Merge branch 'main-v1' into sync-v2-to-v1

src/api/admin/node.rs

@@ -106,17 +106,17 @@ impl RequestHandler for LocalGetNodeStatisticsRequest {
 
 		// Gather block manager statistics
 		writeln!(&mut ret, "\nBlock manager stats:").unwrap();
-		let rc_len = garage.block_manager.rc_len()?.to_string();
+		let rc_len = garage.block_manager.rc_approximate_len()?.to_string();
 
 		ret += &format_table_to_string(vec![
 			format!("  number of RC entries:\t{} (~= number of blocks)", rc_len),
 			format!(
 				"  resync queue length:\t{}",
-				garage.block_manager.resync.queue_len()?
+				garage.block_manager.resync.queue_approximate_len()?
 			),
 			format!(
 				"  blocks with resync errors:\t{}",
-				garage.block_manager.resync.errors_len()?
+				garage.block_manager.resync.errors_approximate_len()?
 			),
 		]);
 
@@ -129,16 +129,21 @@ where
 	F: TableSchema + 'static,
 	R: TableReplication + 'static,
 {
-	let data_len = t.data.store.len().map_err(GarageError::from)?.to_string();
-	let mkl_len = t.merkle_updater.merkle_tree_len()?.to_string();
+	let data_len = t
+		.data
+		.store
+		.approximate_len()
+		.map_err(GarageError::from)?
+		.to_string();
+	let mkl_len = t.merkle_updater.merkle_tree_approximate_len()?.to_string();
 
 	Ok(format!(
 		"  {}\t{}\t{}\t{}\t{}\t{}",
 		F::TABLE_NAME,
 		data_len,
 		mkl_len,
-		t.merkle_updater.todo_len()?,
-		t.data.insert_queue_len()?,
-		t.data.gc_todo_len()?
+		t.merkle_updater.todo_approximate_len()?,
+		t.data.insert_queue_approximate_len()?,
+		t.data.gc_todo_approximate_len()?
 	))
 }

src/api/common/generic_server.rs

@@ -352,7 +352,11 @@ where
 
 	while !*must_exit.borrow() {
 		let (stream, client_addr) = tokio::select! {
-			acc = listener.accept() => acc?,
+			acc = listener.accept() => match acc {
+							Ok(r) => r,
+							Err(e) if e.kind() == std::io::ErrorKind::ConnectionAborted => continue,
+							Err(e) => return Err(e.into()),
+						},
 			_ = must_exit.changed() => continue,
 		};
 

src/block/manager.rs

@@ -48,6 +48,8 @@ pub const INLINE_THRESHOLD: usize = 3072;
 // to delete the block locally.
 pub(crate) const BLOCK_GC_DELAY: Duration = Duration::from_secs(600);
 
+const BLOCK_READ_SEMAPHORE_TIMEOUT: Duration = Duration::from_secs(15);
+
 /// RPC messages used to share blocks of data between nodes
 #[derive(Debug, Serialize, Deserialize)]
 pub enum BlockRpc {
@@ -85,6 +87,7 @@ pub struct BlockManager {
 	disable_scrub: bool,
 
 	mutation_lock: Vec<Mutex<BlockManagerLocked>>,
+	read_semaphore: Semaphore,
 
 	pub rc: BlockRc,
 	pub resync: BlockResyncManager,
@@ -174,6 +177,8 @@ impl BlockManager {
 				.iter()
 				.map(|_| Mutex::new(BlockManagerLocked()))
 				.collect::<Vec<_>>(),
+
+			read_semaphore: Semaphore::new(config.block_max_concurrent_reads),
 			rc,
 			resync,
 			system,
@@ -416,8 +421,8 @@ impl BlockManager {
 	}
 
 	/// Get number of items in the refcount table
-	pub fn rc_len(&self) -> Result<usize, Error> {
-		Ok(self.rc.rc_table.len()?)
+	pub fn rc_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.rc.rc_table.approximate_len()?)
 	}
 
 	/// Send command to start/stop/manager scrub worker
@@ -435,7 +440,7 @@ impl BlockManager {
 
 	/// List all resync errors
 	pub fn list_resync_errors(&self) -> Result<Vec<BlockResyncErrorInfo>, Error> {
-		let mut blocks = Vec::with_capacity(self.resync.errors.len()?);
+		let mut blocks = Vec::with_capacity(self.resync.errors.approximate_len()?);
 		for ent in self.resync.errors.iter()? {
 			let (hash, cnt) = ent?;
 			let cnt = ErrorCounter::decode(&cnt);
@@ -565,9 +570,6 @@ impl BlockManager {
 			match self.find_block(hash).await {
 				Some(p) => self.read_block_from(hash, &p).await,
 				None => {
-					// Not found but maybe we should have had it ??
-					self.resync
-						.put_to_resync(hash, 2 * self.system.rpc_helper().rpc_timeout())?;
 					return Err(Error::Message(format!(
 						"block {:?} not found on node",
 						hash
@@ -589,6 +591,15 @@ impl BlockManager {
 	) -> Result<DataBlock, Error> {
 		let (header, path) = block_path.as_parts_ref();
 
+		let permit = tokio::select! {
+			sem = self.read_semaphore.acquire() => sem.ok_or_message("acquire read semaphore")?,
+			_ = tokio::time::sleep(BLOCK_READ_SEMAPHORE_TIMEOUT) => {
+				self.metrics.block_read_semaphore_timeouts.add(1);
+				debug!("read block {:?}: read_semaphore acquire timeout", hash);
+				return Err(Error::Message("read block: read_semaphore acquire timeout".into()));
+			}
+		};
+
 		let mut f = fs::File::open(&path).await?;
 		let mut data = vec![];
 		f.read_to_end(&mut data).await?;
@@ -613,6 +624,8 @@ impl BlockManager {
 			return Err(Error::CorruptData(*hash));
 		}
 
+		drop(permit);
+
 		Ok(data)
 	}
 

src/block/metrics.rs

@@ -22,6 +22,7 @@ pub struct BlockManagerMetrics {
 
 	pub(crate) bytes_read: BoundCounter<u64>,
 	pub(crate) block_read_duration: BoundValueRecorder<f64>,
+	pub(crate) block_read_semaphore_timeouts: BoundCounter<u64>,
 	pub(crate) bytes_written: BoundCounter<u64>,
 	pub(crate) block_write_duration: BoundValueRecorder<f64>,
 	pub(crate) delete_counter: BoundCounter<u64>,
@@ -50,7 +51,7 @@ impl BlockManagerMetrics {
 				.init(),
 			_rc_size: meter
 				.u64_value_observer("block.rc_size", move |observer| {
-					if let Ok(value) = rc_tree.len() {
+					if let Ok(value) = rc_tree.approximate_len() {
 						observer.observe(value as u64, &[])
 					}
 				})
@@ -58,7 +59,7 @@ impl BlockManagerMetrics {
 				.init(),
 			_resync_queue_len: meter
 				.u64_value_observer("block.resync_queue_length", move |observer| {
-					if let Ok(value) = resync_queue.len() {
+					if let Ok(value) = resync_queue.approximate_len() {
 						observer.observe(value as u64, &[]);
 					}
 				})
@@ -68,7 +69,7 @@ impl BlockManagerMetrics {
 				.init(),
 			_resync_errored_blocks: meter
 				.u64_value_observer("block.resync_errored_blocks", move |observer| {
-					if let Ok(value) = resync_errors.len() {
+					if let Ok(value) = resync_errors.approximate_len() {
 						observer.observe(value as u64, &[]);
 					}
 				})
@@ -119,6 +120,11 @@ impl BlockManagerMetrics {
 				.with_description("Duration of block read operations")
 				.init()
 				.bind(&[]),
+			block_read_semaphore_timeouts: meter
+				.u64_counter("block.read_semaphore_timeouts")
+				.with_description("Number of block reads that failed due to semaphore acquire timeout")
+				.init()
+				.bind(&[]),
 			bytes_written: meter
 				.u64_counter("block.bytes_written")
 				.with_description("Number of bytes written to disk")

src/block/resync.rs

@@ -104,13 +104,13 @@ impl BlockResyncManager {
 	}
 
 	/// Get length of resync queue
-	pub fn queue_len(&self) -> Result<usize, Error> {
-		Ok(self.queue.len()?)
+	pub fn queue_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.queue.approximate_len()?)
 	}
 
 	/// Get number of blocks that have an error
-	pub fn errors_len(&self) -> Result<usize, Error> {
-		Ok(self.errors.len()?)
+	pub fn errors_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.errors.approximate_len()?)
 	}
 
 	/// Clear the error counter for a block and put it in queue immediately
@@ -540,9 +540,11 @@ impl Worker for ResyncWorker {
 		}
 
 		WorkerStatus {
-			queue_length: Some(self.manager.resync.queue_len().unwrap_or(0) as u64),
+			queue_length: Some(self.manager.resync.queue_approximate_len().unwrap_or(0) as u64),
 			tranquility: Some(tranquility),
-			persistent_errors: Some(self.manager.resync.errors_len().unwrap_or(0) as u64),
+			persistent_errors: Some(
+				self.manager.resync.errors_approximate_len().unwrap_or(0) as u64
+			),
 			..Default::default()
 		}
 	}

src/db/Cargo.toml

@@ -16,10 +16,14 @@ err-derive.workspace = true
 tracing.workspace = true
 
 heed = { workspace = true, optional = true }
+
 rusqlite = { workspace = true, optional = true, features = ["backup"] }
 r2d2 = { workspace = true, optional = true }
 r2d2_sqlite = { workspace = true, optional = true }
 
+fjall = { workspace = true, optional = true }
+parking_lot = { workspace = true, optional = true }
+
 [dev-dependencies]
 mktemp.workspace = true
 
@@ -27,4 +31,5 @@ mktemp.workspace = true
 default = [ "lmdb", "sqlite" ]
 bundled-libs = [ "rusqlite?/bundled" ]
 lmdb = [ "heed" ]
+fjall = [ "dep:fjall", "dep:parking_lot" ]
 sqlite = [ "rusqlite", "r2d2", "r2d2_sqlite" ]

src/db/fjall_adapter.rs

@@ -0,0 +1,453 @@
+use core::ops::Bound;
+
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use parking_lot::{MappedRwLockReadGuard, RwLock, RwLockReadGuard};
+
+use fjall::{
+	PartitionCreateOptions, PersistMode, TransactionalKeyspace, TransactionalPartitionHandle,
+	WriteTransaction,
+};
+
+use crate::{
+	open::{Engine, OpenOpt},
+	Db, Error, IDb, ITx, ITxFn, OnCommit, Result, TxError, TxFnResult, TxOpError, TxOpResult,
+	TxResult, TxValueIter, Value, ValueIter,
+};
+
+pub use fjall;
+
+// --
+
+pub(crate) fn open_db(path: &PathBuf, opt: &OpenOpt) -> Result<Db> {
+	info!("Opening Fjall database at: {}", path.display());
+	if opt.fsync {
+		return Err(Error(
+			"metadata_fsync is not supported with the Fjall database engine".into(),
+		));
+	}
+	let mut config = fjall::Config::new(path);
+	if let Some(block_cache_size) = opt.fjall_block_cache_size {
+		config = config.cache_size(block_cache_size as u64);
+	}
+	let keyspace = config.open_transactional()?;
+	Ok(FjallDb::init(keyspace))
+}
+
+// -- err
+
+impl From<fjall::Error> for Error {
+	fn from(e: fjall::Error) -> Error {
+		Error(format!("fjall: {}", e).into())
+	}
+}
+
+impl From<fjall::LsmError> for Error {
+	fn from(e: fjall::LsmError) -> Error {
+		Error(format!("fjall lsm_tree: {}", e).into())
+	}
+}
+
+impl From<fjall::Error> for TxOpError {
+	fn from(e: fjall::Error) -> TxOpError {
+		TxOpError(e.into())
+	}
+}
+
+// -- db
+
+pub struct FjallDb {
+	keyspace: TransactionalKeyspace,
+	trees: RwLock<Vec<(String, TransactionalPartitionHandle)>>,
+}
+
+type ByteRefRangeBound<'r> = (Bound<&'r [u8]>, Bound<&'r [u8]>);
+
+impl FjallDb {
+	pub fn init(keyspace: TransactionalKeyspace) -> Db {
+		let s = Self {
+			keyspace,
+			trees: RwLock::new(Vec::new()),
+		};
+		Db(Arc::new(s))
+	}
+
+	fn get_tree(
+		&self,
+		i: usize,
+	) -> Result<MappedRwLockReadGuard<'_, TransactionalPartitionHandle>> {
+		RwLockReadGuard::try_map(self.trees.read(), |trees: &Vec<_>| {
+			trees.get(i).map(|tup| &tup.1)
+		})
+		.map_err(|_| Error("invalid tree id".into()))
+	}
+}
+
+impl IDb for FjallDb {
+	fn engine(&self) -> String {
+		"Fjall (EXPERIMENTAL!)".into()
+	}
+
+	fn open_tree(&self, name: &str) -> Result<usize> {
+		let mut trees = self.trees.write();
+		let safe_name = encode_name(name)?;
+		if let Some(i) = trees.iter().position(|(name, _)| *name == safe_name) {
+			Ok(i)
+		} else {
+			let tree = self
+				.keyspace
+				.open_partition(&safe_name, PartitionCreateOptions::default())?;
+			let i = trees.len();
+			trees.push((safe_name, tree));
+			Ok(i)
+		}
+	}
+
+	fn list_trees(&self) -> Result<Vec<String>> {
+		Ok(self
+			.keyspace
+			.list_partitions()
+			.iter()
+			.map(|n| decode_name(&n))
+			.collect::<Result<Vec<_>>>()?)
+	}
+
+	fn snapshot(&self, base_path: &PathBuf) -> Result<()> {
+		std::fs::create_dir_all(base_path)?;
+		let path = Engine::Fjall.db_path(base_path);
+
+		let source_state = self.keyspace.read_tx();
+		let copy_keyspace = fjall::Config::new(path).open()?;
+
+		for partition_name in self.keyspace.list_partitions() {
+			let source_partition = self
+				.keyspace
+				.open_partition(&partition_name, PartitionCreateOptions::default())?;
+			let copy_partition =
+				copy_keyspace.open_partition(&partition_name, PartitionCreateOptions::default())?;
+
+			for entry in source_state.iter(&source_partition) {
+				let (key, value) = entry?;
+				copy_partition.insert(key, value)?;
+			}
+		}
+
+		copy_keyspace.persist(PersistMode::SyncAll)?;
+		Ok(())
+	}
+
+	// ----
+
+	fn get(&self, tree_idx: usize, key: &[u8]) -> Result<Option<Value>> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		let val = tx.get(&tree, key)?;
+		match val {
+			None => Ok(None),
+			Some(v) => Ok(Some(v.to_vec())),
+		}
+	}
+
+	fn approximate_len(&self, tree_idx: usize) -> Result<usize> {
+		let tree = self.get_tree(tree_idx)?;
+		Ok(tree.approximate_len())
+	}
+	fn is_empty(&self, tree_idx: usize) -> Result<bool> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		Ok(tx.is_empty(&tree)?)
+	}
+
+	fn insert(&self, tree_idx: usize, key: &[u8], value: &[u8]) -> Result<()> {
+		let tree = self.get_tree(tree_idx)?;
+		let mut tx = self.keyspace.write_tx();
+		tx.insert(&tree, key, value);
+		tx.commit()?;
+		Ok(())
+	}
+
+	fn remove(&self, tree_idx: usize, key: &[u8]) -> Result<()> {
+		let tree = self.get_tree(tree_idx)?;
+		let mut tx = self.keyspace.write_tx();
+		tx.remove(&tree, key);
+		tx.commit()?;
+		Ok(())
+	}
+
+	fn clear(&self, tree_idx: usize) -> Result<()> {
+		let mut trees = self.trees.write();
+
+		if tree_idx >= trees.len() {
+			return Err(Error("invalid tree id".into()));
+		}
+		let (name, tree) = trees.remove(tree_idx);
+
+		self.keyspace.delete_partition(tree)?;
+		let tree = self
+			.keyspace
+			.open_partition(&name, PartitionCreateOptions::default())?;
+		trees.insert(tree_idx, (name, tree));
+
+		Ok(())
+	}
+
+	fn iter(&self, tree_idx: usize) -> Result<ValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		Ok(Box::new(tx.iter(&tree).map(iterator_remap)))
+	}
+
+	fn iter_rev(&self, tree_idx: usize) -> Result<ValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		Ok(Box::new(tx.iter(&tree).rev().map(iterator_remap)))
+	}
+
+	fn range<'r>(
+		&self,
+		tree_idx: usize,
+		low: Bound<&'r [u8]>,
+		high: Bound<&'r [u8]>,
+	) -> Result<ValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		Ok(Box::new(
+			tx.range::<&'r [u8], ByteRefRangeBound>(&tree, (low, high))
+				.map(iterator_remap),
+		))
+	}
+	fn range_rev<'r>(
+		&self,
+		tree_idx: usize,
+		low: Bound<&'r [u8]>,
+		high: Bound<&'r [u8]>,
+	) -> Result<ValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let tx = self.keyspace.read_tx();
+		Ok(Box::new(
+			tx.range::<&'r [u8], ByteRefRangeBound>(&tree, (low, high))
+				.rev()
+				.map(iterator_remap),
+		))
+	}
+
+	// ----
+
+	fn transaction(&self, f: &dyn ITxFn) -> TxResult<OnCommit, ()> {
+		let trees = self.trees.read();
+		let mut tx = FjallTx {
+			trees: &trees[..],
+			tx: self.keyspace.write_tx(),
+		};
+
+		let res = f.try_on(&mut tx);
+		match res {
+			TxFnResult::Ok(on_commit) => {
+				tx.tx.commit().map_err(Error::from).map_err(TxError::Db)?;
+				Ok(on_commit)
+			}
+			TxFnResult::Abort => {
+				tx.tx.rollback();
+				Err(TxError::Abort(()))
+			}
+			TxFnResult::DbErr => {
+				tx.tx.rollback();
+				Err(TxError::Db(Error(
+					"(this message will be discarded)".into(),
+				)))
+			}
+		}
+	}
+}
+
+// ----
+
+struct FjallTx<'a> {
+	trees: &'a [(String, TransactionalPartitionHandle)],
+	tx: WriteTransaction<'a>,
+}
+
+impl<'a> FjallTx<'a> {
+	fn get_tree(&self, i: usize) -> TxOpResult<&TransactionalPartitionHandle> {
+		self.trees.get(i).map(|tup| &tup.1).ok_or_else(|| {
+			TxOpError(Error(
+				"invalid tree id (it might have been openned after the transaction started)".into(),
+			))
+		})
+	}
+}
+
+impl<'a> ITx for FjallTx<'a> {
+	fn get(&self, tree_idx: usize, key: &[u8]) -> TxOpResult<Option<Value>> {
+		let tree = self.get_tree(tree_idx)?;
+		match self.tx.get(tree, key)? {
+			Some(v) => Ok(Some(v.to_vec())),
+			None => Ok(None),
+		}
+	}
+	fn len(&self, tree_idx: usize) -> TxOpResult<usize> {
+		let tree = self.get_tree(tree_idx)?;
+		Ok(self.tx.len(tree)? as usize)
+	}
+
+	fn insert(&mut self, tree_idx: usize, key: &[u8], value: &[u8]) -> TxOpResult<()> {
+		let tree = self.get_tree(tree_idx)?.clone();
+		self.tx.insert(&tree, key, value);
+		Ok(())
+	}
+	fn remove(&mut self, tree_idx: usize, key: &[u8]) -> TxOpResult<()> {
+		let tree = self.get_tree(tree_idx)?.clone();
+		self.tx.remove(&tree, key);
+		Ok(())
+	}
+	fn clear(&mut self, _tree_idx: usize) -> TxOpResult<()> {
+		unimplemented!("LSM tree clearing in cross-partition transaction is not supported")
+	}
+
+	fn iter(&self, tree_idx: usize) -> TxOpResult<TxValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?.clone();
+		Ok(Box::new(self.tx.iter(&tree).map(iterator_remap_tx)))
+	}
+	fn iter_rev(&self, tree_idx: usize) -> TxOpResult<TxValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?.clone();
+		Ok(Box::new(self.tx.iter(&tree).rev().map(iterator_remap_tx)))
+	}
+
+	fn range<'r>(
+		&self,
+		tree_idx: usize,
+		low: Bound<&'r [u8]>,
+		high: Bound<&'r [u8]>,
+	) -> TxOpResult<TxValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let low = clone_bound(low);
+		let high = clone_bound(high);
+		Ok(Box::new(
+			self.tx
+				.range::<Vec<u8>, ByteVecRangeBounds>(&tree, (low, high))
+				.map(iterator_remap_tx),
+		))
+	}
+	fn range_rev<'r>(
+		&self,
+		tree_idx: usize,
+		low: Bound<&'r [u8]>,
+		high: Bound<&'r [u8]>,
+	) -> TxOpResult<TxValueIter<'_>> {
+		let tree = self.get_tree(tree_idx)?;
+		let low = clone_bound(low);
+		let high = clone_bound(high);
+		Ok(Box::new(
+			self.tx
+				.range::<Vec<u8>, ByteVecRangeBounds>(&tree, (low, high))
+				.rev()
+				.map(iterator_remap_tx),
+		))
+	}
+}
+
+// -- maps fjall's (k, v) to ours
+
+fn iterator_remap(r: fjall::Result<(fjall::Slice, fjall::Slice)>) -> Result<(Value, Value)> {
+	r.map(|(k, v)| (k.to_vec(), v.to_vec()))
+		.map_err(|e| e.into())
+}
+
+fn iterator_remap_tx(r: fjall::Result<(fjall::Slice, fjall::Slice)>) -> TxOpResult<(Value, Value)> {
+	r.map(|(k, v)| (k.to_vec(), v.to_vec()))
+		.map_err(|e| e.into())
+}
+
+// -- utils to deal with Garage's tightness on Bound lifetimes
+
+type ByteVecBound = Bound<Vec<u8>>;
+type ByteVecRangeBounds = (ByteVecBound, ByteVecBound);
+
+fn clone_bound(bound: Bound<&[u8]>) -> ByteVecBound {
+	let value = match bound {
+		Bound::Excluded(v) | Bound::Included(v) => v.to_vec(),
+		Bound::Unbounded => vec![],
+	};
+
+	match bound {
+		Bound::Included(_) => Bound::Included(value),
+		Bound::Excluded(_) => Bound::Excluded(value),
+		Bound::Unbounded => Bound::Unbounded,
+	}
+}
+
+// -- utils to encode table names --
+
+fn encode_name(s: &str) -> Result<String> {
+	let base = 'A' as u32;
+
+	let mut ret = String::with_capacity(s.len() + 10);
+	for c in s.chars() {
+		if c.is_alphanumeric() || c == '_' || c == '-' || c == '#' {
+			ret.push(c);
+		} else if c <= u8::MAX as char {
+			ret.push('$');
+			let c_hi = c as u32 / 16;
+			let c_lo = c as u32 % 16;
+			ret.push(char::from_u32(base + c_hi).unwrap());
+			ret.push(char::from_u32(base + c_lo).unwrap());
+		} else {
+			return Err(Error(
+				format!("table name {} could not be safely encoded", s).into(),
+			));
+		}
+	}
+	Ok(ret)
+}
+
+fn decode_name(s: &str) -> Result<String> {
+	use std::convert::TryFrom;
+
+	let errfn = || Error(format!("encoded table name {} is invalid", s).into());
+	let c_map = |c: char| {
+		let c = c as u32;
+		let base = 'A' as u32;
+		if (base..base + 16).contains(&c) {
+			Some(c - base)
+		} else {
+			None
+		}
+	};
+
+	let mut ret = String::with_capacity(s.len());
+	let mut it = s.chars();
+	while let Some(c) = it.next() {
+		if c == '$' {
+			let c_hi = it.next().and_then(c_map).ok_or_else(errfn)?;
+			let c_lo = it.next().and_then(c_map).ok_or_else(errfn)?;
+			let c_dec = char::try_from(c_hi * 16 + c_lo).map_err(|_| errfn())?;
+			ret.push(c_dec);
+		} else {
+			ret.push(c);
+		}
+	}
+	Ok(ret)
+}
+
+#[cfg(test)]
+mod tests {
+	use super::*;
+
+	#[test]
+	fn test_encdec_name() {
+		for name in [
+			"testname",
+			"test_name",
+			"test name",
+			"test$name",
+			"test:name@help.me$get/this**right",
+		] {
+			let encname = encode_name(name).unwrap();
+			assert!(!encname.contains(' '));
+			assert!(!encname.contains('.'));
+			assert!(!encname.contains('*'));
+			assert_eq!(*name, decode_name(&encname).unwrap());
+		}
+	}
+}

src/db/lib.rs

@@ -1,6 +1,8 @@
 #[macro_use]
 extern crate tracing;
 
+#[cfg(feature = "fjall")]
+pub mod fjall_adapter;
 #[cfg(feature = "lmdb")]
 pub mod lmdb_adapter;
 #[cfg(feature = "sqlite")]
@@ -152,7 +154,7 @@ impl Db {
 		let tree_names = other.list_trees()?;
 		for name in tree_names {
 			let tree = self.open_tree(&name)?;
-			if tree.len()? > 0 {
+			if !tree.is_empty()? {
 				return Err(Error(format!("tree {} already contains data", name).into()));
 			}
 
@@ -194,8 +196,12 @@ impl Tree {
 		self.0.get(self.1, key.as_ref())
 	}
 	#[inline]
-	pub fn len(&self) -> Result<usize> {
-		self.0.len(self.1)
+	pub fn approximate_len(&self) -> Result<usize> {
+		self.0.approximate_len(self.1)
+	}
+	#[inline]
+	pub fn is_empty(&self) -> Result<bool> {
+		self.0.is_empty(self.1)
 	}
 
 	#[inline]
@@ -333,7 +339,8 @@ pub(crate) trait IDb: Send + Sync {
 	fn snapshot(&self, path: &PathBuf) -> Result<()>;
 
 	fn get(&self, tree: usize, key: &[u8]) -> Result<Option<Value>>;
-	fn len(&self, tree: usize) -> Result<usize>;
+	fn approximate_len(&self, tree: usize) -> Result<usize>;
+	fn is_empty(&self, tree: usize) -> Result<bool>;
 
 	fn insert(&self, tree: usize, key: &[u8], value: &[u8]) -> Result<()>;
 	fn remove(&self, tree: usize, key: &[u8]) -> Result<()>;

src/db/lmdb_adapter.rs

@@ -1,8 +1,8 @@
 use core::ops::Bound;
-use core::ptr::NonNull;
 
 use std::collections::HashMap;
 use std::convert::TryInto;
+use std::marker::PhantomPinned;
 use std::path::PathBuf;
 use std::pin::Pin;
 use std::sync::{Arc, RwLock};
@@ -11,12 +11,55 @@ use heed::types::ByteSlice;
 use heed::{BytesDecode, Env, RoTxn, RwTxn, UntypedDatabase as Database};
 
 use crate::{
+	open::{Engine, OpenOpt},
 	Db, Error, IDb, ITx, ITxFn, OnCommit, Result, TxError, TxFnResult, TxOpError, TxOpResult,
 	TxResult, TxValueIter, Value, ValueIter,
 };
 
 pub use heed;
 
+// ---- top-level open function
+
+pub(crate) fn open_db(path: &PathBuf, opt: &OpenOpt) -> Result<Db> {
+	info!("Opening LMDB database at: {}", path.display());
+	if let Err(e) = std::fs::create_dir_all(&path) {
+		return Err(Error(
+			format!("Unable to create LMDB data directory: {}", e).into(),
+		));
+	}
+
+	let map_size = match opt.lmdb_map_size {
+		None => recommended_map_size(),
+		Some(v) => v - (v % 4096),
+	};
+
+	let mut env_builder = heed::EnvOpenOptions::new();
+	env_builder.max_dbs(100);
+	env_builder.map_size(map_size);
+	env_builder.max_readers(2048);
+	unsafe {
+		env_builder.flag(heed::flags::Flags::MdbNoRdAhead);
+		env_builder.flag(heed::flags::Flags::MdbNoMetaSync);
+		if !opt.fsync {
+			env_builder.flag(heed::flags::Flags::MdbNoSync);
+		}
+	}
+	match env_builder.open(&path) {
+		Err(heed::Error::Io(e)) if e.kind() == std::io::ErrorKind::OutOfMemory => {
+			return Err(Error(
+				"OutOfMemory error while trying to open LMDB database. This can happen \
+                if your operating system is not allowing you to use sufficient virtual \
+                memory address space. Please check that no limit is set (ulimit -v). \
+                You may also try to set a smaller `lmdb_map_size` configuration parameter. \
+                On 32-bit machines, you should probably switch to another database engine."
+					.into(),
+			))
+		}
+		Err(e) => Err(Error(format!("Cannot open LMDB database: {}", e).into())),
+		Ok(db) => Ok(LmdbDb::init(db)),
+	}
+}
+
 // -- err
 
 impl From<heed::Error> for Error {
@@ -104,10 +147,9 @@ impl IDb for LmdbDb {
 		Ok(ret2)
 	}
 
-	fn snapshot(&self, to: &PathBuf) -> Result<()> {
-		std::fs::create_dir_all(to)?;
-		let mut path = to.clone();
-		path.push("data.mdb");
+	fn snapshot(&self, base_path: &PathBuf) -> Result<()> {
+		std::fs::create_dir_all(base_path)?;
+		let path = Engine::Lmdb.db_path(base_path);
 		self.db
 			.copy_to_path(path, heed::CompactionOption::Enabled)?;
 		Ok(())
@@ -126,11 +168,16 @@ impl IDb for LmdbDb {
 		}
 	}
 
-	fn len(&self, tree: usize) -> Result<usize> {
+	fn approximate_len(&self, tree: usize) -> Result<usize> {
 		let tree = self.get_tree(tree)?;
 		let tx = self.db.read_txn()?;
 		Ok(tree.len(&tx)?.try_into().unwrap())
 	}
+	fn is_empty(&self, tree: usize) -> Result<bool> {
+		let tree = self.get_tree(tree)?;
+		let tx = self.db.read_txn()?;
+		Ok(tree.is_empty(&tx)?)
+	}
 
 	fn insert(&self, tree: usize, key: &[u8], value: &[u8]) -> Result<()> {
 		let tree = self.get_tree(tree)?;
@@ -159,13 +206,15 @@ impl IDb for LmdbDb {
 	fn iter(&self, tree: usize) -> Result<ValueIter<'_>> {
 		let tree = self.get_tree(tree)?;
 		let tx = self.db.read_txn()?;
-		TxAndIterator::make(tx, |tx| Ok(tree.iter(tx)?))
+		// Safety: the cloture does not store its argument anywhere,
+		unsafe { TxAndIterator::make(tx, |tx| Ok(tree.iter(tx)?)) }
 	}
 
 	fn iter_rev(&self, tree: usize) -> Result<ValueIter<'_>> {
 		let tree = self.get_tree(tree)?;
 		let tx = self.db.read_txn()?;
-		TxAndIterator::make(tx, |tx| Ok(tree.rev_iter(tx)?))
+		// Safety: the cloture does not store its argument anywhere,
+		unsafe { TxAndIterator::make(tx, |tx| Ok(tree.rev_iter(tx)?)) }
 	}
 
 	fn range<'r>(
@@ -176,7 +225,8 @@ impl IDb for LmdbDb {
 	) -> Result<ValueIter<'_>> {
 		let tree = self.get_tree(tree)?;
 		let tx = self.db.read_txn()?;
-		TxAndIterator::make(tx, |tx| Ok(tree.range(tx, &(low, high))?))
+		// Safety: the cloture does not store its argument anywhere,
+		unsafe { TxAndIterator::make(tx, |tx| Ok(tree.range(tx, &(low, high))?)) }
 	}
 	fn range_rev<'r>(
 		&self,
@@ -186,7 +236,8 @@ impl IDb for LmdbDb {
 	) -> Result<ValueIter<'_>> {
 		let tree = self.get_tree(tree)?;
 		let tx = self.db.read_txn()?;
-		TxAndIterator::make(tx, |tx| Ok(tree.rev_range(tx, &(low, high))?))
+		// Safety: the cloture does not store its argument anywhere,
+		unsafe { TxAndIterator::make(tx, |tx| Ok(tree.rev_range(tx, &(low, high))?)) }
 	}
 
 	// ----
@@ -316,28 +367,41 @@ where
 {
 	tx: RoTxn<'a>,
 	iter: Option<I>,
+	_pin: PhantomPinned,
 }
 
 impl<'a, I> TxAndIterator<'a, I>
 where
 	I: Iterator<Item = IteratorItem<'a>> + 'a,
 {
-	fn make<F>(tx: RoTxn<'a>, iterfun: F) -> Result<ValueIter<'a>>
+	fn iter(self: Pin<&mut Self>) -> &mut Option<I> {
+		// Safety: iter is not structural
+		unsafe { &mut self.get_unchecked_mut().iter }
+	}
+
+	/// Safety: iterfun must not store its argument anywhere but in its result.
+	unsafe fn make<F>(tx: RoTxn<'a>, iterfun: F) -> Result<ValueIter<'a>>
 	where
 		F: FnOnce(&'a RoTxn<'a>) -> Result<I>,
 	{
-		let res = TxAndIterator { tx, iter: None };
+		let res = TxAndIterator {
+			tx,
+			iter: None,
+			_pin: PhantomPinned,
+		};
 		let mut boxed = Box::pin(res);
 
-		// This unsafe allows us to bypass lifetime checks
-		let tx = unsafe { NonNull::from(&boxed.tx).as_ref() };
-		let iter = iterfun(tx)?;
+		let tx_lifetime_overextended: &'a RoTxn<'a> = {
+			let tx = &boxed.tx;
+			// Safety: Artificially extending the lifetime because
+			// this reference will only be stored and accessed from the
+			// returned ValueIter which guarantees that it is destroyed
+			// before the tx it is pointing  to.
+			unsafe { &*&raw const *tx }
+		};
+		let iter = iterfun(&tx_lifetime_overextended)?;
 
-		let mut_ref = Pin::as_mut(&mut boxed);
-		// This unsafe allows us to write in a field of the pinned struct
-		unsafe {
-			Pin::get_unchecked_mut(mut_ref).iter = Some(iter);
-		}
+		*boxed.as_mut().iter() = Some(iter);
 
 		Ok(Box::new(TxAndIteratorPin(boxed)))
 	}
@@ -348,8 +412,10 @@ where
 	I: Iterator<Item = IteratorItem<'a>> + 'a,
 {
 	fn drop(&mut self) {
-		// ensure the iterator is dropped before the RoTxn it references
-		drop(self.iter.take());
+		// Safety: `new_unchecked` is okay because we know this value is never
+		// used again after being dropped.
+		let this = unsafe { Pin::new_unchecked(self) };
+		drop(this.iter().take());
 	}
 }
 
@@ -365,13 +431,12 @@ where
 
 	fn next(&mut self) -> Option<Self::Item> {
 		let mut_ref = Pin::as_mut(&mut self.0);
-		// This unsafe allows us to mutably access the iterator field
-		let next = unsafe { Pin::get_unchecked_mut(mut_ref).iter.as_mut()?.next() };
-		match next {
-			None => None,
-			Some(Err(e)) => Some(Err(e.into())),
-			Some(Ok((k, v))) => Some(Ok((k.to_vec(), v.to_vec()))),
-		}
+		let next = mut_ref.iter().as_mut()?.next()?;
+		let res = match next {
+			Err(e) => Err(e.into()),
+			Ok((k, v)) => Ok((k.to_vec(), v.to_vec())),
+		};
+		Some(res)
 	}
 }
 

src/db/open.rs

@@ -11,6 +11,7 @@ use crate::{Db, Error, Result};
 pub enum Engine {
 	Lmdb,
 	Sqlite,
+	Fjall,
 }
 
 impl Engine {
@@ -19,8 +20,26 @@ impl Engine {
 		match self {
 			Self::Lmdb => "lmdb",
 			Self::Sqlite => "sqlite",
+			Self::Fjall => "fjall",
 		}
 	}
+
+	/// Return engine-specific DB path from base path
+	pub fn db_path(&self, base_path: &PathBuf) -> PathBuf {
+		let mut ret = base_path.clone();
+		match self {
+			Self::Lmdb => {
+				ret.push("db.lmdb");
+			}
+			Self::Sqlite => {
+				ret.push("db.sqlite");
+			}
+			Self::Fjall => {
+				ret.push("db.fjall");
+			}
+		}
+		ret
+	}
 }
 
 impl std::fmt::Display for Engine {
@@ -36,10 +55,11 @@ impl std::str::FromStr for Engine {
 		match text {
 			"lmdb" | "heed" => Ok(Self::Lmdb),
 			"sqlite" | "sqlite3" | "rusqlite" => Ok(Self::Sqlite),
+            "fjall" => Ok(Self::Fjall),
 			"sled" => Err(Error("Sled is no longer supported as a database engine. Converting your old metadata db can be done using an older Garage binary (e.g. v0.9.4).".into())),
 			kind => Err(Error(
 				format!(
-					"Invalid DB engine: {} (options are: lmdb, sqlite)",
+					"Invalid DB engine: {} (options are: lmdb, sqlite, fjall)",
 					kind
 				)
 				.into(),
@@ -51,6 +71,7 @@ impl std::str::FromStr for Engine {
 pub struct OpenOpt {
 	pub fsync: bool,
 	pub lmdb_map_size: Option<usize>,
+	pub fjall_block_cache_size: Option<usize>,
 }
 
 impl Default for OpenOpt {
@@ -58,6 +79,7 @@ impl Default for OpenOpt {
 		Self {
 			fsync: false,
 			lmdb_map_size: None,
+			fjall_block_cache_size: None,
 		}
 	}
 }
@@ -66,53 +88,15 @@ pub fn open_db(path: &PathBuf, engine: Engine, opt: &OpenOpt) -> Result<Db> {
 	match engine {
 		// ---- Sqlite DB ----
 		#[cfg(feature = "sqlite")]
-		Engine::Sqlite => {
-			info!("Opening Sqlite database at: {}", path.display());
-			let manager = r2d2_sqlite::SqliteConnectionManager::file(path);
-			Ok(crate::sqlite_adapter::SqliteDb::new(manager, opt.fsync)?)
-		}
+		Engine::Sqlite => crate::sqlite_adapter::open_db(path, opt),
 
 		// ---- LMDB DB ----
 		#[cfg(feature = "lmdb")]
-		Engine::Lmdb => {
-			info!("Opening LMDB database at: {}", path.display());
-			if let Err(e) = std::fs::create_dir_all(&path) {
-				return Err(Error(
-					format!("Unable to create LMDB data directory: {}", e).into(),
-				));
-			}
+		Engine::Lmdb => crate::lmdb_adapter::open_db(path, opt),
 
-			let map_size = match opt.lmdb_map_size {
-				None => crate::lmdb_adapter::recommended_map_size(),
-				Some(v) => v - (v % 4096),
-			};
-
-			let mut env_builder = heed::EnvOpenOptions::new();
-			env_builder.max_dbs(100);
-			env_builder.map_size(map_size);
-			env_builder.max_readers(2048);
-			unsafe {
-				env_builder.flag(crate::lmdb_adapter::heed::flags::Flags::MdbNoRdAhead);
-				env_builder.flag(crate::lmdb_adapter::heed::flags::Flags::MdbNoMetaSync);
-				if !opt.fsync {
-					env_builder.flag(heed::flags::Flags::MdbNoSync);
-				}
-			}
-			match env_builder.open(&path) {
-				Err(heed::Error::Io(e)) if e.kind() == std::io::ErrorKind::OutOfMemory => {
-					return Err(Error(
-						"OutOfMemory error while trying to open LMDB database. This can happen \
-                        if your operating system is not allowing you to use sufficient virtual \
-                        memory address space. Please check that no limit is set (ulimit -v). \
-                        You may also try to set a smaller `lmdb_map_size` configuration parameter. \
-                        On 32-bit machines, you should probably switch to another database engine."
-							.into(),
-					))
-				}
-				Err(e) => Err(Error(format!("Cannot open LMDB database: {}", e).into())),
-				Ok(db) => Ok(crate::lmdb_adapter::LmdbDb::init(db)),
-			}
-		}
+		// ---- Fjall DB ----
+		#[cfg(feature = "fjall")]
+		Engine::Fjall => crate::fjall_adapter::open_db(path, opt),
 
 		// Pattern is unreachable when all supported DB engines are compiled into binary. The allow
 		// attribute is added so that we won't have to change this match in case stop building

src/db/sqlite_adapter.rs

@@ -11,12 +11,23 @@ use r2d2_sqlite::SqliteConnectionManager;
 use rusqlite::{params, Rows, Statement, Transaction};
 
 use crate::{
+	open::{Engine, OpenOpt},
 	Db, Error, IDb, ITx, ITxFn, OnCommit, Result, TxError, TxFnResult, TxOpError, TxOpResult,
 	TxResult, TxValueIter, Value, ValueIter,
 };
 
 pub use rusqlite;
 
+// ---- top-level open function
+
+pub(crate) fn open_db(path: &PathBuf, opt: &OpenOpt) -> Result<Db> {
+	info!("Opening Sqlite database at: {}", path.display());
+	let manager = r2d2_sqlite::SqliteConnectionManager::file(path);
+	Ok(SqliteDb::new(manager, opt.fsync)?)
+}
+
+// ----
+
 type Connection = r2d2::PooledConnection<SqliteConnectionManager>;
 
 // --- err
@@ -139,17 +150,32 @@ impl IDb for SqliteDb {
 		Ok(trees)
 	}
 
-	fn snapshot(&self, to: &PathBuf) -> Result<()> {
+	fn snapshot(&self, base_path: &PathBuf) -> Result<()> {
 		fn progress(p: rusqlite::backup::Progress) {
-			let percent = (p.pagecount - p.remaining) * 100 / p.pagecount;
-			info!("Sqlite snapshot progress: {}%", percent);
+			use std::sync::atomic::{AtomicU64, Ordering};
+			use std::time::{SystemTime, UNIX_EPOCH};
+
+			static LAST_LOG_TIME: AtomicU64 = AtomicU64::new(0);
+
+			let now = SystemTime::now()
+				.duration_since(UNIX_EPOCH)
+				.expect("Fix your clock :o")
+				.as_millis() as u64;
+			if now >= LAST_LOG_TIME.load(Ordering::Relaxed) + 10 * 1000 {
+				let percent = (p.pagecount - p.remaining) * 100 / p.pagecount;
+				info!("Sqlite snapshot progress: {}%", percent);
+
+				LAST_LOG_TIME.fetch_max(now, Ordering::Relaxed);
+			}
 		}
-		std::fs::create_dir_all(to)?;
-		let mut path = to.clone();
-		path.push("db.sqlite");
+
+		std::fs::create_dir_all(base_path)?;
+		let path = Engine::Sqlite.db_path(&base_path);
+
 		self.db
 			.get()?
 			.backup(rusqlite::DatabaseName::Main, path, Some(progress))?;
+
 		Ok(())
 	}
 
@@ -160,7 +186,7 @@ impl IDb for SqliteDb {
 		self.internal_get(&self.db.get()?, &tree, key)
 	}
 
-	fn len(&self, tree: usize) -> Result<usize> {
+	fn approximate_len(&self, tree: usize) -> Result<usize> {
 		let tree = self.get_tree(tree)?;
 		let db = self.db.get()?;
 
@@ -172,6 +198,10 @@ impl IDb for SqliteDb {
 		}
 	}
 
+	fn is_empty(&self, tree: usize) -> Result<bool> {
+		Ok(self.approximate_len(tree)? == 0)
+	}
+
 	fn insert(&self, tree: usize, key: &[u8], value: &[u8]) -> Result<()> {
 		let tree = self.get_tree(tree)?;
 		let db = self.db.get()?;

src/db/test.rs

@@ -1,7 +1,7 @@
 use crate::*;
 
 fn test_suite(db: Db) {
-	let tree = db.open_tree("tree").unwrap();
+	let tree = db.open_tree("tree:this_is_a_tree").unwrap();
 
 	let ka: &[u8] = &b"test"[..];
 	let kb: &[u8] = &b"zwello"[..];
@@ -14,7 +14,7 @@ fn test_suite(db: Db) {
 
 	assert!(tree.insert(ka, va).is_ok());
 	assert_eq!(tree.get(ka).unwrap().unwrap(), va);
-	assert_eq!(tree.len().unwrap(), 1);
+	assert_eq!(tree.iter().unwrap().count(), 1);
 
 	// ---- test transaction logic ----
 
@@ -148,3 +148,15 @@ fn test_sqlite_db() {
 	let db = SqliteDb::new(manager, false).unwrap();
 	test_suite(db);
 }
+
+#[test]
+#[cfg(feature = "fjall")]
+fn test_fjall_db() {
+	use crate::fjall_adapter::{fjall, FjallDb};
+
+	let path = mktemp::Temp::new_dir().unwrap();
+	let config = fjall::Config::new(path).temporary(true);
+	let keyspace = config.open_transactional().unwrap();
+	let db = FjallDb::init(keyspace);
+	test_suite(db);
+}

src/garage/Cargo.toml

@@ -93,6 +93,7 @@ k2v = [ "garage_util/k2v", "garage_api_k2v", "garage_api_admin/k2v" ]
 # Database engines
 lmdb = [ "garage_model/lmdb" ]
 sqlite = [ "garage_model/sqlite" ]
+fjall = [ "garage_model/fjall" ]
 
 # Automatic registration and discovery via Consul API
 consul-discovery = [ "garage_rpc/consul-discovery" ]

src/garage/server.rs

@@ -180,10 +180,21 @@ fn watch_shutdown_signal() -> watch::Receiver<bool> {
 		let mut sigterm =
 			signal(SignalKind::terminate()).expect("Failed to install SIGTERM handler");
 		let mut sighup = signal(SignalKind::hangup()).expect("Failed to install SIGHUP handler");
-		tokio::select! {
-			_ = sigint.recv() => info!("Received SIGINT, shutting down."),
-			_ = sigterm.recv() => info!("Received SIGTERM, shutting down."),
-			_ = sighup.recv() => info!("Received SIGHUP, shutting down."),
+		loop {
+			tokio::select! {
+					_ = sigint.recv() => {
+						info!("Received SIGINT, shutting down.");
+						break
+					}
+					_ = sigterm.recv() => {
+						info!("Received SIGTERM, shutting down.");
+						break
+					}
+					_ = sighup.recv() => {
+						info!("Received SIGHUP, reload not supported.");
+						continue
+					}
+			}
 		}
 		send_cancel.send(true).unwrap();
 	});

src/k2v-client/lib.rs

@@ -72,6 +72,16 @@ impl K2vClient {
 			.enable_http2()
 			.build();
 		let client = HttpClient::builder(TokioExecutor::new()).build(connector);
+		Self::new_with_client(config, client)
+	}
+
+	/// Create a new K2V client with an external client.
+	/// Useful for example if you plan on creating many clients but you want to mutualize the
+	/// underlying thread pools & co.
+	pub fn new_with_client(
+		config: K2vClientConfig,
+		client: HttpClient<HttpsConnector<HttpConnector>, Body>,
+	) -> Result<Self, Error> {
 		let user_agent: std::borrow::Cow<str> = match &config.user_agent {
 			Some(ua) => ua.into(),
 			None => format!("k2v/{}", env!("CARGO_PKG_VERSION")).into(),

src/model/Cargo.toml

@@ -45,3 +45,4 @@ default = [ "lmdb", "sqlite" ]
 k2v = [ "garage_util/k2v" ]
 lmdb = [ "garage_db/lmdb" ]
 sqlite = [ "garage_db/sqlite" ]
+fjall = [ "garage_db/fjall" ]

src/model/garage.rs

@@ -119,21 +119,17 @@ impl Garage {
 		info!("Opening database...");
 		let db_engine = db::Engine::from_str(&config.db_engine)
 			.ok_or_message("Invalid `db_engine` value in configuration file")?;
-		let mut db_path = config.metadata_dir.clone();
-		match db_engine {
-			db::Engine::Sqlite => {
-				db_path.push("db.sqlite");
-			}
-			db::Engine::Lmdb => {
-				db_path.push("db.lmdb");
-			}
-		}
+		let db_path = db_engine.db_path(&config.metadata_dir);
 		let db_opt = db::OpenOpt {
 			fsync: config.metadata_fsync,
 			lmdb_map_size: match config.lmdb_map_size {
 				v if v == usize::default() => None,
 				v => Some(v),
 			},
+			fjall_block_cache_size: match config.fjall_block_cache_size {
+				v if v == usize::default() => None,
+				v => Some(v),
+			},
 		};
 		let db = db::open_db(&db_path, db_engine, &db_opt)
 			.ok_or_message("Unable to open metadata db")?;

src/model/s3/lifecycle_worker.rs

@@ -121,13 +121,13 @@ impl Worker for LifecycleWorker {
 				mpu_aborted,
 				..
 			} => {
-				let n_objects = self.garage.object_table.data.store.len().ok();
+				let n_objects = self.garage.object_table.data.store.approximate_len().ok();
 				let progress = match n_objects {
-					None => "...".to_string(),
-					Some(total) => format!(
+					Some(total) if total > 0 => format!(
 						"~{:.2}%",
 						100. * std::cmp::min(*counter, total) as f32 / total as f32
 					),
+					_ => "...".to_string(),
 				};
 				WorkerStatus {
 					progress: Some(progress),

src/rpc/layout/manager.rs

@@ -229,13 +229,11 @@ impl LayoutManager {
 	}
 
 	/// Save cluster layout data to disk
-	async fn save_cluster_layout(&self) -> Result<(), Error> {
+	async fn save_cluster_layout(&self) {
 		let layout = self.layout.read().unwrap().inner().clone();
-		self.persist_cluster_layout
-			.save_async(&layout)
-			.await
-			.expect("Cannot save current cluster layout");
-		Ok(())
+		if let Err(e) = self.persist_cluster_layout.save_async(&layout).await {
+			error!("Failed to save cluster_layout: {}", e);
+		}
 	}
 
 	fn broadcast_update(self: &Arc<Self>, rpc: SystemRpc) {
@@ -313,7 +311,7 @@ impl LayoutManager {
 
 			self.change_notify.notify_waiters();
 			self.broadcast_update(SystemRpc::AdvertiseClusterLayout(new_layout));
-			self.save_cluster_layout().await?;
+			self.save_cluster_layout().await;
 		}
 
 		Ok(SystemRpc::Ok)
@@ -328,7 +326,7 @@ impl LayoutManager {
 		if let Some(new_trackers) = self.merge_layout_trackers(trackers) {
 			self.change_notify.notify_waiters();
 			self.broadcast_update(SystemRpc::AdvertiseClusterLayoutTrackers(new_trackers));
-			self.save_cluster_layout().await?;
+			self.save_cluster_layout().await;
 		}
 
 		Ok(SystemRpc::Ok)

src/table/data.rs

@@ -368,11 +368,11 @@ impl<F: TableSchema, R: TableReplication> TableData<F, R> {
 		}
 	}
 
-	pub fn insert_queue_len(&self) -> Result<usize, Error> {
-		Ok(self.insert_queue.len()?)
+	pub fn insert_queue_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.insert_queue.approximate_len()?)
 	}
 
-	pub fn gc_todo_len(&self) -> Result<usize, Error> {
-		Ok(self.gc_todo.len()?)
+	pub fn gc_todo_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.gc_todo.approximate_len()?)
 	}
 }

src/table/gc.rs

@@ -313,7 +313,7 @@ impl<F: TableSchema, R: TableReplication> Worker for GcWorker<F, R> {
 
 	fn status(&self) -> WorkerStatus {
 		WorkerStatus {
-			queue_length: Some(self.gc.data.gc_todo_len().unwrap_or(0) as u64),
+			queue_length: Some(self.gc.data.gc_todo_approximate_len().unwrap_or(0) as u64),
 			..Default::default()
 		}
 	}

src/table/merkle.rs

@@ -287,12 +287,12 @@ impl<F: TableSchema, R: TableReplication> MerkleUpdater<F, R> {
 		MerkleNode::decode_opt(&ent)
 	}
 
-	pub fn merkle_tree_len(&self) -> Result<usize, Error> {
-		Ok(self.data.merkle_tree.len()?)
+	pub fn merkle_tree_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.data.merkle_tree.approximate_len()?)
 	}
 
-	pub fn todo_len(&self) -> Result<usize, Error> {
-		Ok(self.data.merkle_todo.len()?)
+	pub fn todo_approximate_len(&self) -> Result<usize, Error> {
+		Ok(self.data.merkle_todo.approximate_len()?)
 	}
 }
 
@@ -306,7 +306,7 @@ impl<F: TableSchema, R: TableReplication> Worker for MerkleWorker<F, R> {
 
 	fn status(&self) -> WorkerStatus {
 		WorkerStatus {
-			queue_length: Some(self.0.todo_len().unwrap_or(0) as u64),
+			queue_length: Some(self.0.todo_approximate_len().unwrap_or(0) as u64),
 			..Default::default()
 		}
 	}

src/table/metrics.rs

@@ -36,7 +36,7 @@ impl TableMetrics {
 				.u64_value_observer(
 					"table.size",
 					move |observer| {
-						if let Ok(value) = store.len() {
+						if let Ok(value) = store.approximate_len() {
 							observer.observe(
 								value as u64,
 								&[KeyValue::new("table_name", table_name)],
@@ -50,7 +50,7 @@ impl TableMetrics {
 				.u64_value_observer(
 					"table.merkle_tree_size",
 					move |observer| {
-						if let Ok(value) = merkle_tree.len() {
+						if let Ok(value) = merkle_tree.approximate_len() {
 							observer.observe(
 								value as u64,
 								&[KeyValue::new("table_name", table_name)],
@@ -64,7 +64,7 @@ impl TableMetrics {
 				.u64_value_observer(
 					"table.merkle_updater_todo_queue_length",
 					move |observer| {
-						if let Ok(v) = merkle_todo.len() {
+						if let Ok(v) = merkle_todo.approximate_len() {
 							observer.observe(
 								v as u64,
 								&[KeyValue::new("table_name", table_name)],
@@ -78,7 +78,7 @@ impl TableMetrics {
 				.u64_value_observer(
 					"table.insert_queue_length",
 					move |observer| {
-						if let Ok(v) = insert_queue.len() {
+						if let Ok(v) = insert_queue.approximate_len() {
 							observer.observe(
 								v as u64,
 								&[KeyValue::new("table_name", table_name)],
@@ -92,7 +92,7 @@ impl TableMetrics {
 				.u64_value_observer(
 					"table.gc_todo_queue_length",
 					move |observer| {
-                        if let Ok(value) = gc_todo.len() {
+                        if let Ok(value) = gc_todo.approximate_len() {
                             observer.observe(
                                 value as u64,
                                 &[KeyValue::new("table_name", table_name)],

src/table/queue.rs

@@ -27,7 +27,7 @@ impl<F: TableSchema, R: TableReplication> Worker for InsertQueueWorker<F, R> {
 
 	fn status(&self) -> WorkerStatus {
 		WorkerStatus {
-			queue_length: Some(self.0.data.insert_queue.len().unwrap_or(0) as u64),
+			queue_length: Some(self.0.data.insert_queue.approximate_len().unwrap_or(0) as u64),
 			..Default::default()
 		}
 	}

src/util/config.rs

@@ -75,6 +75,10 @@ pub struct Config {
 	)]
 	pub block_ram_buffer_max: usize,
 
+	/// Maximum number of concurrent reads of block files on disk
+	#[serde(default = "default_block_max_concurrent_reads")]
+	pub block_max_concurrent_reads: usize,
+
 	/// Skip the permission check of secret files. Useful when
 	/// POSIX ACLs (or more complex chmods) are used.
 	#[serde(default)]
@@ -122,6 +126,10 @@ pub struct Config {
 	#[serde(deserialize_with = "deserialize_capacity", default)]
 	pub lmdb_map_size: usize,
 
+	/// Fjall block cache size
+	#[serde(deserialize_with = "deserialize_capacity", default)]
+	pub fjall_block_cache_size: usize,
+
 	// -- APIs
 	/// Configuration for S3 api
 	pub s3_api: S3ApiConfig,
@@ -279,6 +287,9 @@ fn default_block_size() -> usize {
 fn default_block_ram_buffer_max() -> usize {
 	256 * 1024 * 1024
 }
+fn default_block_max_concurrent_reads() -> usize {
+	16
+}
 
 fn default_consistency_mode() -> String {
 	"consistent".into()