pageserver_api: migrate keyspace related functions from pgdatadir_mapping (#6406)

The idea is to achieve separation between keyspace layout definition
and operating on said keyspace. I've inlined all these function since
they're small and we don't use LTO in the storage release builds
at the moment.

Closes https://github.com/neondatabase/neon/issues/6347
This commit is contained in:
Vlad Lazar
2024-01-22 19:16:38 +00:00
committed by GitHub
parent b41ee81308
commit f1901833a6
6 changed files with 394 additions and 368 deletions

View File

@@ -1,9 +1,11 @@
use anyhow::{bail, Result};
use byteorder::{ByteOrder, BE};
use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
use postgres_ffi::{Oid, TransactionId};
use serde::{Deserialize, Serialize};
use std::fmt;
use std::{fmt, ops::Range};
use crate::reltag::{BlockNumber, RelTag};
use crate::reltag::{BlockNumber, RelTag, SlruKind};
/// Key used in the Repository kv-store.
///
@@ -143,12 +145,390 @@ impl Key {
}
}
// Layout of the Key address space
//
// The Key struct, used to address the underlying key-value store, consists of
// 18 bytes, split into six fields. See 'Key' in repository.rs. We need to map
// all the data and metadata keys into those 18 bytes.
//
// Principles for the mapping:
//
// - Things that are often accessed or modified together, should be close to
// each other in the key space. For example, if a relation is extended by one
// block, we create a new key-value pair for the block data, and update the
// relation size entry. Because of that, the RelSize key comes after all the
// RelBlocks of a relation: the RelSize and the last RelBlock are always next
// to each other.
//
// The key space is divided into four major sections, identified by the first
// byte, and the form a hierarchy:
//
// 00 Relation data and metadata
//
// DbDir () -> (dbnode, spcnode)
// Filenodemap
// RelDir -> relnode forknum
// RelBlocks
// RelSize
//
// 01 SLRUs
//
// SlruDir kind
// SlruSegBlocks segno
// SlruSegSize
//
// 02 pg_twophase
//
// 03 misc
// Controlfile
// checkpoint
// pg_version
//
// 04 aux files
//
// Below is a full list of the keyspace allocation:
//
// DbDir:
// 00 00000000 00000000 00000000 00 00000000
//
// Filenodemap:
// 00 SPCNODE DBNODE 00000000 00 00000000
//
// RelDir:
// 00 SPCNODE DBNODE 00000000 00 00000001 (Postgres never uses relfilenode 0)
//
// RelBlock:
// 00 SPCNODE DBNODE RELNODE FORK BLKNUM
//
// RelSize:
// 00 SPCNODE DBNODE RELNODE FORK FFFFFFFF
//
// SlruDir:
// 01 kind 00000000 00000000 00 00000000
//
// SlruSegBlock:
// 01 kind 00000001 SEGNO 00 BLKNUM
//
// SlruSegSize:
// 01 kind 00000001 SEGNO 00 FFFFFFFF
//
// TwoPhaseDir:
// 02 00000000 00000000 00000000 00 00000000
//
// TwoPhaseFile:
// 02 00000000 00000000 00000000 00 XID
//
// ControlFile:
// 03 00000000 00000000 00000000 00 00000000
//
// Checkpoint:
// 03 00000000 00000000 00000000 00 00000001
//
// AuxFiles:
// 03 00000000 00000000 00000000 00 00000002
//
//-- Section 01: relation data and metadata
pub const DBDIR_KEY: Key = Key {
field1: 0x00,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
#[inline(always)]
pub fn dbdir_key_range(spcnode: Oid, dbnode: Oid) -> Range<Key> {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 0,
}..Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0xffffffff,
field5: 0xff,
field6: 0xffffffff,
}
}
#[inline(always)]
pub fn relmap_file_key(spcnode: Oid, dbnode: Oid) -> Key {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 0,
}
}
#[inline(always)]
pub fn rel_dir_to_key(spcnode: Oid, dbnode: Oid) -> Key {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 1,
}
}
#[inline(always)]
pub fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: blknum,
}
}
#[inline(always)]
pub fn rel_size_to_key(rel: RelTag) -> Key {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: 0xffffffff,
}
}
#[inline(always)]
pub fn rel_key_range(rel: RelTag) -> Range<Key> {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: 0,
}..Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum + 1,
field6: 0,
}
}
//-- Section 02: SLRUs
#[inline(always)]
pub fn slru_dir_to_key(kind: SlruKind) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 0,
field4: 0,
field5: 0,
field6: 0,
}
}
#[inline(always)]
pub fn slru_block_to_key(kind: SlruKind, segno: u32, blknum: BlockNumber) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 1,
field4: segno,
field5: 0,
field6: blknum,
}
}
#[inline(always)]
pub fn slru_segment_size_to_key(kind: SlruKind, segno: u32) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 1,
field4: segno,
field5: 0,
field6: 0xffffffff,
}
}
#[inline(always)]
pub fn slru_segment_key_range(kind: SlruKind, segno: u32) -> Range<Key> {
let field2 = match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
};
Key {
field1: 0x01,
field2,
field3: 1,
field4: segno,
field5: 0,
field6: 0,
}..Key {
field1: 0x01,
field2,
field3: 1,
field4: segno,
field5: 1,
field6: 0,
}
}
//-- Section 03: pg_twophase
pub const TWOPHASEDIR_KEY: Key = Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
#[inline(always)]
pub fn twophase_file_key(xid: TransactionId) -> Key {
Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: xid,
}
}
#[inline(always)]
pub fn twophase_key_range(xid: TransactionId) -> Range<Key> {
let (next_xid, overflowed) = xid.overflowing_add(1);
Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: xid,
}..Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: u8::from(overflowed),
field6: next_xid,
}
}
//-- Section 03: Control file
pub const CONTROLFILE_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
pub const CHECKPOINT_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 1,
};
pub const AUX_FILES_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 2,
};
// Reverse mappings for a few Keys.
// These are needed by WAL redo manager.
// AUX_FILES currently stores only data for logical replication (slots etc), and
// we don't preserve these on a branch because safekeepers can't follow timeline
// switch (and generally it likely should be optional), so ignore these.
#[inline(always)]
pub fn is_inherited_key(key: Key) -> bool {
key != AUX_FILES_KEY
}
#[inline(always)]
pub fn is_rel_fsm_block_key(key: Key) -> bool {
key.field1 == 0x00 && key.field4 != 0 && key.field5 == FSM_FORKNUM && key.field6 != 0xffffffff
}
#[inline(always)]
pub fn is_rel_vm_block_key(key: Key) -> bool {
key.field1 == 0x00
&& key.field4 != 0
&& key.field5 == VISIBILITYMAP_FORKNUM
&& key.field6 != 0xffffffff
}
#[inline(always)]
pub fn key_to_slru_block(key: Key) -> anyhow::Result<(SlruKind, u32, BlockNumber)> {
Ok(match key.field1 {
0x01 => {
let kind = match key.field2 {
0x00 => SlruKind::Clog,
0x01 => SlruKind::MultiXactMembers,
0x02 => SlruKind::MultiXactOffsets,
_ => anyhow::bail!("unrecognized slru kind 0x{:02x}", key.field2),
};
let segno = key.field4;
let blknum = key.field6;
(kind, segno, blknum)
}
_ => anyhow::bail!("unexpected value kind 0x{:02x}", key.field1),
})
}
#[inline(always)]
pub fn is_slru_block_key(key: Key) -> bool {
key.field1 == 0x01 // SLRU-related
&& key.field3 == 0x00000001 // but not SlruDir
&& key.field6 != 0xffffffff // and not SlruSegSize
}
#[inline(always)]
pub fn is_rel_block_key(key: &Key) -> bool {
key.field1 == 0x00 && key.field4 != 0 && key.field6 != 0xffffffff
}
/// Guaranteed to return `Ok()` if [[is_rel_block_key]] returns `true` for `key`.
#[inline(always)]
pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
Ok(match key.field1 {
0x00 => (

View File

@@ -61,7 +61,7 @@ use crate::context::{DownloadBehavior, RequestContext};
use crate::import_datadir::import_wal_from_tar;
use crate::metrics;
use crate::metrics::LIVE_CONNECTIONS_COUNT;
use crate::pgdatadir_mapping::{rel_block_to_key, Version};
use crate::pgdatadir_mapping::Version;
use crate::task_mgr;
use crate::task_mgr::TaskKind;
use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
@@ -75,6 +75,7 @@ use crate::tenant::PageReconstructError;
use crate::tenant::Timeline;
use crate::trace::Tracer;
use pageserver_api::key::rel_block_to_key;
use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
use postgres_ffi::BLCKSZ;

View File

@@ -13,7 +13,12 @@ use crate::repository::*;
use crate::walrecord::NeonWalRecord;
use anyhow::{ensure, Context};
use bytes::{Buf, Bytes};
use pageserver_api::key::is_rel_block_key;
use pageserver_api::key::{
dbdir_key_range, is_rel_block_key, is_slru_block_key, rel_block_to_key, rel_dir_to_key,
rel_key_range, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
slru_segment_key_range, slru_segment_size_to_key, twophase_file_key, twophase_key_range,
AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, DBDIR_KEY, TWOPHASEDIR_KEY,
};
use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
use postgres_ffi::BLCKSZ;
@@ -1535,366 +1540,6 @@ struct SlruSegmentDirectory {
static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
// Layout of the Key address space
//
// The Key struct, used to address the underlying key-value store, consists of
// 18 bytes, split into six fields. See 'Key' in repository.rs. We need to map
// all the data and metadata keys into those 18 bytes.
//
// Principles for the mapping:
//
// - Things that are often accessed or modified together, should be close to
// each other in the key space. For example, if a relation is extended by one
// block, we create a new key-value pair for the block data, and update the
// relation size entry. Because of that, the RelSize key comes after all the
// RelBlocks of a relation: the RelSize and the last RelBlock are always next
// to each other.
//
// The key space is divided into four major sections, identified by the first
// byte, and the form a hierarchy:
//
// 00 Relation data and metadata
//
// DbDir () -> (dbnode, spcnode)
// Filenodemap
// RelDir -> relnode forknum
// RelBlocks
// RelSize
//
// 01 SLRUs
//
// SlruDir kind
// SlruSegBlocks segno
// SlruSegSize
//
// 02 pg_twophase
//
// 03 misc
// Controlfile
// checkpoint
// pg_version
//
// 04 aux files
//
// Below is a full list of the keyspace allocation:
//
// DbDir:
// 00 00000000 00000000 00000000 00 00000000
//
// Filenodemap:
// 00 SPCNODE DBNODE 00000000 00 00000000
//
// RelDir:
// 00 SPCNODE DBNODE 00000000 00 00000001 (Postgres never uses relfilenode 0)
//
// RelBlock:
// 00 SPCNODE DBNODE RELNODE FORK BLKNUM
//
// RelSize:
// 00 SPCNODE DBNODE RELNODE FORK FFFFFFFF
//
// SlruDir:
// 01 kind 00000000 00000000 00 00000000
//
// SlruSegBlock:
// 01 kind 00000001 SEGNO 00 BLKNUM
//
// SlruSegSize:
// 01 kind 00000001 SEGNO 00 FFFFFFFF
//
// TwoPhaseDir:
// 02 00000000 00000000 00000000 00 00000000
//
// TwoPhaseFile:
// 02 00000000 00000000 00000000 00 XID
//
// ControlFile:
// 03 00000000 00000000 00000000 00 00000000
//
// Checkpoint:
// 03 00000000 00000000 00000000 00 00000001
//
// AuxFiles:
// 03 00000000 00000000 00000000 00 00000002
//
//-- Section 01: relation data and metadata
const DBDIR_KEY: Key = Key {
field1: 0x00,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
fn dbdir_key_range(spcnode: Oid, dbnode: Oid) -> Range<Key> {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 0,
}..Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0xffffffff,
field5: 0xff,
field6: 0xffffffff,
}
}
fn relmap_file_key(spcnode: Oid, dbnode: Oid) -> Key {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 0,
}
}
fn rel_dir_to_key(spcnode: Oid, dbnode: Oid) -> Key {
Key {
field1: 0x00,
field2: spcnode,
field3: dbnode,
field4: 0,
field5: 0,
field6: 1,
}
}
pub(crate) fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: blknum,
}
}
fn rel_size_to_key(rel: RelTag) -> Key {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: 0xffffffff,
}
}
fn rel_key_range(rel: RelTag) -> Range<Key> {
Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum,
field6: 0,
}..Key {
field1: 0x00,
field2: rel.spcnode,
field3: rel.dbnode,
field4: rel.relnode,
field5: rel.forknum + 1,
field6: 0,
}
}
//-- Section 02: SLRUs
fn slru_dir_to_key(kind: SlruKind) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 0,
field4: 0,
field5: 0,
field6: 0,
}
}
fn slru_block_to_key(kind: SlruKind, segno: u32, blknum: BlockNumber) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 1,
field4: segno,
field5: 0,
field6: blknum,
}
}
fn slru_segment_size_to_key(kind: SlruKind, segno: u32) -> Key {
Key {
field1: 0x01,
field2: match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
},
field3: 1,
field4: segno,
field5: 0,
field6: 0xffffffff,
}
}
fn slru_segment_key_range(kind: SlruKind, segno: u32) -> Range<Key> {
let field2 = match kind {
SlruKind::Clog => 0x00,
SlruKind::MultiXactMembers => 0x01,
SlruKind::MultiXactOffsets => 0x02,
};
Key {
field1: 0x01,
field2,
field3: 1,
field4: segno,
field5: 0,
field6: 0,
}..Key {
field1: 0x01,
field2,
field3: 1,
field4: segno,
field5: 1,
field6: 0,
}
}
//-- Section 03: pg_twophase
const TWOPHASEDIR_KEY: Key = Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
fn twophase_file_key(xid: TransactionId) -> Key {
Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: xid,
}
}
fn twophase_key_range(xid: TransactionId) -> Range<Key> {
let (next_xid, overflowed) = xid.overflowing_add(1);
Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: xid,
}..Key {
field1: 0x02,
field2: 0,
field3: 0,
field4: 0,
field5: u8::from(overflowed),
field6: next_xid,
}
}
//-- Section 03: Control file
const CONTROLFILE_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 0,
};
const CHECKPOINT_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 1,
};
const AUX_FILES_KEY: Key = Key {
field1: 0x03,
field2: 0,
field3: 0,
field4: 0,
field5: 0,
field6: 2,
};
// Reverse mappings for a few Keys.
// These are needed by WAL redo manager.
// AUX_FILES currently stores only data for logical replication (slots etc), and
// we don't preserve these on a branch because safekeepers can't follow timeline
// switch (and generally it likely should be optional), so ignore these.
pub fn is_inherited_key(key: Key) -> bool {
key != AUX_FILES_KEY
}
pub fn is_rel_fsm_block_key(key: Key) -> bool {
key.field1 == 0x00 && key.field4 != 0 && key.field5 == FSM_FORKNUM && key.field6 != 0xffffffff
}
pub fn is_rel_vm_block_key(key: Key) -> bool {
key.field1 == 0x00
&& key.field4 != 0
&& key.field5 == VISIBILITYMAP_FORKNUM
&& key.field6 != 0xffffffff
}
pub fn key_to_slru_block(key: Key) -> anyhow::Result<(SlruKind, u32, BlockNumber)> {
Ok(match key.field1 {
0x01 => {
let kind = match key.field2 {
0x00 => SlruKind::Clog,
0x01 => SlruKind::MultiXactMembers,
0x02 => SlruKind::MultiXactOffsets,
_ => anyhow::bail!("unrecognized slru kind 0x{:02x}", key.field2),
};
let segno = key.field4;
let blknum = key.field6;
(kind, segno, blknum)
}
_ => anyhow::bail!("unexpected value kind 0x{:02x}", key.field1),
})
}
fn is_slru_block_key(key: Key) -> bool {
key.field1 == 0x01 // SLRU-related
&& key.field3 == 0x00000001 // but not SlruDir
&& key.field6 != 0xffffffff // and not SlruSegSize
}
#[allow(clippy::bool_assert_comparison)]
#[cfg(test)]
mod tests {

View File

@@ -73,8 +73,8 @@ use crate::metrics::{
TimelineMetrics, MATERIALIZED_PAGE_CACHE_HIT, MATERIALIZED_PAGE_CACHE_HIT_DIRECT,
};
use crate::pgdatadir_mapping::CalculateLogicalSizeError;
use crate::pgdatadir_mapping::{is_inherited_key, is_rel_fsm_block_key, is_rel_vm_block_key};
use crate::tenant::config::TenantConfOpt;
use pageserver_api::key::{is_inherited_key, is_rel_fsm_block_key, is_rel_vm_block_key};
use pageserver_api::reltag::RelTag;
use pageserver_api::shard::ShardIndex;

View File

@@ -33,11 +33,12 @@ use utils::failpoint_support;
use crate::context::RequestContext;
use crate::metrics::WAL_INGEST;
use crate::pgdatadir_mapping::*;
use crate::pgdatadir_mapping::{DatadirModification, Version};
use crate::tenant::PageReconstructError;
use crate::tenant::Timeline;
use crate::walrecord::*;
use crate::ZERO_PAGE;
use pageserver_api::key::rel_block_to_key;
use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
use postgres_ffi::pg_constants;
use postgres_ffi::relfile_utils::{FSM_FORKNUM, INIT_FORKNUM, MAIN_FORKNUM, VISIBILITYMAP_FORKNUM};

View File

@@ -47,11 +47,10 @@ use crate::metrics::{
WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM, WAL_REDO_RECORDS_HISTOGRAM,
WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
};
use crate::pgdatadir_mapping::key_to_slru_block;
use crate::repository::Key;
use crate::walrecord::NeonWalRecord;
use pageserver_api::key::key_to_rel_block;
use pageserver_api::key::{key_to_rel_block, key_to_slru_block};
use pageserver_api::reltag::{RelTag, SlruKind};
use postgres_ffi::pg_constants;
use postgres_ffi::relfile_utils::VISIBILITYMAP_FORKNUM;