Mercurial > lbo > hg > leveldb-rs
view src/table_reader.rs @ 387:2ac6eedcc9dc
db_impl: Use a BufWriter inside LogWriter
Unbuffered write-a-lot with 32768 entries written, first four iterations on a new DB:
usr=0.32 sys=0.12 elap=0.46
usr=0.47 sys=0.12 elap=0.60
usr=0.59 sys=0.11 elap=0.72
usr=0.50 sys=0.15 elap=0.67
Buffered write-a-lot with 32768 entries written, first four iterations on a new DB:
usr=0.22 sys=0.01 elap=0.24 (delta = -0.1/-0.11/-0.22)
usr=0.43 sys=0.03 elap=0.48 (delta = -0.04/-0.09/-0.12)
usr=0.54 sys=0.05 elap=0.62 (delta = -0.05/-0.06/-0.1)
usr=0.42 sys=0.03 elap=0.47 (delta = (-0.08/-0.12/-0.2)
author | Lewin Bormann <lbo@spheniscida.de> |
---|---|
date | Mon, 09 Oct 2017 05:51:43 +0000 |
parents | a76ff74f0c0f |
children | d12b13664ad6 |
line wrap: on
line source
use block::{Block, BlockIter}; use blockhandle::BlockHandle; use cache; use cmp::{Cmp, InternalKeyCmp}; use env::RandomAccess; use error::{err, StatusCode, Result}; use filter; use filter_block::FilterBlockReader; use key_types::InternalKey; use options::{self, CompressionType, Options}; use table_builder::{self, Footer}; use types::{current_key_val, LdbIterator}; use std::cmp::Ordering; use std::rc::Rc; use integer_encoding::{FixedInt, FixedIntWriter}; use crc::crc32::{self, Hasher32}; /// Reads the table footer. fn read_footer(f: &RandomAccess, size: usize) -> Result<Footer> { let mut buf = vec![0; table_builder::FULL_FOOTER_LENGTH]; f.read_at(size - table_builder::FULL_FOOTER_LENGTH, &mut buf)?; Ok(Footer::decode(&buf)) } fn read_bytes(f: &RandomAccess, location: &BlockHandle) -> Result<Vec<u8>> { let mut buf = vec![0; location.size()]; f.read_at(location.offset(), &mut buf).map(|_| buf) } /// A TableBlock consists of [bytes..., compress (1B), checksum (4B)]. #[derive(Clone)] pub struct TableBlock { block: Block, checksum: u32, compression: CompressionType, } impl TableBlock { /// Reads a block at location. fn read_block(opt: Options, f: &RandomAccess, location: &BlockHandle) -> Result<TableBlock> { // The block is denoted by offset and length in BlockHandle. A block in an encoded // table is followed by 1B compression type and 4B checksum. let buf = try!(read_bytes(f, location)); let compress = try!(read_bytes(f, &BlockHandle::new(location.offset() + location.size(), table_builder::TABLE_BLOCK_COMPRESS_LEN))); let cksum = try!(read_bytes(f, &BlockHandle::new(location.offset() + location.size() + table_builder::TABLE_BLOCK_COMPRESS_LEN, table_builder::TABLE_BLOCK_CKSUM_LEN))); Ok(TableBlock { block: Block::new(opt, buf), checksum: u32::decode_fixed(&cksum), compression: options::int_to_compressiontype(compress[0] as u32) .unwrap_or(CompressionType::CompressionNone), }) } /// Verify checksum of block fn verify(&self) -> bool { let mut digest = crc32::Digest::new(crc32::CASTAGNOLI); digest.write(&self.block.contents()); digest.write(&[self.compression as u8; 1]); digest.sum32() == self.checksum } } #[derive(Clone)] pub struct Table { file: Rc<Box<RandomAccess>>, file_size: usize, cache_id: cache::CacheID, opt: Options, footer: Footer, indexblock: Block, filters: Option<FilterBlockReader>, } impl Table { /// Creates a new table reader operating on unformatted keys (i.e., UserKey). fn new_raw(opt: Options, file: Rc<Box<RandomAccess>>, size: usize) -> Result<Table> { let footer = try!(read_footer(file.as_ref().as_ref(), size)); let indexblock = try!(TableBlock::read_block(opt.clone(), file.as_ref().as_ref(), &footer.index)); let metaindexblock = try!(TableBlock::read_block(opt.clone(), file.as_ref().as_ref(), &footer.meta_index)); if !indexblock.verify() || !metaindexblock.verify() { return err(StatusCode::InvalidData, "Indexblock/Metaindexblock failed verification"); } // Open filter block for reading let mut filter_block_reader = None; let filter_name = format!("filter.{}", opt.filter_policy.name()).as_bytes().to_vec(); let mut metaindexiter = metaindexblock.block.iter(); metaindexiter.seek(&filter_name); if let Some((_key, val)) = current_key_val(&metaindexiter) { let filter_block_location = BlockHandle::decode(&val).0; if filter_block_location.size() > 0 { let buf = try!(read_bytes(file.as_ref().as_ref(), &filter_block_location)); filter_block_reader = Some(FilterBlockReader::new_owned(opt.filter_policy.clone(), buf)); } } metaindexiter.reset(); let cache_id = opt.block_cache.borrow_mut().new_cache_id(); Ok(Table { file: file, file_size: size, cache_id: cache_id, opt: opt, footer: footer, filters: filter_block_reader, indexblock: indexblock.block, }) } /// Creates a new table reader operating on internal keys (i.e., InternalKey). This means that /// a different comparator (internal_key_cmp) and a different filter policy /// (InternalFilterPolicy) are used. pub fn new(mut opt: Options, file: Rc<Box<RandomAccess>>, size: usize) -> Result<Table> { opt.cmp = Rc::new(Box::new(InternalKeyCmp(opt.cmp.clone()))); opt.filter_policy = Rc::new(Box::new(filter::InternalFilterPolicy::new(opt.filter_policy))); Table::new_raw(opt, file, size) } /// block_cache_handle creates a CacheKey for a block with a given offset to be used in the /// block cache. fn block_cache_handle(&self, block_off: usize) -> cache::CacheKey { let mut dst = [0; 2 * 8]; (&mut dst[..8]).write_fixedint(self.cache_id).expect("error writing to vec"); (&mut dst[8..]).write_fixedint(block_off as u64).expect("error writing to vec"); dst } /// Read a block from the current table at `location`, and cache it in the options' block /// cache. fn read_block(&self, location: &BlockHandle) -> Result<TableBlock> { let cachekey = self.block_cache_handle(location.offset()); if let Some(block) = self.opt.block_cache.borrow_mut().get(&cachekey) { return Ok(block.clone()); } // Two times as_ref(): First time to get a ref from Rc<>, then one from Box<>. let b = try!(TableBlock::read_block(self.opt.clone(), self.file.as_ref().as_ref(), location)); if !b.verify() { return err(StatusCode::InvalidData, "Data block failed verification"); } // insert a cheap copy (Rc). self.opt.block_cache.borrow_mut().insert(&cachekey, b.clone()); Ok(b) } /// Returns the offset of the block that contains `key`. pub fn approx_offset_of(&self, key: &[u8]) -> usize { let mut iter = self.indexblock.iter(); iter.seek(key); if let Some((_, val)) = current_key_val(&iter) { let location = BlockHandle::decode(&val).0; return location.offset(); } return self.footer.meta_index.offset(); } /// Iterators read from the file; thus only one iterator can be borrowed (mutably) per scope pub fn iter(&self) -> TableIterator { let iter = TableIterator { current_block: None, current_block_off: 0, index_block: self.indexblock.iter(), table: self.clone(), }; iter } /// Retrieve next-biggest entry for key from table. This function uses the attached filters, so /// is better suited if you frequently look for non-existing values (as it will detect the /// non-existence of an entry in a block without having to load the block). /// /// The caller must check if the returned key, which is the raw key (not e.g. the user portion /// of an InternalKey) is acceptable (e.g. correct value type or sequence number), as it may /// not be an exact match for key. /// /// This is done this way because some key types, like internal keys, will not result in an /// exact match; it depends on other comparators than the one that the table reader knows /// whether a match is acceptable. pub fn get<'a>(&self, key: InternalKey<'a>) -> Result<Option<(Vec<u8>, Vec<u8>)>> { let mut index_iter = self.indexblock.iter(); index_iter.seek(key); let handle; if let Some((last_in_block, h)) = current_key_val(&index_iter) { if self.opt.cmp.cmp(key, &last_in_block) == Ordering::Less { handle = BlockHandle::decode(&h).0; } else { return Ok(None); } } else { return Ok(None); } // found correct block. // Check bloom (or whatever) filter if let Some(ref filters) = self.filters { if !filters.key_may_match(handle.offset(), key) { return Ok(None); } } // Read block (potentially from cache) let tb = self.read_block(&handle)?; let mut iter = tb.block.iter(); // Go to entry and check if it's the wanted entry. iter.seek(key); if let Some((k, v)) = current_key_val(&iter) { if self.opt.cmp.cmp(&k, key) >= Ordering::Equal { return Ok(Some((k, v))); } } Ok(None) } } /// This iterator is a "TwoLevelIterator"; it uses an index block in order to get an offset hint /// into the data blocks. pub struct TableIterator { // A TableIterator is independent of its table (on the syntax level -- it doesn't know its // Table's lifetime). This is mainly required by the dynamic iterators used everywhere, where a // lifetime makes things like returning an iterator from a function neigh-impossible. // // Instead, reference-counted pointers and locks inside the Table ensure that all // TableIterators still share a table. table: Table, current_block: Option<BlockIter>, current_block_off: usize, index_block: BlockIter, } impl TableIterator { // Skips to the entry referenced by the next entry in the index block. // This is called once a block has run out of entries. // Err means corruption or I/O error; Ok(true) means a new block was loaded; Ok(false) means // tht there's no more entries. fn skip_to_next_entry(&mut self) -> Result<bool> { if let Some((_key, val)) = self.index_block.next() { self.load_block(&val).map(|_| true) } else { Ok(false) } } // Load the block at `handle` into `self.current_block` fn load_block(&mut self, handle: &[u8]) -> Result<()> { let (new_block_handle, _) = BlockHandle::decode(handle); let block = self.table.read_block(&new_block_handle)?; self.current_block = Some(block.block.iter()); self.current_block_off = new_block_handle.offset(); Ok(()) } } impl LdbIterator for TableIterator { fn advance(&mut self) -> bool { // Uninitialized case. if self.current_block.is_none() { match self.skip_to_next_entry() { Ok(true) => return self.advance(), Ok(false) => { self.reset(); return false; } // try next block from index, this might be corruption Err(_) => return self.advance(), } } // Initialized case -- does the current block have more entries? if let Some(ref mut cb) = self.current_block { if cb.advance() { return true; } } // If the current block is exhausted, try loading the next block. self.current_block = None; match self.skip_to_next_entry() { Ok(true) => self.advance(), Ok(false) => { self.reset(); false } // try next block, this might be corruption Err(_) => self.advance(), } } // A call to valid() after seeking is necessary to ensure that the seek worked (e.g., no error // while reading from disk) fn seek(&mut self, to: &[u8]) { // first seek in index block, rewind by one entry (so we get the next smaller index entry), // then set current_block and seek there self.index_block.seek(to); // It's possible that this is a seek past-last; reset in that case. if let Some((past_block, handle)) = current_key_val(&self.index_block) { if self.table.opt.cmp.cmp(to, &past_block) <= Ordering::Equal { // ok, found right block: continue if let Ok(()) = self.load_block(&handle) { // current_block is always set if load_block() returned Ok. self.current_block.as_mut().unwrap().seek(to); return; } } } // Reached in case of failure. self.reset(); } fn prev(&mut self) -> bool { // happy path: current block contains previous entry if let Some(ref mut cb) = self.current_block { if cb.prev() { return true; } } // Go back one block and look for the last entry in the previous block if self.index_block.prev() { if let Some((_, handle)) = current_key_val(&self.index_block) { if self.load_block(&handle).is_ok() { self.current_block.as_mut().unwrap().seek_to_last(); self.current_block.as_ref().unwrap().valid() } else { self.reset(); false } } else { false } } else { false } } fn reset(&mut self) { self.index_block.reset(); self.current_block = None; } // This iterator is special in that it's valid even before the first call to advance(). It // behaves correctly, though. fn valid(&self) -> bool { self.current_block.is_some() && (self.current_block.as_ref().unwrap().valid() || self.index_block.valid()) } fn current(&self, key: &mut Vec<u8>, val: &mut Vec<u8>) -> bool { if let Some(ref cb) = self.current_block { cb.current(key, val) } else { false } } } #[cfg(test)] mod tests { use filter::BloomPolicy; use options; use table_builder::TableBuilder; use test_util::{test_iterator_properties, LdbIteratorIter}; use types::{current_key_val, LdbIterator}; use key_types::LookupKey; use super::*; fn build_data() -> Vec<(&'static str, &'static str)> { vec![// block 1 ("abc", "def"), ("abd", "dee"), ("bcd", "asa"), // block 2 ("bsr", "a00"), ("xyz", "xxx"), ("xzz", "yyy"), // block 3 ("zzz", "111")] } // Build a table containing raw keys (no format). It returns (vector, length) for convenience // reason, a call f(v, v.len()) doesn't work for borrowing reasons. fn build_table(data: Vec<(&'static str, &'static str)>) -> (Vec<u8>, usize) { let mut d = Vec::with_capacity(512); let mut opt = options::for_test(); opt.block_restart_interval = 2; opt.block_size = 32; { // Uses the standard comparator in opt. let mut b = TableBuilder::new_raw(opt, &mut d); for &(k, v) in data.iter() { b.add(k.as_bytes(), v.as_bytes()).unwrap(); } b.finish().unwrap(); } let size = d.len(); (d, size) } // Build a table containing keys in InternalKey format. fn build_internal_table() -> (Vec<u8>, usize) { let mut d = Vec::with_capacity(512); let mut opt = options::for_test(); opt.block_restart_interval = 1; opt.block_size = 32; opt.filter_policy = Rc::new(Box::new(BloomPolicy::new(4))); let mut i = 1 as u64; let data: Vec<(Vec<u8>, &'static str)> = build_data() .into_iter() .map(|(k, v)| { i += 1; (LookupKey::new(k.as_bytes(), i).internal_key().to_vec(), v) }) .collect(); { // Uses InternalKeyCmp let mut b = TableBuilder::new(opt, &mut d); for &(ref k, ref v) in data.iter() { b.add(k.as_slice(), v.as_bytes()).unwrap(); } b.finish().unwrap(); } let size = d.len(); (d, size) } fn wrap_buffer(src: Vec<u8>) -> Rc<Box<RandomAccess>> { Rc::new(Box::new(src)) } #[test] fn test_table_approximate_offset() { let (src, size) = build_table(build_data()); let mut opt = options::for_test(); opt.block_size = 32; let table = Table::new_raw(opt, wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); let expected_offsets = vec![0, 0, 0, 42, 42, 42, 86]; let mut i = 0; for (k, _) in LdbIteratorIter::wrap(&mut iter) { assert_eq!(expected_offsets[i], table.approx_offset_of(&k)); i += 1; } // Key-past-last returns offset of metaindex block. assert_eq!(132, table.approx_offset_of("{aa".as_bytes())); } #[test] fn test_table_block_cache_use() { let (src, size) = build_table(build_data()); let mut opt = options::for_test(); opt.block_size = 32; let table = Table::new_raw(opt.clone(), wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); // index/metaindex blocks are not cached. That'd be a waste of memory. assert_eq!(opt.block_cache.borrow().count(), 0); iter.next(); assert_eq!(opt.block_cache.borrow().count(), 1); // This may fail if block parameters or data change. In that case, adapt it. iter.next(); iter.next(); iter.next(); iter.next(); assert_eq!(opt.block_cache.borrow().count(), 2); } #[test] fn test_table_iterator_fwd_bwd() { let (src, size) = build_table(build_data()); let data = build_data(); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); let mut i = 0; while let Some((k, v)) = iter.next() { assert_eq!((data[i].0.as_bytes(), data[i].1.as_bytes()), (k.as_ref(), v.as_ref())); i += 1; } assert_eq!(i, data.len()); assert!(!iter.valid()); // Go forward again, to last entry. while let Some((key, _)) = iter.next() { if key.as_slice() == b"zzz" { break; } } assert!(iter.valid()); // backwards count let mut j = 0; while iter.prev() { if let Some((k, v)) = current_key_val(&iter) { j += 1; assert_eq!((data[data.len() - 1 - j].0.as_bytes(), data[data.len() - 1 - j].1.as_bytes()), (k.as_ref(), v.as_ref())); } else { break; } } // expecting 7 - 1, because the last entry that the iterator stopped on is the last entry // in the table; that is, it needs to go back over 6 entries. assert_eq!(j, 6); } #[test] fn test_table_iterator_filter() { let (src, size) = build_table(build_data()); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); assert!(table.filters.is_some()); let filter_reader = table.filters.clone().unwrap(); let mut iter = table.iter(); loop { if let Some((k, _)) = iter.next() { assert!(filter_reader.key_may_match(iter.current_block_off, &k)); assert!(!filter_reader.key_may_match(iter.current_block_off, b"somerandomkey")); } else { break; } } } #[test] fn test_table_iterator_state_behavior() { let (src, size) = build_table(build_data()); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); // behavior test // See comment on valid() assert!(!iter.valid()); assert!(current_key_val(&iter).is_none()); assert!(!iter.prev()); assert!(iter.advance()); let first = current_key_val(&iter); assert!(iter.valid()); assert!(current_key_val(&iter).is_some()); assert!(iter.advance()); assert!(iter.prev()); assert!(iter.valid()); iter.reset(); assert!(!iter.valid()); assert!(current_key_val(&iter).is_none()); assert_eq!(first, iter.next()); } #[test] fn test_table_iterator_behavior_standard() { let mut data = build_data(); data.truncate(4); let (src, size) = build_table(data); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); test_iterator_properties(table.iter()); } #[test] fn test_table_iterator_values() { let (src, size) = build_table(build_data()); let data = build_data(); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); let mut i = 0; iter.next(); iter.next(); // Go back to previous entry, check, go forward two entries, repeat // Verifies that prev/next works well. loop { iter.prev(); if let Some((k, v)) = current_key_val(&iter) { assert_eq!((data[i].0.as_bytes(), data[i].1.as_bytes()), (k.as_ref(), v.as_ref())); } else { break; } i += 1; if iter.next().is_none() || iter.next().is_none() { break; } } // Skipping the last value because the second next() above will break the loop assert_eq!(i, 6); } #[test] fn test_table_iterator_seek() { let (src, size) = build_table(build_data()); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); let mut iter = table.iter(); iter.seek(b"bcd"); assert!(iter.valid()); assert_eq!(current_key_val(&iter), Some((b"bcd".to_vec(), b"asa".to_vec()))); iter.seek(b"abc"); assert!(iter.valid()); assert_eq!(current_key_val(&iter), Some((b"abc".to_vec(), b"def".to_vec()))); // Seek-past-last invalidates. iter.seek("{{{".as_bytes()); assert!(!iter.valid()); iter.seek(b"bbb"); assert!(iter.valid()); } #[test] fn test_table_get() { let (src, size) = build_table(build_data()); let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); let table2 = table.clone(); let mut _iter = table.iter(); // Test that all of the table's entries are reachable via get() for (k, v) in LdbIteratorIter::wrap(&mut _iter) { let r = table2.get(&k); assert_eq!(Ok(Some((k, v))), r); } assert_eq!(table.opt.block_cache.borrow().count(), 3); // test that filters work and don't return anything at all. assert!(table.get(b"aaa").unwrap().is_none()); assert!(table.get(b"aaaa").unwrap().is_none()); assert!(table.get(b"aa").unwrap().is_none()); assert!(table.get(b"abcd").unwrap().is_none()); assert!(table.get(b"abb").unwrap().is_none()); assert!(table.get(b"zzy").unwrap().is_none()); assert!(table.get(b"zz1").unwrap().is_none()); assert!(table.get("zz{".as_bytes()).unwrap().is_none()); } // This test verifies that the table and filters work with internal keys. This means: // The table contains keys in InternalKey format and it uses a filter wrapped by // InternalFilterPolicy. // All the other tests use raw keys that don't have any internal structure; this is fine in // general, but here we want to see that the other infrastructure works too. #[test] fn test_table_internal_keys() { use key_types::LookupKey; let (src, size) = build_internal_table(); let table = Table::new(options::for_test(), wrap_buffer(src), size).unwrap(); let filter_reader = table.filters.clone().unwrap(); // Check that we're actually using internal keys let mut _iter = table.iter(); for (ref k, ref v) in LdbIteratorIter::wrap(&mut _iter) { assert_eq!(k.len(), 3 + 8); assert_eq!((k.to_vec(), v.to_vec()), table.get(k).unwrap().unwrap()); } assert!(table.get(LookupKey::new(b"abc", 1000).internal_key()).unwrap().is_some()); let mut iter = table.iter(); loop { if let Some((k, _)) = iter.next() { let lk = LookupKey::new(&k, 123); let userkey = lk.user_key(); assert!(filter_reader.key_may_match(iter.current_block_off, userkey)); assert!(!filter_reader.key_may_match(iter.current_block_off, b"somerandomkey")); } else { break; } } } #[test] fn test_table_reader_checksum() { let (mut src, size) = build_table(build_data()); src[10] += 1; let table = Table::new_raw(options::for_test(), wrap_buffer(src), size).unwrap(); assert!(table.filters.is_some()); assert_eq!(table.filters.as_ref().unwrap().num(), 1); { let mut _iter = table.iter(); let iter = LdbIteratorIter::wrap(&mut _iter); // first block is skipped assert_eq!(iter.count(), 4); } { let mut _iter = table.iter(); let iter = LdbIteratorIter::wrap(&mut _iter); for (k, _) in iter { if k == build_data()[5].0.as_bytes() { return; } } panic!("Should have hit 5th record in table!"); } } }