pbs-datastore/src/chunk_store.rs

   1 use std::io::Write;
   2 use std::os::unix::io::AsRawFd;
   3 use std::path::{Path, PathBuf};
   4 use std::sync::{Arc, Mutex};
   5
   6 use anyhow::{bail, format_err, Error};
   7
   8 use proxmox::tools::fs::{CreateOptions, create_path, create_dir};
   9 use proxmox_sys::process_locker::{ProcessLocker, ProcessLockSharedGuard, ProcessLockExclusiveGuard};
  10 use proxmox_sys::worker_task_context::WorkerTaskContext;
  11 use proxmox_sys::task_log;
  12 use pbs_api_types::GarbageCollectionStatus;
  13
  14
  15 use crate::DataBlob;
  16
  17 /// File system based chunk store
  18 pub struct ChunkStore {
  19     name: String, // used for error reporting
  20     pub (crate) base: PathBuf,
  21     chunk_dir: PathBuf,
  22     mutex: Mutex<()>,
  23     locker: Arc<Mutex<ProcessLocker>>,
  24 }
  25
  26 // TODO: what about sysctl setting vm.vfs_cache_pressure (0 - 100) ?
  27
  28 pub fn verify_chunk_size(size: usize) -> Result<(), Error> {
  29
  30     static SIZES: [usize; 7] = [64*1024, 128*1024, 256*1024, 512*1024, 1024*1024, 2048*1024, 4096*1024];
  31
  32     if !SIZES.contains(&size) {
  33         bail!("Got unsupported chunk size '{}'", size);
  34     }
  35     Ok(())
  36 }
  37
  38 fn digest_to_prefix(digest: &[u8]) -> PathBuf {
  39
  40     let mut buf = Vec::<u8>::with_capacity(2+1+2+1);
  41
  42     const HEX_CHARS: &[u8; 16] = b"0123456789abcdef";
  43
  44     buf.push(HEX_CHARS[(digest[0] as usize) >> 4]);
  45     buf.push(HEX_CHARS[(digest[0] as usize) &0xf]);
  46     buf.push(HEX_CHARS[(digest[1] as usize) >> 4]);
  47     buf.push(HEX_CHARS[(digest[1] as usize) & 0xf]);
  48     buf.push(b'/');
  49
  50     let path = unsafe { String::from_utf8_unchecked(buf)};
  51
  52     path.into()
  53 }
  54
  55 impl ChunkStore {
  56
  57     fn chunk_dir<P: AsRef<Path>>(path: P) -> PathBuf {
  58
  59         let mut chunk_dir: PathBuf = PathBuf::from(path.as_ref());
  60         chunk_dir.push(".chunks");
  61
  62         chunk_dir
  63     }
  64
  65     pub fn base(&self) -> &Path {
  66         &self.base
  67     }
  68
  69     pub fn create<P>(name: &str, path: P, uid: nix::unistd::Uid, gid: nix::unistd::Gid, worker: Option<&dyn WorkerTaskContext>) -> Result<Self, Error>
  70     where
  71         P: Into<PathBuf>,
  72     {
  73
  74         let base: PathBuf = path.into();
  75
  76         if !base.is_absolute() {
  77             bail!("expected absolute path - got {:?}", base);
  78         }
  79
  80         let chunk_dir = Self::chunk_dir(&base);
  81
  82         let options = CreateOptions::new()
  83             .owner(uid)
  84             .group(gid);
  85
  86         let default_options = CreateOptions::new();
  87
  88         match create_path(&base, Some(default_options), Some(options.clone())) {
  89             Err(err) => bail!("unable to create chunk store '{}' at {:?} - {}", name, base, err),
  90             Ok(res) => if ! res  { nix::unistd::chown(&base, Some(uid), Some(gid))? },
  91         }
  92
  93         if let Err(err) = create_dir(&chunk_dir, options.clone()) {
  94             bail!("unable to create chunk store '{}' subdir {:?} - {}", name, chunk_dir, err);
  95         }
  96
  97         // create lock file with correct owner/group
  98         let lockfile_path = Self::lockfile_path(&base);
  99         proxmox::tools::fs::replace_file(lockfile_path, b"", options.clone(), false)?;
 100
 101         // create 64*1024 subdirs
 102         let mut last_percentage = 0;
 103
 104         for i in 0..64*1024 {
 105             let mut l1path = chunk_dir.clone();
 106             l1path.push(format!("{:04x}", i));
 107             if let Err(err) = create_dir(&l1path, options.clone()) {
 108                 bail!("unable to create chunk store '{}' subdir {:?} - {}", name, l1path, err);
 109             }
 110             let percentage = (i*100)/(64*1024);
 111             if percentage != last_percentage {
 112                 if let Some(worker) = worker {
 113                     task_log!(worker, "Chunkstore create: {}%", percentage)
 114                 }
 115                 last_percentage = percentage;
 116             }
 117         }
 118
 119         Self::open(name, base)
 120     }
 121
 122     fn lockfile_path<P: Into<PathBuf>>(base: P) -> PathBuf {
 123         let mut lockfile_path: PathBuf = base.into();
 124
 125         lockfile_path.push(".lock");
 126
 127         lockfile_path
 128     }
 129
 130     pub fn open<P: Into<PathBuf>>(name: &str, base: P) -> Result<Self, Error> {
 131
 132         let base: PathBuf = base.into();
 133
 134         if !base.is_absolute() {
 135             bail!("expected absolute path - got {:?}", base);
 136         }
 137
 138         let chunk_dir = Self::chunk_dir(&base);
 139
 140         if let Err(err) = std::fs::metadata(&chunk_dir) {
 141             bail!("unable to open chunk store '{}' at {:?} - {}", name, chunk_dir, err);
 142         }
 143
 144         let lockfile_path = Self::lockfile_path(&base);
 145
 146         let locker = ProcessLocker::new(&lockfile_path)?;
 147
 148         Ok(ChunkStore {
 149             name: name.to_owned(),
 150             base,
 151             chunk_dir,
 152             locker,
 153             mutex: Mutex::new(())
 154         })
 155     }
 156
 157     pub fn touch_chunk(&self, digest: &[u8; 32]) -> Result<(), Error> {
 158         self.cond_touch_chunk(digest, true)?;
 159         Ok(())
 160     }
 161
 162     pub fn cond_touch_chunk(&self, digest: &[u8; 32], fail_if_not_exist: bool) -> Result<bool, Error> {
 163         let (chunk_path, _digest_str) = self.chunk_path(digest);
 164         self.cond_touch_path(&chunk_path, fail_if_not_exist)
 165     }
 166
 167     pub fn cond_touch_path(&self, path: &Path, fail_if_not_exist: bool) -> Result<bool, Error> {
 168         const UTIME_NOW: i64 = (1 << 30) - 1;
 169         const UTIME_OMIT: i64 = (1 << 30) - 2;
 170
 171         let times: [libc::timespec; 2] = [
 172             libc::timespec { tv_sec: 0, tv_nsec: UTIME_NOW },
 173             libc::timespec { tv_sec: 0, tv_nsec: UTIME_OMIT }
 174         ];
 175
 176         use nix::NixPath;
 177
 178         let res = path.with_nix_path(|cstr| unsafe {
 179             let tmp = libc::utimensat(-1, cstr.as_ptr(), &times[0], libc::AT_SYMLINK_NOFOLLOW);
 180             nix::errno::Errno::result(tmp)
 181         })?;
 182
 183         if let Err(err) = res {
 184             if !fail_if_not_exist && err.as_errno() == Some(nix::errno::Errno::ENOENT) {
 185                 return Ok(false);
 186             }
 187
 188             bail!("update atime failed for chunk/file {:?} - {}", path, err);
 189         }
 190
 191         Ok(true)
 192     }
 193
 194     pub fn get_chunk_iterator(
 195         &self,
 196     ) -> Result<
 197         impl Iterator<Item = (Result<pbs_tools::fs::ReadDirEntry, Error>, usize, bool)> + std::iter::FusedIterator,
 198         Error
 199     > {
 200         use nix::dir::Dir;
 201         use nix::fcntl::OFlag;
 202         use nix::sys::stat::Mode;
 203
 204         let base_handle = Dir::open(&self.chunk_dir, OFlag::O_RDONLY, Mode::empty())
 205             .map_err(|err| {
 206                 format_err!(
 207                     "unable to open store '{}' chunk dir {:?} - {}",
 208                     self.name,
 209                     self.chunk_dir,
 210                     err,
 211                 )
 212             })?;
 213
 214         let mut done = false;
 215         let mut inner: Option<pbs_tools::fs::ReadDir> = None;
 216         let mut at = 0;
 217         let mut percentage = 0;
 218         Ok(std::iter::from_fn(move || {
 219             if done {
 220                 return None;
 221             }
 222
 223             loop {
 224                 if let Some(ref mut inner) = inner {
 225                     match inner.next() {
 226                         Some(Ok(entry)) => {
 227                             // skip files if they're not a hash
 228                             let bytes = entry.file_name().to_bytes();
 229                             if bytes.len() != 64 && bytes.len() != 64 + ".0.bad".len() {
 230                                 continue;
 231                             }
 232                             if !bytes.iter().take(64).all(u8::is_ascii_hexdigit) {
 233                                 continue;
 234                             }
 235
 236                             let bad = bytes.ends_with(b".bad");
 237                             return Some((Ok(entry), percentage, bad));
 238                         }
 239                         Some(Err(err)) => {
 240                             // stop after first error
 241                             done = true;
 242                             // and pass the error through:
 243                             return Some((Err(err), percentage, false));
 244                         }
 245                         None => (), // open next directory
 246                     }
 247                 }
 248
 249                 inner = None;
 250
 251                 if at == 0x10000 {
 252                     done = true;
 253                     return None;
 254                 }
 255
 256                 let subdir: &str = &format!("{:04x}", at);
 257                 percentage = (at * 100) / 0x10000;
 258                 at += 1;
 259                 match pbs_tools::fs::read_subdir(base_handle.as_raw_fd(), subdir) {
 260                     Ok(dir) => {
 261                         inner = Some(dir);
 262                         // start reading:
 263                         continue;
 264                     }
 265                     Err(ref err) if err.as_errno() == Some(nix::errno::Errno::ENOENT) => {
 266                         // non-existing directories are okay, just keep going:
 267                         continue;
 268                     }
 269                     Err(err) => {
 270                         // other errors are fatal, so end our iteration
 271                         done = true;
 272                         // and pass the error through:
 273                         return Some((Err(format_err!("unable to read subdir '{}' - {}", subdir, err)), percentage, false));
 274                     }
 275                 }
 276             }
 277         }).fuse())
 278     }
 279
 280     pub fn oldest_writer(&self) -> Option<i64> {
 281         ProcessLocker::oldest_shared_lock(self.locker.clone())
 282     }
 283
 284     pub fn sweep_unused_chunks(
 285         &self,
 286         oldest_writer: i64,
 287         phase1_start_time: i64,
 288         status: &mut GarbageCollectionStatus,
 289         worker: &dyn WorkerTaskContext,
 290     ) -> Result<(), Error> {
 291         use nix::sys::stat::fstatat;
 292         use nix::unistd::{unlinkat, UnlinkatFlags};
 293
 294         let mut min_atime = phase1_start_time - 3600*24; // at least 24h (see mount option relatime)
 295
 296         if oldest_writer < min_atime {
 297             min_atime = oldest_writer;
 298         }
 299
 300         min_atime -= 300; // add 5 mins gap for safety
 301
 302         let mut last_percentage = 0;
 303         let mut chunk_count = 0;
 304
 305         for (entry, percentage, bad) in self.get_chunk_iterator()? {
 306             if last_percentage != percentage {
 307                 last_percentage = percentage;
 308                 task_log!(
 309                     worker,
 310                     "processed {}% ({} chunks)",
 311                     percentage,
 312                     chunk_count,
 313                 );
 314             }
 315
 316             worker.check_abort()?;
 317             worker.fail_on_shutdown()?;
 318
 319             let (dirfd, entry) = match entry {
 320                 Ok(entry) => (entry.parent_fd(), entry),
 321                 Err(err) => bail!("chunk iterator on chunk store '{}' failed - {}", self.name, err),
 322             };
 323
 324             let file_type = match entry.file_type() {
 325                 Some(file_type) => file_type,
 326                 None => bail!("unsupported file system type on chunk store '{}'", self.name),
 327             };
 328             if file_type != nix::dir::Type::File {
 329                 continue;
 330             }
 331
 332             chunk_count += 1;
 333
 334             let filename = entry.file_name();
 335
 336             let lock = self.mutex.lock();
 337
 338             if let Ok(stat) = fstatat(dirfd, filename, nix::fcntl::AtFlags::AT_SYMLINK_NOFOLLOW) {
 339                 if stat.st_atime < min_atime {
 340                     //let age = now - stat.st_atime;
 341                     //println!("UNLINK {}  {:?}", age/(3600*24), filename);
 342                     if let Err(err) = unlinkat(Some(dirfd), filename, UnlinkatFlags::NoRemoveDir) {
 343                         if bad {
 344                             status.still_bad += 1;
 345                         }
 346                         bail!(
 347                             "unlinking chunk {:?} failed on store '{}' - {}",
 348                             filename,
 349                             self.name,
 350                             err,
 351                         );
 352                     }
 353                     if bad {
 354                         status.removed_bad += 1;
 355                     } else {
 356                         status.removed_chunks += 1;
 357                     }
 358                     status.removed_bytes += stat.st_size as u64;
 359                 } else if stat.st_atime < oldest_writer {
 360                     if bad {
 361                         status.still_bad += 1;
 362                     } else {
 363                         status.pending_chunks += 1;
 364                     }
 365                     status.pending_bytes += stat.st_size as u64;
 366                 } else {
 367                     if !bad {
 368                         status.disk_chunks += 1;
 369                     }
 370                     status.disk_bytes += stat.st_size as u64;
 371                 }
 372             }
 373             drop(lock);
 374         }
 375
 376         Ok(())
 377     }
 378
 379     pub fn insert_chunk(
 380         &self,
 381         chunk: &DataBlob,
 382         digest: &[u8; 32],
 383     ) -> Result<(bool, u64), Error> {
 384
 385         //println!("DIGEST {}", proxmox::tools::digest_to_hex(digest));
 386
 387         let (chunk_path, digest_str) = self.chunk_path(digest);
 388
 389         let lock = self.mutex.lock();
 390
 391         if let Ok(metadata) = std::fs::metadata(&chunk_path) {
 392             if metadata.is_file() {
 393                 self.touch_chunk(digest)?;
 394                 return Ok((true, metadata.len()));
 395             } else {
 396                 bail!("Got unexpected file type on store '{}' for chunk {}", self.name, digest_str);
 397             }
 398         }
 399
 400         let mut tmp_path = chunk_path.clone();
 401         tmp_path.set_extension("tmp");
 402
 403         let mut file = std::fs::File::create(&tmp_path).map_err(|err| {
 404             format_err!(
 405                 "creating temporary chunk on store '{}' failed for {} - {}",
 406                 self.name,
 407                 digest_str,
 408                 err
 409             )
 410         })?;
 411
 412         let raw_data = chunk.raw_data();
 413         let encoded_size = raw_data.len() as u64;
 414
 415         file.write_all(raw_data).map_err(|err| {
 416             format_err!(
 417                 "writing temporary chunk on store '{}' failed for {} - {}",
 418                 self.name,
 419                 digest_str,
 420                 err
 421             )
 422         })?;
 423
 424         if let Err(err) = std::fs::rename(&tmp_path, &chunk_path) {
 425             if std::fs::remove_file(&tmp_path).is_err()  { /* ignore */ }
 426             bail!(
 427                 "Atomic rename on store '{}' failed for chunk {} - {}",
 428                 self.name,
 429                 digest_str,
 430                 err,
 431             );
 432         }
 433
 434         drop(lock);
 435
 436         Ok((false, encoded_size))
 437     }
 438
 439     pub fn chunk_path(&self, digest:&[u8; 32]) -> (PathBuf, String) {
 440         let mut chunk_path = self.chunk_dir.clone();
 441         let prefix = digest_to_prefix(digest);
 442         chunk_path.push(&prefix);
 443         let digest_str = proxmox::tools::digest_to_hex(digest);
 444         chunk_path.push(&digest_str);
 445         (chunk_path, digest_str)
 446     }
 447
 448     pub fn relative_path(&self, path: &Path) -> PathBuf {
 449
 450         let mut full_path = self.base.clone();
 451         full_path.push(path);
 452         full_path
 453     }
 454
 455     pub fn name(&self) -> &str {
 456         &self.name
 457     }
 458
 459     pub fn base_path(&self) -> PathBuf {
 460         self.base.clone()
 461     }
 462
 463     pub fn try_shared_lock(&self) -> Result<ProcessLockSharedGuard, Error> {
 464         ProcessLocker::try_shared_lock(self.locker.clone())
 465     }
 466
 467     pub fn try_exclusive_lock(&self) -> Result<ProcessLockExclusiveGuard, Error> {
 468         ProcessLocker::try_exclusive_lock(self.locker.clone())
 469     }
 470 }
 471
 472
 473 #[test]
 474 fn test_chunk_store1() {
 475
 476     let mut path = std::fs::canonicalize(".").unwrap(); // we need absolute path
 477     path.push(".testdir");
 478
 479     if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ }
 480
 481     let chunk_store = ChunkStore::open("test", &path);
 482     assert!(chunk_store.is_err());
 483
 484     let user = nix::unistd::User::from_uid(nix::unistd::Uid::current()).unwrap().unwrap();
 485     let chunk_store = ChunkStore::create("test", &path, user.uid, user.gid, None).unwrap();
 486
 487     let (chunk, digest) = crate::data_blob::DataChunkBuilder::new(&[0u8, 1u8]).build().unwrap();
 488
 489     let (exists, _) = chunk_store.insert_chunk(&chunk, &digest).unwrap();
 490     assert!(!exists);
 491
 492     let (exists, _) = chunk_store.insert_chunk(&chunk, &digest).unwrap();
 493     assert!(exists);
 494
 495
 496     let chunk_store = ChunkStore::create("test", &path, user.uid, user.gid, None);
 497     assert!(chunk_store.is_err());
 498
 499     if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ }
 500 }