marble/

lib.rs

//! # Marble
//!
//! Marble is a low-level object store that can be used
//! to build your own storage engines and databases on
//! top of.
//!
//! At a high-level, it supports atomic batch writes and
//! single-object reads. Garbage collection is manual.
//! All operations are blocking. Nothing is cached
//! in-memory. Objects may be sharded upon GC by providing
//! a custom `Config::partition_function`. Partitioning
//! is not performed on the write batch when it
//! is initially written, because the write batch
//! must be stored in a single file for it to be atomic.
//! But when future calls to `Marble::maintenance`
//! defragment the storage files by rewriting objects
//! that are still live, it will use this function
//! to assign the rewritten objects into a particular
//! partition.
//!
//! You should think of Marble as the heap that you
//! flush your write-ahead logs into periodically.
//! It will create a new file for each write batch,
//! and this might actually expand to more files after
//! garbage collection if the batch is significantly
//! larger than the `Config::target_file_size`.
//!
//! Marble does not create any threads or call
//! `Marble::maintenance` automatically under any
//! conditions. You should probably create a background
//! thread that calls this periodically.
//!
//! # Examples
//!
//! ```
//! let marble = marble::open("heap").unwrap();
//!
//! // Write new data keyed by a `u64` object ID.
//! // Batches contain insertions and deletions
//! // based on whether the value is a Some or None.
//! marble.write_batch([
//!     (0_u64, Some(&[32_u8] as &[u8])),
//!     (4_u64, None),
//! ]).unwrap();
//!
//! // read it back
//! assert_eq!(marble.read(0).unwrap(), Some(vec![32].into_boxed_slice()));
//! assert_eq!(marble.read(4).unwrap(), None);
//! assert_eq!(marble.read(6).unwrap(), None);
//!
//! // after a few more batches that may have caused fragmentation
//! // by overwriting previous objects, perform maintenance which
//! // will defragment the object store based on `Config` settings.
//! let objects_defragmented = marble.maintenance().unwrap();
//!
//! // print out system statistics
//! dbg!(marble.stats());
//! # drop(marble);
//! # std::fs::remove_dir_all("heap").unwrap();
//! ```
//!
//! which prints out something like
//! ```txt,no_run
//! marble.stats() = Stats {
//!     live_objects: 1048576,
//!     stored_objects: 1181100,
//!     dead_objects: 132524,
//!     live_percent: 88,
//!     files: 11,
//! }
//! ```
//!
//! If you want to customize the settings passed to Marble,
//! you may specify your own `Config`:
//!
//! ```
//! let config = marble::Config {
//!     path: "my_path".into(),
//!     fsync_each_batch: true,
//!     target_file_size: 64 * 1024 * 1024,
//!     file_compaction_percent: 50,
//!     ..Default::default()
//! };
//!
//! let marble = config.open().unwrap();
//! # drop(marble);
//! # std::fs::remove_dir_all("my_path").unwrap();
//! ```
//!
//! A custom GC sharding function may be provided
//! for partitioning objects based on the object ID
//! and size. This may be useful if your higher-level
//! system allocates certain ranges of object IDs for
//! certain types of objects that you would like to
//! group together in the hope of grouping items together
//! that have similar fragmentation properties (similar
//! expected lifespan etc...). This will only shard
//! objects when they are defragmented through the
//! `Marble::maintenance` method, because each new
//! write batch must be written together in one
//! file to retain write batch atomicity in the
//! face of crashes.
//!
//! ```
//! // This function shards objects into partitions
//! // similarly to a slab allocator that groups objects
//! // into size buckets based on powers of two.
//! fn shard_by_size(object_id: u64, object_size: usize) -> u8 {
//!     let next_po2 = object_size.next_power_of_two();
//!     u8::try_from(next_po2.trailing_zeros()).unwrap()
//! }
//!
//! let config = marble::Config {
//!     path: "my_sharded_path".into(),
//!     partition_function: shard_by_size,
//!     ..Default::default()
//! };
//!
//! let marble = config.open().unwrap();
//! # drop(marble);
//! # std::fs::remove_dir_all("my_sharded_path").unwrap();
//! ```
use std::fs::File;
use std::io;
use std::path::{Path, PathBuf};
use std::sync::{
    atomic::{
        AtomicBool, AtomicPtr, AtomicU64,
        Ordering::{Acquire, SeqCst},
    },
    Arc,
};

use fault_injection::fallible;

#[derive(Clone, Copy)]
pub struct LocationHasher(u64);

impl Default for LocationHasher {
    #[inline]
    fn default() -> LocationHasher {
        LocationHasher(0)
    }
}

impl std::hash::Hasher for LocationHasher {
    #[inline]
    fn finish(&self) -> u64 {
        self.0
    }

    #[inline]
    fn write_u8(&mut self, n: u8) {
        self.0 = u64::from(n);
    }

    #[inline]
    fn write_u64(&mut self, n: u64) {
        self.0 = n;
    }

    #[inline]
    fn write(&mut self, _: &[u8]) {
        panic!("trying to use LocationHasher with incorrect type");
    }
}

type Map<K, V> = std::collections::HashMap<K, V, std::hash::BuildHasherDefault<LocationHasher>>;

mod config;
mod debug_delay;
#[cfg(feature = "runtime_validation")]
mod debug_history;
mod disk_location;
mod file_map;
mod gc;
mod location_table;
mod readpath;
mod recovery;
mod trailer;
mod writepath;

pub use config::Config;
use debug_delay::debug_delay;
use disk_location::{DiskLocation, RelativeDiskLocation};
use file_map::FileMap;
use location_table::LocationTable;
use trailer::{read_trailer, read_trailer_from_buf, write_trailer};

const HEADER_LEN: usize = 20;
const NEW_WRITE_BATCH_BIT: u64 = 1 << 62;
const NEW_WRITE_BATCH_MASK: u64 = u64::MAX - NEW_WRITE_BATCH_BIT;

type ObjectId = u64;

fn read_range_at(file: &File, start: u64, end: u64) -> io::Result<Vec<u8>> {
    use std::os::unix::fs::FileExt;

    let buf_sz: usize = (end - start).try_into().unwrap();

    let mut buf = Vec::with_capacity(buf_sz);

    unsafe {
        buf.set_len(buf_sz);
    }

    fallible!(file.read_exact_at(&mut buf, start));

    Ok(buf)
}

fn uninit_boxed_slice(len: usize) -> Box<[u8]> {
    use std::alloc::{alloc, Layout};

    let layout = Layout::array::<u8>(len).unwrap();

    unsafe {
        let ptr = alloc(layout);
        let slice = std::slice::from_raw_parts_mut(ptr, len);
        Box::from_raw(slice)
    }
}

fn hash(len_buf: [u8; 8], pid_buf: [u8; 8], object_buf: &[u8]) -> [u8; 4] {
    let mut hasher = crc32fast::Hasher::new();
    hasher.update(&len_buf);
    hasher.update(&pid_buf);
    hasher.update(&object_buf);
    let crc: u32 = hasher.finalize();
    crc.to_le_bytes()
}

/// Statistics for file contents, to base decisions around
/// calls to `maintenance`.
#[derive(Debug, Copy, Clone)]
pub struct Stats {
    /// The number of live objects stored in the backing
    /// storage files.
    pub live_objects: u64,
    /// The total number of (potentially duplicated)
    /// objects stored in the backing storage files.
    pub stored_objects: u64,
    /// The number of dead objects that have been replaced
    /// or removed in other storage files, contributing
    /// to fragmentation.
    pub dead_objects: u64,
    /// The ratio of all objects on disk that are
    /// live to all objects in total. This is another way of expressing fragmentation.
    pub live_ratio: f32,
    /// The number of backing storage files that exist and are
    /// being held open.
    pub files: usize,
    /// The sum of the sizes of all files currently on-disk.
    pub total_file_size: u64,
    /// The number of bytes that have been read since
    /// this instance of `Marble` was recovered.
    pub bytes_read: u64,
    /// The number of bytes that have been written due
    /// to calls to both `write_batch` and rewrites caused by
    /// calls to `maintenance` since this instance of `Marble` was recovered.
    pub bytes_written: u64,
    /// This is the number of bytes that are written from user
    /// calls to [`crate::Marble::write_batch`] since this instance
    /// was recovered.
    pub high_level_user_bytes_written: u64,
    /// The ratio of all bytes written to high-level user data
    /// since this instance of `Marble` was recovered. This is basically the
    /// maintenance overhead of on-disk GC in response to objects being rewritten
    /// and defragmentation maintenance copying old data to new homes. 1.0 is "perfect".
    /// If all data needs to be copied once, this will be 2.0, etc... For reference,
    /// many LSM tries will see write amplifications of a few dozen, and b-trees can often
    /// see write amplifications of several hundred. So, if you're under 10 for serious workloads,
    /// you're doing much better than most industrial systems.
    pub write_amplification: f32,
    /// The ratio of the sum of the size of all bytes written to the sum of the size of all high-level user data written
    /// since this instance of `Marble` was recovered. This goes up with fragmentation, and is
    /// brought back down with calls to `maintenance` that defragment storage files.
    pub space_amplification: f32,
}

#[derive(Default, Debug, Clone, Copy)]
struct Metadata {
    lsn: u64,
    trailer_offset: u64,
    present_objects: u64,
    generation: u8,
    file_size: u64,
}

impl Metadata {
    fn parse(name: &str, file_size: u64) -> Option<Metadata> {
        let mut splits = name.split("-");

        Some(Metadata {
            lsn: u64::from_str_radix(&splits.next()?, 16).ok()?,
            trailer_offset: u64::from_str_radix(&splits.next()?, 16).ok()?,
            present_objects: u64::from_str_radix(&splits.next()?, 16).ok()?,
            generation: u8::from_str_radix(splits.next()?, 16).ok()?,
            file_size,
        })
    }

    fn to_file_name(&self) -> String {
        let ret = format!(
            "{:016x}-{:016x}-{:016x}-{:01x}",
            self.lsn, self.trailer_offset, self.present_objects, self.generation
        );
        ret
    }
}

#[derive(Debug)]
struct FileAndMetadata {
    file: File,
    location: DiskLocation,
    path: AtomicPtr<PathBuf>,
    metadata: AtomicPtr<Metadata>,
    live_objects: AtomicU64,
    generation: u8,
    rewrite_claim: AtomicBool,
    synced: AtomicBool,
}

impl Drop for FileAndMetadata {
    fn drop(&mut self) {
        let empty = self.live_objects.load(Acquire) == 0;
        if empty {
            if let Err(e) = std::fs::remove_file(self.path().unwrap()) {
                eprintln!("failed to remove empty FileAndMetadata on drop: {:?}", e);
            }
        }

        let path_ptr = self.path.load(Acquire);
        if !path_ptr.is_null() {
            drop(unsafe { Box::from_raw(path_ptr) });
        }
        let metadata_ptr = self.metadata.load(Acquire);
        if !metadata_ptr.is_null() {
            drop(unsafe { Box::from_raw(metadata_ptr) });
        }
    }
}

impl FileAndMetadata {
    fn metadata(&self) -> Option<&Metadata> {
        let metadata_ptr = self.metadata.load(Acquire);
        if metadata_ptr.is_null() {
            // metadata not yet initialized
            None
        } else {
            Some(unsafe { &*metadata_ptr })
        }
    }

    fn install_metadata_and_path(&self, metadata: Metadata, path: PathBuf) {
        // NB: install path first because later on we
        // want to be able to assume that if metadata
        // is present, then so is path.
        let path_ptr = Box::into_raw(Box::new(path));
        let old_path_ptr = self.path.swap(path_ptr, SeqCst);
        assert!(old_path_ptr.is_null());

        let meta_ptr = Box::into_raw(Box::new(metadata));
        let old_meta_ptr = self.metadata.swap(meta_ptr, SeqCst);
        assert!(old_meta_ptr.is_null());
    }

    fn path(&self) -> Option<&PathBuf> {
        let path_ptr = self.path.load(Acquire);
        if path_ptr.is_null() {
            // metadata not yet initialized
            None
        } else {
            Some(unsafe { &*path_ptr })
        }
    }
}

/// Shard based on rough size ranges corresponding to SSD
/// page and block sizes
pub fn default_partition_function(_object_id: u64, size: usize) -> u8 {
    const SUBPAGE_MAX: usize = PAGE_MIN - 1;
    const PAGE_MIN: usize = 2048;
    const PAGE_MAX: usize = 16 * 1024;
    const BLOCK_MIN: usize = PAGE_MAX + 1;
    const BLOCK_MAX: usize = 4 * 1024 * 1024;

    match size {
        // items smaller than known SSD page sizes go to shard 0
        0..=SUBPAGE_MAX => 0,
        // items that fall roughly within the range of SSD page sizes go to shard 1
        PAGE_MIN..=PAGE_MAX => 1,
        // items that fall roughly within the size of an SSD block go to shard 2
        BLOCK_MIN..=BLOCK_MAX => 2,
        // large items that are larger than typical SSD block sizes go to shard 3
        _ => 3,
    }
}

/// Open the system with default configuration at the
/// provided path.
pub fn open<P: AsRef<Path>>(path: P) -> io::Result<Marble> {
    let config = Config {
        path: path.as_ref().into(),
        ..Config::default()
    };

    config.open()
}

/// Garbage-collecting object store. A nice solution to back
/// a pagecache, for people building their own databases.
///
/// Writes should generally be performed by some background
/// process whose job it is to clean logs etc...
#[derive(Clone)]
pub struct Marble {
    // maps from ObjectId to DiskLocation
    location_table: LocationTable,
    max_object_id: Arc<AtomicU64>,
    file_map: FileMap,
    config: Config,
    directory_lock: Arc<File>,
    #[cfg(feature = "runtime_validation")]
    debug_history: Arc<std::sync::Mutex<debug_history::DebugHistory>>,
    bytes_read: Arc<AtomicU64>,
    bytes_written: Arc<AtomicU64>,
    high_level_user_bytes_written: Arc<AtomicU64>,
}

impl std::fmt::Debug for Marble {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("Marble")
            .field("stats", &self.stats())
            .finish()
    }
}

impl std::fmt::Display for Marble {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "Marble {{ ... }}")
    }
}

impl Marble {
    /// Statistics about current files, intended to inform
    /// decisions about when to call `maintenance` based on
    /// desired write and space amplification
    /// characteristics.
    #[doc(alias = "file_statistics")]
    #[doc(alias = "statistics")]
    #[doc(alias = "metrics")]
    #[doc(alias = "info")]
    pub fn stats(&self) -> Stats {
        let (fams_len, total_file_size, stored_objects, live_objects) = self.file_map.stats();

        let bytes_read = self.bytes_read.load(Acquire);

        let bytes_written = self.bytes_written.load(Acquire);

        let high_level_user_bytes_written = self.high_level_user_bytes_written.load(Acquire);

        let live_ratio = live_objects as f32 / stored_objects.max(1) as f32;
        let approximate_live_data = live_ratio * total_file_size as f32;

        let write_amplification = bytes_written as f32 / high_level_user_bytes_written as f32;
        let space_amplification = total_file_size as f32 / approximate_live_data;

        Stats {
            live_objects,
            stored_objects,
            dead_objects: stored_objects - live_objects,
            live_ratio,
            files: fams_len,
            total_file_size,
            bytes_read,
            bytes_written,
            high_level_user_bytes_written,
            write_amplification,
            space_amplification,
        }
    }

    fn prune_empty_files(&self) -> io::Result<()> {
        self.file_map.prune_empty_files(&self.location_table)
    }

    /// If `Config::fsync_each_batch` is `false`, this
    /// method can be called at a desired interval to
    /// ensure that the written batches are durable on
    /// disk.
    pub fn sync_all(&self) -> io::Result<()> {
        let synced_files = self.file_map.sync_all()?;
        if synced_files {
            fallible!(self.directory_lock.sync_all());
        }
        Ok(())
    }

    /// Intended for use in recovery, to bootstrap a higher level object ID allocator.
    ///
    /// Returns a tuple of 1 higher than the current max allocated object ID,
    /// and an iterator over all object IDs beneath that which are
    /// currently deleted (due to being stored as a `None` in a write batch).
    pub fn free_object_ids<'a>(&'a self) -> (u64, impl 'a + Iterator<Item = u64>) {
        let max = self.max_object_id.load(Acquire);

        let iter = (0..=max).filter_map(|oid| {
            if self.location_table.load(oid).is_none() {
                Some(oid)
            } else {
                None
            }
        });
        (max + 1, iter)
    }

    /// Returns an Iterator over all currently allocated object IDs.
    pub fn allocated_object_ids<'a>(&'a self) -> impl 'a + Iterator<Item = u64> {
        let max = self.max_object_id.load(Acquire);
        (0..=max).filter_map(|oid| {
            if self.location_table.load(oid).is_some() {
                Some(oid)
            } else {
                None
            }
        })
    }
}