : let nested_child_dir = existing_dir_path.join("child1").join("child2");
249 1 : let err = create_dir(nested_child_dir).unwrap_err();
250 1 : assert_eq!(err.kind(), io::ErrorKind::NotFound);
251 1 : }
252 :
253 : #[test]
254 1 : fn test_create_dir_all_fsyncd() {
255 1 : let dir = camino_tempfile::tempdir().unwrap();
256 1 :
257 1 : let existing_dir_path = dir.path();
258 1 : create_dir_all(existing_dir_path).unwrap();
259 1 :
260 1 : let child_dir = existing_dir_path.join("child");
261 1 : assert!(!child_dir.exists());
262 1 : create_dir_all(&child_dir).unwrap();
263 1 : assert!(child_dir.exists());
264 :
265 1 : let nested_child_dir = existing_dir_path.join("child1").join("child2");
266 1 : assert!(!nested_child_dir.exists());
267 1 : create_dir_all(&nested_child_dir).unwrap();
268 1 : assert!(nested_child_dir.exists());
269 :
270 1 : let file_path = existing_dir_path.join("file");
271 1 : std::fs::write(&file_path, b"").unwrap();
272 1 :
273 1 : let err = create_dir_all(&file_path).unwrap_err();
274 1 : assert_eq!(err.kind(), io::ErrorKind::AlreadyExists);
275 :
276 1 : let invalid_dir_path = file_path.join("folder");
277 1 : create_dir_all(invalid_dir_path).unwrap_err();
278 1 : }
279 :
280 : #[test]
281 1 : fn test_path_with_suffix_extension() {
282 1 : let p = Utf8PathBuf::from("/foo/bar");
283 1 : assert_eq!(
284 1 : &path_with_suffix_extension(p, "temp").to_string(),
285 1 : "/foo/bar.temp"
286 1 : );
287 1 : let p = Utf8PathBuf::from("/foo/bar");
288 1 : assert_eq!(
289 1 : &path_with_suffix_extension(p, "temp.temp").to_string(),
290 1 : "/foo/bar.temp.temp"
291 1 : );
292 1 : let p = Utf8PathBuf::from("/foo/bar.baz");
293 1 : assert_eq!(
294 1 : &path_with_suffix_extension(p, "temp.temp").to_string(),
295 1 : "/foo/bar.baz.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").to_string(),
300 1 : "/foo/bar.baz..temp"
301 1 : );
302 1 : let p = Utf8PathBuf::from("/foo/bar/dir/");
303 1 : assert_eq!(
304 1 : &path_with_suffix_extension(p, ".temp").to_string(),
305 1 : "/foo/bar/dir..temp"
306 1 : );
307 1 : }
308 : }
|
