LCOV - code coverage report
Current view: top level - pageserver/src/tenant - storage_layer.rs (source / functions) Coverage Total Hit
Test: a43a77853355b937a79c57b07a8f05607cf29e6c.info Lines: 74.8 % 341 255
Test Date: 2024-09-19 12:04:32 Functions: 76.0 % 50 38

            Line data    Source code
       1              : //! Common traits and structs for layers
       2              : 
       3              : pub mod delta_layer;
       4              : pub mod image_layer;
       5              : pub mod inmemory_layer;
       6              : pub(crate) mod layer;
       7              : mod layer_desc;
       8              : mod layer_name;
       9              : pub mod merge_iterator;
      10              : 
      11              : pub mod split_writer;
      12              : 
      13              : use crate::context::{AccessStatsBehavior, RequestContext};
      14              : use crate::repository::Value;
      15              : use crate::walrecord::NeonWalRecord;
      16              : use bytes::Bytes;
      17              : use pageserver_api::key::Key;
      18              : use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
      19              : use std::cmp::{Ordering, Reverse};
      20              : use std::collections::hash_map::Entry;
      21              : use std::collections::{BinaryHeap, HashMap};
      22              : use std::ops::Range;
      23              : use std::sync::Arc;
      24              : use std::time::{Duration, SystemTime, UNIX_EPOCH};
      25              : 
      26              : use utils::lsn::Lsn;
      27              : 
      28              : pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
      29              : pub use image_layer::{ImageLayer, ImageLayerWriter};
      30              : pub use inmemory_layer::InMemoryLayer;
      31              : pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
      32              : pub use layer_name::{DeltaLayerName, ImageLayerName, LayerName};
      33              : 
      34              : pub(crate) use layer::{EvictionError, Layer, ResidentLayer};
      35              : 
      36              : use self::inmemory_layer::InMemoryLayerFileId;
      37              : 
      38              : use super::timeline::GetVectoredError;
      39              : use super::PageReconstructError;
      40              : 
      41            0 : pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
      42            0 : where
      43            0 :     T: PartialOrd<T>,
      44            0 : {
      45            0 :     if a.start < b.start {
      46            0 :         a.end > b.start
      47              :     } else {
      48            0 :         b.end > a.start
      49              :     }
      50            0 : }
      51              : 
      52              : /// Struct used to communicate across calls to 'get_value_reconstruct_data'.
      53              : ///
      54              : /// Before first call, you can fill in 'page_img' if you have an older cached
      55              : /// version of the page available. That can save work in
      56              : /// 'get_value_reconstruct_data', as it can stop searching for page versions
      57              : /// when all the WAL records going back to the cached image have been collected.
      58              : ///
      59              : /// When get_value_reconstruct_data returns Complete, 'img' is set to an image
      60              : /// of the page, or the oldest WAL record in 'records' is a will_init-type
      61              : /// record that initializes the page without requiring a previous image.
      62              : ///
      63              : /// If 'get_page_reconstruct_data' returns Continue, some 'records' may have
      64              : /// been collected, but there are more records outside the current layer. Pass
      65              : /// the same ValueReconstructState struct in the next 'get_value_reconstruct_data'
      66              : /// call, to collect more records.
      67              : ///
      68              : #[derive(Debug, Default)]
      69              : pub(crate) struct ValueReconstructState {
      70              :     pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
      71              :     pub(crate) img: Option<(Lsn, Bytes)>,
      72              : }
      73              : 
      74              : #[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
      75              : pub(crate) enum ValueReconstructSituation {
      76              :     Complete,
      77              :     #[default]
      78              :     Continue,
      79              : }
      80              : 
      81              : /// Reconstruct data accumulated for a single key during a vectored get
      82              : #[derive(Debug, Default, Clone)]
      83              : pub(crate) struct VectoredValueReconstructState {
      84              :     pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
      85              :     pub(crate) img: Option<(Lsn, Bytes)>,
      86              : 
      87              :     situation: ValueReconstructSituation,
      88              : }
      89              : 
      90              : impl VectoredValueReconstructState {
      91       121338 :     fn get_cached_lsn(&self) -> Option<Lsn> {
      92       121338 :         self.img.as_ref().map(|img| img.0)
      93       121338 :     }
      94              : }
      95              : 
      96              : impl From<VectoredValueReconstructState> for ValueReconstructState {
      97      2000400 :     fn from(mut state: VectoredValueReconstructState) -> Self {
      98      2000400 :         // walredo expects the records to be descending in terms of Lsn
      99      2000400 :         state.records.sort_by_key(|(lsn, _)| Reverse(*lsn));
     100      2000400 : 
     101      2000400 :         ValueReconstructState {
     102      2000400 :             records: state.records,
     103      2000400 :             img: state.img,
     104      2000400 :         }
     105      2000400 :     }
     106              : }
     107              : 
     108              : /// Bag of data accumulated during a vectored get..
     109              : pub(crate) struct ValuesReconstructState {
     110              :     /// The keys will be removed after `get_vectored` completes. The caller outside `Timeline`
     111              :     /// should not expect to get anything from this hashmap.
     112              :     pub(crate) keys: HashMap<Key, Result<VectoredValueReconstructState, PageReconstructError>>,
     113              :     /// The keys which are already retrieved
     114              :     keys_done: KeySpaceRandomAccum,
     115              : 
     116              :     /// The keys covered by the image layers
     117              :     keys_with_image_coverage: Option<Range<Key>>,
     118              : 
     119              :     // Statistics that are still accessible as a caller of `get_vectored_impl`.
     120              :     layers_visited: u32,
     121              :     delta_layers_visited: u32,
     122              : }
     123              : 
     124              : impl ValuesReconstructState {
     125      1881030 :     pub(crate) fn new() -> Self {
     126      1881030 :         Self {
     127      1881030 :             keys: HashMap::new(),
     128      1881030 :             keys_done: KeySpaceRandomAccum::new(),
     129      1881030 :             keys_with_image_coverage: None,
     130      1881030 :             layers_visited: 0,
     131      1881030 :             delta_layers_visited: 0,
     132      1881030 :         }
     133      1881030 :     }
     134              : 
     135              :     /// Associate a key with the error which it encountered and mark it as done
     136            0 :     pub(crate) fn on_key_error(&mut self, key: Key, err: PageReconstructError) {
     137            0 :         let previous = self.keys.insert(key, Err(err));
     138            0 :         if let Some(Ok(state)) = previous {
     139            0 :             if state.situation == ValueReconstructSituation::Continue {
     140            0 :                 self.keys_done.add_key(key);
     141            0 :             }
     142            0 :         }
     143            0 :     }
     144              : 
     145      2459034 :     pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
     146      2459034 :         self.layers_visited += 1;
     147      2459034 :         if let ReadableLayer::PersistentLayer(layer) = layer {
     148       639622 :             if layer.layer_desc().is_delta() {
     149       615597 :                 self.delta_layers_visited += 1;
     150       615597 :             }
     151      1819412 :         }
     152      2459034 :     }
     153              : 
     154          588 :     pub(crate) fn get_delta_layers_visited(&self) -> u32 {
     155          588 :         self.delta_layers_visited
     156          588 :     }
     157              : 
     158      1880340 :     pub(crate) fn get_layers_visited(&self) -> u32 {
     159      1880340 :         self.layers_visited
     160      1880340 :     }
     161              : 
     162              :     /// This function is called after reading a keyspace from a layer.
     163              :     /// It checks if the read path has now moved past the cached Lsn for any keys.
     164              :     ///
     165              :     /// Implementation note: We intentionally iterate over the keys for which we've
     166              :     /// already collected some reconstruct data. This avoids scaling complexity with
     167              :     /// the size of the search space.
     168      2435009 :     pub(crate) fn on_lsn_advanced(&mut self, keyspace: &KeySpace, advanced_to: Lsn) {
     169      2435009 :         for (key, value) in self.keys.iter_mut() {
     170      2073259 :             if !keyspace.contains(key) {
     171       126711 :                 continue;
     172      1946548 :             }
     173              : 
     174      1946548 :             if let Ok(state) = value {
     175      1946548 :                 if state.situation != ValueReconstructSituation::Complete
     176          930 :                     && state.get_cached_lsn() >= Some(advanced_to)
     177            0 :                 {
     178            0 :                     state.situation = ValueReconstructSituation::Complete;
     179            0 :                     self.keys_done.add_key(*key);
     180      1946548 :                 }
     181            0 :             }
     182              :         }
     183      2435009 :     }
     184              : 
     185              :     /// On hitting image layer, we can mark all keys in this range as done, because
     186              :     /// if the image layer does not contain a key, it is deleted/never added.
     187        24061 :     pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
     188        24061 :         let prev_val = self.keys_with_image_coverage.replace(key_range.clone());
     189        24061 :         assert_eq!(
     190              :             prev_val, None,
     191            0 :             "should consume the keyspace before the next iteration"
     192              :         );
     193        24061 :     }
     194              : 
     195              :     /// Update the state collected for a given key.
     196              :     /// Returns true if this was the last value needed for the key and false otherwise.
     197              :     ///
     198              :     /// If the key is done after the update, mark it as such.
     199      2001834 :     pub(crate) fn update_key(
     200      2001834 :         &mut self,
     201      2001834 :         key: &Key,
     202      2001834 :         lsn: Lsn,
     203      2001834 :         value: Value,
     204      2001834 :     ) -> ValueReconstructSituation {
     205      2001834 :         let state = self
     206      2001834 :             .keys
     207      2001834 :             .entry(*key)
     208      2001834 :             .or_insert(Ok(VectoredValueReconstructState::default()));
     209              : 
     210      2001834 :         if let Ok(state) = state {
     211      2001834 :             let key_done = match state.situation {
     212            0 :                 ValueReconstructSituation::Complete => unreachable!(),
     213      2001834 :                 ValueReconstructSituation::Continue => match value {
     214      2000610 :                     Value::Image(img) => {
     215      2000610 :                         state.img = Some((lsn, img));
     216      2000610 :                         true
     217              :                     }
     218         1224 :                     Value::WalRecord(rec) => {
     219         1224 :                         debug_assert!(
     220         1224 :                             Some(lsn) > state.get_cached_lsn(),
     221            0 :                             "Attempt to collect a record below cached LSN for walredo: {} < {}",
     222            0 :                             lsn,
     223            0 :                             state
     224            0 :                                 .get_cached_lsn()
     225            0 :                                 .expect("Assertion can only fire if a cached lsn is present")
     226              :                         );
     227              : 
     228         1224 :                         let will_init = rec.will_init();
     229         1224 :                         state.records.push((lsn, rec));
     230         1224 :                         will_init
     231              :                     }
     232              :                 },
     233              :             };
     234              : 
     235      2001834 :             if key_done && state.situation == ValueReconstructSituation::Continue {
     236      2000610 :                 state.situation = ValueReconstructSituation::Complete;
     237      2000610 :                 self.keys_done.add_key(*key);
     238      2000610 :             }
     239              : 
     240      2001834 :             state.situation
     241              :         } else {
     242            0 :             ValueReconstructSituation::Complete
     243              :         }
     244      2001834 :     }
     245              : 
     246              :     /// Returns the Lsn at which this key is cached if one exists.
     247              :     /// The read path should go no further than this Lsn for the given key.
     248      3250335 :     pub(crate) fn get_cached_lsn(&self, key: &Key) -> Option<Lsn> {
     249      3250335 :         self.keys
     250      3250335 :             .get(key)
     251      3250335 :             .and_then(|k| k.as_ref().ok())
     252      3250335 :             .and_then(|state| state.get_cached_lsn())
     253      3250335 :     }
     254              : 
     255              :     /// Returns the key space describing the keys that have
     256              :     /// been marked as completed since the last call to this function.
     257              :     /// Returns individual keys done, and the image layer coverage.
     258      5017840 :     pub(crate) fn consume_done_keys(&mut self) -> (KeySpace, Option<Range<Key>>) {
     259      5017840 :         (
     260      5017840 :             self.keys_done.consume_keyspace(),
     261      5017840 :             self.keys_with_image_coverage.take(),
     262      5017840 :         )
     263      5017840 :     }
     264              : }
     265              : 
     266              : impl Default for ValuesReconstructState {
     267          762 :     fn default() -> Self {
     268          762 :         Self::new()
     269          762 :     }
     270              : }
     271              : 
     272              : /// A key that uniquely identifies a layer in a timeline
     273              : #[derive(Debug, PartialEq, Eq, Clone, Hash)]
     274              : pub(crate) enum LayerId {
     275              :     PersitentLayerId(PersistentLayerKey),
     276              :     InMemoryLayerId(InMemoryLayerFileId),
     277              : }
     278              : 
     279              : /// Layer wrapper for the read path. Note that it is valid
     280              : /// to use these layers even after external operations have
     281              : /// been performed on them (compaction, freeze, etc.).
     282              : #[derive(Debug)]
     283              : pub(crate) enum ReadableLayer {
     284              :     PersistentLayer(Layer),
     285              :     InMemoryLayer(Arc<InMemoryLayer>),
     286              : }
     287              : 
     288              : /// A partial description of a read to be done.
     289              : #[derive(Debug, Clone)]
     290              : struct ReadDesc {
     291              :     /// An id used to resolve the readable layer within the fringe
     292              :     layer_id: LayerId,
     293              :     /// Lsn range for the read, used for selecting the next read
     294              :     lsn_range: Range<Lsn>,
     295              : }
     296              : 
     297              : /// Data structure which maintains a fringe of layers for the
     298              : /// read path. The fringe is the set of layers which intersects
     299              : /// the current keyspace that the search is descending on.
     300              : /// Each layer tracks the keyspace that intersects it.
     301              : ///
     302              : /// The fringe must appear sorted by Lsn. Hence, it uses
     303              : /// a two layer indexing scheme.
     304              : #[derive(Debug)]
     305              : pub(crate) struct LayerFringe {
     306              :     planned_reads_by_lsn: BinaryHeap<ReadDesc>,
     307              :     layers: HashMap<LayerId, LayerKeyspace>,
     308              : }
     309              : 
     310              : #[derive(Debug)]
     311              : struct LayerKeyspace {
     312              :     layer: ReadableLayer,
     313              :     target_keyspace: KeySpaceRandomAccum,
     314              : }
     315              : 
     316              : impl LayerFringe {
     317      2558806 :     pub(crate) fn new() -> Self {
     318      2558806 :         LayerFringe {
     319      2558806 :             planned_reads_by_lsn: BinaryHeap::new(),
     320      2558806 :             layers: HashMap::new(),
     321      2558806 :         }
     322      2558806 :     }
     323              : 
     324      5017840 :     pub(crate) fn next_layer(&mut self) -> Option<(ReadableLayer, KeySpace, Range<Lsn>)> {
     325      5017840 :         let read_desc = match self.planned_reads_by_lsn.pop() {
     326      2459034 :             Some(desc) => desc,
     327      2558806 :             None => return None,
     328              :         };
     329              : 
     330      2459034 :         let removed = self.layers.remove_entry(&read_desc.layer_id);
     331      2459034 : 
     332      2459034 :         match removed {
     333              :             Some((
     334              :                 _,
     335              :                 LayerKeyspace {
     336      2459034 :                     layer,
     337      2459034 :                     mut target_keyspace,
     338      2459034 :                 },
     339      2459034 :             )) => Some((
     340      2459034 :                 layer,
     341      2459034 :                 target_keyspace.consume_keyspace(),
     342      2459034 :                 read_desc.lsn_range,
     343      2459034 :             )),
     344            0 :             None => unreachable!("fringe internals are always consistent"),
     345              :         }
     346      5017840 :     }
     347              : 
     348      2650412 :     pub(crate) fn update(
     349      2650412 :         &mut self,
     350      2650412 :         layer: ReadableLayer,
     351      2650412 :         keyspace: KeySpace,
     352      2650412 :         lsn_range: Range<Lsn>,
     353      2650412 :     ) {
     354      2650412 :         let layer_id = layer.id();
     355      2650412 :         let entry = self.layers.entry(layer_id.clone());
     356      2650412 :         match entry {
     357       191378 :             Entry::Occupied(mut entry) => {
     358       191378 :                 entry.get_mut().target_keyspace.add_keyspace(keyspace);
     359       191378 :             }
     360      2459034 :             Entry::Vacant(entry) => {
     361      2459034 :                 self.planned_reads_by_lsn.push(ReadDesc {
     362      2459034 :                     lsn_range,
     363      2459034 :                     layer_id: layer_id.clone(),
     364      2459034 :                 });
     365      2459034 :                 let mut accum = KeySpaceRandomAccum::new();
     366      2459034 :                 accum.add_keyspace(keyspace);
     367      2459034 :                 entry.insert(LayerKeyspace {
     368      2459034 :                     layer,
     369      2459034 :                     target_keyspace: accum,
     370      2459034 :                 });
     371      2459034 :             }
     372              :         }
     373      2650412 :     }
     374              : }
     375              : 
     376              : impl Default for LayerFringe {
     377            0 :     fn default() -> Self {
     378            0 :         Self::new()
     379            0 :     }
     380              : }
     381              : 
     382              : impl Ord for ReadDesc {
     383           48 :     fn cmp(&self, other: &Self) -> Ordering {
     384           48 :         let ord = self.lsn_range.end.cmp(&other.lsn_range.end);
     385           48 :         if ord == std::cmp::Ordering::Equal {
     386           48 :             self.lsn_range.start.cmp(&other.lsn_range.start).reverse()
     387              :         } else {
     388            0 :             ord
     389              :         }
     390           48 :     }
     391              : }
     392              : 
     393              : impl PartialOrd for ReadDesc {
     394           48 :     fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
     395           48 :         Some(self.cmp(other))
     396           48 :     }
     397              : }
     398              : 
     399              : impl PartialEq for ReadDesc {
     400            0 :     fn eq(&self, other: &Self) -> bool {
     401            0 :         self.lsn_range == other.lsn_range
     402            0 :     }
     403              : }
     404              : 
     405              : impl Eq for ReadDesc {}
     406              : 
     407              : impl ReadableLayer {
     408      2650412 :     pub(crate) fn id(&self) -> LayerId {
     409      2650412 :         match self {
     410       831000 :             Self::PersistentLayer(layer) => LayerId::PersitentLayerId(layer.layer_desc().key()),
     411      1819412 :             Self::InMemoryLayer(layer) => LayerId::InMemoryLayerId(layer.file_id()),
     412              :         }
     413      2650412 :     }
     414              : 
     415      2459034 :     pub(crate) async fn get_values_reconstruct_data(
     416      2459034 :         &self,
     417      2459034 :         keyspace: KeySpace,
     418      2459034 :         lsn_range: Range<Lsn>,
     419      2459034 :         reconstruct_state: &mut ValuesReconstructState,
     420      2459034 :         ctx: &RequestContext,
     421      2459034 :     ) -> Result<(), GetVectoredError> {
     422      2459034 :         match self {
     423       639622 :             ReadableLayer::PersistentLayer(layer) => {
     424       639622 :                 layer
     425       639622 :                     .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_state, ctx)
     426       286580 :                     .await
     427              :             }
     428      1819412 :             ReadableLayer::InMemoryLayer(layer) => {
     429      1819412 :                 layer
     430      1819412 :                     .get_values_reconstruct_data(keyspace, lsn_range.end, reconstruct_state, ctx)
     431       260688 :                     .await
     432              :             }
     433              :         }
     434      2459034 :     }
     435              : }
     436              : 
     437              : /// Layers contain a hint indicating whether they are likely to be used for reads.
     438              : ///
     439              : /// This is a hint rather than an authoritative value, so that we do not have to update it synchronously
     440              : /// when changing the visibility of layers (for example when creating a branch that makes some previously
     441              : /// covered layers visible).  It should be used for cache management but not for correctness-critical checks.
     442              : #[derive(Debug, Clone, PartialEq, Eq)]
     443              : pub enum LayerVisibilityHint {
     444              :     /// A Visible layer might be read while serving a read, because there is not an image layer between it
     445              :     /// and a readable LSN (the tip of the branch or a child's branch point)
     446              :     Visible,
     447              :     /// A Covered layer probably won't be read right now, but _can_ be read in future if someone creates
     448              :     /// a branch or ephemeral endpoint at an LSN below the layer that covers this.
     449              :     Covered,
     450              : }
     451              : 
     452              : pub(crate) struct LayerAccessStats(std::sync::atomic::AtomicU64);
     453              : 
     454            0 : #[derive(Clone, Copy, strum_macros::EnumString)]
     455              : pub(crate) enum LayerAccessStatsReset {
     456              :     NoReset,
     457              :     AllStats,
     458              : }
     459              : 
     460              : impl Default for LayerAccessStats {
     461         5178 :     fn default() -> Self {
     462         5178 :         // Default value is to assume resident since creation time, and visible.
     463         5178 :         let (_mask, mut value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, SystemTime::now());
     464         5178 :         value |= 0x1 << Self::VISIBILITY_SHIFT;
     465         5178 : 
     466         5178 :         Self(std::sync::atomic::AtomicU64::new(value))
     467         5178 :     }
     468              : }
     469              : 
     470              : // Efficient store of two very-low-resolution timestamps and some bits.  Used for storing last access time and
     471              : // last residence change time.
     472              : impl LayerAccessStats {
     473              :     // How many high bits to drop from a u32 timestamp?
     474              :     // - Only storing up to a u32 timestamp will work fine until 2038 (if this code is still in use
     475              :     //   after that, this software has been very successful!)
     476              :     // - Dropping the top bit is implicitly safe because unix timestamps are meant to be
     477              :     // stored in an i32, so they never used it.
     478              :     // - Dropping the next two bits is safe because this code is only running on systems in
     479              :     // years >= 2024, and these bits have been 1 since 2021
     480              :     //
     481              :     // Therefore we may store only 28 bits for a timestamp with one second resolution.  We do
     482              :     // this truncation to make space for some flags in the high bits of our u64.
     483              :     const TS_DROP_HIGH_BITS: u32 = u32::count_ones(Self::TS_ONES) + 1;
     484              :     const TS_MASK: u32 = 0x1f_ff_ff_ff;
     485              :     const TS_ONES: u32 = 0x60_00_00_00;
     486              : 
     487              :     const ATIME_SHIFT: u32 = 0;
     488              :     const RTIME_SHIFT: u32 = 32 - Self::TS_DROP_HIGH_BITS;
     489              :     const VISIBILITY_SHIFT: u32 = 64 - 2 * Self::TS_DROP_HIGH_BITS;
     490              : 
     491       648214 :     fn write_bits(&self, mask: u64, value: u64) -> u64 {
     492       648214 :         self.0
     493       648214 :             .fetch_update(
     494       648214 :                 // TODO: decide what orderings are correct
     495       648214 :                 std::sync::atomic::Ordering::Relaxed,
     496       648214 :                 std::sync::atomic::Ordering::Relaxed,
     497       648214 :                 |v| Some((v & !mask) | (value & mask)),
     498       648214 :             )
     499       648214 :             .expect("Inner function is infallible")
     500       648214 :     }
     501              : 
     502       645298 :     fn to_low_res_timestamp(shift: u32, time: SystemTime) -> (u64, u64) {
     503       645298 :         // Drop the low three bits of the timestamp, for an ~8s accuracy
     504       645298 :         let timestamp = time.duration_since(UNIX_EPOCH).unwrap().as_secs() & (Self::TS_MASK as u64);
     505       645298 : 
     506       645298 :         ((Self::TS_MASK as u64) << shift, timestamp << shift)
     507       645298 :     }
     508              : 
     509          186 :     fn read_low_res_timestamp(&self, shift: u32) -> Option<SystemTime> {
     510          186 :         let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
     511          186 : 
     512          186 :         let ts_bits = (read & ((Self::TS_MASK as u64) << shift)) >> shift;
     513          186 :         if ts_bits == 0 {
     514           72 :             None
     515              :         } else {
     516          114 :             Some(UNIX_EPOCH + Duration::from_secs(ts_bits | (Self::TS_ONES as u64)))
     517              :         }
     518          186 :     }
     519              : 
     520              :     /// Record a change in layer residency.
     521              :     ///
     522              :     /// Recording the event must happen while holding the layer map lock to
     523              :     /// ensure that latest-activity-threshold-based layer eviction (eviction_task.rs)
     524              :     /// can do an "imitate access" to this layer, before it observes `now-latest_activity() > threshold`.
     525              :     ///
     526              :     /// If we instead recorded the residence event with a timestamp from before grabbing the layer map lock,
     527              :     /// the following race could happen:
     528              :     ///
     529              :     /// - Compact: Write out an L1 layer from several L0 layers. This records residence event LayerCreate with the current timestamp.
     530              :     /// - Eviction: imitate access logical size calculation. This accesses the L0 layers because the L1 layer is not yet in the layer map.
     531              :     /// - Compact: Grab layer map lock, add the new L1 to layer map and remove the L0s, release layer map lock.
     532              :     /// - Eviction: observes the new L1 layer whose only activity timestamp is the LayerCreate event.
     533           78 :     pub(crate) fn record_residence_event_at(&self, now: SystemTime) {
     534           78 :         let (mask, value) = Self::to_low_res_timestamp(Self::RTIME_SHIFT, now);
     535           78 :         self.write_bits(mask, value);
     536           78 :     }
     537              : 
     538           72 :     pub(crate) fn record_residence_event(&self) {
     539           72 :         self.record_residence_event_at(SystemTime::now())
     540           72 :     }
     541              : 
     542       640042 :     fn record_access_at(&self, now: SystemTime) -> bool {
     543       640042 :         let (mut mask, mut value) = Self::to_low_res_timestamp(Self::ATIME_SHIFT, now);
     544       640042 : 
     545       640042 :         // A layer which is accessed must be visible.
     546       640042 :         mask |= 0x1 << Self::VISIBILITY_SHIFT;
     547       640042 :         value |= 0x1 << Self::VISIBILITY_SHIFT;
     548       640042 : 
     549       640042 :         let old_bits = self.write_bits(mask, value);
     550            0 :         !matches!(
     551       640042 :             self.decode_visibility(old_bits),
     552              :             LayerVisibilityHint::Visible
     553              :         )
     554       640042 :     }
     555              : 
     556              :     /// Returns true if we modified the layer's visibility to set it to Visible implicitly
     557              :     /// as a result of this access
     558       640864 :     pub(crate) fn record_access(&self, ctx: &RequestContext) -> bool {
     559       640864 :         if ctx.access_stats_behavior() == AccessStatsBehavior::Skip {
     560          828 :             return false;
     561       640036 :         }
     562       640036 : 
     563       640036 :         self.record_access_at(SystemTime::now())
     564       640864 :     }
     565              : 
     566            0 :     fn as_api_model(
     567            0 :         &self,
     568            0 :         reset: LayerAccessStatsReset,
     569            0 :     ) -> pageserver_api::models::LayerAccessStats {
     570            0 :         let ret = pageserver_api::models::LayerAccessStats {
     571            0 :             access_time: self
     572            0 :                 .read_low_res_timestamp(Self::ATIME_SHIFT)
     573            0 :                 .unwrap_or(UNIX_EPOCH),
     574            0 :             residence_time: self
     575            0 :                 .read_low_res_timestamp(Self::RTIME_SHIFT)
     576            0 :                 .unwrap_or(UNIX_EPOCH),
     577            0 :             visible: matches!(self.visibility(), LayerVisibilityHint::Visible),
     578              :         };
     579            0 :         match reset {
     580            0 :             LayerAccessStatsReset::NoReset => {}
     581            0 :             LayerAccessStatsReset::AllStats => {
     582            0 :                 self.write_bits((Self::TS_MASK as u64) << Self::ATIME_SHIFT, 0x0);
     583            0 :                 self.write_bits((Self::TS_MASK as u64) << Self::RTIME_SHIFT, 0x0);
     584            0 :             }
     585              :         }
     586            0 :         ret
     587            0 :     }
     588              : 
     589              :     /// Get the latest access timestamp, falling back to latest residence event.  The latest residence event
     590              :     /// will be this Layer's construction time, if its residence hasn't changed since then.
     591           48 :     pub(crate) fn latest_activity(&self) -> SystemTime {
     592           48 :         if let Some(t) = self.read_low_res_timestamp(Self::ATIME_SHIFT) {
     593           18 :             t
     594              :         } else {
     595           30 :             self.read_low_res_timestamp(Self::RTIME_SHIFT)
     596           30 :                 .expect("Residence time is set on construction")
     597              :         }
     598           48 :     }
     599              : 
     600              :     /// Whether this layer has been accessed (excluding in [`AccessStatsBehavior::Skip`]).
     601              :     ///
     602              :     /// This indicates whether the layer has been used for some purpose that would motivate
     603              :     /// us to keep it on disk, such as for serving a getpage request.
     604           54 :     fn accessed(&self) -> bool {
     605           54 :         // Consider it accessed if the most recent access is more recent than
     606           54 :         // the most recent change in residence status.
     607           54 :         match (
     608           54 :             self.read_low_res_timestamp(Self::ATIME_SHIFT),
     609           54 :             self.read_low_res_timestamp(Self::RTIME_SHIFT),
     610              :         ) {
     611           42 :             (None, _) => false,
     612            0 :             (Some(_), None) => true,
     613           12 :             (Some(a), Some(r)) => a >= r,
     614              :         }
     615           54 :     }
     616              : 
     617              :     /// Helper for extracting the visibility hint from the literal value of our inner u64
     618       649872 :     fn decode_visibility(&self, bits: u64) -> LayerVisibilityHint {
     619       649872 :         match (bits >> Self::VISIBILITY_SHIFT) & 0x1 {
     620       649770 :             1 => LayerVisibilityHint::Visible,
     621          102 :             0 => LayerVisibilityHint::Covered,
     622            0 :             _ => unreachable!(),
     623              :         }
     624       649872 :     }
     625              : 
     626              :     /// Returns the old value which has been replaced
     627         8094 :     pub(crate) fn set_visibility(&self, visibility: LayerVisibilityHint) -> LayerVisibilityHint {
     628         8094 :         let value = match visibility {
     629         7914 :             LayerVisibilityHint::Visible => 0x1 << Self::VISIBILITY_SHIFT,
     630          180 :             LayerVisibilityHint::Covered => 0x0,
     631              :         };
     632              : 
     633         8094 :         let old_bits = self.write_bits(0x1 << Self::VISIBILITY_SHIFT, value);
     634         8094 :         self.decode_visibility(old_bits)
     635         8094 :     }
     636              : 
     637         1736 :     pub(crate) fn visibility(&self) -> LayerVisibilityHint {
     638         1736 :         let read = self.0.load(std::sync::atomic::Ordering::Relaxed);
     639         1736 :         self.decode_visibility(read)
     640         1736 :     }
     641              : }
     642              : 
     643              : /// Get a layer descriptor from a layer.
     644              : pub(crate) trait AsLayerDesc {
     645              :     /// Get the layer descriptor.
     646              :     fn layer_desc(&self) -> &PersistentLayerDesc;
     647              : }
     648              : 
     649              : pub mod tests {
     650              :     use pageserver_api::shard::TenantShardId;
     651              :     use utils::id::TimelineId;
     652              : 
     653              :     use super::*;
     654              : 
     655              :     impl From<DeltaLayerName> for PersistentLayerDesc {
     656            0 :         fn from(value: DeltaLayerName) -> Self {
     657            0 :             PersistentLayerDesc::new_delta(
     658            0 :                 TenantShardId::from([0; 18]),
     659            0 :                 TimelineId::from_array([0; 16]),
     660            0 :                 value.key_range,
     661            0 :                 value.lsn_range,
     662            0 :                 233,
     663            0 :             )
     664            0 :         }
     665              :     }
     666              : 
     667              :     impl From<ImageLayerName> for PersistentLayerDesc {
     668            0 :         fn from(value: ImageLayerName) -> Self {
     669            0 :             PersistentLayerDesc::new_img(
     670            0 :                 TenantShardId::from([0; 18]),
     671            0 :                 TimelineId::from_array([0; 16]),
     672            0 :                 value.key_range,
     673            0 :                 value.lsn,
     674            0 :                 233,
     675            0 :             )
     676            0 :         }
     677              :     }
     678              : 
     679              :     impl From<LayerName> for PersistentLayerDesc {
     680            0 :         fn from(value: LayerName) -> Self {
     681            0 :             match value {
     682            0 :                 LayerName::Delta(d) => Self::from(d),
     683            0 :                 LayerName::Image(i) => Self::from(i),
     684              :             }
     685            0 :         }
     686              :     }
     687              : }
     688              : 
     689              : /// Range wrapping newtype, which uses display to render Debug.
     690              : ///
     691              : /// Useful with `Key`, which has too verbose `{:?}` for printing multiple layers.
     692              : struct RangeDisplayDebug<'a, T: std::fmt::Display>(&'a Range<T>);
     693              : 
     694              : impl<'a, T: std::fmt::Display> std::fmt::Debug for RangeDisplayDebug<'a, T> {
     695            0 :     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
     696            0 :         write!(f, "{}..{}", self.0.start, self.0.end)
     697            0 :     }
     698              : }
        

Generated by: LCOV version 2.1-beta