Line data Source code
1 : //!
2 : //! Utilities for vectored reading of variable-sized "blobs".
3 : //!
4 : //! The "blob" api is an abstraction on top of the "block" api,
5 : //! with the main difference being that blobs do not have a fixed
6 : //! size (each blob is prefixed with 1 or 4 byte length field)
7 : //!
8 : //! The vectored apis provided in this module allow for planning
9 : //! and executing disk IO which covers multiple blobs.
10 : //!
11 : //! Reads are planned with [`VectoredReadPlanner`] which will coalesce
12 : //! adjacent blocks into a single disk IO request and exectuted by
13 : //! [`VectoredBlobReader`] which does all the required offset juggling
14 : //! and returns a buffer housing all the blobs and a list of offsets.
15 : //!
16 : //! Note that the vectored blob api does *not* go through the page cache.
17 :
18 : use std::collections::BTreeMap;
19 : use std::num::NonZeroUsize;
20 :
21 : use bytes::BytesMut;
22 : use pageserver_api::key::Key;
23 : use utils::lsn::Lsn;
24 : use utils::vec_map::VecMap;
25 :
26 : use crate::virtual_file::VirtualFile;
27 :
28 : #[derive(Copy, Clone, Debug, PartialEq, Eq)]
29 : pub struct MaxVectoredReadBytes(pub NonZeroUsize);
30 :
31 : /// Metadata bundled with the start and end offset of a blob.
32 : #[derive(Copy, Clone, Debug)]
33 : pub struct BlobMeta {
34 : pub key: Key,
35 : pub lsn: Lsn,
36 : }
37 :
38 : /// Blob offsets into [`VectoredBlobsBuf::buf`]
39 : pub struct VectoredBlob {
40 : pub start: usize,
41 : pub end: usize,
42 : pub meta: BlobMeta,
43 : }
44 :
45 : /// Return type of [`VectoredBlobReader::read_blobs`]
46 : pub struct VectoredBlobsBuf {
47 : /// Buffer for all blobs in this read
48 : pub buf: BytesMut,
49 : /// Offsets into the buffer and metadata for all blobs in this read
50 : pub blobs: Vec<VectoredBlob>,
51 : }
52 :
53 : /// Description of one disk read for multiple blobs.
54 : /// Used as the argument form [`VectoredBlobReader::read_blobs`]
55 : #[derive(Debug)]
56 : pub struct VectoredRead {
57 : pub start: u64,
58 : pub end: u64,
59 : /// Starting offsets and metadata for each blob in this read
60 : pub blobs_at: VecMap<u64, BlobMeta>,
61 : }
62 :
63 : impl VectoredRead {
64 79186 : pub fn size(&self) -> usize {
65 79186 : (self.end - self.start) as usize
66 79186 : }
67 : }
68 :
69 : #[derive(Eq, PartialEq)]
70 : enum VectoredReadExtended {
71 : Yes,
72 : No,
73 : }
74 :
75 : struct VectoredReadBuilder {
76 : start: u64,
77 : end: u64,
78 : blobs_at: VecMap<u64, BlobMeta>,
79 : max_read_size: usize,
80 : }
81 :
82 : impl VectoredReadBuilder {
83 19764 : fn new(start_offset: u64, end_offset: u64, meta: BlobMeta, max_read_size: usize) -> Self {
84 19764 : let mut blobs_at = VecMap::default();
85 19764 : blobs_at
86 19764 : .append(start_offset, meta)
87 19764 : .expect("First insertion always succeeds");
88 19764 :
89 19764 : Self {
90 19764 : start: start_offset,
91 19764 : end: end_offset,
92 19764 : blobs_at,
93 19764 : max_read_size,
94 19764 : }
95 19764 : }
96 :
97 : /// Attempt to extend the current read with a new blob if the start
98 : /// offset matches with the current end of the vectored read
99 : /// and the resuting size is below the max read size
100 57494 : fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
101 57494 : let size = (end - start) as usize;
102 57494 : if self.end == start && self.size() + size <= self.max_read_size {
103 37954 : self.end = end;
104 37954 : self.blobs_at
105 37954 : .append(start, meta)
106 37954 : .expect("LSNs are ordered within vectored reads");
107 37954 :
108 37954 : return VectoredReadExtended::Yes;
109 19540 : }
110 19540 :
111 19540 : VectoredReadExtended::No
112 57494 : }
113 :
114 57306 : fn size(&self) -> usize {
115 57306 : (self.end - self.start) as usize
116 57306 : }
117 :
118 19764 : fn build(self) -> VectoredRead {
119 19764 : VectoredRead {
120 19764 : start: self.start,
121 19764 : end: self.end,
122 19764 : blobs_at: self.blobs_at,
123 19764 : }
124 19764 : }
125 : }
126 :
127 : #[derive(Copy, Clone, Debug)]
128 : pub enum BlobFlag {
129 : None,
130 : Ignore,
131 : ReplaceAll,
132 : }
133 :
134 : /// Planner for vectored blob reads.
135 : ///
136 : /// Blob offsets are received via [`VectoredReadPlanner::handle`]
137 : /// and coalesced into disk reads.
138 : ///
139 : /// The implementation is very simple:
140 : /// * Collect all blob offsets in an ordered structure
141 : /// * Iterate over the collected blobs and coalesce them into reads at the end
142 : pub struct VectoredReadPlanner {
143 : // Track all the blob offsets. Start offsets must be ordered.
144 : blobs: BTreeMap<Key, Vec<(Lsn, u64, u64)>>,
145 : // Arguments for previous blob passed into [`VectoredReadPlanner::handle`]
146 : prev: Option<(Key, Lsn, u64, BlobFlag)>,
147 :
148 : max_read_size: usize,
149 : }
150 :
151 : impl VectoredReadPlanner {
152 224 : pub fn new(max_read_size: usize) -> Self {
153 224 : Self {
154 224 : blobs: BTreeMap::new(),
155 224 : prev: None,
156 224 : max_read_size,
157 224 : }
158 224 : }
159 :
160 : /// Include a new blob in the read plan.
161 : ///
162 : /// This function is called from a B-Tree index visitor (see `DeltaLayerInner::plan_reads`
163 : /// and `ImageLayerInner::plan_reads`). Said visitor wants to collect blob offsets for all
164 : /// keys in a given keyspace. This function must be called for each key in the desired
165 : /// keyspace (monotonically continuous). [`Self::handle_range_end`] must
166 : /// be called after every range in the offset.
167 : ///
168 : /// In the event that keys are skipped, the behaviour is undefined and can lead to an
169 : /// incorrect read plan. We can end up asserting, erroring in wal redo or returning
170 : /// incorrect data to the user.
171 : ///
172 : /// The `flag` argument has two interesting values:
173 : /// * [`BlobFlag::ReplaceAll`]: The blob for this key should replace all existing blobs.
174 : /// This is used for WAL records that `will_init`.
175 : /// * [`BlobFlag::Ignore`]: This blob should not be included in the read. This happens
176 : /// if the blob is cached.
177 57724 : pub fn handle(&mut self, key: Key, lsn: Lsn, offset: u64, flag: BlobFlag) {
178 57298 : // Implementation note: internally lag behind by one blob such that
179 57298 : // we have a start and end offset when initialising [`VectoredRead`]
180 57724 : let (prev_key, prev_lsn, prev_offset, prev_flag) = match self.prev {
181 : None => {
182 426 : self.prev = Some((key, lsn, offset, flag));
183 426 : return;
184 : }
185 57298 : Some(prev) => prev,
186 57298 : };
187 57298 :
188 57298 : self.add_blob(prev_key, prev_lsn, prev_offset, offset, prev_flag);
189 57298 :
190 57298 : self.prev = Some((key, lsn, offset, flag));
191 57724 : }
192 :
193 426 : pub fn handle_range_end(&mut self, offset: u64) {
194 426 : if let Some((prev_key, prev_lsn, prev_offset, prev_flag)) = self.prev {
195 426 : self.add_blob(prev_key, prev_lsn, prev_offset, offset, prev_flag);
196 426 : }
197 :
198 426 : self.prev = None;
199 426 : }
200 :
201 57724 : fn add_blob(&mut self, key: Key, lsn: Lsn, start_offset: u64, end_offset: u64, flag: BlobFlag) {
202 57724 : match flag {
203 57356 : BlobFlag::None => {
204 57356 : let blobs_for_key = self.blobs.entry(key).or_default();
205 57356 : blobs_for_key.push((lsn, start_offset, end_offset));
206 57356 : }
207 366 : BlobFlag::ReplaceAll => {
208 366 : let blobs_for_key = self.blobs.entry(key).or_default();
209 366 : blobs_for_key.clear();
210 366 : blobs_for_key.push((lsn, start_offset, end_offset));
211 366 : }
212 2 : BlobFlag::Ignore => {}
213 : }
214 57724 : }
215 :
216 224 : pub fn finish(self) -> Vec<VectoredRead> {
217 224 : let mut current_read_builder: Option<VectoredReadBuilder> = None;
218 224 : let mut reads = Vec::new();
219 :
220 6902 : for (key, blobs_for_key) in self.blobs {
221 64396 : for (lsn, start_offset, end_offset) in blobs_for_key {
222 57718 : let extended = match &mut current_read_builder {
223 57494 : Some(read_builder) => {
224 57494 : read_builder.extend(start_offset, end_offset, BlobMeta { key, lsn })
225 : }
226 224 : None => VectoredReadExtended::No,
227 : };
228 :
229 57718 : if extended == VectoredReadExtended::No {
230 19764 : let next_read_builder = VectoredReadBuilder::new(
231 19764 : start_offset,
232 19764 : end_offset,
233 19764 : BlobMeta { key, lsn },
234 19764 : self.max_read_size,
235 19764 : );
236 19764 :
237 19764 : let prev_read_builder = current_read_builder.replace(next_read_builder);
238 :
239 : // `current_read_builder` is None in the first iteration of the outer loop
240 19764 : if let Some(read_builder) = prev_read_builder {
241 19540 : reads.push(read_builder.build());
242 19540 : }
243 37954 : }
244 : }
245 : }
246 :
247 224 : if let Some(read_builder) = current_read_builder {
248 224 : reads.push(read_builder.build());
249 224 : }
250 :
251 224 : reads
252 224 : }
253 : }
254 :
255 : /// Disk reader for vectored blob spans (does not go through the page cache)
256 : pub struct VectoredBlobReader<'a> {
257 : file: &'a VirtualFile,
258 : }
259 :
260 : impl<'a> VectoredBlobReader<'a> {
261 218 : pub fn new(file: &'a VirtualFile) -> Self {
262 218 : Self { file }
263 218 : }
264 :
265 : /// Read the requested blobs into the buffer.
266 : ///
267 : /// We have to deal with the fact that blobs are not fixed size.
268 : /// Each blob is prefixed by a size header.
269 : ///
270 : /// The success return value is a struct which contains the buffer
271 : /// filled from disk and a list of offsets at which each blob lies
272 : /// in the buffer.
273 19742 : pub async fn read_blobs(
274 19742 : &self,
275 19742 : read: &VectoredRead,
276 19742 : buf: BytesMut,
277 19742 : ) -> Result<VectoredBlobsBuf, std::io::Error> {
278 19742 : assert!(read.size() > 0);
279 19742 : assert!(
280 19742 : read.size() <= buf.capacity(),
281 0 : "{} > {}",
282 0 : read.size(),
283 0 : buf.capacity()
284 : );
285 19742 : let buf = self
286 19742 : .file
287 19742 : .read_exact_at_n(buf, read.start, read.size())
288 10026 : .await?;
289 :
290 19742 : let blobs_at = read.blobs_at.as_slice();
291 19742 : let start_offset = blobs_at.first().expect("VectoredRead is never empty").0;
292 19742 :
293 19742 : let mut metas = Vec::with_capacity(blobs_at.len());
294 19742 :
295 19742 : // Blobs in `read` only provide their starting offset. The end offset
296 19742 : // of a blob is implicit: the start of the next blob if one exists
297 19742 : // or the end of the read.
298 19742 : let pairs = blobs_at.iter().zip(
299 19742 : blobs_at
300 19742 : .iter()
301 19742 : .map(Some)
302 19742 : .skip(1)
303 19742 : .chain(std::iter::once(None)),
304 19742 : );
305 :
306 77408 : for ((offset, meta), next) in pairs {
307 57666 : let offset_in_buf = offset - start_offset;
308 57666 : let first_len_byte = buf[offset_in_buf as usize];
309 :
310 : // Each blob is prefixed by a header containing it's size.
311 : // Extract the size and skip that header to find the start of the data.
312 : // The size can be 1 or 4 bytes. The most significant bit is 0 in the
313 : // 1 byte case and 1 in the 4 byte case.
314 57666 : let (size_length, blob_size) = if first_len_byte < 0x80 {
315 24052 : (1, first_len_byte as u64)
316 : } else {
317 33614 : let mut blob_size_buf = [0u8; 4];
318 33614 : let offset_in_buf = offset_in_buf as usize;
319 33614 :
320 33614 : blob_size_buf.copy_from_slice(&buf[offset_in_buf..offset_in_buf + 4]);
321 33614 : blob_size_buf[0] &= 0x7f;
322 33614 : (4, u32::from_be_bytes(blob_size_buf) as u64)
323 : };
324 :
325 57666 : let start = offset_in_buf + size_length;
326 57666 : let end = match next {
327 37924 : Some((next_blob_start_offset, _)) => next_blob_start_offset - start_offset,
328 19742 : None => start + blob_size,
329 : };
330 :
331 57666 : assert_eq!(end - start, blob_size);
332 :
333 57666 : metas.push(VectoredBlob {
334 57666 : start: start as usize,
335 57666 : end: end as usize,
336 57666 : meta: *meta,
337 57666 : })
338 : }
339 :
340 19742 : Ok(VectoredBlobsBuf { buf, blobs: metas })
341 19742 : }
342 : }
343 :
344 : #[cfg(test)]
345 : mod tests {
346 : use super::*;
347 :
348 16 : fn validate_read(read: &VectoredRead, offset_range: &[(Key, Lsn, u64, BlobFlag)]) {
349 16 : assert_eq!(read.start, offset_range.first().unwrap().2);
350 :
351 24 : let expected_offsets_in_read: Vec<_> = offset_range.iter().map(|o| o.2).collect();
352 16 :
353 16 : let offsets_in_read: Vec<_> = read
354 16 : .blobs_at
355 16 : .as_slice()
356 16 : .iter()
357 24 : .map(|(offset, _)| *offset)
358 16 : .collect();
359 16 :
360 16 : assert_eq!(expected_offsets_in_read, offsets_in_read);
361 16 : }
362 :
363 : #[test]
364 2 : fn planner_max_read_size_test() {
365 2 : let max_read_size = 128 * 1024;
366 2 : let key = Key::MIN;
367 2 : let lsn = Lsn(0);
368 2 :
369 2 : let blob_descriptions = vec![
370 2 : (key, lsn, 0, BlobFlag::None),
371 2 : (key, lsn, 32 * 1024, BlobFlag::None),
372 2 : (key, lsn, 96 * 1024, BlobFlag::None), // Last in read 1
373 2 : (key, lsn, 128 * 1024, BlobFlag::None), // Last in read 2
374 2 : (key, lsn, 198 * 1024, BlobFlag::None), // Last in read 3
375 2 : (key, lsn, 268 * 1024, BlobFlag::None), // Last in read 4
376 2 : (key, lsn, 396 * 1024, BlobFlag::None), // Last in read 5
377 2 : (key, lsn, 652 * 1024, BlobFlag::None), // Last in read 6
378 2 : ];
379 2 :
380 2 : let ranges = [
381 2 : &blob_descriptions[0..3],
382 2 : &blob_descriptions[3..4],
383 2 : &blob_descriptions[4..5],
384 2 : &blob_descriptions[5..6],
385 2 : &blob_descriptions[6..7],
386 2 : &blob_descriptions[7..],
387 2 : ];
388 2 :
389 2 : let mut planner = VectoredReadPlanner::new(max_read_size);
390 16 : for (key, lsn, offset, flag) in blob_descriptions.clone() {
391 16 : planner.handle(key, lsn, offset, flag);
392 16 : }
393 :
394 2 : planner.handle_range_end(652 * 1024);
395 2 :
396 2 : let reads = planner.finish();
397 2 : assert_eq!(reads.len(), 6);
398 :
399 12 : for (idx, read) in reads.iter().enumerate() {
400 12 : validate_read(read, ranges[idx]);
401 12 : }
402 2 : }
403 :
404 : #[test]
405 2 : fn planner_replacement_test() {
406 2 : let max_read_size = 128 * 1024;
407 2 : let first_key = Key::MIN;
408 2 : let second_key = first_key.next();
409 2 : let lsn = Lsn(0);
410 2 :
411 2 : let blob_descriptions = vec![
412 2 : (first_key, lsn, 0, BlobFlag::None), // First in read 1
413 2 : (first_key, lsn, 1024, BlobFlag::None), // Last in read 1
414 2 : (second_key, lsn, 2 * 1024, BlobFlag::ReplaceAll),
415 2 : (second_key, lsn, 3 * 1024, BlobFlag::None),
416 2 : (second_key, lsn, 4 * 1024, BlobFlag::ReplaceAll), // First in read 2
417 2 : (second_key, lsn, 5 * 1024, BlobFlag::None), // Last in read 2
418 2 : ];
419 2 :
420 2 : let ranges = [&blob_descriptions[0..2], &blob_descriptions[4..]];
421 2 :
422 2 : let mut planner = VectoredReadPlanner::new(max_read_size);
423 12 : for (key, lsn, offset, flag) in blob_descriptions.clone() {
424 12 : planner.handle(key, lsn, offset, flag);
425 12 : }
426 :
427 2 : planner.handle_range_end(6 * 1024);
428 2 :
429 2 : let reads = planner.finish();
430 2 : assert_eq!(reads.len(), 2);
431 :
432 4 : for (idx, read) in reads.iter().enumerate() {
433 4 : validate_read(read, ranges[idx]);
434 4 : }
435 2 : }
436 : }
|
