Some work in actually storing things

2026-08-10 14:36:51 +00:00 · 2020-04-08 22:00:41 +02:00
parent d50edcdb4f
commit bacc76a057
11 changed files with 302 additions and 110 deletions
@@ -9,21 +9,21 @@ use hyper::{Body, Method, Request, Response, Server, StatusCode};
 use futures::future::Future;

 use crate::error::Error;
-use crate::membership::System;
 use crate::data::*;
 use crate::proto::*;
 use crate::rpc_client::*;
+use crate::server::Garage;

-pub async fn run_api_server(sys: Arc<System>, shutdown_signal: impl Future<Output=()>) -> Result<(), hyper::Error> {
-    let addr = ([0, 0, 0, 0], sys.config.api_port).into();
+pub async fn run_api_server(garage: Arc<Garage>, shutdown_signal: impl Future<Output=()>) -> Result<(), hyper::Error> {
+    let addr = ([0, 0, 0, 0], garage.system.config.api_port).into();

    let service = make_service_fn(|conn: &AddrStream| {
-		let sys = sys.clone();
+		let garage = garage.clone();
 		let client_addr = conn.remote_addr();
 		async move {
 			Ok::<_, Error>(service_fn(move |req: Request<Body>| {
-				let sys = sys.clone();
-				handler(sys, req, client_addr)
+				let garage = garage.clone();
+				handler(garage, req, client_addr)
 			}))
 		}
 	});
@@ -36,8 +36,8 @@ pub async fn run_api_server(sys: Arc<System>, shutdown_signal: impl Future<Outpu
 	graceful.await
 }

-async fn handler(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
-	match handler_inner(sys, req, addr).await {
+async fn handler(garage: Arc<Garage>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
+	match handler_inner(garage, req, addr).await {
 		Ok(x) => Ok(x),
 		Err(Error::BadRequest(e)) => {
 			let mut bad_request = Response::new(Body::from(format!("{}\n", e)));
@@ -53,7 +53,7 @@ async fn handler(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Resu
 	}
 }

-async fn handler_inner(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
+async fn handler_inner(garage: Arc<Garage>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
 	eprintln!("{} {} {}", addr, req.method(), req.uri());

 	let bucket = req.headers()
@@ -75,7 +75,7 @@ async fn handler_inner(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -
 				.map(|x| x.to_str())
 				.unwrap_or(Ok("blob"))?
 				.to_string();
-			let version_uuid = handle_put(sys, &mime_type, &bucket, &key, req.into_body()).await?;
+			let version_uuid = handle_put(garage, &mime_type, &bucket, &key, req.into_body()).await?;
 			Ok(Response::new(Body::from(
 				format!("Version UUID: {:?}", version_uuid),
 			)))
@@ -84,22 +84,24 @@ async fn handler_inner(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -
    }
 }

-async fn handle_put(sys: Arc<System>,
+async fn handle_put(garage: Arc<Garage>,
 					mime_type: &str,
 					bucket: &str, key: &str, body: Body)
 	-> Result<UUID, Error> 
 {
 	let version_uuid = gen_uuid();

-	let mut chunker = BodyChunker::new(body, sys.config.block_size);
+	let mut chunker = BodyChunker::new(body, garage.system.config.block_size);
 	let first_block = match chunker.next().await? {
 		Some(x) => x,
 		None => return Err(Error::BadRequest(format!("Empty body"))),
 	};

-	let mut version = VersionMeta{
+	let version_key = VersionMetaKey{
 		bucket: bucket.to_string(),
 		key: key.to_string(),
+	};
+	let mut version_value = VersionMetaValue {
 		timestamp: now_msec(),
 		uuid: version_uuid.clone(),
 		mime_type: mime_type.to_string(),
@@ -107,27 +109,17 @@ async fn handle_put(sys: Arc<System>,
 		is_complete: false,
 		data: VersionData::DeleteMarker,
 	};
-	let version_who = sys.members.read().await
-		.walk_ring(&version_uuid, sys.config.meta_replication_factor);

 	if first_block.len() < INLINE_THRESHOLD {
-		version.data = VersionData::Inline(first_block);
-		version.is_complete = true;
-		rpc_try_call_many(sys.clone(),
-						  &version_who[..],
-						  &Message::AdvertiseVersion(version),
-						  (sys.config.meta_replication_factor+1)/2,
-						  DEFAULT_TIMEOUT).await?;
+		version_value.data = VersionData::Inline(first_block);
+		version_value.is_complete = true;
+		garage.version_table.insert(&version_key, &version_value).await?;
 		return Ok(version_uuid)
 	}

 	let first_block_hash = hash(&first_block[..]);
-	version.data = VersionData::FirstBlock(first_block_hash);
-	rpc_try_call_many(sys.clone(),
-					  &version_who[..],
-					  &Message::AdvertiseVersion(version.clone()),
-					  (sys.config.meta_replication_factor+1)/2,
-					  DEFAULT_TIMEOUT).await?;
+	version_value.data = VersionData::FirstBlock(first_block_hash);
+	garage.version_table.insert(&version_key, &version_value).await?;

 	let block_meta = BlockMeta{
 		version_uuid: version_uuid.clone(),
@@ -135,7 +127,7 @@ async fn handle_put(sys: Arc<System>,
 		hash: hash(&first_block[..]),
 	};
 	let mut next_offset = first_block.len();
-	let mut put_curr_block = put_block(sys.clone(), block_meta, first_block);
+	let mut put_curr_block = put_block(garage.clone(), block_meta, first_block);
 	loop {
 		let (_, next_block) = futures::try_join!(put_curr_block, chunker.next())?;
 		if let Some(block) = next_block {
@@ -145,7 +137,7 @@ async fn handle_put(sys: Arc<System>,
 				hash: hash(&block[..]),
 			};
 			next_offset += block.len();
-			put_curr_block = put_block(sys.clone(), block_meta, block);
+			put_curr_block = put_block(garage.clone(), block_meta, block);
 		} else {
 			break;
 		}
@@ -153,25 +145,21 @@ async fn handle_put(sys: Arc<System>,

 	// TODO: if at any step we have an error, we should undo everything we did

-	version.is_complete = true;
-	rpc_try_call_many(sys.clone(),
-					  &version_who[..],
-					  &Message::AdvertiseVersion(version),
-					  (sys.config.meta_replication_factor+1)/2,
-					  DEFAULT_TIMEOUT).await?;
+	version_value.is_complete = true;
+	garage.version_table.insert(&version_key, &version_value).await?;
 	Ok(version_uuid)
 }

-async fn put_block(sys: Arc<System>, meta: BlockMeta, data: Vec<u8>) -> Result<(), Error> {
-	let who = sys.members.read().await
-		.walk_ring(&meta.hash, sys.config.meta_replication_factor);
-	rpc_try_call_many(sys.clone(),
+async fn put_block(garage: Arc<Garage>, meta: BlockMeta, data: Vec<u8>) -> Result<(), Error> {
+	let who = garage.system.members.read().await
+		.walk_ring(&meta.hash, garage.system.config.meta_replication_factor);
+	rpc_try_call_many(garage.system.clone(),
 					  &who[..],
 					  &Message::PutBlock(PutBlockMessage{
 						  meta,
 						  data,
 						}),
-					  (sys.config.meta_replication_factor+1)/2,
+					  (garage.system.config.meta_replication_factor+1)/2,
 					  DEFAULT_TIMEOUT).await?;
 	Ok(())
 }
@@ -106,7 +106,7 @@ pub struct NetworkConfigEntry {

 // Data management

-pub const INLINE_THRESHOLD: usize = 2048;
+pub const INLINE_THRESHOLD: usize = 3072;

 #[derive(Debug, Serialize, Deserialize)]
 pub struct SplitpointMeta {
@@ -118,27 +118,7 @@ pub struct SplitpointMeta {
 	pub deleted: bool,
 }

-#[derive(Clone, Debug, Serialize, Deserialize)]
-pub struct VersionMeta {
-	pub bucket: String,
-	pub key: String,
-
-	pub timestamp: u64,
-	pub uuid: UUID,
-
-	pub mime_type: String,
-	pub size: u64,
-	pub is_complete: bool,
-
-	pub data: VersionData,
-}
-
-#[derive(Clone, Debug, Serialize, Deserialize)]
-pub enum VersionData {
-	DeleteMarker,
-	Inline(#[serde(with="serde_bytes")] Vec<u8>),
-	FirstBlock(Hash),
-}
+pub use crate::version_table::*;

 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct BlockMeta {
@@ -1,7 +1,12 @@
 mod error;
 mod data;
 mod proto;
+
 mod membership;
+mod table;
+
+mod version_table;
+
 mod server;
 mod rpc_server;
 mod rpc_client;
@@ -124,6 +124,11 @@ impl Members {
                i - 1
            }
        };
+
+		self.walk_ring_from_pos(start, n)
+    }
+
+	fn walk_ring_from_pos(&self, start: usize, n: usize) -> Vec<UUID> {
        let mut ret = vec![];
        let mut datacenters = vec![];

@@ -143,7 +148,7 @@ impl Members {
        }

        ret
-    }
+	}
 }

 fn read_network_config(metadata_dir: &PathBuf) -> Result<NetworkConfig, Error> {
@@ -18,7 +18,8 @@ pub enum Message {
 	AdvertiseConfig(NetworkConfig),

 	PutBlock(PutBlockMessage),
-	AdvertiseVersion(VersionMeta),
+
+	TableRPC(String, #[serde(with = "serde_bytes")] Vec<u8>),
 }

 #[derive(Debug, Serialize, Deserialize)]
@@ -1,6 +1,7 @@
 use std::net::SocketAddr;
 use std::sync::Arc;

+use futures_util::future::FutureExt;
 use bytes::IntoBuf;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::server::conn::AddrStream;
@@ -9,7 +10,7 @@ use futures::future::Future;

 use crate::error::Error;
 use crate::proto::Message;
-use crate::membership::System;
+use crate::server::Garage;

 fn err_to_msg(x: Result<Message, Error>) -> Message {
 	match x {
@@ -18,7 +19,7 @@ fn err_to_msg(x: Result<Message, Error>) -> Message {
 	}
 }

-async fn handler(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
+async fn handler(garage: Arc<Garage>, req: Request<Body>, addr: SocketAddr) -> Result<Response<Body>, Error> {
 	if req.method() != &Method::POST {
 		let mut bad_request = Response::default();
 		*bad_request.status_mut() = StatusCode::BAD_REQUEST;
@@ -30,12 +31,21 @@ async fn handler(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Resu

 	eprintln!("RPC from {}: {:?}", addr, msg);

+	let sys = garage.system.clone();
 	let resp = err_to_msg(match &msg {
 		Message::Ping(ping) => sys.handle_ping(&addr, ping).await,
 		Message::PullStatus => sys.handle_pull_status().await,
 		Message::PullConfig => sys.handle_pull_config().await,
 		Message::AdvertiseNodesUp(adv) => sys.handle_advertise_nodes_up(adv).await,
 		Message::AdvertiseConfig(adv) => sys.handle_advertise_config(adv).await,
+		Message::TableRPC(table, msg) => {
+			if let Some(rpc_handler) = garage.table_rpc_handlers.get(table) {
+				rpc_handler.handle(&msg[..]).await
+					.map(|rep| Message::TableRPC(table.to_string(), rep))
+			} else {
+				Ok(Message::Error(format!("Unknown table: {}", table)))
+			}
+		}

 		_ => Ok(Message::Error(format!("Unexpected message: {:?}", msg))),
 	});
@@ -46,16 +56,16 @@ async fn handler(sys: Arc<System>, req: Request<Body>, addr: SocketAddr) -> Resu
 }


-pub async fn run_rpc_server(sys: Arc<System>, shutdown_signal: impl Future<Output=()>) -> Result<(), hyper::Error> {
-    let bind_addr = ([0, 0, 0, 0], sys.config.rpc_port).into();
+pub async fn run_rpc_server(garage: Arc<Garage>, shutdown_signal: impl Future<Output=()>) -> Result<(), hyper::Error> {
+    let bind_addr = ([0, 0, 0, 0], garage.system.config.rpc_port).into();

    let service = make_service_fn(|conn: &AddrStream| {
 		let client_addr = conn.remote_addr();
-		let sys = sys.clone();
+		let garage = garage.clone();
 		async move {
 			Ok::<_, Error>(service_fn(move |req: Request<Body>| {
-				let sys = sys.clone();
-				handler(sys, req, client_addr)
+				let garage = garage.clone();
+				handler(garage, req, client_addr)
 			}))
 		}
 	});
@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::io::{Read, Write};
 use std::sync::Arc;
 use std::net::SocketAddr;
@@ -6,10 +7,51 @@ use futures::channel::oneshot;
 use serde::Deserialize;

 use crate::data::*;
+use crate::proto::*;
 use crate::error::Error;
 use crate::membership::System;
 use crate::api_server;
 use crate::rpc_server;
+use crate::table::*;
+
+pub struct Garage {
+	pub db: sled::Db,
+	pub system: Arc<System>,
+
+	pub table_rpc_handlers: HashMap<String, Box<dyn TableRpcHandler + Sync + Send>>,
+
+	pub version_table: Arc<Table<VersionTable>>,
+}
+
+impl Garage {
+	pub fn new(config: Config, id: UUID, db: sled::Db) -> Self {
+		let system = Arc::new(System::new(config, id));
+
+		let meta_rep_param = TableReplicationParams{
+			replication_factor: system.config.meta_replication_factor,
+			write_quorum: (system.config.meta_replication_factor+1)/2,
+			read_quorum: (system.config.meta_replication_factor+1)/2,
+			timeout: DEFAULT_TIMEOUT,
+		};
+
+		let version_table = Arc::new(Table::new(
+			system.clone(),
+			&db,
+			"version".to_string(),
+			meta_rep_param.clone())); 
+
+		let mut garage = Self{
+			db,
+			system: system.clone(),
+			table_rpc_handlers: HashMap::new(),
+			version_table,
+		};
+		garage.table_rpc_handlers.insert(
+			garage.version_table.name.clone(),
+			garage.version_table.clone().rpc_handler());
+		garage
+	}
+}

 fn default_block_size() -> usize {
 	1048576
@@ -88,20 +130,25 @@ pub async fn run_server(config_file: PathBuf) -> Result<(), Error> {
 	let config = read_config(config_file)
 		.expect("Unable to read config file");

+	let mut db_path = config.metadata_dir.clone();
+	db_path.push("garage_metadata");
+	let db = sled::open(db_path)
+		.expect("Unable to open DB");
+
 	let id = gen_node_id(&config.metadata_dir)
 		.expect("Unable to read or generate node ID");
 	println!("Node ID: {}", hex::encode(&id));

-	let sys = Arc::new(System::new(config, id));
+	let garage = Arc::new(Garage::new(config, id, db));

 	let (tx1, rx1) = oneshot::channel();
 	let (tx2, rx2) = oneshot::channel();

-	let rpc_server = rpc_server::run_rpc_server(sys.clone(), wait_from(rx1));
-	let api_server = api_server::run_api_server(sys.clone(), wait_from(rx2));
+	let rpc_server = rpc_server::run_rpc_server(garage.clone(), wait_from(rx1));
+	let api_server = api_server::run_api_server(garage.clone(), wait_from(rx2));

 	tokio::spawn(shutdown_signal(vec![tx1, tx2]));
-	tokio::spawn(sys.bootstrap());
+	tokio::spawn(garage.system.clone().bootstrap());

 	futures::try_join!(rpc_server, api_server)?;
 	Ok(())
@@ -0,0 +1,116 @@
+use std::marker::PhantomData;
+use std::time::Duration;
+use std::sync::Arc;
+use serde::{Serialize, Deserialize};
+use async_trait::async_trait;
+
+use crate::error::Error;
+use crate::proto::*;
+use crate::data::*;
+use crate::membership::System;
+use crate::rpc_client::*;
+
+
+pub struct Table<F: TableFormat> {
+	phantom: PhantomData<F>,
+
+	pub name: String,
+
+	pub system: Arc<System>,
+	pub store: sled::Tree,
+	pub partitions: Vec<Partition>,
+
+	pub param: TableReplicationParams,
+}
+
+#[derive(Clone)]
+pub struct TableReplicationParams {
+	pub replication_factor: usize,
+	pub read_quorum: usize,
+	pub write_quorum: usize,
+	pub timeout: Duration,
+}
+
+#[async_trait]
+pub trait TableRpcHandler {
+	async fn handle(&self, rpc: &[u8]) -> Result<Vec<u8>, Error>;
+}
+
+struct TableRpcHandlerAdapter<F: TableFormat> {
+	table: Arc<Table<F>>,
+}
+
+#[async_trait]
+impl<F: TableFormat + 'static> TableRpcHandler for TableRpcHandlerAdapter<F> {
+	async fn handle(&self, rpc: &[u8]) -> Result<Vec<u8>, Error> {
+		let msg = rmp_serde::decode::from_read_ref::<_, TableRPC<F>>(rpc)?;
+		let rep = self.table.handle(msg).await?;
+		Ok(rmp_serde::encode::to_vec_named(&rep)?)
+	}
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+pub enum TableRPC<F: TableFormat> {
+	Update(F::K, F::V),
+}
+
+pub struct Partition {
+	pub begin: Hash,
+	pub end: Hash,
+	pub other_nodes: Vec<UUID>,
+}
+
+pub trait KeyHash {
+	fn hash(&self) -> Hash;
+}
+
+pub trait ValueMerge {
+	fn merge(&mut self, other: &Self);
+}
+
+#[async_trait]
+pub trait TableFormat: Send + Sync {
+	type K: Clone + Serialize + for<'de> Deserialize<'de> + KeyHash + Send + Sync;
+	type V: Clone + Serialize + for<'de> Deserialize<'de> + ValueMerge + Send + Sync;
+
+	async fn updated(&self, key: &Self::K, old: Option<&Self::V>, new: &Self::V);
+}
+
+impl<F: TableFormat + 'static> Table<F> {
+	pub fn new(system: Arc<System>, db: &sled::Db, name: String, param: TableReplicationParams) -> Self {
+		let store = db.open_tree(&name)
+				.expect("Unable to open DB tree");
+		Self{
+			phantom: PhantomData::default(),
+			name,
+			system,
+			store,
+			partitions: Vec::new(),
+			param,
+		}
+	}
+
+	pub fn rpc_handler(self: Arc<Self>) -> Box<dyn TableRpcHandler + Send + Sync> {
+		Box::new(TableRpcHandlerAdapter::<F>{ table: self })
+	}
+
+	pub async fn insert(&self, k: &F::K, v: &F::V) -> Result<(), Error> {
+		unimplemented!();
+
+		let hash = k.hash();
+		let who = self.system.members.read().await
+			.walk_ring(&hash, self.param.replication_factor);
+
+		let msg = rmp_serde::encode::to_vec_named(&TableRPC::<F>::Update(k.clone(), v.clone()))?;
+		rpc_try_call_many(self.system.clone(),
+						  &who[..],
+						  &Message::TableRPC(self.name.to_string(), msg),
+						  self.param.write_quorum,
+						  self.param.timeout).await?;
+		Ok(())
+	}
+
+	async fn handle(&self, msg: TableRPC<F>) -> Result<TableRPC<F>, Error> {
+		unimplemented!()
+	}
+}
@@ -0,0 +1,59 @@
+use std::sync::Arc;
+use serde::{Serialize, Deserialize};
+use async_trait::async_trait;
+
+use crate::data::*;
+use crate::table::*;
+use crate::membership::System;
+
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct VersionMetaKey {
+	pub bucket: String,
+	pub key: String,
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct VersionMetaValue {
+	pub timestamp: u64,
+	pub uuid: UUID,
+
+	pub mime_type: String,
+	pub size: u64,
+	pub is_complete: bool,
+
+	pub data: VersionData,
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub enum VersionData {
+	DeleteMarker,
+	Inline(#[serde(with="serde_bytes")] Vec<u8>),
+	FirstBlock(Hash),
+}
+
+pub struct VersionTable {
+	system: Arc<System>,
+}
+
+impl KeyHash for VersionMetaKey {
+	fn hash(&self) -> Hash {
+		hash(self.bucket.as_bytes())
+	}
+}
+
+impl ValueMerge for VersionMetaValue {
+	fn merge(&mut self, other: &Self) {
+		unimplemented!()
+	}
+}
+
+#[async_trait]
+impl TableFormat for VersionTable {
+	type K = VersionMetaKey;
+	type V = VersionMetaValue;
+
+	async fn updated(&self, key: &Self::K, old: Option<&Self::V>, new: &Self::V) {
+		unimplemented!()
+	}
+}