Pass prev_record_ptr through zenith.signal file to compute node

pageserver/src/basebackup.rs

@@ -31,6 +31,7 @@ pub struct Basebackup<'a> {
     ar: Builder<&'a mut dyn Write>,
     timeline: &'a Arc<dyn Timeline>,
     lsn: Lsn,
+    prev_record_lsn: Lsn,
     snappath: String,
     slru_buf: [u8; pg_constants::SLRU_SEG_SIZE],
     slru_segno: u32,
@@ -43,12 +44,14 @@ impl<'a> Basebackup<'a> {
         timelineid: ZTimelineId,
         timeline: &'a Arc<dyn Timeline>,
         lsn: Lsn,
+        prev_record_lsn: Lsn,
         snapshot_lsn: Lsn,
     ) -> Basebackup<'a> {
         Basebackup {
             ar: Builder::new(write),
             timeline,
             lsn,
+            prev_record_lsn,
             snappath: format!("timelines/{}/snapshots/{:016X}", timelineid, snapshot_lsn.0),
             slru_path: "",
             slru_segno: u32::MAX,
@@ -239,8 +242,10 @@ impl<'a> Basebackup<'a> {
         pg_control.state = pg_constants::DB_SHUTDOWNED;
 
         // add zenith.signal file
-        self.ar
-            .append(&new_tar_header("zenith.signal", 0)?, &b""[..])?;
+        self.ar.append(
+            &new_tar_header("zenith.signal", 8)?,
+            &self.prev_record_lsn.0.to_le_bytes()[..],
+        )?;
 
         //send pg_control
         let pg_control_bytes = pg_control.encode();

pageserver/src/object_repository.rs

@@ -110,6 +110,7 @@ impl Repository for ObjectRepository {
         let metadata = MetadataEntry {
             last_valid_lsn: start_lsn,
             last_record_lsn: start_lsn,
+            prev_record_lsn: Lsn(0),
             ancestor_timeline: None,
             ancestor_lsn: start_lsn,
         };
@@ -146,6 +147,7 @@ impl Repository for ObjectRepository {
         let metadata = MetadataEntry {
             last_valid_lsn: at_lsn,
             last_record_lsn: at_lsn,
+            prev_record_lsn: src_timeline.get_prev_record_lsn(),
             ancestor_timeline: Some(src),
             ancestor_lsn: at_lsn,
         };
@@ -216,6 +218,7 @@ pub struct ObjectTimeline {
     //
     last_valid_lsn: SeqWait<Lsn>,
     last_record_lsn: AtomicLsn,
+    prev_record_lsn: AtomicLsn,
 
     ancestor_timeline: Option<ZTimelineId>,
     ancestor_lsn: Lsn,
@@ -244,6 +247,7 @@ impl ObjectTimeline {
             walredo_mgr,
             last_valid_lsn: SeqWait::new(metadata.last_valid_lsn),
             last_record_lsn: AtomicLsn::new(metadata.last_record_lsn.0),
+            prev_record_lsn: AtomicLsn::new(metadata.prev_record_lsn.0),
             ancestor_timeline: metadata.ancestor_timeline,
             ancestor_lsn: metadata.ancestor_lsn,
             rel_meta: RwLock::new(BTreeMap::new()),
@@ -469,7 +473,13 @@ impl Timeline for ObjectTimeline {
     ///
     /// Memorize a full image of a page version
     ///
-    fn put_page_image(&self, tag: ObjectTag, lsn: Lsn, img: Bytes, update_meta: bool) -> Result<()> {
+    fn put_page_image(
+        &self,
+        tag: ObjectTag,
+        lsn: Lsn,
+        img: Bytes,
+        update_meta: bool,
+    ) -> Result<()> {
         self.put_page_entry(&tag, lsn, PageEntry::Page(img))?;
 
         debug!("put_page_image rel {:?} at {}", tag, lsn);
@@ -552,6 +562,7 @@ impl Timeline for ObjectTimeline {
         assert!(old == Lsn(0));
         let old = self.last_record_lsn.fetch_max(lsn);
         assert!(old == Lsn(0));
+        self.prev_record_lsn.store(Lsn(0));
     }
 
     /// Like `advance_last_valid_lsn`, but this always points to the end of
@@ -566,6 +577,8 @@ impl Timeline for ObjectTimeline {
         let old = self.last_record_lsn.fetch_max(lsn);
         assert!(old <= lsn);
 
+        self.prev_record_lsn.fetch_max(old);
+
         // Also advance last_valid_lsn
         let old = self.last_valid_lsn.advance(lsn);
         // Can't move backwards.
@@ -580,6 +593,10 @@ impl Timeline for ObjectTimeline {
         self.last_record_lsn.load()
     }
 
+    fn get_prev_record_lsn(&self) -> Lsn {
+        self.prev_record_lsn.load()
+    }
+
     ///
     /// Flush to disk all data that was written with the put_* functions
     ///
@@ -593,6 +610,7 @@ impl Timeline for ObjectTimeline {
         let metadata = MetadataEntry {
             last_valid_lsn: self.last_valid_lsn.load(),
             last_record_lsn: self.last_record_lsn.load(),
+            prev_record_lsn: self.prev_record_lsn.load(),
             ancestor_timeline: self.ancestor_timeline,
             ancestor_lsn: self.ancestor_lsn,
         };
@@ -1120,6 +1138,7 @@ const fn relation_size_key(timelineid: ZTimelineId, rel: RelTag) -> ObjectKey {
 pub struct MetadataEntry {
     last_valid_lsn: Lsn,
     last_record_lsn: Lsn,
+    prev_record_lsn: Lsn,
     ancestor_timeline: Option<ZTimelineId>,
     ancestor_lsn: Lsn,
 }

pageserver/src/page_service.rs

@@ -303,6 +303,7 @@ impl PageServerHandler {
                 timelineid,
                 &timeline,
                 req_lsn,
+                timeline.get_prev_record_lsn(),
                 snapshot_lsn,
             );
             basebackup.send_tarball()?;

pageserver/src/repository.rs

@@ -89,7 +89,8 @@ pub trait Timeline: Send + Sync {
     fn put_raw_data(&self, tag: ObjectTag, lsn: Lsn, data: &[u8]) -> Result<()>;
 
     /// Like put_wal_record, but with ready-made image of the page.
-    fn put_page_image(&self, tag: ObjectTag, lsn: Lsn, img: Bytes, update_meta: bool) -> Result<()>;
+    fn put_page_image(&self, tag: ObjectTag, lsn: Lsn, img: Bytes, update_meta: bool)
+        -> Result<()>;
 
     /// Truncate relation
     fn put_truncation(&self, rel: RelTag, lsn: Lsn, nblocks: u32) -> Result<()>;
@@ -120,6 +121,9 @@ pub trait Timeline: Send + Sync {
     fn advance_last_record_lsn(&self, lsn: Lsn);
     fn get_last_record_lsn(&self) -> Lsn;
 
+    // Like `advance_last_record_lsn`, but points to the start position of last record
+    fn get_prev_record_lsn(&self) -> Lsn;
+
     ///
     /// Flush to disk all data that was written with the put_* functions
     ///

pageserver/src/restore_local_repo.rs

@@ -256,7 +256,12 @@ fn import_slru_file(
         let r = file.read_exact(&mut buf);
         match r {
             Ok(_) => {
-                timeline.put_page_image(gen_tag(blknum), lsn, Bytes::copy_from_slice(&buf), false)?;
+                timeline.put_page_image(
+                    gen_tag(blknum),
+                    lsn,
+                    Bytes::copy_from_slice(&buf),
+                    false,
+                )?;
             }
 
             // TODO: UnexpectedEof is expected

postgres_ffi/src/xlog_utils.rs

@@ -96,8 +96,8 @@ pub fn get_current_timestamp() -> TimestampTz {
     match SystemTime::now().duration_since(SystemTime::UNIX_EPOCH) {
         Ok(n) => {
             ((n.as_secs() - ((POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY))
-             * USECS_PER_SEC
-             + n.subsec_micros() as u64) as i64
+                * USECS_PER_SEC
+                + n.subsec_micros() as u64) as i64
         }
         Err(_) => panic!("SystemTime before UNIX EPOCH!"),
     }