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 async_stream::try_stream;
22 : use byteorder::{ReadBytesExt, BE};
23 : use bytes::{BufMut, Bytes, BytesMut};
24 : use either::Either;
25 : use futures::{Stream, StreamExt};
26 : use hex;
27 : use std::{
28 : cmp::Ordering,
29 : io,
30 : iter::Rev,
31 : ops::{Range, RangeInclusive},
32 : result,
33 : };
34 : use thiserror::Error;
35 : use tracing::error;
36 :
37 : use crate::{
38 : context::{DownloadBehavior, RequestContext},
39 : task_mgr::TaskKind,
40 : tenant::block_io::{BlockReader, BlockWriter},
41 : };
42 :
43 : // The maximum size of a value stored in the B-tree. 5 bytes is enough currently.
44 : pub const VALUE_SZ: usize = 5;
45 : pub const MAX_VALUE: u64 = 0x007f_ffff_ffff;
46 :
47 : pub const PAGE_SZ: usize = 8192;
48 :
49 : #[derive(Clone, Copy, Debug)]
50 : struct Value([u8; VALUE_SZ]);
51 :
52 : impl Value {
53 14125721 : fn from_slice(slice: &[u8]) -> Value {
54 14125721 : let mut b = [0u8; VALUE_SZ];
55 14125721 : b.copy_from_slice(slice);
56 14125721 : Value(b)
57 14125721 : }
58 :
59 14438455 : fn from_u64(x: u64) -> Value {
60 14438455 : assert!(x <= 0x007f_ffff_ffff);
61 14438455 : Value([
62 14438455 : (x >> 32) as u8,
63 14438455 : (x >> 24) as u8,
64 14438455 : (x >> 16) as u8,
65 14438455 : (x >> 8) as u8,
66 14438455 : x as u8,
67 14438455 : ])
68 14438455 : }
69 :
70 25749 : fn from_blknum(x: u32) -> Value {
71 25749 : Value([
72 25749 : 0x80,
73 25749 : (x >> 24) as u8,
74 25749 : (x >> 16) as u8,
75 25749 : (x >> 8) as u8,
76 25749 : x as u8,
77 25749 : ])
78 25749 : }
79 :
80 : #[allow(dead_code)]
81 0 : fn is_offset(self) -> bool {
82 0 : self.0[0] & 0x80 != 0
83 0 : }
84 :
85 12863217 : fn to_u64(self) -> u64 {
86 12863217 : let b = &self.0;
87 12863217 : ((b[0] as u64) << 32)
88 12863217 : | ((b[1] as u64) << 24)
89 12863217 : | ((b[2] as u64) << 16)
90 12863217 : | ((b[3] as u64) << 8)
91 12863217 : | b[4] as u64
92 12863217 : }
93 :
94 1250456 : fn to_blknum(self) -> u32 {
95 1250456 : let b = &self.0;
96 1250456 : assert!(b[0] == 0x80);
97 1250456 : ((b[1] as u32) << 24) | ((b[2] as u32) << 16) | ((b[3] as u32) << 8) | b[4] as u32
98 1250456 : }
99 : }
100 :
101 : #[derive(Error, Debug)]
102 : pub enum DiskBtreeError {
103 : #[error("Attempt to append a value that is too large {0} > {}", MAX_VALUE)]
104 : AppendOverflow(u64),
105 :
106 : #[error("Unsorted input: key {key:?} is <= last_key {last_key:?}")]
107 : UnsortedInput { key: Box<[u8]>, last_key: Box<[u8]> },
108 :
109 : #[error("Could not push to new leaf node")]
110 : FailedToPushToNewLeafNode,
111 :
112 : #[error("IoError: {0}")]
113 : Io(#[from] io::Error),
114 : }
115 :
116 : pub type Result<T> = result::Result<T, DiskBtreeError>;
117 :
118 : /// This is the on-disk representation.
119 : struct OnDiskNode<'a, const L: usize> {
120 : // Fixed-width fields
121 : num_children: u16,
122 : level: u8,
123 : prefix_len: u8,
124 : suffix_len: u8,
125 :
126 : // Variable-length fields. These are stored on-disk after the fixed-width
127 : // fields, in this order. In the in-memory representation, these point to
128 : // the right parts in the page buffer.
129 : prefix: &'a [u8],
130 : keys: &'a [u8],
131 : values: &'a [u8],
132 : }
133 :
134 : impl<const L: usize> OnDiskNode<'_, L> {
135 : ///
136 : /// Interpret a PAGE_SZ page as a node.
137 : ///
138 3003180 : fn deparse(buf: &[u8]) -> Result<OnDiskNode<L>> {
139 3003180 : let mut cursor = std::io::Cursor::new(buf);
140 3003180 : let num_children = cursor.read_u16::<BE>()?;
141 3003180 : let level = cursor.read_u8()?;
142 3003180 : let prefix_len = cursor.read_u8()?;
143 3003180 : let suffix_len = cursor.read_u8()?;
144 :
145 3003180 : let mut off = cursor.position();
146 3003180 : let prefix_off = off as usize;
147 3003180 : off += prefix_len as u64;
148 3003180 :
149 3003180 : let keys_off = off as usize;
150 3003180 : let keys_len = num_children as usize * suffix_len as usize;
151 3003180 : off += keys_len as u64;
152 3003180 :
153 3003180 : let values_off = off as usize;
154 3003180 : let values_len = num_children as usize * VALUE_SZ;
155 3003180 : //off += values_len as u64;
156 3003180 :
157 3003180 : let prefix = &buf[prefix_off..prefix_off + prefix_len as usize];
158 3003180 : let keys = &buf[keys_off..keys_off + keys_len];
159 3003180 : let values = &buf[values_off..values_off + values_len];
160 3003180 :
161 3003180 : Ok(OnDiskNode {
162 3003180 : num_children,
163 3003180 : level,
164 3003180 : prefix_len,
165 3003180 : suffix_len,
166 3003180 : prefix,
167 3003180 : keys,
168 3003180 : values,
169 3003180 : })
170 3003180 : }
171 :
172 : ///
173 : /// Read a value at 'idx'
174 : ///
175 14125721 : fn value(&self, idx: usize) -> Value {
176 14125721 : let value_off = idx * VALUE_SZ;
177 14125721 : let value_slice = &self.values[value_off..value_off + VALUE_SZ];
178 14125721 : Value::from_slice(value_slice)
179 14125721 : }
180 :
181 2555571 : fn binary_search(
182 2555571 : &self,
183 2555571 : search_key: &[u8; L],
184 2555571 : keybuf: &mut [u8],
185 2555571 : ) -> result::Result<usize, usize> {
186 2555571 : let mut size = self.num_children as usize;
187 2555571 : let mut low = 0;
188 2555571 : let mut high = size;
189 19633316 : while low < high {
190 17534539 : let mid = low + size / 2;
191 17534539 :
192 17534539 : let key_off = mid * self.suffix_len as usize;
193 17534539 : let suffix = &self.keys[key_off..key_off + self.suffix_len as usize];
194 17534539 : // Does this match?
195 17534539 : keybuf[self.prefix_len as usize..].copy_from_slice(suffix);
196 17534539 :
197 17534539 : let cmp = keybuf[..].cmp(search_key);
198 17534539 :
199 17534539 : if cmp == Ordering::Less {
200 11249345 : low = mid + 1;
201 11249345 : } else if cmp == Ordering::Greater {
202 5828400 : high = mid;
203 5828400 : } else {
204 456794 : return Ok(mid);
205 : }
206 17077745 : size = high - low;
207 : }
208 2098777 : Err(low)
209 2555571 : }
210 : }
211 :
212 : ///
213 : /// Public reader object, to search the tree.
214 : ///
215 : #[derive(Clone)]
216 : pub struct DiskBtreeReader<R, const L: usize>
217 : where
218 : R: BlockReader,
219 : {
220 : start_blk: u32,
221 : root_blk: u32,
222 : reader: R,
223 : }
224 :
225 : #[derive(Clone, Copy, Debug, PartialEq, Eq)]
226 : pub enum VisitDirection {
227 : Forwards,
228 : Backwards,
229 : }
230 :
231 : impl<R, const L: usize> DiskBtreeReader<R, L>
232 : where
233 : R: BlockReader,
234 : {
235 482148 : pub fn new(start_blk: u32, root_blk: u32, reader: R) -> Self {
236 482148 : DiskBtreeReader {
237 482148 : start_blk,
238 482148 : root_blk,
239 482148 : reader,
240 482148 : }
241 482148 : }
242 :
243 : ///
244 : /// Read the value for given key. Returns the value, or None if it doesn't exist.
245 : ///
246 806415 : pub async fn get(&self, search_key: &[u8; L], ctx: &RequestContext) -> Result<Option<u64>> {
247 806415 : let mut result: Option<u64> = None;
248 806415 : self.visit(
249 806415 : search_key,
250 806415 : VisitDirection::Forwards,
251 806415 : |key, value| {
252 406367 : if key == search_key {
253 402355 : result = Some(value);
254 402355 : }
255 406367 : false
256 806415 : },
257 806415 : ctx,
258 806415 : )
259 806415 : .await?;
260 806415 : Ok(result)
261 806415 : }
262 :
263 1396 : pub fn iter<'a>(self, start_key: &'a [u8; L], ctx: &'a RequestContext) -> DiskBtreeIterator<'a>
264 1396 : where
265 1396 : R: 'a + Send,
266 1396 : {
267 1396 : DiskBtreeIterator {
268 1396 : stream: Box::pin(self.into_stream(start_key, ctx)),
269 1396 : }
270 1396 : }
271 :
272 : /// Return a stream which yields all key, value pairs from the index
273 : /// starting from the first key greater or equal to `start_key`.
274 : ///
275 : /// Note 1: that this is a copy of [`Self::visit`].
276 : /// TODO: Once the sequential read path is removed this will become
277 : /// the only index traversal method.
278 : ///
279 : /// Note 2: this function used to take `&self` but it now consumes `self`. This is due to
280 : /// the lifetime constraints of the reader and the stream / iterator it creates. Using `&self`
281 : /// requires the reader to be present when the stream is used, and this creates a lifetime
282 : /// dependency between the reader and the stream. Now if we want to create an iterator that
283 : /// holds the stream, someone will need to keep a reference to the reader, which is inconvenient
284 : /// to use from the image/delta layer APIs.
285 : ///
286 : /// Feel free to add the `&self` variant back if it's necessary.
287 481796 : pub fn into_stream<'a>(
288 481796 : self,
289 481796 : start_key: &'a [u8; L],
290 481796 : ctx: &'a RequestContext,
291 481796 : ) -> impl Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a
292 481796 : where
293 481796 : R: 'a,
294 481796 : {
295 481796 : try_stream! {
296 481796 : let mut stack = Vec::new();
297 481796 : stack.push((self.root_blk, None));
298 481796 : let block_cursor = self.reader.block_cursor();
299 481796 : let mut node_buf = [0_u8; PAGE_SZ];
300 481796 : while let Some((node_blknum, opt_iter)) = stack.pop() {
301 481796 : // Read the node, through the PS PageCache, into local variable `node_buf`.
302 481796 : // We could keep the page cache read guard alive, but, at the time of writing,
303 481796 : // we run quite small PS PageCache s => can't risk running out of
304 481796 : // PageCache space because this stream isn't consumed fast enough.
305 481796 : let page_read_guard = block_cursor
306 481796 : .read_blk(self.start_blk + node_blknum, ctx)
307 481796 : .await?;
308 481796 : node_buf.copy_from_slice(page_read_guard.as_ref());
309 481796 : drop(page_read_guard); // drop page cache read guard early
310 481796 :
311 481796 : let node = OnDiskNode::deparse(&node_buf)?;
312 481796 : let prefix_len = node.prefix_len as usize;
313 481796 : let suffix_len = node.suffix_len as usize;
314 481796 :
315 481796 : assert!(node.num_children > 0);
316 481796 :
317 481796 : let mut keybuf = Vec::new();
318 481796 : keybuf.extend(node.prefix);
319 481796 : keybuf.resize(prefix_len + suffix_len, 0);
320 481796 :
321 481796 : let mut iter: Either<Range<usize>, Rev<RangeInclusive<usize>>> = if let Some(iter) = opt_iter {
322 481796 : iter
323 481796 : } else {
324 481796 : // Locate the first match
325 481796 : let idx = match node.binary_search(start_key, keybuf.as_mut_slice()) {
326 481796 : Ok(idx) => idx,
327 481796 : Err(idx) => {
328 481796 : if node.level == 0 {
329 481796 : // Imagine that the node contains the following keys:
330 481796 : //
331 481796 : // 1
332 481796 : // 3 <-- idx
333 481796 : // 5
334 481796 : //
335 481796 : // If the search key is '2' and there is exact match,
336 481796 : // the binary search would return the index of key
337 481796 : // '3'. That's cool, '3' is the first key to return.
338 481796 : idx
339 481796 : } else {
340 481796 : // This is an internal page, so each key represents a lower
341 481796 : // bound for what's in the child page. If there is no exact
342 481796 : // match, we have to return the *previous* entry.
343 481796 : //
344 481796 : // 1 <-- return this
345 481796 : // 3 <-- idx
346 481796 : // 5
347 481796 : idx.saturating_sub(1)
348 481796 : }
349 481796 : }
350 481796 : };
351 481796 : Either::Left(idx..node.num_children.into())
352 481796 : };
353 481796 :
354 481796 :
355 481796 : // idx points to the first match now. Keep going from there
356 481796 : while let Some(idx) = iter.next() {
357 481796 : let key_off = idx * suffix_len;
358 481796 : let suffix = &node.keys[key_off..key_off + suffix_len];
359 481796 : keybuf[prefix_len..].copy_from_slice(suffix);
360 481796 : let value = node.value(idx);
361 481796 : #[allow(clippy::collapsible_if)]
362 481796 : if node.level == 0 {
363 481796 : // leaf
364 481796 : yield (keybuf.clone(), value.to_u64());
365 481796 : } else {
366 481796 : stack.push((node_blknum, Some(iter)));
367 481796 : stack.push((value.to_blknum(), None));
368 481796 : break;
369 481796 : }
370 481796 : }
371 481796 : }
372 481796 : }
373 481796 : }
374 :
375 : ///
376 : /// Scan the tree, starting from 'search_key', in the given direction. 'visitor'
377 : /// will be called for every key >= 'search_key' (or <= 'search_key', if scanning
378 : /// backwards)
379 : ///
380 823335 : pub async fn visit<V>(
381 823335 : &self,
382 823335 : search_key: &[u8; L],
383 823335 : dir: VisitDirection,
384 823335 : mut visitor: V,
385 823335 : ctx: &RequestContext,
386 823335 : ) -> Result<bool>
387 823335 : where
388 823335 : V: FnMut(&[u8], u64) -> bool,
389 823335 : {
390 823335 : let mut stack = Vec::new();
391 823335 : stack.push((self.root_blk, None));
392 823335 : let block_cursor = self.reader.block_cursor();
393 2438322 : while let Some((node_blknum, opt_iter)) = stack.pop() {
394 : // Locate the node.
395 2037254 : let node_buf = block_cursor
396 2037254 : .read_blk(self.start_blk + node_blknum, ctx)
397 2037254 : .await?;
398 :
399 2037254 : let node = OnDiskNode::deparse(node_buf.as_ref())?;
400 2037254 : let prefix_len = node.prefix_len as usize;
401 2037254 : let suffix_len = node.suffix_len as usize;
402 2037254 :
403 2037254 : assert!(node.num_children > 0);
404 :
405 2037254 : let mut keybuf = Vec::new();
406 2037254 : keybuf.extend(node.prefix);
407 2037254 : keybuf.resize(prefix_len + suffix_len, 0);
408 :
409 2037254 : let mut iter = if let Some(iter) = opt_iter {
410 407796 : iter
411 1629458 : } else if dir == VisitDirection::Forwards {
412 : // Locate the first match
413 1621406 : let idx = match node.binary_search(search_key, keybuf.as_mut_slice()) {
414 406770 : Ok(idx) => idx,
415 1214636 : Err(idx) => {
416 1214636 : if node.level == 0 {
417 : // Imagine that the node contains the following keys:
418 : //
419 : // 1
420 : // 3 <-- idx
421 : // 5
422 : //
423 : // If the search key is '2' and there is exact match,
424 : // the binary search would return the index of key
425 : // '3'. That's cool, '3' is the first key to return.
426 416232 : idx
427 : } else {
428 : // This is an internal page, so each key represents a lower
429 : // bound for what's in the child page. If there is no exact
430 : // match, we have to return the *previous* entry.
431 : //
432 : // 1 <-- return this
433 : // 3 <-- idx
434 : // 5
435 798404 : idx.saturating_sub(1)
436 : }
437 : }
438 : };
439 1621406 : Either::Left(idx..node.num_children.into())
440 : } else {
441 8052 : let idx = match node.binary_search(search_key, keybuf.as_mut_slice()) {
442 4004 : Ok(idx) => {
443 4004 : // Exact match. That's the first entry to return, and walk
444 4004 : // backwards from there.
445 4004 : idx
446 : }
447 4048 : Err(idx) => {
448 : // No exact match. The binary search returned the index of the
449 : // first key that's > search_key. Back off by one, and walk
450 : // backwards from there.
451 4048 : if let Some(idx) = idx.checked_sub(1) {
452 4040 : idx
453 : } else {
454 8 : return Ok(false);
455 : }
456 : }
457 : };
458 8044 : Either::Right((0..=idx).rev())
459 : };
460 :
461 : // idx points to the first match now. Keep going from there
462 6325154 : while let Some(idx) = iter.next() {
463 5516290 : let key_off = idx * suffix_len;
464 5516290 : let suffix = &node.keys[key_off..key_off + suffix_len];
465 5516290 : keybuf[prefix_len..].copy_from_slice(suffix);
466 5516290 : let value = node.value(idx);
467 5516290 : #[allow(clippy::collapsible_if)]
468 5516290 : if node.level == 0 {
469 : // leaf
470 4710167 : if !visitor(&keybuf, value.to_u64()) {
471 422259 : return Ok(false);
472 4287908 : }
473 : } else {
474 806123 : stack.push((node_blknum, Some(iter)));
475 806123 : stack.push((value.to_blknum(), None));
476 806123 : break;
477 : }
478 : }
479 : }
480 401068 : Ok(true)
481 823335 : }
482 :
483 : #[allow(dead_code)]
484 20 : pub async fn dump(&self) -> Result<()> {
485 20 : let mut stack = Vec::new();
486 20 : let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
487 20 :
488 20 : stack.push((self.root_blk, String::new(), 0, 0, 0));
489 20 :
490 20 : let block_cursor = self.reader.block_cursor();
491 :
492 12084 : while let Some((blknum, path, depth, child_idx, key_off)) = stack.pop() {
493 12064 : let blk = block_cursor.read_blk(self.start_blk + blknum, &ctx).await?;
494 12064 : let buf: &[u8] = blk.as_ref();
495 12064 : let node = OnDiskNode::<L>::deparse(buf)?;
496 :
497 12064 : if child_idx == 0 {
498 36 : print!("{:indent$}", "", indent = depth * 2);
499 36 : let path_prefix = stack
500 36 : .iter()
501 36 : .map(|(_blknum, path, ..)| path.as_str())
502 36 : .collect::<String>();
503 36 : println!(
504 36 : "blk #{blknum}: path {path_prefix}{path}: prefix {}, suffix_len {}",
505 36 : hex::encode(node.prefix),
506 36 : node.suffix_len
507 36 : );
508 12028 : }
509 :
510 12064 : if child_idx + 1 < node.num_children {
511 12028 : let key_off = key_off + node.suffix_len as usize;
512 12028 : stack.push((blknum, path.clone(), depth, child_idx + 1, key_off));
513 12028 : }
514 12064 : let key = &node.keys[key_off..key_off + node.suffix_len as usize];
515 12064 : let val = node.value(child_idx as usize);
516 12064 :
517 12064 : print!("{:indent$}", "", indent = depth * 2 + 2);
518 12064 : println!("{}: {}", hex::encode(key), hex::encode(val.0));
519 12064 :
520 12064 : if node.level > 0 {
521 16 : stack.push((val.to_blknum(), hex::encode(node.prefix), depth + 1, 0, 0));
522 12048 : }
523 : }
524 20 : Ok(())
525 20 : }
526 : }
527 :
528 : pub struct DiskBtreeIterator<'a> {
529 : #[allow(clippy::type_complexity)]
530 : stream: std::pin::Pin<
531 : Box<dyn Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a + Send>,
532 : >,
533 : }
534 :
535 : impl DiskBtreeIterator<'_> {
536 4647643 : pub async fn next(&mut self) -> Option<std::result::Result<(Vec<u8>, u64), DiskBtreeError>> {
537 4647643 : self.stream.next().await
538 4647643 : }
539 : }
540 :
541 : ///
542 : /// Public builder object, for creating a new tree.
543 : ///
544 : /// Usage: Create a builder object by calling 'new', load all the data into the
545 : /// tree by calling 'append' for each key-value pair, and then call 'finish'
546 : ///
547 : /// 'L' is the key length in bytes
548 : pub struct DiskBtreeBuilder<W, const L: usize>
549 : where
550 : W: BlockWriter,
551 : {
552 : writer: W,
553 :
554 : ///
555 : /// `stack[0]` is the current root page, `stack.last()` is the leaf.
556 : ///
557 : /// We maintain the length of the stack to be always greater than zero.
558 : /// Two exceptions are:
559 : /// 1. `Self::flush_node`. The method will push the new node if it extracted the last one.
560 : /// So because other methods cannot see the intermediate state invariant still holds.
561 : /// 2. `Self::finish`. It consumes self and does not return it back,
562 : /// which means that this is where the structure is destroyed.
563 : /// Thus stack of zero length cannot be observed by other methods.
564 : stack: Vec<BuildNode<L>>,
565 :
566 : /// Last key that was appended to the tree. Used to sanity check that append
567 : /// is called in increasing key order.
568 : last_key: Option<[u8; L]>,
569 : }
570 :
571 : impl<W, const L: usize> DiskBtreeBuilder<W, L>
572 : where
573 : W: BlockWriter,
574 : {
575 4140 : pub fn new(writer: W) -> Self {
576 4140 : DiskBtreeBuilder {
577 4140 : writer,
578 4140 : last_key: None,
579 4140 : stack: vec![BuildNode::new(0)],
580 4140 : }
581 4140 : }
582 :
583 14438459 : pub fn append(&mut self, key: &[u8; L], value: u64) -> Result<()> {
584 14438459 : if value > MAX_VALUE {
585 0 : return Err(DiskBtreeError::AppendOverflow(value));
586 14438459 : }
587 14438459 : if let Some(last_key) = &self.last_key {
588 14434743 : if key <= last_key {
589 4 : return Err(DiskBtreeError::UnsortedInput {
590 4 : key: key.as_slice().into(),
591 4 : last_key: last_key.as_slice().into(),
592 4 : });
593 14434739 : }
594 3716 : }
595 14438455 : self.last_key = Some(*key);
596 14438455 :
597 14438455 : self.append_internal(key, Value::from_u64(value))
598 14438459 : }
599 :
600 14464204 : fn append_internal(&mut self, key: &[u8; L], value: Value) -> Result<()> {
601 14464204 : // Try to append to the current leaf buffer
602 14464204 : let last = self
603 14464204 : .stack
604 14464204 : .last_mut()
605 14464204 : .expect("should always have at least one item");
606 14464204 : let level = last.level;
607 14464204 : if last.push(key, value) {
608 14415197 : return Ok(());
609 49007 : }
610 49007 :
611 49007 : // It did not fit. Try to compress, and if it succeeds to make
612 49007 : // some room on the node, try appending to it again.
613 49007 : #[allow(clippy::collapsible_if)]
614 49007 : if last.compress() {
615 24871 : if last.push(key, value) {
616 24854 : return Ok(());
617 17 : }
618 24136 : }
619 :
620 : // Could not append to the current leaf. Flush it and create a new one.
621 24153 : self.flush_node()?;
622 :
623 : // Replace the node we flushed with an empty one and append the new
624 : // key to it.
625 24153 : let mut last = BuildNode::new(level);
626 24153 : if !last.push(key, value) {
627 0 : return Err(DiskBtreeError::FailedToPushToNewLeafNode);
628 24153 : }
629 24153 :
630 24153 : self.stack.push(last);
631 24153 :
632 24153 : Ok(())
633 14464204 : }
634 :
635 : /// Flush the bottommost node in the stack to disk. Appends a downlink to its parent,
636 : /// and recursively flushes the parent too, if it becomes full. If the root page becomes full,
637 : /// creates a new root page, increasing the height of the tree.
638 25749 : fn flush_node(&mut self) -> Result<()> {
639 25749 : // Get the current bottommost node in the stack and flush it to disk.
640 25749 : let last = self
641 25749 : .stack
642 25749 : .pop()
643 25749 : .expect("should always have at least one item");
644 25749 : let buf = last.pack();
645 25749 : let downlink_key = last.first_key();
646 25749 : let downlink_ptr = self.writer.write_blk(buf)?;
647 :
648 : // Append the downlink to the parent. If there is no parent, ie. this was the root page,
649 : // create a new root page, increasing the height of the tree.
650 25749 : if self.stack.is_empty() {
651 1596 : self.stack.push(BuildNode::new(last.level + 1));
652 24153 : }
653 25749 : self.append_internal(&downlink_key, Value::from_blknum(downlink_ptr))
654 25749 : }
655 :
656 : ///
657 : /// Flushes everything to disk, and returns the block number of the root page.
658 : /// The caller must store the root block number "out-of-band", and pass it
659 : /// to the DiskBtreeReader::new() when you want to read the tree again.
660 : /// (In the image and delta layers, it is stored in the beginning of the file,
661 : /// in the summary header)
662 : ///
663 3620 : pub fn finish(mut self) -> Result<(u32, W)> {
664 : // flush all levels, except the root.
665 5216 : while self.stack.len() > 1 {
666 1596 : self.flush_node()?;
667 : }
668 :
669 3620 : let root = self
670 3620 : .stack
671 3620 : .first()
672 3620 : .expect("by the check above we left one item there");
673 3620 : let buf = root.pack();
674 3620 : let root_blknum = self.writer.write_blk(buf)?;
675 :
676 3620 : Ok((root_blknum, self.writer))
677 3620 : }
678 :
679 4095304 : pub fn borrow_writer(&self) -> &W {
680 4095304 : &self.writer
681 4095304 : }
682 : }
683 :
684 : ///
685 : /// BuildNode represesnts an incomplete page that we are appending to.
686 : ///
687 : #[derive(Clone, Debug)]
688 : struct BuildNode<const L: usize> {
689 : num_children: u16,
690 : level: u8,
691 : prefix: Vec<u8>,
692 : suffix_len: usize,
693 :
694 : keys: Vec<u8>,
695 : values: Vec<u8>,
696 :
697 : size: usize, // physical size of this node, if it was written to disk like this
698 : }
699 :
700 : const NODE_SIZE: usize = PAGE_SZ;
701 :
702 : const NODE_HDR_SIZE: usize = 2 + 1 + 1 + 1;
703 :
704 : impl<const L: usize> BuildNode<L> {
705 29889 : fn new(level: u8) -> Self {
706 29889 : BuildNode {
707 29889 : num_children: 0,
708 29889 : level,
709 29889 : prefix: Vec::new(),
710 29889 : suffix_len: 0,
711 29889 : keys: Vec::new(),
712 29889 : values: Vec::new(),
713 29889 : size: NODE_HDR_SIZE,
714 29889 : }
715 29889 : }
716 :
717 : /// Try to append a key-value pair to this node. Returns 'true' on
718 : /// success, 'false' if the page was full or the key was
719 : /// incompatible with the prefix of the existing keys.
720 14513228 : fn push(&mut self, key: &[u8; L], value: Value) -> bool {
721 14513228 : // If we have already performed prefix-compression on the page,
722 14513228 : // check that the incoming key has the same prefix.
723 14513228 : if self.num_children > 0 {
724 : // does the prefix allow it?
725 14483763 : if !key.starts_with(&self.prefix) {
726 415 : return false;
727 14483348 : }
728 29465 : } else {
729 29465 : self.suffix_len = key.len();
730 29465 : }
731 :
732 : // Is the node too full?
733 14512813 : if self.size + self.suffix_len + VALUE_SZ >= NODE_SIZE {
734 48609 : return false;
735 14464204 : }
736 14464204 :
737 14464204 : // All clear
738 14464204 : self.num_children += 1;
739 14464204 : self.keys.extend(&key[self.prefix.len()..]);
740 14464204 : self.values.extend(value.0);
741 14464204 :
742 14464204 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
743 14464204 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
744 :
745 14464204 : self.size += self.suffix_len + VALUE_SZ;
746 14464204 :
747 14464204 : true
748 14513228 : }
749 :
750 : ///
751 : /// Perform prefix-compression.
752 : ///
753 : /// Returns 'true' on success, 'false' if no compression was possible.
754 : ///
755 49007 : fn compress(&mut self) -> bool {
756 49007 : let first_suffix = self.first_suffix();
757 49007 : let last_suffix = self.last_suffix();
758 49007 :
759 49007 : // Find the common prefix among all keys
760 49007 : let mut prefix_len = 0;
761 445661 : while prefix_len < self.suffix_len {
762 445661 : if first_suffix[prefix_len] != last_suffix[prefix_len] {
763 49007 : break;
764 396654 : }
765 396654 : prefix_len += 1;
766 : }
767 49007 : if prefix_len == 0 {
768 24136 : return false;
769 24871 : }
770 24871 :
771 24871 : // Can compress. Rewrite the keys without the common prefix.
772 24871 : self.prefix.extend(&self.keys[..prefix_len]);
773 24871 :
774 24871 : let mut new_keys = Vec::new();
775 24871 : let mut key_off = 0;
776 6728646 : while key_off < self.keys.len() {
777 6703775 : let next_key_off = key_off + self.suffix_len;
778 6703775 : new_keys.extend(&self.keys[key_off + prefix_len..next_key_off]);
779 6703775 : key_off = next_key_off;
780 6703775 : }
781 24871 : self.keys = new_keys;
782 24871 : self.suffix_len -= prefix_len;
783 24871 :
784 24871 : self.size -= prefix_len * self.num_children as usize;
785 24871 : self.size += prefix_len;
786 24871 :
787 24871 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
788 24871 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
789 :
790 24871 : true
791 49007 : }
792 :
793 : ///
794 : /// Serialize the node to on-disk format.
795 : ///
796 29369 : fn pack(&self) -> Bytes {
797 29369 : assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
798 29369 : assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
799 29369 : assert!(self.num_children > 0);
800 :
801 29369 : let mut buf = BytesMut::new();
802 29369 :
803 29369 : buf.put_u16(self.num_children);
804 29369 : buf.put_u8(self.level);
805 29369 : buf.put_u8(self.prefix.len() as u8);
806 29369 : buf.put_u8(self.suffix_len as u8);
807 29369 : buf.put(&self.prefix[..]);
808 29369 : buf.put(&self.keys[..]);
809 29369 : buf.put(&self.values[..]);
810 29369 :
811 29369 : assert!(buf.len() == self.size);
812 :
813 29369 : assert!(buf.len() <= PAGE_SZ);
814 29369 : buf.resize(PAGE_SZ, 0);
815 29369 : buf.freeze()
816 29369 : }
817 :
818 74756 : fn first_suffix(&self) -> &[u8] {
819 74756 : &self.keys[..self.suffix_len]
820 74756 : }
821 49007 : fn last_suffix(&self) -> &[u8] {
822 49007 : &self.keys[self.keys.len() - self.suffix_len..]
823 49007 : }
824 :
825 : /// Return the full first key of the page, including the prefix
826 25749 : fn first_key(&self) -> [u8; L] {
827 25749 : let mut key = [0u8; L];
828 25749 : key[..self.prefix.len()].copy_from_slice(&self.prefix);
829 25749 : key[self.prefix.len()..].copy_from_slice(self.first_suffix());
830 25749 : key
831 25749 : }
832 : }
833 :
834 : #[cfg(test)]
835 : pub(crate) mod tests {
836 : use super::*;
837 : use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReaderRef};
838 : use rand::Rng;
839 : use std::collections::BTreeMap;
840 : use std::sync::atomic::{AtomicUsize, Ordering};
841 :
842 : #[derive(Clone, Default)]
843 : pub(crate) struct TestDisk {
844 : blocks: Vec<Bytes>,
845 : }
846 : impl TestDisk {
847 20 : fn new() -> Self {
848 20 : Self::default()
849 20 : }
850 2033740 : pub(crate) fn read_blk(&self, blknum: u32) -> io::Result<BlockLease> {
851 2033740 : let mut buf = [0u8; PAGE_SZ];
852 2033740 : buf.copy_from_slice(&self.blocks[blknum as usize]);
853 2033740 : Ok(std::sync::Arc::new(buf).into())
854 2033740 : }
855 : }
856 : impl BlockReader for TestDisk {
857 822547 : fn block_cursor(&self) -> BlockCursor<'_> {
858 822547 : BlockCursor::new(BlockReaderRef::TestDisk(self))
859 822547 : }
860 : }
861 : impl BlockWriter for &mut TestDisk {
862 431 : fn write_blk(&mut self, buf: Bytes) -> io::Result<u32> {
863 431 : let blknum = self.blocks.len();
864 431 : self.blocks.push(buf);
865 431 : Ok(blknum as u32)
866 431 : }
867 : }
868 :
869 : #[tokio::test]
870 4 : async fn basic() -> Result<()> {
871 4 : let mut disk = TestDisk::new();
872 4 : let mut writer = DiskBtreeBuilder::<_, 6>::new(&mut disk);
873 4 :
874 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
875 4 :
876 4 : let all_keys: Vec<&[u8; 6]> = vec![
877 4 : b"xaaaaa", b"xaaaba", b"xaaaca", b"xabaaa", b"xababa", b"xabaca", b"xabada", b"xabadb",
878 4 : ];
879 4 : let all_data: Vec<(&[u8; 6], u64)> = all_keys
880 4 : .iter()
881 4 : .enumerate()
882 32 : .map(|(idx, key)| (*key, idx as u64))
883 4 : .collect();
884 32 : for (key, val) in all_data.iter() {
885 32 : writer.append(key, *val)?;
886 4 : }
887 4 :
888 4 : let (root_offset, _writer) = writer.finish()?;
889 4 :
890 4 : let reader = DiskBtreeReader::new(0, root_offset, disk);
891 4 :
892 4 : reader.dump().await?;
893 4 :
894 4 : // Test the `get` function on all the keys.
895 32 : for (key, val) in all_data.iter() {
896 32 : assert_eq!(reader.get(key, &ctx).await?, Some(*val));
897 4 : }
898 4 : // And on some keys that don't exist
899 4 : assert_eq!(reader.get(b"aaaaaa", &ctx).await?, None);
900 4 : assert_eq!(reader.get(b"zzzzzz", &ctx).await?, None);
901 4 : assert_eq!(reader.get(b"xaaabx", &ctx).await?, None);
902 4 :
903 4 : // Test search with `visit` function
904 4 : let search_key = b"xabaaa";
905 4 : let expected: Vec<(Vec<u8>, u64)> = all_data
906 4 : .iter()
907 32 : .filter(|(key, _value)| key[..] >= search_key[..])
908 20 : .map(|(key, value)| (key.to_vec(), *value))
909 4 : .collect();
910 4 :
911 4 : let mut data = Vec::new();
912 4 : reader
913 4 : .visit(
914 4 : search_key,
915 4 : VisitDirection::Forwards,
916 20 : |key, value| {
917 20 : data.push((key.to_vec(), value));
918 20 : true
919 20 : },
920 4 : &ctx,
921 4 : )
922 4 : .await?;
923 4 : assert_eq!(data, expected);
924 4 :
925 4 : // Test a backwards scan
926 4 : let mut expected: Vec<(Vec<u8>, u64)> = all_data
927 4 : .iter()
928 32 : .filter(|(key, _value)| key[..] <= search_key[..])
929 16 : .map(|(key, value)| (key.to_vec(), *value))
930 4 : .collect();
931 4 : expected.reverse();
932 4 : let mut data = Vec::new();
933 4 : reader
934 4 : .visit(
935 4 : search_key,
936 4 : VisitDirection::Backwards,
937 16 : |key, value| {
938 16 : data.push((key.to_vec(), value));
939 16 : true
940 16 : },
941 4 : &ctx,
942 4 : )
943 4 : .await?;
944 4 : assert_eq!(data, expected);
945 4 :
946 4 : // Backward scan where nothing matches
947 4 : reader
948 4 : .visit(
949 4 : b"aaaaaa",
950 4 : VisitDirection::Backwards,
951 4 : |key, value| {
952 0 : panic!("found unexpected key {}: {}", hex::encode(key), value);
953 4 : },
954 4 : &ctx,
955 4 : )
956 4 : .await?;
957 4 :
958 4 : // Full scan
959 4 : let expected: Vec<(Vec<u8>, u64)> = all_data
960 4 : .iter()
961 32 : .map(|(key, value)| (key.to_vec(), *value))
962 4 : .collect();
963 4 : let mut data = Vec::new();
964 4 : reader
965 4 : .visit(
966 4 : &[0u8; 6],
967 4 : VisitDirection::Forwards,
968 32 : |key, value| {
969 32 : data.push((key.to_vec(), value));
970 32 : true
971 32 : },
972 4 : &ctx,
973 4 : )
974 4 : .await?;
975 4 : assert_eq!(data, expected);
976 4 :
977 4 : Ok(())
978 4 : }
979 :
980 : #[tokio::test]
981 4 : async fn lots_of_keys() -> Result<()> {
982 4 : let mut disk = TestDisk::new();
983 4 : let mut writer = DiskBtreeBuilder::<_, 8>::new(&mut disk);
984 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
985 4 :
986 4 : const NUM_KEYS: u64 = 1000;
987 4 :
988 4 : let mut all_data: BTreeMap<u64, u64> = BTreeMap::new();
989 4 :
990 4004 : for idx in 0..NUM_KEYS {
991 4000 : let key_int: u64 = 1 + idx * 2;
992 4000 : let key = u64::to_be_bytes(key_int);
993 4000 : writer.append(&key, idx)?;
994 4 :
995 4000 : all_data.insert(key_int, idx);
996 4 : }
997 4 :
998 4 : let (root_offset, _writer) = writer.finish()?;
999 4 :
1000 4 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1001 4 :
1002 4 : reader.dump().await?;
1003 4 :
1004 4 : use std::sync::Mutex;
1005 4 :
1006 4 : let result = Mutex::new(Vec::new());
1007 4 : let limit: AtomicUsize = AtomicUsize::new(10);
1008 167640 : let take_ten = |key: &[u8], value: u64| {
1009 167640 : let mut keybuf = [0u8; 8];
1010 167640 : keybuf.copy_from_slice(key);
1011 167640 : let key_int = u64::from_be_bytes(keybuf);
1012 167640 :
1013 167640 : let mut result = result.lock().unwrap();
1014 167640 : result.push((key_int, value));
1015 167640 :
1016 167640 : // keep going until we have 10 matches
1017 167640 : result.len() < limit.load(Ordering::Relaxed)
1018 167640 : };
1019 4 :
1020 8040 : for search_key_int in 0..(NUM_KEYS * 2 + 10) {
1021 8040 : let search_key = u64::to_be_bytes(search_key_int);
1022 8040 : assert_eq!(
1023 8040 : reader.get(&search_key, &ctx).await?,
1024 8040 : all_data.get(&search_key_int).cloned()
1025 4 : );
1026 4 :
1027 4 : // Test a forward scan starting with this key
1028 8040 : result.lock().unwrap().clear();
1029 8040 : reader
1030 8040 : .visit(&search_key, VisitDirection::Forwards, take_ten, &ctx)
1031 8040 : .await?;
1032 8040 : let expected = all_data
1033 8040 : .range(search_key_int..)
1034 8040 : .take(10)
1035 79640 : .map(|(&key, &val)| (key, val))
1036 8040 : .collect::<Vec<(u64, u64)>>();
1037 8040 : assert_eq!(*result.lock().unwrap(), expected);
1038 4 :
1039 4 : // And a backwards scan
1040 8040 : result.lock().unwrap().clear();
1041 8040 : reader
1042 8040 : .visit(&search_key, VisitDirection::Backwards, take_ten, &ctx)
1043 8040 : .await?;
1044 8040 : let expected = all_data
1045 8040 : .range(..=search_key_int)
1046 8040 : .rev()
1047 8040 : .take(10)
1048 80000 : .map(|(&key, &val)| (key, val))
1049 8040 : .collect::<Vec<(u64, u64)>>();
1050 8040 : assert_eq!(*result.lock().unwrap(), expected);
1051 4 : }
1052 4 :
1053 4 : // full scan
1054 4 : let search_key = u64::to_be_bytes(0);
1055 4 : limit.store(usize::MAX, Ordering::Relaxed);
1056 4 : result.lock().unwrap().clear();
1057 4 : reader
1058 4 : .visit(&search_key, VisitDirection::Forwards, take_ten, &ctx)
1059 4 : .await?;
1060 4 : let expected = all_data
1061 4 : .iter()
1062 4000 : .map(|(&key, &val)| (key, val))
1063 4 : .collect::<Vec<(u64, u64)>>();
1064 4 : assert_eq!(*result.lock().unwrap(), expected);
1065 4 :
1066 4 : // full scan
1067 4 : let search_key = u64::to_be_bytes(u64::MAX);
1068 4 : limit.store(usize::MAX, Ordering::Relaxed);
1069 4 : result.lock().unwrap().clear();
1070 4 : reader
1071 4 : .visit(&search_key, VisitDirection::Backwards, take_ten, &ctx)
1072 4 : .await?;
1073 4 : let expected = all_data
1074 4 : .iter()
1075 4 : .rev()
1076 4000 : .map(|(&key, &val)| (key, val))
1077 4 : .collect::<Vec<(u64, u64)>>();
1078 4 : assert_eq!(*result.lock().unwrap(), expected);
1079 4 :
1080 4 : Ok(())
1081 4 : }
1082 :
1083 : #[tokio::test]
1084 4 : async fn random_data() -> Result<()> {
1085 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1086 4 :
1087 4 : // Generate random keys with exponential distribution, to
1088 4 : // exercise the prefix compression
1089 4 : const NUM_KEYS: usize = 100000;
1090 4 : let mut all_data: BTreeMap<u128, u64> = BTreeMap::new();
1091 400004 : for idx in 0..NUM_KEYS {
1092 400000 : let u: f64 = rand::thread_rng().gen_range(0.0..1.0);
1093 400000 : let t = -(f64::ln(u));
1094 400000 : let key_int = (t * 1000000.0) as u128;
1095 400000 :
1096 400000 : all_data.insert(key_int, idx as u64);
1097 400000 : }
1098 4 :
1099 4 : // Build a tree from it
1100 4 : let mut disk = TestDisk::new();
1101 4 : let mut writer = DiskBtreeBuilder::<_, 16>::new(&mut disk);
1102 4 :
1103 390323 : for (&key, &val) in all_data.iter() {
1104 390323 : writer.append(&u128::to_be_bytes(key), val)?;
1105 4 : }
1106 4 : let (root_offset, _writer) = writer.finish()?;
1107 4 :
1108 4 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1109 4 :
1110 4 : // Test get() operation on all the keys
1111 390323 : for (&key, &val) in all_data.iter() {
1112 390323 : let search_key = u128::to_be_bytes(key);
1113 390323 : assert_eq!(reader.get(&search_key, &ctx).await?, Some(val));
1114 4 : }
1115 4 :
1116 4 : // Test get() operations on random keys, most of which will not exist
1117 400004 : for _ in 0..100000 {
1118 400000 : let key_int = rand::thread_rng().gen::<u128>();
1119 400000 : let search_key = u128::to_be_bytes(key_int);
1120 400000 : assert!(reader.get(&search_key, &ctx).await? == all_data.get(&key_int).cloned());
1121 4 : }
1122 4 :
1123 4 : // Test boundary cases
1124 4 : assert!(
1125 4 : reader.get(&u128::to_be_bytes(u128::MIN), &ctx).await?
1126 4 : == all_data.get(&u128::MIN).cloned()
1127 4 : );
1128 4 : assert!(
1129 4 : reader.get(&u128::to_be_bytes(u128::MAX), &ctx).await?
1130 4 : == all_data.get(&u128::MAX).cloned()
1131 4 : );
1132 4 :
1133 4 : // Test iterator and get_stream API
1134 4 : let mut iter = reader.iter(&[0; 16], &ctx);
1135 4 : let mut cnt = 0;
1136 390327 : while let Some(res) = iter.next().await {
1137 390323 : let (key, val) = res?;
1138 390323 : let key = u128::from_be_bytes(key.as_slice().try_into().unwrap());
1139 390323 : assert_eq!(val, *all_data.get(&key).unwrap());
1140 390323 : cnt += 1;
1141 4 : }
1142 4 : assert_eq!(cnt, all_data.len());
1143 4 :
1144 4 : Ok(())
1145 4 : }
1146 :
1147 : #[test]
1148 4 : fn unsorted_input() {
1149 4 : let mut disk = TestDisk::new();
1150 4 : let mut writer = DiskBtreeBuilder::<_, 2>::new(&mut disk);
1151 4 :
1152 4 : let _ = writer.append(b"ba", 1);
1153 4 : let _ = writer.append(b"bb", 2);
1154 4 : let err = writer.append(b"aa", 3).expect_err("should've failed");
1155 4 : match err {
1156 4 : DiskBtreeError::UnsortedInput { key, last_key } => {
1157 4 : assert_eq!(key.as_ref(), b"aa".as_slice());
1158 4 : assert_eq!(last_key.as_ref(), b"bb".as_slice());
1159 : }
1160 0 : _ => panic!("unexpected error variant, expected DiskBtreeError::UnsortedInput"),
1161 : }
1162 4 : }
1163 :
1164 : ///
1165 : /// This test contains a particular data set, see disk_btree_test_data.rs
1166 : ///
1167 : #[tokio::test]
1168 4 : async fn particular_data() -> Result<()> {
1169 4 : // Build a tree from it
1170 4 : let mut disk = TestDisk::new();
1171 4 : let mut writer = DiskBtreeBuilder::<_, 26>::new(&mut disk);
1172 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1173 4 :
1174 8004 : for (key, val) in disk_btree_test_data::TEST_DATA {
1175 8000 : writer.append(&key, val)?;
1176 4 : }
1177 4 : let (root_offset, writer) = writer.finish()?;
1178 4 :
1179 4 : println!("SIZE: {} blocks", writer.blocks.len());
1180 4 :
1181 4 : let reader = DiskBtreeReader::new(0, root_offset, disk);
1182 4 :
1183 4 : // Test get() operation on all the keys
1184 8004 : for (key, val) in disk_btree_test_data::TEST_DATA {
1185 8000 : assert_eq!(reader.get(&key, &ctx).await?, Some(val));
1186 4 : }
1187 4 :
1188 4 : // Test full scan
1189 4 : let mut count = 0;
1190 4 : reader
1191 4 : .visit(
1192 4 : &[0u8; 26],
1193 4 : VisitDirection::Forwards,
1194 8000 : |_key, _value| {
1195 8000 : count += 1;
1196 8000 : true
1197 8000 : },
1198 4 : &ctx,
1199 4 : )
1200 4 : .await?;
1201 4 : assert_eq!(count, disk_btree_test_data::TEST_DATA.len());
1202 4 :
1203 4 : reader.dump().await?;
1204 4 :
1205 4 : Ok(())
1206 4 : }
1207 : }
1208 :
1209 : #[cfg(test)]
1210 : #[path = "disk_btree_test_data.rs"]
1211 : mod disk_btree_test_data;
|