]>
Commit | Line | Data |
---|---|---|
35cf5daa | 1 | use failure::*; |
43b13033 | 2 | |
35cf5daa | 3 | use std::path::{Path, PathBuf}; |
f98ac774 | 4 | use std::io::Write; |
43b13033 DM |
5 | use std::sync::{Arc, Mutex}; |
6 | use std::os::unix::io::AsRawFd; | |
11515438 | 7 | use serde::Serialize; |
128b37fe | 8 | |
0b97bc61 DM |
9 | use proxmox::tools::fs::{CreateOptions, create_path, create_dir}; |
10 | ||
365bb90f | 11 | use crate::tools; |
4ee8f53d | 12 | use super::DataBlob; |
a5736098 | 13 | use crate::server::WorkerTask; |
365bb90f | 14 | |
f2b99c34 | 15 | #[derive(Clone, Serialize)] |
64e53b28 | 16 | pub struct GarbageCollectionStatus { |
f2b99c34 | 17 | pub upid: Option<String>, |
a660978c DM |
18 | pub index_file_count: usize, |
19 | pub index_data_bytes: u64, | |
20 | pub disk_bytes: u64, | |
64e53b28 | 21 | pub disk_chunks: usize, |
a660978c DM |
22 | pub removed_bytes: u64, |
23 | pub removed_chunks: usize, | |
64e53b28 DM |
24 | } |
25 | ||
26 | impl Default for GarbageCollectionStatus { | |
27 | fn default() -> Self { | |
28 | GarbageCollectionStatus { | |
f2b99c34 | 29 | upid: None, |
a660978c DM |
30 | index_file_count: 0, |
31 | index_data_bytes: 0, | |
64e53b28 DM |
32 | disk_bytes: 0, |
33 | disk_chunks: 0, | |
a660978c DM |
34 | removed_bytes: 0, |
35 | removed_chunks: 0, | |
64e53b28 DM |
36 | } |
37 | } | |
38 | } | |
39 | ||
e5064ba6 | 40 | /// File system based chunk store |
35cf5daa | 41 | pub struct ChunkStore { |
277fc5a3 | 42 | name: String, // used for error reporting |
2c32fdde | 43 | pub (crate) base: PathBuf, |
35cf5daa | 44 | chunk_dir: PathBuf, |
c5d82e5f | 45 | mutex: Mutex<bool>, |
43b13033 | 46 | locker: Arc<Mutex<tools::ProcessLocker>>, |
128b37fe DM |
47 | } |
48 | ||
176e4af9 DM |
49 | // TODO: what about sysctl setting vm.vfs_cache_pressure (0 - 100) ? |
50 | ||
36898ffc | 51 | pub fn verify_chunk_size(size: usize) -> Result<(), Error> { |
247cdbce | 52 | |
36898ffc | 53 | static SIZES: [usize; 7] = [64*1024, 128*1024, 256*1024, 512*1024, 1024*1024, 2048*1024, 4096*1024]; |
247cdbce DM |
54 | |
55 | if !SIZES.contains(&size) { | |
56 | bail!("Got unsupported chunk size '{}'", size); | |
57 | } | |
58 | Ok(()) | |
59 | } | |
60 | ||
08481a0b | 61 | fn digest_to_prefix(digest: &[u8]) -> PathBuf { |
128b37fe | 62 | |
e95950e4 | 63 | let mut buf = Vec::<u8>::with_capacity(2+1+2+1); |
128b37fe | 64 | |
62ee2eb4 | 65 | const HEX_CHARS: &[u8; 16] = b"0123456789abcdef"; |
22968600 | 66 | |
128b37fe DM |
67 | buf.push(HEX_CHARS[(digest[0] as usize) >> 4]); |
68 | buf.push(HEX_CHARS[(digest[0] as usize) &0xf]); | |
e95950e4 | 69 | buf.push(HEX_CHARS[(digest[1] as usize) >> 4]); |
128b37fe | 70 | buf.push(HEX_CHARS[(digest[1] as usize) & 0xf]); |
128b37fe DM |
71 | buf.push('/' as u8); |
72 | ||
73 | let path = unsafe { String::from_utf8_unchecked(buf)}; | |
74 | ||
75 | path.into() | |
35cf5daa DM |
76 | } |
77 | ||
78 | impl ChunkStore { | |
79 | ||
45773720 DM |
80 | fn chunk_dir<P: AsRef<Path>>(path: P) -> PathBuf { |
81 | ||
82 | let mut chunk_dir: PathBuf = PathBuf::from(path.as_ref()); | |
83 | chunk_dir.push(".chunks"); | |
84 | ||
85 | chunk_dir | |
86 | } | |
87 | ||
277fc5a3 | 88 | pub fn create<P: Into<PathBuf>>(name: &str, path: P) -> Result<Self, Error> { |
35cf5daa | 89 | |
45773720 | 90 | let base: PathBuf = path.into(); |
68469eeb DM |
91 | |
92 | if !base.is_absolute() { | |
93 | bail!("expected absolute path - got {:?}", base); | |
94 | } | |
95 | ||
45773720 | 96 | let chunk_dir = Self::chunk_dir(&base); |
35cf5daa | 97 | |
f74a03da | 98 | let backup_user = crate::backup::backup_user()?; |
0b97bc61 DM |
99 | |
100 | let options = CreateOptions::new() | |
f74a03da DM |
101 | .owner(backup_user.uid) |
102 | .group(backup_user.gid); | |
0b97bc61 DM |
103 | |
104 | let default_options = CreateOptions::new(); | |
105 | ||
106 | if let Err(err) = create_path(&base, Some(default_options.clone()), Some(options.clone())) { | |
277fc5a3 | 107 | bail!("unable to create chunk store '{}' at {:?} - {}", name, base, err); |
2989f6bf DM |
108 | } |
109 | ||
0b97bc61 | 110 | if let Err(err) = create_dir(&chunk_dir, options.clone()) { |
277fc5a3 | 111 | bail!("unable to create chunk store '{}' subdir {:?} - {}", name, chunk_dir, err); |
2989f6bf | 112 | } |
35cf5daa | 113 | |
7e210bd0 DM |
114 | // create lock file with correct owner/group |
115 | let lockfile_path = Self::lockfile_path(&base); | |
116 | proxmox::tools::fs::replace_file(lockfile_path, b"", options.clone())?; | |
117 | ||
bc616633 | 118 | // create 64*1024 subdirs |
e95950e4 DM |
119 | let mut last_percentage = 0; |
120 | ||
bc616633 | 121 | for i in 0..64*1024 { |
af3e7d75 | 122 | let mut l1path = chunk_dir.clone(); |
bc616633 | 123 | l1path.push(format!("{:04x}", i)); |
0b97bc61 | 124 | if let Err(err) = create_dir(&l1path, options.clone()) { |
277fc5a3 | 125 | bail!("unable to create chunk store '{}' subdir {:?} - {}", name, l1path, err); |
2989f6bf | 126 | } |
bc616633 DM |
127 | let percentage = (i*100)/(64*1024); |
128 | if percentage != last_percentage { | |
129 | eprintln!("Percentage done: {}", percentage); | |
130 | last_percentage = percentage; | |
e95950e4 | 131 | } |
35cf5daa DM |
132 | } |
133 | ||
7e210bd0 | 134 | |
277fc5a3 | 135 | Self::open(name, base) |
35cf5daa DM |
136 | } |
137 | ||
7e210bd0 DM |
138 | fn lockfile_path<P: Into<PathBuf>>(base: P) -> PathBuf { |
139 | let base: PathBuf = base.into(); | |
35cf5daa | 140 | |
7e210bd0 DM |
141 | let mut lockfile_path = base.clone(); |
142 | lockfile_path.push(".lock"); | |
143 | ||
144 | lockfile_path | |
145 | } | |
146 | ||
147 | pub fn open<P: Into<PathBuf>>(name: &str, base: P) -> Result<Self, Error> { | |
148 | ||
149 | let base: PathBuf = base.into(); | |
68469eeb DM |
150 | |
151 | if !base.is_absolute() { | |
152 | bail!("expected absolute path - got {:?}", base); | |
153 | } | |
154 | ||
45773720 DM |
155 | let chunk_dir = Self::chunk_dir(&base); |
156 | ||
ce55dbbc | 157 | if let Err(err) = std::fs::metadata(&chunk_dir) { |
277fc5a3 | 158 | bail!("unable to open chunk store '{}' at {:?} - {}", name, chunk_dir, err); |
ce55dbbc | 159 | } |
45773720 | 160 | |
7e210bd0 | 161 | let lockfile_path = Self::lockfile_path(&base); |
45773720 | 162 | |
43b13033 | 163 | let locker = tools::ProcessLocker::new(&lockfile_path)?; |
35cf5daa | 164 | |
b8d4766a | 165 | Ok(ChunkStore { |
277fc5a3 | 166 | name: name.to_owned(), |
b8d4766a DM |
167 | base, |
168 | chunk_dir, | |
43b13033 | 169 | locker, |
b8d4766a DM |
170 | mutex: Mutex::new(false) |
171 | }) | |
35cf5daa DM |
172 | } |
173 | ||
a660978c | 174 | pub fn touch_chunk(&self, digest: &[u8; 32]) -> Result<(), Error> { |
3d5c11e5 | 175 | |
a660978c | 176 | let (chunk_path, _digest_str) = self.chunk_path(digest); |
3d5c11e5 | 177 | |
7ee2aa1b DM |
178 | const UTIME_NOW: i64 = ((1 << 30) - 1); |
179 | const UTIME_OMIT: i64 = ((1 << 30) - 2); | |
180 | ||
a198d74f | 181 | let times: [libc::timespec; 2] = [ |
7ee2aa1b DM |
182 | libc::timespec { tv_sec: 0, tv_nsec: UTIME_NOW }, |
183 | libc::timespec { tv_sec: 0, tv_nsec: UTIME_OMIT } | |
184 | ]; | |
185 | ||
186 | use nix::NixPath; | |
187 | ||
188 | let res = chunk_path.with_nix_path(|cstr| unsafe { | |
189 | libc::utimensat(-1, cstr.as_ptr(), ×[0], libc::AT_SYMLINK_NOFOLLOW) | |
190 | })?; | |
191 | ||
192 | if let Err(err) = nix::errno::Errno::result(res) { | |
193 | bail!("updata atime failed for chunk {:?} - {}", chunk_path, err); | |
194 | } | |
195 | ||
3d5c11e5 DM |
196 | Ok(()) |
197 | } | |
198 | ||
4ee8f53d | 199 | pub fn read_chunk(&self, digest: &[u8; 32]) -> Result<DataBlob, Error> { |
96df2fb4 | 200 | |
81a6ce6f | 201 | let (chunk_path, digest_str) = self.chunk_path(digest); |
f98ac774 | 202 | let mut file = std::fs::File::open(&chunk_path) |
a24e3993 DM |
203 | .map_err(|err| { |
204 | format_err!( | |
205 | "store '{}', unable to read chunk '{}' - {}", | |
206 | self.name, | |
207 | digest_str, | |
208 | err, | |
209 | ) | |
210 | })?; | |
96df2fb4 | 211 | |
4ee8f53d | 212 | DataBlob::load(&mut file) |
96df2fb4 DM |
213 | } |
214 | ||
9739aca4 WB |
215 | pub fn get_chunk_iterator( |
216 | &self, | |
217 | ) -> Result< | |
a5736098 | 218 | impl Iterator<Item = (Result<tools::fs::ReadDirEntry, Error>, usize)> + std::iter::FusedIterator, |
9739aca4 WB |
219 | Error |
220 | > { | |
221 | use nix::dir::Dir; | |
222 | use nix::fcntl::OFlag; | |
223 | use nix::sys::stat::Mode; | |
224 | ||
a24e3993 DM |
225 | let base_handle = Dir::open(&self.chunk_dir, OFlag::O_RDONLY, Mode::empty()) |
226 | .map_err(|err| { | |
227 | format_err!( | |
228 | "unable to open store '{}' chunk dir {:?} - {}", | |
229 | self.name, | |
230 | self.chunk_dir, | |
231 | err, | |
232 | ) | |
233 | })?; | |
9739aca4 | 234 | |
a3f3e91d WB |
235 | let mut done = false; |
236 | let mut inner: Option<tools::fs::ReadDir> = None; | |
237 | let mut at = 0; | |
238 | let mut percentage = 0; | |
239 | Ok(std::iter::from_fn(move || { | |
240 | if done { | |
241 | return None; | |
242 | } | |
243 | ||
244 | loop { | |
245 | if let Some(ref mut inner) = inner { | |
246 | match inner.next() { | |
247 | Some(Ok(entry)) => { | |
248 | // skip files if they're not a hash | |
249 | let bytes = entry.file_name().to_bytes(); | |
250 | if bytes.len() != 64 { | |
251 | continue; | |
252 | } | |
253 | if !bytes.iter().all(u8::is_ascii_hexdigit) { | |
254 | continue; | |
255 | } | |
256 | return Some((Ok(entry), percentage)); | |
257 | } | |
258 | Some(Err(err)) => { | |
259 | // stop after first error | |
260 | done = true; | |
261 | // and pass the error through: | |
262 | return Some((Err(err), percentage)); | |
263 | } | |
264 | None => (), // open next directory | |
c7f481b6 | 265 | } |
c7f481b6 | 266 | } |
a3f3e91d WB |
267 | |
268 | inner = None; | |
269 | ||
270 | if at == 0x10000 { | |
271 | done = true; | |
272 | return None; | |
273 | } | |
274 | ||
275 | let subdir: &str = &format!("{:04x}", at); | |
276 | percentage = (at * 100) / 0x10000; | |
277 | at += 1; | |
278 | match tools::fs::read_subdir(base_handle.as_raw_fd(), subdir) { | |
279 | Ok(dir) => { | |
280 | inner = Some(dir); | |
281 | // start reading: | |
282 | continue; | |
283 | } | |
284 | Err(ref err) if err.as_errno() == Some(nix::errno::Errno::ENOENT) => { | |
285 | // non-existing directories are okay, just keep going: | |
286 | continue; | |
287 | } | |
288 | Err(err) => { | |
289 | // other errors are fatal, so end our iteration | |
290 | done = true; | |
291 | // and pass the error through: | |
9850bcdf | 292 | return Some((Err(format_err!("unable to read subdir '{}' - {}", subdir, err)), percentage)); |
a3f3e91d | 293 | } |
62f2422f WB |
294 | } |
295 | } | |
a3f3e91d | 296 | }).fuse()) |
c7f481b6 WB |
297 | } |
298 | ||
11861a48 DM |
299 | pub fn oldest_writer(&self) -> Option<i64> { |
300 | tools::ProcessLocker::oldest_shared_lock(self.locker.clone()) | |
301 | } | |
302 | ||
303 | pub fn sweep_unused_chunks( | |
304 | &self, | |
305 | oldest_writer: Option<i64>, | |
a5736098 DM |
306 | status: &mut GarbageCollectionStatus, |
307 | worker: Arc<WorkerTask>, | |
11861a48 | 308 | ) -> Result<(), Error> { |
9349d2a1 | 309 | use nix::sys::stat::fstatat; |
08481a0b | 310 | |
fdd71f52 | 311 | let now = unsafe { libc::time(std::ptr::null_mut()) }; |
c7f481b6 | 312 | |
11861a48 DM |
313 | let mut min_atime = now - 3600*24; // at least 24h (see mount option relatime) |
314 | ||
315 | if let Some(stamp) = oldest_writer { | |
316 | if stamp < min_atime { | |
317 | min_atime = stamp; | |
318 | } | |
319 | } | |
320 | ||
321 | min_atime -= 300; // add 5 mins gap for safety | |
322 | ||
a5736098 | 323 | let mut last_percentage = 0; |
e4c2fbf1 DM |
324 | let mut chunk_count = 0; |
325 | ||
a5736098 DM |
326 | for (entry, percentage) in self.get_chunk_iterator()? { |
327 | if last_percentage != percentage { | |
328 | last_percentage = percentage; | |
e4c2fbf1 | 329 | worker.log(format!("percentage done: {}, chunk count: {}", percentage, chunk_count)); |
a5736098 | 330 | } |
92da93b2 DM |
331 | |
332 | tools::fail_on_shutdown()?; | |
333 | ||
fdd71f52 | 334 | let (dirfd, entry) = match entry { |
c7f481b6 | 335 | Ok(entry) => (entry.parent_fd(), entry), |
9850bcdf | 336 | Err(err) => bail!("chunk iterator on chunk store '{}' failed - {}", self.name, err), |
eae8aa3a | 337 | }; |
fdd71f52 | 338 | |
eae8aa3a DM |
339 | let file_type = match entry.file_type() { |
340 | Some(file_type) => file_type, | |
277fc5a3 | 341 | None => bail!("unsupported file system type on chunk store '{}'", self.name), |
eae8aa3a | 342 | }; |
82bc0ad4 WB |
343 | if file_type != nix::dir::Type::File { |
344 | continue; | |
345 | } | |
eae8aa3a | 346 | |
e4c2fbf1 DM |
347 | chunk_count += 1; |
348 | ||
d85987ae DM |
349 | let filename = entry.file_name(); |
350 | ||
15a77c4c DM |
351 | let lock = self.mutex.lock(); |
352 | ||
9349d2a1 | 353 | if let Ok(stat) = fstatat(dirfd, filename, nix::fcntl::AtFlags::AT_SYMLINK_NOFOLLOW) { |
eae8aa3a | 354 | let age = now - stat.st_atime; |
e95950e4 | 355 | //println!("FOUND {} {:?}", age/(3600*24), filename); |
11861a48 | 356 | if stat.st_atime < min_atime { |
eae8aa3a | 357 | println!("UNLINK {} {:?}", age/(3600*24), filename); |
fdd71f52 | 358 | let res = unsafe { libc::unlinkat(dirfd, filename.as_ptr(), 0) }; |
277fc5a3 DM |
359 | if res != 0 { |
360 | let err = nix::Error::last(); | |
9349d2a1 WB |
361 | bail!( |
362 | "unlink chunk {:?} failed on store '{}' - {}", | |
363 | filename, | |
364 | self.name, | |
365 | err, | |
366 | ); | |
277fc5a3 | 367 | } |
a660978c DM |
368 | status.removed_chunks += 1; |
369 | status.removed_bytes += stat.st_size as u64; | |
370 | } else { | |
64e53b28 | 371 | status.disk_chunks += 1; |
a660978c | 372 | status.disk_bytes += stat.st_size as u64; |
08481a0b | 373 | } |
eae8aa3a | 374 | } |
15a77c4c | 375 | drop(lock); |
eae8aa3a | 376 | } |
a5736098 | 377 | |
277fc5a3 | 378 | Ok(()) |
08481a0b DM |
379 | } |
380 | ||
f98ac774 | 381 | pub fn insert_chunk( |
9ac6ec86 | 382 | &self, |
4ee8f53d DM |
383 | chunk: &DataBlob, |
384 | digest: &[u8; 32], | |
9ac6ec86 | 385 | ) -> Result<(bool, u64), Error> { |
7b2b40a8 | 386 | |
bffd40d6 | 387 | //println!("DIGEST {}", proxmox::tools::digest_to_hex(digest)); |
128b37fe | 388 | |
81a6ce6f | 389 | let (chunk_path, digest_str) = self.chunk_path(digest); |
c5d82e5f DM |
390 | |
391 | let lock = self.mutex.lock(); | |
392 | ||
393 | if let Ok(metadata) = std::fs::metadata(&chunk_path) { | |
394 | if metadata.is_file() { | |
9ac6ec86 | 395 | return Ok((true, metadata.len())); |
c5d82e5f | 396 | } else { |
f7dd683b | 397 | bail!("Got unexpected file type on store '{}' for chunk {}", self.name, digest_str); |
c5d82e5f DM |
398 | } |
399 | } | |
128b37fe | 400 | |
c5d82e5f DM |
401 | let mut tmp_path = chunk_path.clone(); |
402 | tmp_path.set_extension("tmp"); | |
78216a5a | 403 | |
f98ac774 | 404 | let mut file = std::fs::File::create(&tmp_path)?; |
78216a5a | 405 | |
f98ac774 DM |
406 | let raw_data = chunk.raw_data(); |
407 | let encoded_size = raw_data.len() as u64; | |
78216a5a | 408 | |
f98ac774 | 409 | file.write_all(raw_data)?; |
128b37fe | 410 | |
c5d82e5f DM |
411 | if let Err(err) = std::fs::rename(&tmp_path, &chunk_path) { |
412 | if let Err(_) = std::fs::remove_file(&tmp_path) { /* ignore */ } | |
9349d2a1 WB |
413 | bail!( |
414 | "Atomic rename on store '{}' failed for chunk {} - {}", | |
415 | self.name, | |
416 | digest_str, | |
417 | err, | |
418 | ); | |
c5d82e5f DM |
419 | } |
420 | ||
c5d82e5f DM |
421 | drop(lock); |
422 | ||
f98ac774 | 423 | Ok((false, encoded_size)) |
128b37fe DM |
424 | } |
425 | ||
81a6ce6f DM |
426 | pub fn chunk_path(&self, digest:&[u8; 32]) -> (PathBuf, String) { |
427 | let mut chunk_path = self.chunk_dir.clone(); | |
428 | let prefix = digest_to_prefix(digest); | |
429 | chunk_path.push(&prefix); | |
430 | let digest_str = proxmox::tools::digest_to_hex(digest); | |
431 | chunk_path.push(&digest_str); | |
432 | (chunk_path, digest_str) | |
433 | } | |
434 | ||
606ce64b DM |
435 | pub fn relative_path(&self, path: &Path) -> PathBuf { |
436 | ||
437 | let mut full_path = self.base.clone(); | |
438 | full_path.push(path); | |
439 | full_path | |
440 | } | |
441 | ||
3d5c11e5 DM |
442 | pub fn base_path(&self) -> PathBuf { |
443 | self.base.clone() | |
444 | } | |
43b13033 DM |
445 | |
446 | pub fn try_shared_lock(&self) -> Result<tools::ProcessLockSharedGuard, Error> { | |
447 | tools::ProcessLocker::try_shared_lock(self.locker.clone()) | |
448 | } | |
449 | ||
450 | pub fn try_exclusive_lock(&self) -> Result<tools::ProcessLockExclusiveGuard, Error> { | |
451 | tools::ProcessLocker::try_exclusive_lock(self.locker.clone()) | |
452 | } | |
35cf5daa DM |
453 | } |
454 | ||
455 | ||
456 | #[test] | |
457 | fn test_chunk_store1() { | |
458 | ||
332dcc22 DM |
459 | let mut path = std::fs::canonicalize(".").unwrap(); // we need absulute path |
460 | path.push(".testdir"); | |
461 | ||
35cf5daa DM |
462 | if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ } |
463 | ||
332dcc22 | 464 | let chunk_store = ChunkStore::open("test", &path); |
35cf5daa DM |
465 | assert!(chunk_store.is_err()); |
466 | ||
332dcc22 | 467 | let chunk_store = ChunkStore::create("test", &path).unwrap(); |
f98ac774 | 468 | |
4ee8f53d | 469 | let (chunk, digest) = super::DataChunkBuilder::new(&[0u8, 1u8]).build().unwrap(); |
f98ac774 | 470 | |
4ee8f53d | 471 | let (exists, _) = chunk_store.insert_chunk(&chunk, &digest).unwrap(); |
391a2e43 DM |
472 | assert!(!exists); |
473 | ||
4ee8f53d | 474 | let (exists, _) = chunk_store.insert_chunk(&chunk, &digest).unwrap(); |
391a2e43 | 475 | assert!(exists); |
128b37fe | 476 | |
35cf5daa | 477 | |
332dcc22 | 478 | let chunk_store = ChunkStore::create("test", &path); |
35cf5daa DM |
479 | assert!(chunk_store.is_err()); |
480 | ||
e0a5d1ca | 481 | if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ } |
35cf5daa | 482 | } |