Line data Source code
1 : use anyhow::{bail, ensure, Context};
2 : use itertools::Itertools;
3 : use pageserver_api::shard::TenantShardId;
4 : use std::{collections::HashMap, sync::Arc};
5 : use tracing::trace;
6 : use utils::{
7 : id::TimelineId,
8 : lsn::{AtomicLsn, Lsn},
9 : };
10 :
11 : use crate::{
12 : config::PageServerConf,
13 : context::RequestContext,
14 : metrics::TimelineMetrics,
15 : tenant::{
16 : layer_map::{BatchedUpdates, LayerMap},
17 : storage_layer::{
18 : AsLayerDesc, InMemoryLayer, Layer, PersistentLayerDesc, PersistentLayerKey,
19 : ResidentLayer,
20 : },
21 : },
22 : };
23 :
24 : use super::TimelineWriterState;
25 :
26 : /// Provides semantic APIs to manipulate the layer map.
27 : pub(crate) enum LayerManager {
28 : /// Open as in not shutdown layer manager; we still have in-memory layers and we can manipulate
29 : /// the layers.
30 : Open(OpenLayerManager),
31 : /// Shutdown layer manager where there are no more in-memory layers and persistent layers are
32 : /// read-only.
33 : Closed {
34 : layers: HashMap<PersistentLayerKey, Layer>,
35 : },
36 : }
37 :
38 : impl Default for LayerManager {
39 422 : fn default() -> Self {
40 422 : LayerManager::Open(OpenLayerManager::default())
41 422 : }
42 : }
43 :
44 : impl LayerManager {
45 242121 : pub(crate) fn get_from_key(&self, key: &PersistentLayerKey) -> Layer {
46 242121 : // The assumption for the `expect()` is that all code maintains the following invariant:
47 242121 : // A layer's descriptor is present in the LayerMap => the LayerFileManager contains a layer for the descriptor.
48 242121 : self.try_get_from_key(key)
49 242121 : .with_context(|| format!("get layer from key: {key}"))
50 242121 : .expect("not found")
51 242121 : .clone()
52 242121 : }
53 :
54 242121 : pub(crate) fn try_get_from_key(&self, key: &PersistentLayerKey) -> Option<&Layer> {
55 242121 : self.layers().get(key)
56 242121 : }
57 :
58 242099 : pub(crate) fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Layer {
59 242099 : self.get_from_key(&desc.key())
60 242099 : }
61 :
62 : /// Get an immutable reference to the layer map.
63 : ///
64 : /// We expect users only to be able to get an immutable layer map. If users want to make modifications,
65 : /// they should use the below semantic APIs. This design makes us step closer to immutable storage state.
66 1081435 : pub(crate) fn layer_map(&self) -> Result<&LayerMap, Shutdown> {
67 : use LayerManager::*;
68 1081435 : match self {
69 1081435 : Open(OpenLayerManager { layer_map, .. }) => Ok(layer_map),
70 0 : Closed { .. } => Err(Shutdown),
71 : }
72 1081435 : }
73 :
74 4972 : pub(crate) fn open_mut(&mut self) -> Result<&mut OpenLayerManager, Shutdown> {
75 : use LayerManager::*;
76 :
77 4972 : match self {
78 4972 : Open(open) => Ok(open),
79 0 : Closed { .. } => Err(Shutdown),
80 : }
81 4972 : }
82 :
83 : /// LayerManager shutdown. The in-memory layers do cleanup on drop, so we must drop them in
84 : /// order to allow shutdown to complete.
85 : ///
86 : /// If there was a want to flush in-memory layers, it must have happened earlier.
87 10 : pub(crate) fn shutdown(&mut self, writer_state: &mut Option<TimelineWriterState>) {
88 : use LayerManager::*;
89 10 : match self {
90 : Open(OpenLayerManager {
91 10 : layer_map,
92 10 : layer_fmgr: LayerFileManager(hashmap),
93 10 : }) => {
94 10 : let open = layer_map.open_layer.take();
95 10 : let frozen = layer_map.frozen_layers.len();
96 10 : let taken_writer_state = writer_state.take();
97 10 : tracing::info!(open = open.is_some(), frozen, "dropped inmemory layers");
98 10 : let layers = std::mem::take(hashmap);
99 10 : *self = Closed { layers };
100 10 : assert_eq!(open.is_some(), taken_writer_state.is_some());
101 : }
102 : Closed { .. } => {
103 0 : tracing::debug!("ignoring multiple shutdowns on layer manager")
104 : }
105 : }
106 10 : }
107 :
108 : /// Sum up the historic layer sizes
109 0 : pub(crate) fn layer_size_sum(&self) -> u64 {
110 0 : self.layers()
111 0 : .values()
112 0 : .map(|l| l.layer_desc().file_size)
113 0 : .sum()
114 0 : }
115 :
116 22 : pub(crate) fn likely_resident_layers(&self) -> impl Iterator<Item = &'_ Layer> + '_ {
117 34 : self.layers().values().filter(|l| l.is_likely_resident())
118 22 : }
119 :
120 308 : pub(crate) fn contains(&self, layer: &Layer) -> bool {
121 308 : self.contains_key(&layer.layer_desc().key())
122 308 : }
123 :
124 410 : pub(crate) fn contains_key(&self, key: &PersistentLayerKey) -> bool {
125 410 : self.layers().contains_key(key)
126 410 : }
127 :
128 0 : pub(crate) fn all_persistent_layers(&self) -> Vec<PersistentLayerKey> {
129 0 : self.layers().keys().cloned().collect_vec()
130 0 : }
131 :
132 242553 : fn layers(&self) -> &HashMap<PersistentLayerKey, Layer> {
133 : use LayerManager::*;
134 242553 : match self {
135 242553 : Open(OpenLayerManager { layer_fmgr, .. }) => &layer_fmgr.0,
136 0 : Closed { layers } => layers,
137 : }
138 242553 : }
139 : }
140 :
141 : #[derive(Default)]
142 : pub(crate) struct OpenLayerManager {
143 : layer_map: LayerMap,
144 : layer_fmgr: LayerFileManager<Layer>,
145 : }
146 :
147 : impl std::fmt::Debug for OpenLayerManager {
148 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
149 0 : f.debug_struct("OpenLayerManager")
150 0 : .field("layer_count", &self.layer_fmgr.0.len())
151 0 : .finish()
152 0 : }
153 : }
154 :
155 : #[derive(Debug, thiserror::Error)]
156 : #[error("layer manager has been shutdown")]
157 : pub(crate) struct Shutdown;
158 :
159 : impl OpenLayerManager {
160 : /// Called from `load_layer_map`. Initialize the layer manager with:
161 : /// 1. all on-disk layers
162 : /// 2. next open layer (with disk disk_consistent_lsn LSN)
163 6 : pub(crate) fn initialize_local_layers(&mut self, layers: Vec<Layer>, next_open_layer_at: Lsn) {
164 6 : let mut updates = self.layer_map.batch_update();
165 22 : for layer in layers {
166 16 : Self::insert_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
167 16 : }
168 6 : updates.flush();
169 6 : self.layer_map.next_open_layer_at = Some(next_open_layer_at);
170 6 : }
171 :
172 : /// Initialize when creating a new timeline, called in `init_empty_layer_map`.
173 416 : pub(crate) fn initialize_empty(&mut self, next_open_layer_at: Lsn) {
174 416 : self.layer_map.next_open_layer_at = Some(next_open_layer_at);
175 416 : }
176 :
177 : /// Open a new writable layer to append data if there is no open layer, otherwise return the
178 : /// current open layer, called within `get_layer_for_write`.
179 1272 : pub(crate) async fn get_layer_for_write(
180 1272 : &mut self,
181 1272 : lsn: Lsn,
182 1272 : conf: &'static PageServerConf,
183 1272 : timeline_id: TimelineId,
184 1272 : tenant_shard_id: TenantShardId,
185 1272 : gate: &utils::sync::gate::Gate,
186 1272 : ctx: &RequestContext,
187 1272 : ) -> anyhow::Result<Arc<InMemoryLayer>> {
188 1272 : ensure!(lsn.is_aligned());
189 :
190 : // Do we have a layer open for writing already?
191 1272 : let layer = if let Some(open_layer) = &self.layer_map.open_layer {
192 0 : if open_layer.get_lsn_range().start > lsn {
193 0 : bail!(
194 0 : "unexpected open layer in the future: open layers starts at {}, write lsn {}",
195 0 : open_layer.get_lsn_range().start,
196 0 : lsn
197 0 : );
198 0 : }
199 0 :
200 0 : Arc::clone(open_layer)
201 : } else {
202 : // No writeable layer yet. Create one.
203 1272 : let start_lsn = self
204 1272 : .layer_map
205 1272 : .next_open_layer_at
206 1272 : .context("No next open layer found")?;
207 :
208 1272 : trace!(
209 0 : "creating in-memory layer at {}/{} for record at {}",
210 : timeline_id,
211 : start_lsn,
212 : lsn
213 : );
214 :
215 1272 : let new_layer =
216 1272 : InMemoryLayer::create(conf, timeline_id, tenant_shard_id, start_lsn, gate, ctx)
217 1272 : .await?;
218 1272 : let layer = Arc::new(new_layer);
219 1272 :
220 1272 : self.layer_map.open_layer = Some(layer.clone());
221 1272 : self.layer_map.next_open_layer_at = None;
222 1272 :
223 1272 : layer
224 : };
225 :
226 1272 : Ok(layer)
227 1272 : }
228 :
229 : /// Tries to freeze an open layer and also manages clearing the TimelineWriterState.
230 : ///
231 : /// Returns true if anything was frozen.
232 1176 : pub(super) async fn try_freeze_in_memory_layer(
233 1176 : &mut self,
234 1176 : lsn: Lsn,
235 1176 : last_freeze_at: &AtomicLsn,
236 1176 : write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
237 1176 : ) -> bool {
238 1176 : let Lsn(last_record_lsn) = lsn;
239 1176 : let end_lsn = Lsn(last_record_lsn + 1);
240 :
241 1176 : let froze = if let Some(open_layer) = &self.layer_map.open_layer {
242 1148 : let open_layer_rc = Arc::clone(open_layer);
243 1148 : open_layer.freeze(end_lsn).await;
244 :
245 : // The layer is no longer open, update the layer map to reflect this.
246 : // We will replace it with on-disk historics below.
247 1148 : self.layer_map.frozen_layers.push_back(open_layer_rc);
248 1148 : self.layer_map.open_layer = None;
249 1148 : self.layer_map.next_open_layer_at = Some(end_lsn);
250 1148 :
251 1148 : true
252 : } else {
253 28 : false
254 : };
255 :
256 : // Even if there was no layer to freeze, advance last_freeze_at to last_record_lsn+1: this
257 : // accounts for regions in the LSN range where we might have ingested no data due to sharding.
258 1176 : last_freeze_at.store(end_lsn);
259 1176 :
260 1176 : // the writer state must no longer have a reference to the frozen layer
261 1176 : let taken = write_lock.take();
262 1176 : assert_eq!(
263 1176 : froze,
264 1176 : taken.is_some(),
265 0 : "should only had frozen a layer when TimelineWriterState existed"
266 : );
267 :
268 1176 : froze
269 1176 : }
270 :
271 : /// Add image layers to the layer map, called from [`super::Timeline::create_image_layers`].
272 724 : pub(crate) fn track_new_image_layers(
273 724 : &mut self,
274 724 : image_layers: &[ResidentLayer],
275 724 : metrics: &TimelineMetrics,
276 724 : ) {
277 724 : let mut updates = self.layer_map.batch_update();
278 942 : for layer in image_layers {
279 218 : Self::insert_historic_layer(layer.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
280 218 :
281 218 : // record these here instead of Layer::finish_creating because otherwise partial
282 218 : // failure with create_image_layers would balloon up the physical size gauge. downside
283 218 : // is that all layers need to be created before metrics are updated.
284 218 : metrics.record_new_file_metrics(layer.layer_desc().file_size);
285 218 : }
286 724 : updates.flush();
287 724 : }
288 :
289 : /// Flush a frozen layer and add the written delta layer to the layer map.
290 1148 : pub(crate) fn finish_flush_l0_layer(
291 1148 : &mut self,
292 1148 : delta_layer: Option<&ResidentLayer>,
293 1148 : frozen_layer_for_check: &Arc<InMemoryLayer>,
294 1148 : metrics: &TimelineMetrics,
295 1148 : ) {
296 1148 : let inmem = self
297 1148 : .layer_map
298 1148 : .frozen_layers
299 1148 : .pop_front()
300 1148 : .expect("there must be a inmem layer to flush");
301 1148 :
302 1148 : // Only one task may call this function at a time (for this
303 1148 : // timeline). If two tasks tried to flush the same frozen
304 1148 : // layer to disk at the same time, that would not work.
305 1148 : assert_eq!(Arc::as_ptr(&inmem), Arc::as_ptr(frozen_layer_for_check));
306 :
307 1148 : if let Some(l) = delta_layer {
308 968 : let mut updates = self.layer_map.batch_update();
309 968 : Self::insert_historic_layer(l.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
310 968 : metrics.record_new_file_metrics(l.layer_desc().file_size);
311 968 : updates.flush();
312 968 : }
313 1148 : }
314 :
315 : /// Called when compaction is completed.
316 76 : pub(crate) fn finish_compact_l0(
317 76 : &mut self,
318 76 : compact_from: &[Layer],
319 76 : compact_to: &[ResidentLayer],
320 76 : metrics: &TimelineMetrics,
321 76 : ) {
322 76 : let mut updates = self.layer_map.batch_update();
323 438 : for l in compact_to {
324 362 : Self::insert_historic_layer(l.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
325 362 : metrics.record_new_file_metrics(l.layer_desc().file_size);
326 362 : }
327 582 : for l in compact_from {
328 506 : Self::delete_historic_layer(l, &mut updates, &mut self.layer_fmgr);
329 506 : }
330 76 : updates.flush();
331 76 : }
332 :
333 : /// Called when a GC-compaction is completed.
334 48 : pub(crate) fn finish_gc_compaction(
335 48 : &mut self,
336 48 : compact_from: &[Layer],
337 48 : compact_to: &[ResidentLayer],
338 48 : metrics: &TimelineMetrics,
339 48 : ) {
340 48 : // We can simply reuse compact l0 logic. Use a different function name to indicate a different type of layer map modification.
341 48 : self.finish_compact_l0(compact_from, compact_to, metrics)
342 48 : }
343 :
344 : /// Called post-compaction when some previous generation image layers were trimmed.
345 0 : pub(crate) fn rewrite_layers(
346 0 : &mut self,
347 0 : rewrite_layers: &[(Layer, ResidentLayer)],
348 0 : drop_layers: &[Layer],
349 0 : metrics: &TimelineMetrics,
350 0 : ) {
351 0 : let mut updates = self.layer_map.batch_update();
352 0 : for (old_layer, new_layer) in rewrite_layers {
353 0 : debug_assert_eq!(
354 0 : old_layer.layer_desc().key_range,
355 0 : new_layer.layer_desc().key_range
356 : );
357 0 : debug_assert_eq!(
358 0 : old_layer.layer_desc().lsn_range,
359 0 : new_layer.layer_desc().lsn_range
360 : );
361 :
362 : // Transfer visibility hint from old to new layer, since the new layer covers the same key space. This is not guaranteed to
363 : // be accurate (as the new layer may cover a different subset of the key range), but is a sensible default, and prevents
364 : // always marking rewritten layers as visible.
365 0 : new_layer.as_ref().set_visibility(old_layer.visibility());
366 0 :
367 0 : // Safety: we may never rewrite the same file in-place. Callers are responsible
368 0 : // for ensuring that they only rewrite layers after something changes the path,
369 0 : // such as an increment in the generation number.
370 0 : assert_ne!(old_layer.local_path(), new_layer.local_path());
371 :
372 0 : Self::delete_historic_layer(old_layer, &mut updates, &mut self.layer_fmgr);
373 0 :
374 0 : Self::insert_historic_layer(
375 0 : new_layer.as_ref().clone(),
376 0 : &mut updates,
377 0 : &mut self.layer_fmgr,
378 0 : );
379 0 :
380 0 : metrics.record_new_file_metrics(new_layer.layer_desc().file_size);
381 : }
382 0 : for l in drop_layers {
383 0 : Self::delete_historic_layer(l, &mut updates, &mut self.layer_fmgr);
384 0 : }
385 0 : updates.flush();
386 0 : }
387 :
388 : /// Called when garbage collect has selected the layers to be removed.
389 8 : pub(crate) fn finish_gc_timeline(&mut self, gc_layers: &[Layer]) {
390 8 : let mut updates = self.layer_map.batch_update();
391 18 : for doomed_layer in gc_layers {
392 10 : Self::delete_historic_layer(doomed_layer, &mut updates, &mut self.layer_fmgr);
393 10 : }
394 8 : updates.flush()
395 8 : }
396 :
397 : #[cfg(test)]
398 146 : pub(crate) fn force_insert_layer(&mut self, layer: ResidentLayer) {
399 146 : let mut updates = self.layer_map.batch_update();
400 146 : Self::insert_historic_layer(layer.as_ref().clone(), &mut updates, &mut self.layer_fmgr);
401 146 : updates.flush()
402 146 : }
403 :
404 : /// Helper function to insert a layer into the layer map and file manager.
405 1710 : fn insert_historic_layer(
406 1710 : layer: Layer,
407 1710 : updates: &mut BatchedUpdates<'_>,
408 1710 : mapping: &mut LayerFileManager<Layer>,
409 1710 : ) {
410 1710 : updates.insert_historic(layer.layer_desc().clone());
411 1710 : mapping.insert(layer);
412 1710 : }
413 :
414 : /// Removes the layer from local FS (if present) and from memory.
415 : /// Remote storage is not affected by this operation.
416 516 : fn delete_historic_layer(
417 516 : // we cannot remove layers otherwise, since gc and compaction will race
418 516 : layer: &Layer,
419 516 : updates: &mut BatchedUpdates<'_>,
420 516 : mapping: &mut LayerFileManager<Layer>,
421 516 : ) {
422 516 : let desc = layer.layer_desc();
423 516 :
424 516 : // TODO Removing from the bottom of the layer map is expensive.
425 516 : // Maybe instead discard all layer map historic versions that
426 516 : // won't be needed for page reconstruction for this timeline,
427 516 : // and mark what we can't delete yet as deleted from the layer
428 516 : // map index without actually rebuilding the index.
429 516 : updates.remove_historic(desc);
430 516 : mapping.remove(layer);
431 516 : layer.delete_on_drop();
432 516 : }
433 : }
434 :
435 : pub(crate) struct LayerFileManager<T>(HashMap<PersistentLayerKey, T>);
436 :
437 : impl<T> Default for LayerFileManager<T> {
438 422 : fn default() -> Self {
439 422 : Self(HashMap::default())
440 422 : }
441 : }
442 :
443 : impl<T: AsLayerDesc + Clone> LayerFileManager<T> {
444 1710 : pub(crate) fn insert(&mut self, layer: T) {
445 1710 : let present = self.0.insert(layer.layer_desc().key(), layer.clone());
446 1710 : if present.is_some() && cfg!(debug_assertions) {
447 0 : panic!("overwriting a layer: {:?}", layer.layer_desc())
448 1710 : }
449 1710 : }
450 :
451 516 : pub(crate) fn remove(&mut self, layer: &T) {
452 516 : let present = self.0.remove(&layer.layer_desc().key());
453 516 : if present.is_none() && cfg!(debug_assertions) {
454 0 : panic!(
455 0 : "removing layer that is not present in layer mapping: {:?}",
456 0 : layer.layer_desc()
457 0 : )
458 516 : }
459 516 : }
460 : }
|
