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