Merge pull request 'rewrite read_and_put_block as a series of steps with channels' (#734) from refactor-put into main

Reviewed-on: Deuxfleurs/garage#734

src/api/s3/copy.rs

@@ -387,7 +387,10 @@ pub async fn handle_upload_part_copy(
 			// we need to insert that data as a new block.
 			async move {
 				if must_upload {
-					garage2.block_manager.rpc_put_block(final_hash, data).await
+					garage2
+						.block_manager
+						.rpc_put_block(final_hash, data, None)
+						.await
 				} else {
 					Ok(())
 				}

src/api/s3/multipart.rs

@@ -6,7 +6,6 @@ use hyper::{Request, Response};
 use md5::{Digest as Md5Digest, Md5};
 
 use garage_table::*;
-use garage_util::async_hash::*;
 use garage_util::data::*;
 
 use garage_model::bucket_table::Bucket;
@@ -135,17 +134,8 @@ pub async fn handle_put_part(
 	garage.version_table.insert(&version).await?;
 
 	// Copy data to version
-	let first_block_hash = async_blake2sum(first_block.clone()).await;
+	let (total_size, data_md5sum, data_sha256sum, _) =
-
+		read_and_put_blocks(&garage, &version, part_number, first_block, &mut chunker).await?;
-	let (total_size, data_md5sum, data_sha256sum) = read_and_put_blocks(
-		&garage,
-		&version,
-		part_number,
-		first_block,
-		first_block_hash,
-		&mut chunker,
-	)
-	.await?;
 
 	// Verify that checksums map
 	ensure_checksum_matches(

src/api/s3/put.rs

@@ -3,10 +3,13 @@ use std::sync::Arc;
 
 use base64::prelude::*;
 use futures::prelude::*;
+use futures::stream::FuturesOrdered;
 use futures::try_join;
 use md5::{digest::generic_array::*, Digest as Md5Digest, Md5};
 use sha2::Sha256;
 
+use tokio::sync::mpsc;
+
 use hyper::body::Bytes;
 use hyper::header::{HeaderMap, HeaderValue};
 use hyper::{Request, Response};
@@ -17,6 +20,7 @@ use opentelemetry::{
 };
 
 use garage_net::bytes_buf::BytesBuf;
+use garage_rpc::rpc_helper::OrderTag;
 use garage_table::*;
 use garage_util::async_hash::*;
 use garage_util::data::*;
@@ -35,6 +39,8 @@ use crate::helpers::*;
 use crate::s3::api_server::{ReqBody, ResBody};
 use crate::s3::error::*;
 
+const PUT_BLOCKS_MAX_PARALLEL: usize = 3;
+
 pub async fn handle_put(
 	garage: Arc<Garage>,
 	req: Request<ReqBody>,
@@ -168,17 +174,8 @@ pub(crate) async fn save_stream<S: Stream<Item = Result<Bytes, Error>> + Unpin>(
 	garage.version_table.insert(&version).await?;
 
 	// Transfer data and verify checksum
-	let first_block_hash = async_blake2sum(first_block.clone()).await;
+	let (total_size, data_md5sum, data_sha256sum, first_block_hash) =
-
+		read_and_put_blocks(&garage, &version, 1, first_block, &mut chunker).await?;
-	let (total_size, data_md5sum, data_sha256sum) = read_and_put_blocks(
-		&garage,
-		&version,
-		1,
-		first_block,
-		first_block_hash,
-		&mut chunker,
-	)
-	.await?;
 
 	ensure_checksum_matches(
 		data_md5sum.as_slice(),
@@ -299,84 +296,164 @@ pub(crate) async fn read_and_put_blocks<S: Stream<Item = Result<Bytes, Error>> +
 	version: &Version,
 	part_number: u64,
 	first_block: Bytes,
-	first_block_hash: Hash,
 	chunker: &mut StreamChunker<S>,
-) -> Result<(u64, GenericArray<u8, typenum::U16>, Hash), Error> {
+) -> Result<(u64, GenericArray<u8, typenum::U16>, Hash, Hash), Error> {
 	let tracer = opentelemetry::global::tracer("garage");
 
-	let md5hasher = AsyncHasher::<Md5>::new();
+	let (block_tx, mut block_rx) = mpsc::channel::<Result<Bytes, Error>>(2);
-	let sha256hasher = AsyncHasher::<Sha256>::new();
+	let read_blocks = async {
+		block_tx.send(Ok(first_block)).await?;
+		loop {
+			let res = chunker
+				.next()
+				.with_context(Context::current_with_span(
+					tracer.start("Read block from client"),
+				))
+				.await;
+			match res {
+				Ok(Some(block)) => block_tx.send(Ok(block)).await?,
+				Ok(None) => break,
+				Err(e) => {
+					block_tx.send(Err(e)).await?;
+					break;
+				}
+			}
+		}
+		drop(block_tx);
+		Ok::<_, mpsc::error::SendError<_>>(())
+	};
 
-	futures::future::join(
+	let (block_tx2, mut block_rx2) = mpsc::channel::<Result<Bytes, Error>>(1);
-		md5hasher.update(first_block.clone()),
+	let hash_stream = async {
-		sha256hasher.update(first_block.clone()),
+		let md5hasher = AsyncHasher::<Md5>::new();
-	)
+		let sha256hasher = AsyncHasher::<Sha256>::new();
-	.with_context(Context::current_with_span(
+		while let Some(next) = block_rx.recv().await {
-		tracer.start("Hash first block (md5, sha256)"),
+			match next {
-	))
+				Ok(block) => {
-	.await;
+					block_tx2.send(Ok(block.clone())).await?;
+					futures::future::join(
+						md5hasher.update(block.clone()),
+						sha256hasher.update(block.clone()),
+					)
+					.with_context(Context::current_with_span(
+						tracer.start("Hash block (md5, sha256)"),
+					))
+					.await;
+				}
+				Err(e) => {
+					block_tx2.send(Err(e)).await?;
+					break;
+				}
+			}
+		}
+		drop(block_tx2);
+		Ok::<_, mpsc::error::SendError<_>>(futures::join!(
+			md5hasher.finalize(),
+			sha256hasher.finalize()
+		))
+	};
 
-	let mut next_offset = first_block.len();
+	let (block_tx3, mut block_rx3) = mpsc::channel::<Result<(Bytes, Hash), Error>>(1);
-	let mut put_curr_version_block = put_block_meta(
+	let hash_blocks = async {
-		garage,
+		let mut first_block_hash = None;
-		version,
+		while let Some(next) = block_rx2.recv().await {
-		part_number,
+			match next {
-		0,
+				Ok(block) => {
-		first_block_hash,
+					let hash = async_blake2sum(block.clone())
-		first_block.len() as u64,
+						.with_context(Context::current_with_span(
-	);
+							tracer.start("Hash block (blake2)"),
-	let mut put_curr_block = garage
+						))
-		.block_manager
+						.await;
-		.rpc_put_block(first_block_hash, first_block);
+					if first_block_hash.is_none() {
+						first_block_hash = Some(hash);
+					}
+					block_tx3.send(Ok((block, hash))).await?;
+				}
+				Err(e) => {
+					block_tx3.send(Err(e)).await?;
+					break;
+				}
+			}
+		}
+		drop(block_tx3);
+		Ok::<_, mpsc::error::SendError<_>>(first_block_hash.unwrap())
+	};
 
-	loop {
+	let put_blocks = async {
-		let (_, _, next_block) = futures::try_join!(
+		// Structure for handling several concurrent writes to storage nodes
-			put_curr_block.map_err(Error::from),
+		let order_stream = OrderTag::stream();
-			put_curr_version_block.map_err(Error::from),
+		let mut write_futs = FuturesOrdered::new();
-			chunker.next(),
+		let mut written_bytes = 0u64;
-		)?;
+		loop {
-		if let Some(block) = next_block {
+			// Simultaneously write blocks to storage nodes & await for next block to be written
-			let (_, _, block_hash) = futures::future::join3(
+			let currently_running = write_futs.len();
-				md5hasher.update(block.clone()),
+			let write_futs_next = async {
-				sha256hasher.update(block.clone()),
+				if write_futs.is_empty() {
-				async_blake2sum(block.clone()),
+					futures::future::pending().await
-			)
+				} else {
-			.with_context(Context::current_with_span(
+					write_futs.next().await.unwrap()
-				tracer.start("Hash block (md5, sha256, blake2)"),
+				}
-			))
+			};
-			.await;
+			let recv_next = async {
-			let block_len = block.len();
+				// If more than a maximum number of writes are in progress, don't add more for now
-			put_curr_version_block = put_block_meta(
+				if currently_running >= PUT_BLOCKS_MAX_PARALLEL {
+					futures::future::pending().await
+				} else {
+					block_rx3.recv().await
+				}
+			};
+			let (block, hash) = tokio::select! {
+				result = write_futs_next => {
+					result?;
+					continue;
+				},
+				recv = recv_next => match recv {
+					Some(next) => next?,
+					None => break,
+				},
+			};
+
+			// For next block to be written: count its size and spawn future to write it
+			let offset = written_bytes;
+			written_bytes += block.len() as u64;
+			write_futs.push_back(put_block_and_meta(
 				garage,
 				version,
 				part_number,
-				next_offset as u64,
+				offset,
-				block_hash,
+				hash,
-				block_len as u64,
+				block,
-			);
+				order_stream.order(written_bytes),
-			put_curr_block = garage.block_manager.rpc_put_block(block_hash, block);
+			));
-			next_offset += block_len;
-		} else {
-			break;
 		}
-	}
+		while let Some(res) = write_futs.next().await {
+			res?;
+		}
+		Ok::<_, Error>(written_bytes)
+	};
 
-	let total_size = next_offset as u64;
+	let (_, stream_hash_result, block_hash_result, final_result) =
-	let data_md5sum = md5hasher.finalize().await;
+		futures::join!(read_blocks, hash_stream, hash_blocks, put_blocks);
+
+	let total_size = final_result?;
+	// unwrap here is ok, because if hasher failed, it is because something failed
+	// later in the pipeline which already caused a return at the ? on previous line
+	let (data_md5sum, data_sha256sum) = stream_hash_result.unwrap();
+	let first_block_hash = block_hash_result.unwrap();
 
-	let data_sha256sum = sha256hasher.finalize().await;
 	let data_sha256sum = Hash::try_from(&data_sha256sum[..]).unwrap();
 
-	Ok((total_size, data_md5sum, data_sha256sum))
+	Ok((total_size, data_md5sum, data_sha256sum, first_block_hash))
 }
 
-async fn put_block_meta(
+async fn put_block_and_meta(
 	garage: &Garage,
 	version: &Version,
 	part_number: u64,
 	offset: u64,
 	hash: Hash,
-	size: u64,
+	block: Bytes,
+	order_tag: OrderTag,
 ) -> Result<(), GarageError> {
 	let mut version = version.clone();
 	version.blocks.put(
@@ -384,7 +461,10 @@ async fn put_block_meta(
 			part_number,
 			offset,
 		},
-		VersionBlock { hash, size },
+		VersionBlock {
+			hash,
+			size: block.len() as u64,
+		},
 	);
 
 	let block_ref = BlockRef {
@@ -394,6 +474,9 @@ async fn put_block_meta(
 	};
 
 	futures::try_join!(
+		garage
+			.block_manager
+			.rpc_put_block(hash, block, Some(order_tag)),
 		garage.version_table.insert(&version),
 		garage.block_ref_table.insert(&block_ref),
 	)?;

src/block/manager.rs

@@ -346,7 +346,12 @@ impl BlockManager {
 	}
 
 	/// Send block to nodes that should have it
-	pub async fn rpc_put_block(&self, hash: Hash, data: Bytes) -> Result<(), Error> {
+	pub async fn rpc_put_block(
+		&self,
+		hash: Hash,
+		data: Bytes,
+		order_tag: Option<OrderTag>,
+	) -> Result<(), Error> {
 		let who = self.replication.write_nodes(&hash);
 
 		let (header, bytes) = DataBlock::from_buffer(data, self.compression_level)
@@ -354,6 +359,11 @@ impl BlockManager {
 			.into_parts();
 		let put_block_rpc =
 			Req::new(BlockRpc::PutBlock { hash, header })?.with_stream_from_buffer(bytes);
+		let put_block_rpc = if let Some(tag) = order_tag {
+			put_block_rpc.with_order_tag(tag)
+		} else {
+			put_block_rpc
+		};
 
 		self.system
 			.rpc