Line data Source code
1 : //!
2 : //! Simple on-disk B-tree implementation
3 : //!
4 : //! This is used as the index structure within image and delta layers
5 : //!
6 : //! Features:
7 : //! - Fixed-width keys
8 : //! - Fixed-width values (VALUE_SZ)
9 : //! - The tree is created in a bulk operation. Insert/deletion after creation
10 : //! is not supported
11 : //! - page-oriented
12 : //!
13 : //! TODO:
14 : //! - maybe something like an Adaptive Radix Tree would be more efficient?
15 : //! - the values stored by image and delta layers are offsets into the file,
16 : //! and they are in monotonically increasing order. Prefix compression would
17 : //! be very useful for them, too.
18 : //! - An Iterator interface would be more convenient for the callers than the
19 : //! 'visit' function
20 : //!
21 : use std::cmp::Ordering;
22 : use std::iter::Rev;
23 : use std::ops::{Range, RangeInclusive};
24 : use std::{io, result};
25 :
26 : use async_stream::try_stream;
27 : use byteorder::{BE, ReadBytesExt};
28 : use bytes::BufMut;
29 : use either::Either;
30 : use futures::{Stream, StreamExt};
31 : use hex;
32 : use thiserror::Error;
33 : use tracing::error;
34 :
35 : use crate::context::RequestContext;
36 : use crate::tenant::block_io::{BlockReader, BlockWriter};
37 : use crate::virtual_file::{IoBuffer, IoBufferMut, owned_buffers_io::write::Buffer};
38 :
39 : // The maximum size of a value stored in the B-tree. 5 bytes is enough currently.
40 : pub const VALUE_SZ: usize = 5;
41 : pub const MAX_VALUE: u64 = 0x007f_ffff_ffff;
42 :
43 : pub const PAGE_SZ: usize = 8192;
44 :
45 : #[derive(Clone, Copy, Debug)]
46 : struct Value([u8; VALUE_SZ]);
47 :
48 : impl Value {
49 4573264 : fn from_slice(slice: &[u8]) -> Value {
50 4573264 : let mut b = [0u8; VALUE_SZ];
51 4573264 : b.copy_from_slice(slice);
52 4573264 : Value(b)
53 4573264 : }
54 :
55 3374337 : fn from_u64(x: u64) -> Value {
56 3374337 : assert!(x <= 0x007f_ffff_ffff);
57 3374337 : Value([
58 3374337 : (x >> 32) as u8,
59 3374337 : (x >> 24) as u8,
60 3374337 : (x >> 16) as u8,
61 3374337 : (x >> 8) as u8,
62 3374337 : x as u8,
63 3374337 : ])
64 3374337 : }
65 :
66 6246 : fn from_blknum(x: u32) -> Value {
67 6246 : Value([
68 6246 : 0x80,
69 6246 : (x >> 24) as u8,
70 6246 : (x >> 16) as u8,
71 6246 : (x >> 8) as u8,
72 6246 : x as u8,
73 6246 : ])
74 6246 : }
75 :
76 : #[allow(dead_code)]
77 0 : fn is_offset(self) -> bool {
78 0 : self.0[0] & 0x80 != 0
79 0 : }
80 :
81 4237253 : fn to_u64(self) -> u64 {
82 4237253 : let b = &self.0;
83 4237253 : ((b[0] as u64) << 32)
84 4237253 : | ((b[1] as u64) << 24)
85 4237253 : | ((b[2] as u64) << 16)
86 4237253 : | ((b[3] as u64) << 8)
87 4237253 : | b[4] as u64
88 4237253 : }
89 :
90 332999 : fn to_blknum(self) -> u32 {
91 332999 : let b = &self.0;
92 332999 : assert!(b[0] == 0x80);
93 332999 : ((b[1] as u32) << 24) | ((b[2] as u32) << 16) | ((b[3] as u32) << 8) | b[4] as u32
94 332999 : }
95 : }
96 :
97 : #[derive(Error, Debug)]
98 : pub enum DiskBtreeError {
99 : #[error("Attempt to append a value that is too large {0} > {}", MAX_VALUE)]
100 : AppendOverflow(u64),
101 :
102 : #[error("Unsorted input: key {key:?} is <= last_key {last_key:?}")]
103 : UnsortedInput { key: Box<[u8]>, last_key: Box<[u8]> },
104 :
105 : #[error("Could not push to new leaf node")]
106 : FailedToPushToNewLeafNode,
107 :
108 : #[error("IoError: {0}")]
109 : Io(#[from] io::Error),
110 : }
111 :
112 : pub type Result<T> = result::Result<T, DiskBtreeError>;
113 :
114 : /// This is the on-disk representation.
115 : struct OnDiskNode<'a, const L: usize> {
116 : // Fixed-width fields
117 : num_children: u16,
118 : level: u8,
119 : prefix_len: u8,
120 : suffix_len: u8,
121 :
122 : // Variable-length fields. These are stored on-disk after the fixed-width
123 : // fields, in this order. In the in-memory representation, these point to
124 : // the right parts in the page buffer.
125 : prefix: &'a [u8],
126 : keys: &'a [u8],
127 : values: &'a [u8],
128 : }
129 :
130 : impl<const L: usize> OnDiskNode<'_, L> {
131 : ///
132 : /// Interpret a PAGE_SZ page as a node.
133 : ///
134 803221 : fn deparse(buf: &[u8]) -> Result<OnDiskNode<L>> {
135 803221 : let mut cursor = std::io::Cursor::new(buf);
136 803221 : let num_children = cursor.read_u16::<BE>()?;
137 803221 : let level = cursor.read_u8()?;
138 803221 : let prefix_len = cursor.read_u8()?;
139 803221 : let suffix_len = cursor.read_u8()?;
140 :
141 803221 : let mut off = cursor.position();
142 803221 : let prefix_off = off as usize;
143 803221 : off += prefix_len as u64;
144 :
145 803221 : let keys_off = off as usize;
146 803221 : let keys_len = num_children as usize * suffix_len as usize;
147 803221 : off += keys_len as u64;
148 :
149 803221 : let values_off = off as usize;
150 803221 : let values_len = num_children as usize * VALUE_SZ;
151 : //off += values_len as u64;
152 :
153 803221 : let prefix = &buf[prefix_off..prefix_off + prefix_len as usize];
154 803221 : let keys = &buf[keys_off..keys_off + keys_len];
155 803221 : let values = &buf[values_off..values_off + values_len];
156 :
157 803221 : Ok(OnDiskNode {
158 803221 : num_children,
159 803221 : level,
160 803221 : prefix_len,
161 803221 : suffix_len,
162 803221 : prefix,
163 803221 : keys,
164 803221 : values,
165 803221 : })
166 803221 : }
167 :
168 : ///
169 : /// Read a value at 'idx'
170 : ///
171 4573264 : fn value(&self, idx: usize) -> Value {
172 4573264 : let value_off = idx * VALUE_SZ;
173 4573264 : let value_slice = &self.values[value_off..value_off + VALUE_SZ];
174 4573264 : Value::from_slice(value_slice)
175 4573264 : }
176 :
177 688012 : fn binary_search(
178 688012 : &self,
179 688012 : search_key: &[u8; L],
180 688012 : keybuf: &mut [u8],
181 688012 : ) -> result::Result<usize, usize> {
182 688012 : let mut size = self.num_children as usize;
183 688012 : let mut low = 0;
184 688012 : let mut high = size;
185 5130761 : while low < high {
186 4579222 : let mid = low + size / 2;
187 :
188 4579222 : let key_off = mid * self.suffix_len as usize;
189 4579222 : let suffix = &self.keys[key_off..key_off + self.suffix_len as usize];
190 : // Does this match?
191 4579222 : keybuf[self.prefix_len as usize..].copy_from_slice(suffix);
192 :
193 4579222 : let cmp = keybuf[..].cmp(search_key);
194 :
195 4579222 : if cmp == Ordering::Less {
196 2856854 : low = mid + 1;
197 2856854 : } else if cmp == Ordering::Greater {
198 1585895 : high = mid;
199 1585895 : } else {
200 136473 : return Ok(mid);
201 : }
202 4442749 : size = high - low;
203 : }
204 551539 : Err(low)
205 688012 : }
206 : }
207 :
208 : ///
209 : /// Public reader object, to search the tree.
210 : ///
211 : #[derive(Clone)]
212 : pub struct DiskBtreeReader<R, const L: usize>
213 : where
214 : R: BlockReader,
215 : {
216 : start_blk: u32,
217 : root_blk: u32,
218 : reader: R,
219 : }
220 :
221 : #[derive(Clone, Copy, Debug, PartialEq, Eq)]
222 : pub enum VisitDirection {
223 : Forwards,
224 : Backwards,
225 : }
226 :
227 : impl<R, const L: usize> DiskBtreeReader<R, L>
228 : where
229 : R: BlockReader,
230 : {
231 139531 : pub fn new(start_blk: u32, root_blk: u32, reader: R) -> Self {
232 139531 : DiskBtreeReader {
233 139531 : start_blk,
234 139531 : root_blk,
235 139531 : reader,
236 139531 : }
237 139531 : }
238 :
239 : ///
240 : /// Read the value for given key. Returns the value, or None if it doesn't exist.
241 : ///
242 201560 : pub async fn get(&self, search_key: &[u8; L], ctx: &RequestContext) -> Result<Option<u64>> {
243 201560 : let mut result: Option<u64> = None;
244 201560 : self.visit(
245 201560 : search_key,
246 201560 : VisitDirection::Forwards,
247 101548 : |key, value| {
248 101548 : if key == search_key {
249 100545 : result = Some(value);
250 100545 : }
251 101548 : false
252 101548 : },
253 201560 : ctx,
254 : )
255 201560 : .await?;
256 201560 : Ok(result)
257 201560 : }
258 :
259 352 : pub fn iter<'a>(self, start_key: &'a [u8; L], ctx: &'a RequestContext) -> DiskBtreeIterator<'a>
260 352 : where
261 352 : R: 'a + Send,
262 : {
263 352 : DiskBtreeIterator {
264 352 : stream: Box::pin(self.into_stream(start_key, ctx)),
265 352 : }
266 352 : }
267 :
268 : /// Return a stream which yields all key, value pairs from the index
269 : /// starting from the first key greater or equal to `start_key`.
270 : ///
271 : /// Note 1: that this is a copy of [`Self::visit`].
272 : /// TODO: Once the sequential read path is removed this will become
273 : /// the only index traversal method.
274 : ///
275 : /// Note 2: this function used to take `&self` but it now consumes `self`. This is due to
276 : /// the lifetime constraints of the reader and the stream / iterator it creates. Using `&self`
277 : /// requires the reader to be present when the stream is used, and this creates a lifetime
278 : /// dependency between the reader and the stream. Now if we want to create an iterator that
279 : /// holds the stream, someone will need to keep a reference to the reader, which is inconvenient
280 : /// to use from the image/delta layer APIs.
281 : ///
282 : /// Feel free to add the `&self` variant back if it's necessary.
283 149227 : pub fn into_stream<'a>(
284 149227 : self,
285 149227 : start_key: &'a [u8; L],
286 149227 : ctx: &'a RequestContext,
287 149227 : ) -> impl Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a
288 149227 : where
289 149227 : R: 'a,
290 : {
291 149227 : try_stream! {
292 : let mut stack = Vec::new();
293 : stack.push((self.root_blk, None));
294 : let block_cursor = self.reader.block_cursor();
295 : let mut node_buf = [0_u8; PAGE_SZ];
296 : while let Some((node_blknum, opt_iter)) = stack.pop() {
297 : // Read the node, through the PS PageCache, into local variable `node_buf`.
298 : // We could keep the page cache read guard alive, but, at the time of writing,
299 : // we run quite small PS PageCache s => can't risk running out of
300 : // PageCache space because this stream isn't consumed fast enough.
301 : let page_read_guard = block_cursor
302 : .read_blk(self.start_blk + node_blknum, ctx)
303 : .await?;
304 : node_buf.copy_from_slice(page_read_guard.as_ref());
305 : drop(page_read_guard); // drop page cache read guard early
306 :
307 : let node = OnDiskNode::deparse(&node_buf)?;
308 : let prefix_len = node.prefix_len as usize;
309 : let suffix_len = node.suffix_len as usize;
310 :
311 : assert!(node.num_children > 0);
312 :
313 : let mut keybuf = Vec::new();
314 : keybuf.extend(node.prefix);
315 : keybuf.resize(prefix_len + suffix_len, 0);
316 :
317 : let mut iter: Either<Range<usize>, Rev<RangeInclusive<usize>>> = if let Some(iter) = opt_iter {
318 : iter
319 : } else {
320 : // Locate the first match
321 : let idx = match node.binary_search(start_key, keybuf.as_mut_slice()) {
322 : Ok(idx) => idx,
323 : Err(idx) => {
324 : if node.level == 0 {
325 : // Imagine that the node contains the following keys:
326 : //
327 : // 1
328 : // 3 <-- idx
329 : // 5
330 : //
331 : // If the search key is '2' and there is exact match,
332 : // the binary search would return the index of key
333 : // '3'. That's cool, '3' is the first key to return.
334 : idx
335 : } else {
336 : // This is an internal page, so each key represents a lower
337 : // bound for what's in the child page. If there is no exact
338 : // match, we have to return the *previous* entry.
339 : //
340 : // 1 <-- return this
341 : // 3 <-- idx
342 : // 5
343 : idx.saturating_sub(1)
344 : }
345 : }
346 : };
347 : Either::Left(idx..node.num_children.into())
348 : };
349 :
350 :
351 : // idx points to the first match now. Keep going from there
352 : while let Some(idx) = iter.next() {
353 : let key_off = idx * suffix_len;
354 : let suffix = &node.keys[key_off..key_off + suffix_len];
355 : keybuf[prefix_len..].copy_from_slice(suffix);
356 : let value = node.value(idx);
357 : #[allow(clippy::collapsible_if)]
358 : if node.level == 0 {
359 : // leaf
360 : yield (keybuf.clone(), value.to_u64());
361 : } else {
362 : stack.push((node_blknum, Some(iter)));
363 : stack.push((value.to_blknum(), None));
364 : break;
365 : }
366 : }
367 : }
368 : }
369 149227 : }
370 :
371 : ///
372 : /// Scan the tree, starting from 'search_key', in the given direction. 'visitor'
373 : /// will be called for every key >= 'search_key' (or <= 'search_key', if scanning
374 : /// backwards)
375 : ///
376 205790 : pub async fn visit<V>(
377 205790 : &self,
378 205790 : search_key: &[u8; L],
379 205790 : dir: VisitDirection,
380 205790 : mut visitor: V,
381 205790 : ctx: &RequestContext,
382 205790 : ) -> Result<bool>
383 205790 : where
384 205790 : V: FnMut(&[u8], u64) -> bool,
385 205790 : {
386 205790 : let mut stack = Vec::new();
387 205790 : stack.push((self.root_blk, None));
388 205790 : let block_cursor = self.reader.block_cursor();
389 609493 : while let Some((node_blknum, opt_iter)) = stack.pop() {
390 : // Locate the node.
391 509226 : let node_buf = block_cursor
392 509226 : .read_blk(self.start_blk + node_blknum, ctx)
393 509226 : .await?;
394 :
395 509226 : let node = OnDiskNode::deparse(node_buf.as_ref())?;
396 509226 : let prefix_len = node.prefix_len as usize;
397 509226 : let suffix_len = node.suffix_len as usize;
398 :
399 509226 : assert!(node.num_children > 0);
400 :
401 509226 : let mut keybuf = Vec::new();
402 509226 : keybuf.extend(node.prefix);
403 509226 : keybuf.resize(prefix_len + suffix_len, 0);
404 :
405 509226 : let mut iter = if let Some(iter) = opt_iter {
406 101949 : iter
407 407277 : } else if dir == VisitDirection::Forwards {
408 : // Locate the first match
409 405264 : let idx = match node.binary_search(search_key, keybuf.as_mut_slice()) {
410 101648 : Ok(idx) => idx,
411 303616 : Err(idx) => {
412 303616 : if node.level == 0 {
413 : // Imagine that the node contains the following keys:
414 : //
415 : // 1
416 : // 3 <-- idx
417 : // 5
418 : //
419 : // If the search key is '2' and there is exact match,
420 : // the binary search would return the index of key
421 : // '3'. That's cool, '3' is the first key to return.
422 104058 : idx
423 : } else {
424 : // This is an internal page, so each key represents a lower
425 : // bound for what's in the child page. If there is no exact
426 : // match, we have to return the *previous* entry.
427 : //
428 : // 1 <-- return this
429 : // 3 <-- idx
430 : // 5
431 199558 : idx.saturating_sub(1)
432 : }
433 : }
434 : };
435 405264 : Either::Left(idx..node.num_children.into())
436 : } else {
437 2013 : let idx = match node.binary_search(search_key, keybuf.as_mut_slice()) {
438 1001 : Ok(idx) => {
439 : // Exact match. That's the first entry to return, and walk
440 : // backwards from there.
441 1001 : idx
442 : }
443 1012 : Err(idx) => {
444 : // No exact match. The binary search returned the index of the
445 : // first key that's > search_key. Back off by one, and walk
446 : // backwards from there.
447 1012 : if let Some(idx) = idx.checked_sub(1) {
448 1010 : idx
449 : } else {
450 2 : return Ok(false);
451 : }
452 : }
453 : };
454 2011 : Either::Right((0..=idx).rev())
455 : };
456 :
457 : // idx points to the first match now. Keep going from there
458 1581201 : while let Some(idx) = iter.next() {
459 1378985 : let key_off = idx * suffix_len;
460 1378985 : let suffix = &node.keys[key_off..key_off + suffix_len];
461 1378985 : keybuf[prefix_len..].copy_from_slice(suffix);
462 1378985 : let value = node.value(idx);
463 : #[allow(clippy::collapsible_if)]
464 1378985 : if node.level == 0 {
465 : // leaf
466 1177498 : if !visitor(&keybuf, value.to_u64()) {
467 105521 : return Ok(false);
468 1071977 : }
469 : } else {
470 201487 : stack.push((node_blknum, Some(iter)));
471 201487 : stack.push((value.to_blknum(), None));
472 201487 : break;
473 : }
474 : }
475 : }
476 100267 : Ok(true)
477 205790 : }
478 :
479 : #[allow(dead_code)]
480 5 : pub async fn dump(&self, ctx: &RequestContext) -> Result<()> {
481 5 : let mut stack = Vec::new();
482 :
483 5 : stack.push((self.root_blk, String::new(), 0, 0, 0));
484 :
485 5 : let block_cursor = self.reader.block_cursor();
486 :
487 3021 : while let Some((blknum, path, depth, child_idx, key_off)) = stack.pop() {
488 3016 : let blk = block_cursor.read_blk(self.start_blk + blknum, ctx).await?;
489 3016 : let buf: &[u8] = blk.as_ref();
490 3016 : let node = OnDiskNode::<L>::deparse(buf)?;
491 :
492 3016 : if child_idx == 0 {
493 9 : print!("{:indent$}", "", indent = depth * 2);
494 9 : let path_prefix = stack
495 9 : .iter()
496 9 : .map(|(_blknum, path, ..)| path.as_str())
497 9 : .collect::<String>();
498 9 : println!(
499 9 : "blk #{blknum}: path {path_prefix}{path}: prefix {}, suffix_len {}",
500 9 : hex::encode(node.prefix),
501 : node.suffix_len
502 : );
503 3007 : }
504 :
505 3016 : if child_idx + 1 < node.num_children {
506 3007 : let key_off = key_off + node.suffix_len as usize;
507 3007 : stack.push((blknum, path.clone(), depth, child_idx + 1, key_off));
508 3007 : }
509 3016 : let key = &node.keys[key_off..key_off + node.suffix_len as usize];
510 3016 : let val = node.value(child_idx as usize);
511 :
512 3016 : print!("{:indent$}", "", indent = depth * 2 + 2);
513 3016 : println!("{}: {}", hex::encode(key), hex::encode(val.0));
514 :
515 3016 : if node.level > 0 {
516 4 : stack.push((val.to_blknum(), hex::encode(node.prefix), depth + 1, 0, 0));
517 3012 : }
518 : }
519 5 : Ok(())
520 5 : }
521 : }
522 :
523 : pub struct DiskBtreeIterator<'a> {
524 : #[allow(clippy::type_complexity)]
525 : stream: std::pin::Pin<
526 : Box<dyn Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a + Send>,
527 : >,
528 : }
529 :
530 : impl DiskBtreeIterator<'_> {
531 1161890 : pub async fn next(&mut self) -> Option<std::result::Result<(Vec<u8>, u64), DiskBtreeError>> {
532 1161890 : self.stream.next().await
533 1161890 : }
534 : }
535 :
536 : ///
537 : /// Public builder object, for creating a new tree.
538 : ///
539 : /// Usage: Create a builder object by calling 'new', load all the data into the
540 : /// tree by calling 'append' for each key-value pair, and then call 'finish'
541 : ///
542 : /// 'L' is the key length in bytes
543 : pub struct DiskBtreeBuilder<W, const L: usize>
544 : where
545 : W: BlockWriter,
546 : {
547 : writer: W,
548 :
549 : ///
550 : /// `stack[0]` is the current root page, `stack.last()` is the leaf.
551 : ///
552 : /// We maintain the length of the stack to be always greater than zero.
553 : /// Two exceptions are:
554 : /// 1. `Self::flush_node`. The method will push the new node if it extracted the last one.
555 : /// So because other methods cannot see the intermediate state invariant still holds.
556 : /// 2. `Self::finish`. It consumes self and does not return it back,
557 : /// which means that this is where the structure is destroyed.
558 : /// Thus stack of zero length cannot be observed by other methods.
559 : stack: Vec<BuildNode<L>>,
560 :
561 : /// Last key that was appended to the tree. Used to sanity check that append
562 : /// is called in increasing key order.
563 : last_key: Option<[u8; L]>,
564 : }
565 :
566 : impl<W, const L: usize> DiskBtreeBuilder<W, L>
567 : where
568 : W: BlockWriter,
569 : {
570 1123 : pub fn new(writer: W) -> Self {
571 1123 : DiskBtreeBuilder {
572 1123 : writer,
573 1123 : last_key: None,
574 1123 : stack: vec![BuildNode::new(0)],
575 1123 : }
576 1123 : }
577 :
578 3374338 : pub fn append(&mut self, key: &[u8; L], value: u64) -> Result<()> {
579 3374338 : if value > MAX_VALUE {
580 0 : return Err(DiskBtreeError::AppendOverflow(value));
581 3374338 : }
582 3374338 : if let Some(last_key) = &self.last_key {
583 3373379 : if key <= last_key {
584 1 : return Err(DiskBtreeError::UnsortedInput {
585 1 : key: key.as_slice().into(),
586 1 : last_key: last_key.as_slice().into(),
587 1 : });
588 3373378 : }
589 959 : }
590 3374337 : self.last_key = Some(*key);
591 :
592 3374337 : self.append_internal(key, Value::from_u64(value))
593 3374338 : }
594 :
595 3380583 : fn append_internal(&mut self, key: &[u8; L], value: Value) -> Result<()> {
596 : // Try to append to the current leaf buffer
597 3380583 : let last = self
598 3380583 : .stack
599 3380583 : .last_mut()
600 3380583 : .expect("should always have at least one item");
601 3380583 : let level = last.level;
602 3380583 : if last.push(key, value) {
603 3368725 : return Ok(());
604 11858 : }
605 :
606 : // It did not fit. Try to compress, and if it succeeds to make
607 : // some room on the node, try appending to it again.
608 : #[allow(clippy::collapsible_if)]
609 11858 : if last.compress() {
610 6024 : if last.push(key, value) {
611 6023 : return Ok(());
612 1 : }
613 5834 : }
614 :
615 : // Could not append to the current leaf. Flush it and create a new one.
616 5835 : self.flush_node()?;
617 :
618 : // Replace the node we flushed with an empty one and append the new
619 : // key to it.
620 5835 : let mut last = BuildNode::new(level);
621 5835 : if !last.push(key, value) {
622 0 : return Err(DiskBtreeError::FailedToPushToNewLeafNode);
623 5835 : }
624 :
625 5835 : self.stack.push(last);
626 :
627 5835 : Ok(())
628 3380583 : }
629 :
630 : /// Flush the bottommost node in the stack to disk. Appends a downlink to its parent,
631 : /// and recursively flushes the parent too, if it becomes full. If the root page becomes full,
632 : /// creates a new root page, increasing the height of the tree.
633 6246 : fn flush_node(&mut self) -> Result<()> {
634 : // Get the current bottommost node in the stack and flush it to disk.
635 6246 : let last = self
636 6246 : .stack
637 6246 : .pop()
638 6246 : .expect("should always have at least one item");
639 6246 : let buf = last.pack();
640 6246 : let downlink_key = last.first_key();
641 6246 : let downlink_ptr = self.writer.write_blk(buf)?;
642 :
643 : // Append the downlink to the parent. If there is no parent, ie. this was the root page,
644 : // create a new root page, increasing the height of the tree.
645 6246 : if self.stack.is_empty() {
646 411 : self.stack.push(BuildNode::new(last.level + 1));
647 5835 : }
648 6246 : self.append_internal(&downlink_key, Value::from_blknum(downlink_ptr))
649 6246 : }
650 :
651 : ///
652 : /// Flushes everything to disk, and returns the block number of the root page.
653 : /// The caller must store the root block number "out-of-band", and pass it
654 : /// to the DiskBtreeReader::new() when you want to read the tree again.
655 : /// (In the image and delta layers, it is stored in the beginning of the file,
656 : /// in the summary header)
657 : ///
658 934 : pub fn finish(mut self) -> Result<(u32, W)> {
659 : // flush all levels, except the root.
660 1345 : while self.stack.len() > 1 {
661 411 : self.flush_node()?;
662 : }
663 :
664 934 : let root = self
665 934 : .stack
666 934 : .first()
667 934 : .expect("by the check above we left one item there");
668 934 : let buf = root.pack();
669 934 : let root_blknum = self.writer.write_blk(buf)?;
670 :
671 934 : Ok((root_blknum, self.writer))
672 934 : }
673 :
674 1029527 : pub fn borrow_writer(&self) -> &W {
675 1029527 : &self.writer
676 1029527 : }
677 : }
678 :
679 : ///
680 : /// BuildNode represesnts an incomplete page that we are appending to.
681 : ///
682 : #[derive(Clone, Debug)]
683 : struct BuildNode<const L: usize> {
684 : num_children: u16,
685 : level: u8,
686 : prefix: Vec<u8>,
687 : suffix_len: usize,
688 :
689 : keys: Vec<u8>,
690 : values: Vec<u8>,
691 :
692 : size: usize, // physical size of this node, if it was written to disk like this
693 : }
694 :
695 : const NODE_SIZE: usize = PAGE_SZ;
696 :
697 : const NODE_HDR_SIZE: usize = 2 + 1 + 1 + 1;
698 :
699 : impl<const L: usize> BuildNode<L> {
700 7369 : fn new(level: u8) -> Self {
701 7369 : BuildNode {
702 7369 : num_children: 0,
703 7369 : level,
704 7369 : prefix: Vec::new(),
705 7369 : suffix_len: 0,
706 7369 : keys: Vec::new(),
707 7369 : values: Vec::new(),
708 7369 : size: NODE_HDR_SIZE,
709 7369 : }
710 7369 : }
711 :
712 : /// Try to append a key-value pair to this node. Returns 'true' on
713 : /// success, 'false' if the page was full or the key was
714 : /// incompatible with the prefix of the existing keys.
715 3392442 : fn push(&mut self, key: &[u8; L], value: Value) -> bool {
716 : // If we have already performed prefix-compression on the page,
717 : // check that the incoming key has the same prefix.
718 3392442 : if self.num_children > 0 {
719 : // does the prefix allow it?
720 3385237 : if !key.starts_with(&self.prefix) {
721 104 : return false;
722 3385133 : }
723 7205 : } else {
724 7205 : self.suffix_len = key.len();
725 7205 : }
726 :
727 : // Is the node too full?
728 3392338 : if self.size + self.suffix_len + VALUE_SZ >= NODE_SIZE {
729 11755 : return false;
730 3380583 : }
731 :
732 : // All clear
733 3380583 : self.num_children += 1;
734 3380583 : self.keys.extend(&key[self.prefix.len()..]);
735 3380583 : self.values.extend(value.0);
736 :
737 3380583 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
738 3380583 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
739 :
740 3380583 : self.size += self.suffix_len + VALUE_SZ;
741 :
742 3380583 : true
743 3392442 : }
744 :
745 : ///
746 : /// Perform prefix-compression.
747 : ///
748 : /// Returns 'true' on success, 'false' if no compression was possible.
749 : ///
750 11858 : fn compress(&mut self) -> bool {
751 11858 : let first_suffix = self.first_suffix();
752 11858 : let last_suffix = self.last_suffix();
753 :
754 : // Find the common prefix among all keys
755 11858 : let mut prefix_len = 0;
756 107947 : while prefix_len < self.suffix_len {
757 107947 : if first_suffix[prefix_len] != last_suffix[prefix_len] {
758 11858 : break;
759 96089 : }
760 96089 : prefix_len += 1;
761 : }
762 11858 : if prefix_len == 0 {
763 5834 : return false;
764 6024 : }
765 :
766 : // Can compress. Rewrite the keys without the common prefix.
767 6024 : self.prefix.extend(&self.keys[..prefix_len]);
768 :
769 6024 : let mut new_keys = Vec::new();
770 6024 : let mut key_off = 0;
771 1611489 : while key_off < self.keys.len() {
772 1605465 : let next_key_off = key_off + self.suffix_len;
773 1605465 : new_keys.extend(&self.keys[key_off + prefix_len..next_key_off]);
774 1605465 : key_off = next_key_off;
775 1605465 : }
776 6024 : self.keys = new_keys;
777 6024 : self.suffix_len -= prefix_len;
778 :
779 6024 : self.size -= prefix_len * self.num_children as usize;
780 6024 : self.size += prefix_len;
781 :
782 6024 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
783 6024 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
784 :
785 6024 : true
786 11858 : }
787 :
788 : ///
789 : /// Serialize the node to on-disk format.
790 : ///
791 7180 : fn pack(&self) -> IoBuffer {
792 7180 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
793 7180 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
794 7180 : assert!(self.num_children > 0);
795 :
796 7180 : let mut buf = IoBufferMut::with_capacity(PAGE_SZ);
797 :
798 7180 : buf.put_u16(self.num_children);
799 7180 : buf.put_u8(self.level);
800 7180 : buf.put_u8(self.prefix.len() as u8);
801 7180 : buf.put_u8(self.suffix_len as u8);
802 7180 : buf.put(&self.prefix[..]);
803 7180 : buf.put(&self.keys[..]);
804 7180 : buf.put(&self.values[..]);
805 :
806 7180 : assert!(buf.len() == self.size);
807 :
808 7180 : assert!(buf.len() <= PAGE_SZ);
809 7180 : buf.extend_with(0, PAGE_SZ - buf.len());
810 7180 : buf.freeze()
811 7180 : }
812 :
813 18104 : fn first_suffix(&self) -> &[u8] {
814 18104 : &self.keys[..self.suffix_len]
815 18104 : }
816 11858 : fn last_suffix(&self) -> &[u8] {
817 11858 : &self.keys[self.keys.len() - self.suffix_len..]
818 11858 : }
819 :
820 : /// Return the full first key of the page, including the prefix
821 6246 : fn first_key(&self) -> [u8; L] {
822 6246 : let mut key = [0u8; L];
823 6246 : key[..self.prefix.len()].copy_from_slice(&self.prefix);
824 6246 : key[self.prefix.len()..].copy_from_slice(self.first_suffix());
825 6246 : key
826 6246 : }
827 : }
828 :
829 : #[cfg(test)]
830 : pub(crate) mod tests {
831 : use std::collections::BTreeMap;
832 : use std::sync::atomic::{AtomicUsize, Ordering};
833 :
834 : use rand::Rng;
835 :
836 : use super::*;
837 : use crate::context::DownloadBehavior;
838 : use crate::task_mgr::TaskKind;
839 : use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReaderRef};
840 :
841 : #[derive(Clone, Default)]
842 : pub(crate) struct TestDisk {
843 : blocks: Vec<IoBuffer>,
844 : }
845 : impl TestDisk {
846 5 : fn new() -> Self {
847 5 : Self::default()
848 5 : }
849 508346 : pub(crate) fn read_blk(&self, blknum: u32) -> io::Result<BlockLease> {
850 508346 : let mut buf = [0u8; PAGE_SZ];
851 508346 : buf.copy_from_slice(&self.blocks[blknum as usize]);
852 508346 : Ok(std::sync::Arc::new(buf).into())
853 508346 : }
854 : }
855 : impl BlockReader for TestDisk {
856 205593 : fn block_cursor(&self) -> BlockCursor<'_> {
857 205593 : BlockCursor::new(BlockReaderRef::TestDisk(self))
858 205593 : }
859 : }
860 : impl BlockWriter for &mut TestDisk {
861 107 : fn write_blk(&mut self, buf: IoBuffer) -> io::Result<u32> {
862 107 : let blknum = self.blocks.len();
863 107 : self.blocks.push(buf);
864 107 : Ok(blknum as u32)
865 107 : }
866 : }
867 :
868 : #[tokio::test]
869 1 : async fn basic() -> Result<()> {
870 1 : let mut disk = TestDisk::new();
871 1 : let mut writer = DiskBtreeBuilder::<_, 6>::new(&mut disk);
872 :
873 1 : let ctx =
874 1 : RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error).with_scope_unit_test();
875 :
876 1 : let all_keys: Vec<&[u8; 6]> = vec![
877 1 : b"xaaaaa", b"xaaaba", b"xaaaca", b"xabaaa", b"xababa", b"xabaca", b"xabada", b"xabadb",
878 : ];
879 1 : let all_data: Vec<(&[u8; 6], u64)> = all_keys
880 1 : .iter()
881 1 : .enumerate()
882 8 : .map(|(idx, key)| (*key, idx as u64))
883 1 : .collect();
884 8 : for (key, val) in all_data.iter() {
885 8 : writer.append(key, *val)?;
886 : }
887 :
888 1 : let (root_offset, _writer) = writer.finish()?;
889 :
890 1 : let reader = DiskBtreeReader::new(0, root_offset, disk);
891 :
892 1 : reader.dump(&ctx).await?;
893 :
894 : // Test the `get` function on all the keys.
895 8 : for (key, val) in all_data.iter() {
896 8 : assert_eq!(reader.get(key, &ctx).await?, Some(*val));
897 : }
898 : // And on some keys that don't exist
899 1 : assert_eq!(reader.get(b"aaaaaa", &ctx).await?, None);
900 1 : assert_eq!(reader.get(b"zzzzzz", &ctx).await?, None);
901 1 : assert_eq!(reader.get(b"xaaabx", &ctx).await?, None);
902 :
903 : // Test search with `visit` function
904 1 : let search_key = b"xabaaa";
905 1 : let expected: Vec<(Vec<u8>, u64)> = all_data
906 1 : .iter()
907 8 : .filter(|(key, _value)| key[..] >= search_key[..])
908 5 : .map(|(key, value)| (key.to_vec(), *value))
909 1 : .collect();
910 :
911 1 : let mut data = Vec::new();
912 1 : reader
913 1 : .visit(
914 1 : search_key,
915 1 : VisitDirection::Forwards,
916 5 : |key, value| {
917 5 : data.push((key.to_vec(), value));
918 5 : true
919 5 : },
920 1 : &ctx,
921 : )
922 1 : .await?;
923 1 : assert_eq!(data, expected);
924 :
925 : // Test a backwards scan
926 1 : let mut expected: Vec<(Vec<u8>, u64)> = all_data
927 1 : .iter()
928 8 : .filter(|(key, _value)| key[..] <= search_key[..])
929 4 : .map(|(key, value)| (key.to_vec(), *value))
930 1 : .collect();
931 1 : expected.reverse();
932 1 : let mut data = Vec::new();
933 1 : reader
934 1 : .visit(
935 1 : search_key,
936 1 : VisitDirection::Backwards,
937 4 : |key, value| {
938 4 : data.push((key.to_vec(), value));
939 4 : true
940 4 : },
941 1 : &ctx,
942 : )
943 1 : .await?;
944 1 : assert_eq!(data, expected);
945 :
946 : // Backward scan where nothing matches
947 1 : reader
948 1 : .visit(
949 1 : b"aaaaaa",
950 1 : VisitDirection::Backwards,
951 0 : |key, value| {
952 0 : panic!("found unexpected key {}: {}", hex::encode(key), value);
953 : },
954 1 : &ctx,
955 : )
956 1 : .await?;
957 :
958 : // Full scan
959 1 : let expected: Vec<(Vec<u8>, u64)> = all_data
960 1 : .iter()
961 8 : .map(|(key, value)| (key.to_vec(), *value))
962 1 : .collect();
963 1 : let mut data = Vec::new();
964 1 : reader
965 1 : .visit(
966 1 : &[0u8; 6],
967 1 : VisitDirection::Forwards,
968 8 : |key, value| {
969 8 : data.push((key.to_vec(), value));
970 8 : true
971 8 : },
972 1 : &ctx,
973 : )
974 1 : .await?;
975 1 : assert_eq!(data, expected);
976 :
977 2 : Ok(())
978 1 : }
979 :
980 : #[tokio::test]
981 1 : async fn lots_of_keys() -> Result<()> {
982 1 : let mut disk = TestDisk::new();
983 1 : let mut writer = DiskBtreeBuilder::<_, 8>::new(&mut disk);
984 1 : let ctx =
985 1 : RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error).with_scope_unit_test();
986 :
987 : const NUM_KEYS: u64 = 1000;
988 :
989 1 : let mut all_data: BTreeMap<u64, u64> = BTreeMap::new();
990 :
991 1001 : for idx in 0..NUM_KEYS {
992 1000 : let key_int: u64 = 1 + idx * 2;
993 1000 : let key = u64::to_be_bytes(key_int);
994 1000 : writer.append(&key, idx)?;
995 :
996 1000 : all_data.insert(key_int, idx);
997 : }
998 :
999 1 : let (root_offset, _writer) = writer.finish()?;
1000 :
1001 1 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1002 :
1003 1 : reader.dump(&ctx).await?;
1004 :
1005 : use std::sync::Mutex;
1006 :
1007 1 : let result = Mutex::new(Vec::new());
1008 1 : let limit: AtomicUsize = AtomicUsize::new(10);
1009 41910 : let take_ten = |key: &[u8], value: u64| {
1010 41910 : let mut keybuf = [0u8; 8];
1011 41910 : keybuf.copy_from_slice(key);
1012 41910 : let key_int = u64::from_be_bytes(keybuf);
1013 :
1014 41910 : let mut result = result.lock().unwrap();
1015 41910 : result.push((key_int, value));
1016 :
1017 : // keep going until we have 10 matches
1018 41910 : result.len() < limit.load(Ordering::Relaxed)
1019 41910 : };
1020 :
1021 2010 : for search_key_int in 0..(NUM_KEYS * 2 + 10) {
1022 2010 : let search_key = u64::to_be_bytes(search_key_int);
1023 2010 : assert_eq!(
1024 2010 : reader.get(&search_key, &ctx).await?,
1025 2010 : all_data.get(&search_key_int).cloned()
1026 : );
1027 :
1028 : // Test a forward scan starting with this key
1029 2010 : result.lock().unwrap().clear();
1030 2010 : reader
1031 2010 : .visit(&search_key, VisitDirection::Forwards, take_ten, &ctx)
1032 2010 : .await?;
1033 2010 : let expected = all_data
1034 2010 : .range(search_key_int..)
1035 2010 : .take(10)
1036 19910 : .map(|(&key, &val)| (key, val))
1037 2010 : .collect::<Vec<(u64, u64)>>();
1038 2010 : assert_eq!(*result.lock().unwrap(), expected);
1039 :
1040 : // And a backwards scan
1041 2010 : result.lock().unwrap().clear();
1042 2010 : reader
1043 2010 : .visit(&search_key, VisitDirection::Backwards, take_ten, &ctx)
1044 2010 : .await?;
1045 2010 : let expected = all_data
1046 2010 : .range(..=search_key_int)
1047 2010 : .rev()
1048 2010 : .take(10)
1049 20000 : .map(|(&key, &val)| (key, val))
1050 2010 : .collect::<Vec<(u64, u64)>>();
1051 2010 : assert_eq!(*result.lock().unwrap(), expected);
1052 : }
1053 :
1054 : // full scan
1055 1 : let search_key = u64::to_be_bytes(0);
1056 1 : limit.store(usize::MAX, Ordering::Relaxed);
1057 1 : result.lock().unwrap().clear();
1058 1 : reader
1059 1 : .visit(&search_key, VisitDirection::Forwards, take_ten, &ctx)
1060 1 : .await?;
1061 1 : let expected = all_data
1062 1 : .iter()
1063 1000 : .map(|(&key, &val)| (key, val))
1064 1 : .collect::<Vec<(u64, u64)>>();
1065 1 : assert_eq!(*result.lock().unwrap(), expected);
1066 :
1067 : // full scan
1068 1 : let search_key = u64::to_be_bytes(u64::MAX);
1069 1 : limit.store(usize::MAX, Ordering::Relaxed);
1070 1 : result.lock().unwrap().clear();
1071 1 : reader
1072 1 : .visit(&search_key, VisitDirection::Backwards, take_ten, &ctx)
1073 1 : .await?;
1074 1 : let expected = all_data
1075 1 : .iter()
1076 1 : .rev()
1077 1000 : .map(|(&key, &val)| (key, val))
1078 1 : .collect::<Vec<(u64, u64)>>();
1079 1 : assert_eq!(*result.lock().unwrap(), expected);
1080 :
1081 2 : Ok(())
1082 1 : }
1083 :
1084 : #[tokio::test]
1085 1 : async fn random_data() -> Result<()> {
1086 1 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1087 :
1088 : // Generate random keys with exponential distribution, to
1089 : // exercise the prefix compression
1090 : const NUM_KEYS: usize = 100000;
1091 1 : let mut all_data: BTreeMap<u128, u64> = BTreeMap::new();
1092 100001 : for idx in 0..NUM_KEYS {
1093 100000 : let u: f64 = rand::thread_rng().gen_range(0.0..1.0);
1094 100000 : let t = -(f64::ln(u));
1095 100000 : let key_int = (t * 1000000.0) as u128;
1096 100000 :
1097 100000 : all_data.insert(key_int, idx as u64);
1098 100000 : }
1099 :
1100 : // Build a tree from it
1101 1 : let mut disk = TestDisk::new();
1102 1 : let mut writer = DiskBtreeBuilder::<_, 16>::new(&mut disk);
1103 :
1104 97537 : for (&key, &val) in all_data.iter() {
1105 97537 : writer.append(&u128::to_be_bytes(key), val)?;
1106 : }
1107 1 : let (root_offset, _writer) = writer.finish()?;
1108 :
1109 1 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1110 :
1111 : // Test get() operation on all the keys
1112 97537 : for (&key, &val) in all_data.iter() {
1113 97537 : let search_key = u128::to_be_bytes(key);
1114 97537 : assert_eq!(reader.get(&search_key, &ctx).await?, Some(val));
1115 : }
1116 :
1117 : // Test get() operations on random keys, most of which will not exist
1118 100001 : for _ in 0..100000 {
1119 100000 : let key_int = rand::thread_rng().r#gen::<u128>();
1120 100000 : let search_key = u128::to_be_bytes(key_int);
1121 100000 : assert!(reader.get(&search_key, &ctx).await? == all_data.get(&key_int).cloned());
1122 : }
1123 :
1124 : // Test boundary cases
1125 1 : assert!(
1126 1 : reader.get(&u128::to_be_bytes(u128::MIN), &ctx).await?
1127 1 : == all_data.get(&u128::MIN).cloned()
1128 : );
1129 1 : assert!(
1130 1 : reader.get(&u128::to_be_bytes(u128::MAX), &ctx).await?
1131 1 : == all_data.get(&u128::MAX).cloned()
1132 : );
1133 :
1134 : // Test iterator and get_stream API
1135 1 : let mut iter = reader.iter(&[0; 16], &ctx);
1136 1 : let mut cnt = 0;
1137 97538 : while let Some(res) = iter.next().await {
1138 97537 : let (key, val) = res?;
1139 97537 : let key = u128::from_be_bytes(key.as_slice().try_into().unwrap());
1140 97537 : assert_eq!(val, *all_data.get(&key).unwrap());
1141 97537 : cnt += 1;
1142 : }
1143 1 : assert_eq!(cnt, all_data.len());
1144 :
1145 2 : Ok(())
1146 1 : }
1147 :
1148 : #[test]
1149 1 : fn unsorted_input() {
1150 1 : let mut disk = TestDisk::new();
1151 1 : let mut writer = DiskBtreeBuilder::<_, 2>::new(&mut disk);
1152 :
1153 1 : let _ = writer.append(b"ba", 1);
1154 1 : let _ = writer.append(b"bb", 2);
1155 1 : let err = writer.append(b"aa", 3).expect_err("should've failed");
1156 1 : match err {
1157 1 : DiskBtreeError::UnsortedInput { key, last_key } => {
1158 1 : assert_eq!(key.as_ref(), b"aa".as_slice());
1159 1 : assert_eq!(last_key.as_ref(), b"bb".as_slice());
1160 : }
1161 0 : _ => panic!("unexpected error variant, expected DiskBtreeError::UnsortedInput"),
1162 : }
1163 1 : }
1164 :
1165 : ///
1166 : /// This test contains a particular data set, see disk_btree_test_data.rs
1167 : ///
1168 : #[tokio::test]
1169 1 : async fn particular_data() -> Result<()> {
1170 : // Build a tree from it
1171 1 : let mut disk = TestDisk::new();
1172 1 : let mut writer = DiskBtreeBuilder::<_, 26>::new(&mut disk);
1173 1 : let ctx =
1174 1 : RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error).with_scope_unit_test();
1175 :
1176 2001 : for (key, val) in disk_btree_test_data::TEST_DATA {
1177 2000 : writer.append(&key, val)?;
1178 : }
1179 1 : let (root_offset, writer) = writer.finish()?;
1180 :
1181 1 : println!("SIZE: {} blocks", writer.blocks.len());
1182 :
1183 1 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1184 :
1185 : // Test get() operation on all the keys
1186 2001 : for (key, val) in disk_btree_test_data::TEST_DATA {
1187 2000 : assert_eq!(reader.get(&key, &ctx).await?, Some(val));
1188 : }
1189 :
1190 : // Test full scan
1191 1 : let mut count = 0;
1192 1 : reader
1193 1 : .visit(
1194 1 : &[0u8; 26],
1195 1 : VisitDirection::Forwards,
1196 2000 : |_key, _value| {
1197 2000 : count += 1;
1198 2000 : true
1199 2000 : },
1200 1 : &ctx,
1201 : )
1202 1 : .await?;
1203 1 : assert_eq!(count, disk_btree_test_data::TEST_DATA.len());
1204 :
1205 1 : reader.dump(&ctx).await?;
1206 :
1207 2 : Ok(())
1208 1 : }
1209 : }
1210 :
1211 : #[cfg(test)]
1212 : #[path = "disk_btree_test_data.rs"]
1213 : mod disk_btree_test_data;
|
