diff options
Diffstat (limited to 'src/block')
-rw-r--r-- | src/block/Cargo.toml | 2 | ||||
-rw-r--r-- | src/block/block.rs | 2 | ||||
-rw-r--r-- | src/block/lib.rs | 3 | ||||
-rw-r--r-- | src/block/manager.rs | 59 | ||||
-rw-r--r-- | src/block/metrics.rs | 17 | ||||
-rw-r--r-- | src/block/rc.rs | 83 | ||||
-rw-r--r-- | src/block/repair.rs | 2 | ||||
-rw-r--r-- | src/block/resync.rs | 46 |
8 files changed, 142 insertions, 72 deletions
diff --git a/src/block/Cargo.toml b/src/block/Cargo.toml index 0f6fb5bb..b5763120 100644 --- a/src/block/Cargo.toml +++ b/src/block/Cargo.toml @@ -1,6 +1,6 @@ [package] name = "garage_block" -version = "0.9.3" +version = "0.10.0" authors = ["Alex Auvolat <alex@adnab.me>"] edition = "2018" license = "AGPL-3.0" diff --git a/src/block/block.rs b/src/block/block.rs index 504d11f8..bd95680e 100644 --- a/src/block/block.rs +++ b/src/block/block.rs @@ -96,7 +96,7 @@ impl DataBlock { } } -fn zstd_encode<R: std::io::Read>(mut source: R, level: i32) -> std::io::Result<Vec<u8>> { +pub fn zstd_encode<R: std::io::Read>(mut source: R, level: i32) -> std::io::Result<Vec<u8>> { let mut result = Vec::<u8>::new(); let mut encoder = Encoder::new(&mut result, level)?; encoder.include_checksum(true)?; diff --git a/src/block/lib.rs b/src/block/lib.rs index c9ff2845..944f0d83 100644 --- a/src/block/lib.rs +++ b/src/block/lib.rs @@ -9,3 +9,6 @@ mod block; mod layout; mod metrics; mod rc; + +pub use block::zstd_encode; +pub use rc::CalculateRefcount; diff --git a/src/block/manager.rs b/src/block/manager.rs index 62829a24..40b177a2 100644 --- a/src/block/manager.rs +++ b/src/block/manager.rs @@ -89,7 +89,7 @@ pub struct BlockManager { mutation_lock: Vec<Mutex<BlockManagerLocked>>, - pub(crate) rc: BlockRc, + pub rc: BlockRc, pub resync: BlockResyncManager, pub(crate) system: Arc<System>, @@ -158,7 +158,7 @@ impl BlockManager { let metrics = BlockManagerMetrics::new( config.compression_level, - rc.rc.clone(), + rc.rc_table.clone(), resync.queue.clone(), resync.errors.clone(), buffer_kb_semaphore.clone(), @@ -233,6 +233,12 @@ impl BlockManager { } } + /// Initialization: set how block references are recalculated + /// for repair operations + pub fn set_recalc_rc(&self, recalc: Vec<CalculateRefcount>) { + self.rc.recalc_rc.store(Some(Arc::new(recalc))); + } + /// Ask nodes that might have a (possibly compressed) block for it /// Return it as a stream with a header async fn rpc_get_raw_block_streaming( @@ -279,8 +285,10 @@ impl BlockManager { F: Fn(DataBlockStream) -> Fut, Fut: futures::Future<Output = Result<T, Error>>, { - let who = self.replication.read_nodes(hash); - let who = self.system.rpc.request_order(&who); + let who = self + .system + .rpc_helper() + .block_read_nodes_of(hash, self.system.rpc_helper()); for node in who.iter() { let node_id = NodeID::from(*node); @@ -320,15 +328,15 @@ impl BlockManager { // if the first one doesn't succeed rapidly // TODO: keep first request running when initiating a new one and take the // one that finishes earlier - _ = tokio::time::sleep(self.system.rpc.rpc_timeout()) => { + _ = tokio::time::sleep(self.system.rpc_helper().rpc_timeout()) => { debug!("Get block {:?}: node {:?} didn't return block in time, trying next.", hash, node); } }; } - let msg = format!("Get block {:?}: no node returned a valid block", hash); - debug!("{}", msg); - Err(Error::Message(msg)) + let err = Error::MissingBlock(*hash); + debug!("{}", err); + Err(err) } // ---- Public interface ---- @@ -355,26 +363,18 @@ impl BlockManager { } } - /// Ask nodes that might have a block for it, return it as one big Bytes - pub async fn rpc_get_block( - &self, - hash: &Hash, - order_tag: Option<OrderTag>, - ) -> Result<Bytes, Error> { - let stream = self.rpc_get_block_streaming(hash, order_tag).await?; - Ok(read_stream_to_end(stream).await?.into_bytes()) - } - /// Send block to nodes that should have it pub async fn rpc_put_block( &self, hash: Hash, data: Bytes, + prevent_compression: bool, order_tag: Option<OrderTag>, ) -> Result<(), Error> { - let who = self.replication.write_nodes(&hash); + let who = self.replication.write_sets(&hash); - let (header, bytes) = DataBlock::from_buffer(data, self.compression_level) + let compression_level = self.compression_level.filter(|_| !prevent_compression); + let (header, bytes) = DataBlock::from_buffer(data, compression_level) .await .into_parts(); @@ -394,10 +394,10 @@ impl BlockManager { }; self.system - .rpc - .try_call_many( + .rpc_helper() + .try_write_many_sets( &self.endpoint, - &who[..], + who.as_ref(), put_block_rpc, RequestStrategy::with_priority(PRIO_NORMAL | PRIO_SECONDARY) .with_drop_on_completion(permit) @@ -410,12 +410,7 @@ impl BlockManager { /// Get number of items in the refcount table pub fn rc_len(&self) -> Result<usize, Error> { - Ok(self.rc.rc.len()?) - } - - /// Get number of items in the refcount table - pub fn rc_fast_len(&self) -> Result<Option<usize>, Error> { - Ok(self.rc.rc.fast_len()?) + Ok(self.rc.rc_table.len()?) } /// Send command to start/stop/manager scrub worker @@ -433,7 +428,7 @@ 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.len()?); for ent in self.resync.errors.iter()? { let (hash, cnt) = ent?; let cnt = ErrorCounter::decode(&cnt); @@ -471,7 +466,7 @@ impl BlockManager { tokio::spawn(async move { if let Err(e) = this .resync - .put_to_resync(&hash, 2 * this.system.rpc.rpc_timeout()) + .put_to_resync(&hash, 2 * this.system.rpc_helper().rpc_timeout()) { error!("Block {:?} could not be put in resync queue: {}.", hash, e); } @@ -565,7 +560,7 @@ impl BlockManager { None => { // Not found but maybe we should have had it ?? self.resync - .put_to_resync(hash, 2 * self.system.rpc.rpc_timeout())?; + .put_to_resync(hash, 2 * self.system.rpc_helper().rpc_timeout())?; return Err(Error::Message(format!( "block {:?} not found on node", hash diff --git a/src/block/metrics.rs b/src/block/metrics.rs index c989f940..2d41e365 100644 --- a/src/block/metrics.rs +++ b/src/block/metrics.rs @@ -5,7 +5,6 @@ use tokio::sync::Semaphore; use opentelemetry::{global, metrics::*}; use garage_db as db; -use garage_db::counted_tree_hack::CountedTree; /// TableMetrics reference all counter used for metrics pub struct BlockManagerMetrics { @@ -34,8 +33,8 @@ impl BlockManagerMetrics { pub fn new( compression_level: Option<i32>, rc_tree: db::Tree, - resync_queue: CountedTree, - resync_errors: CountedTree, + resync_queue: db::Tree, + resync_errors: db::Tree, buffer_semaphore: Arc<Semaphore>, ) -> Self { let meter = global::meter("garage_model/block"); @@ -51,15 +50,17 @@ impl BlockManagerMetrics { .init(), _rc_size: meter .u64_value_observer("block.rc_size", move |observer| { - if let Ok(Some(v)) = rc_tree.fast_len() { - observer.observe(v as u64, &[]) + if let Ok(value) = rc_tree.len() { + observer.observe(value as u64, &[]) } }) .with_description("Number of blocks known to the reference counter") .init(), _resync_queue_len: meter .u64_value_observer("block.resync_queue_length", move |observer| { - observer.observe(resync_queue.len() as u64, &[]) + if let Ok(value) = resync_queue.len() { + observer.observe(value as u64, &[]); + } }) .with_description( "Number of block hashes queued for local check and possible resync", @@ -67,7 +68,9 @@ impl BlockManagerMetrics { .init(), _resync_errored_blocks: meter .u64_value_observer("block.resync_errored_blocks", move |observer| { - observer.observe(resync_errors.len() as u64, &[]) + if let Ok(value) = resync_errors.len() { + observer.observe(value as u64, &[]); + } }) .with_description("Number of block hashes whose last resync resulted in an error") .init(), diff --git a/src/block/rc.rs b/src/block/rc.rs index b6afb277..4a55ee29 100644 --- a/src/block/rc.rs +++ b/src/block/rc.rs @@ -1,5 +1,7 @@ use std::convert::TryInto; +use arc_swap::ArcSwapOption; + use garage_db as db; use garage_util::data::*; @@ -8,13 +10,20 @@ use garage_util::time::*; use crate::manager::BLOCK_GC_DELAY; +pub type CalculateRefcount = + Box<dyn Fn(&db::Transaction, &Hash) -> db::TxResult<usize, Error> + Send + Sync>; + pub struct BlockRc { - pub(crate) rc: db::Tree, + pub rc_table: db::Tree, + pub(crate) recalc_rc: ArcSwapOption<Vec<CalculateRefcount>>, } impl BlockRc { pub(crate) fn new(rc: db::Tree) -> Self { - Self { rc } + Self { + rc_table: rc, + recalc_rc: ArcSwapOption::new(None), + } } /// Increment the reference counter associated to a hash. @@ -24,9 +33,9 @@ impl BlockRc { tx: &mut db::Transaction, hash: &Hash, ) -> db::TxOpResult<bool> { - let old_rc = RcEntry::parse_opt(tx.get(&self.rc, hash)?); + let old_rc = RcEntry::parse_opt(tx.get(&self.rc_table, hash)?); match old_rc.increment().serialize() { - Some(x) => tx.insert(&self.rc, hash, x)?, + Some(x) => tx.insert(&self.rc_table, hash, x)?, None => unreachable!(), }; Ok(old_rc.is_zero()) @@ -39,28 +48,28 @@ impl BlockRc { tx: &mut db::Transaction, hash: &Hash, ) -> db::TxOpResult<bool> { - let new_rc = RcEntry::parse_opt(tx.get(&self.rc, hash)?).decrement(); + let new_rc = RcEntry::parse_opt(tx.get(&self.rc_table, hash)?).decrement(); match new_rc.serialize() { - Some(x) => tx.insert(&self.rc, hash, x)?, - None => tx.remove(&self.rc, hash)?, + Some(x) => tx.insert(&self.rc_table, hash, x)?, + None => tx.remove(&self.rc_table, hash)?, }; Ok(matches!(new_rc, RcEntry::Deletable { .. })) } /// Read a block's reference count pub(crate) fn get_block_rc(&self, hash: &Hash) -> Result<RcEntry, Error> { - Ok(RcEntry::parse_opt(self.rc.get(hash.as_ref())?)) + Ok(RcEntry::parse_opt(self.rc_table.get(hash.as_ref())?)) } /// Delete an entry in the RC table if it is deletable and the /// deletion time has passed pub(crate) fn clear_deleted_block_rc(&self, hash: &Hash) -> Result<(), Error> { let now = now_msec(); - self.rc.db().transaction(|tx| { - let rcval = RcEntry::parse_opt(tx.get(&self.rc, hash)?); + self.rc_table.db().transaction(|tx| { + let rcval = RcEntry::parse_opt(tx.get(&self.rc_table, hash)?); match rcval { RcEntry::Deletable { at_time } if now > at_time => { - tx.remove(&self.rc, hash)?; + tx.remove(&self.rc_table, hash)?; } _ => (), }; @@ -68,6 +77,58 @@ impl BlockRc { })?; Ok(()) } + + /// Recalculate the reference counter of a block + /// to fix potential inconsistencies + pub fn recalculate_rc(&self, hash: &Hash) -> Result<(usize, bool), Error> { + if let Some(recalc_fns) = self.recalc_rc.load().as_ref() { + trace!("Repair block RC for {:?}", hash); + let res = self + .rc_table + .db() + .transaction(|tx| { + let mut cnt = 0; + for f in recalc_fns.iter() { + cnt += f(&tx, hash)?; + } + let old_rc = RcEntry::parse_opt(tx.get(&self.rc_table, hash)?); + trace!( + "Block RC for {:?}: stored={}, calculated={}", + hash, + old_rc.as_u64(), + cnt + ); + if cnt as u64 != old_rc.as_u64() { + warn!( + "Fixing inconsistent block RC for {:?}: was {}, should be {}", + hash, + old_rc.as_u64(), + cnt + ); + let new_rc = if cnt > 0 { + RcEntry::Present { count: cnt as u64 } + } else { + RcEntry::Deletable { + at_time: now_msec() + BLOCK_GC_DELAY.as_millis() as u64, + } + }; + tx.insert(&self.rc_table, hash, new_rc.serialize().unwrap())?; + Ok((cnt, true)) + } else { + Ok((cnt, false)) + } + }) + .map_err(Error::from); + if let Err(e) = &res { + error!("Failed to fix RC for block {:?}: {}", hash, e); + } + res + } else { + Err(Error::Message( + "Block RC recalculation is not available at this point".into(), + )) + } + } } /// Describes the state of the reference counter for a block diff --git a/src/block/repair.rs b/src/block/repair.rs index 2c8acbc9..ef271094 100644 --- a/src/block/repair.rs +++ b/src/block/repair.rs @@ -107,7 +107,7 @@ impl Worker for RepairWorker { for entry in self .manager .rc - .rc + .rc_table .range::<&[u8], _>((start_bound, Bound::Unbounded))? { let (hash, _) = entry?; diff --git a/src/block/resync.rs b/src/block/resync.rs index 7221b093..ab4604ad 100644 --- a/src/block/resync.rs +++ b/src/block/resync.rs @@ -15,7 +15,6 @@ use opentelemetry::{ }; use garage_db as db; -use garage_db::counted_tree_hack::CountedTree; use garage_util::background::*; use garage_util::data::*; @@ -47,9 +46,9 @@ pub(crate) const MAX_RESYNC_WORKERS: usize = 8; const INITIAL_RESYNC_TRANQUILITY: u32 = 2; pub struct BlockResyncManager { - pub(crate) queue: CountedTree, + pub(crate) queue: db::Tree, pub(crate) notify: Arc<Notify>, - pub(crate) errors: CountedTree, + pub(crate) errors: db::Tree, busy_set: BusySet, @@ -90,12 +89,10 @@ impl BlockResyncManager { let queue = db .open_tree("block_local_resync_queue") .expect("Unable to open block_local_resync_queue tree"); - let queue = CountedTree::new(queue).expect("Could not count block_local_resync_queue"); let errors = db .open_tree("block_local_resync_errors") .expect("Unable to open block_local_resync_errors tree"); - let errors = CountedTree::new(errors).expect("Could not count block_local_resync_errors"); let persister = PersisterShared::new(&system.metadata_dir, "resync_cfg"); @@ -110,16 +107,12 @@ impl BlockResyncManager { /// Get lenght of resync queue pub fn queue_len(&self) -> Result<usize, Error> { - // This currently can't return an error because the CountedTree hack - // doesn't error on .len(), but this will change when we remove the hack - // (hopefully someday!) - Ok(self.queue.len()) + Ok(self.queue.len()?) } /// Get number of blocks that have an error pub fn errors_len(&self) -> Result<usize, Error> { - // (see queue_len comment) - Ok(self.errors.len()) + Ok(self.errors.len()?) } /// Clear the error counter for a block and put it in queue immediately @@ -180,7 +173,7 @@ impl BlockResyncManager { // deleted once the garbage collection delay has passed. // // Here are some explanations on how the resync queue works. - // There are two Sled trees that are used to have information + // There are two db trees that are used to have information // about the status of blocks that need to be resynchronized: // // - resync.queue: a tree that is ordered first by a timestamp @@ -374,10 +367,17 @@ impl BlockResyncManager { } if exists && rc.is_deletable() { + if manager.rc.recalculate_rc(hash)?.0 > 0 { + return Err(Error::Message(format!( + "Refcount for block {:?} was inconsistent, retrying later", + hash + ))); + } + info!("Resync block {:?}: offloading and deleting", hash); let existing_path = existing_path.unwrap(); - let mut who = manager.replication.write_nodes(hash); + let mut who = manager.replication.storage_nodes(hash); if who.len() < manager.replication.write_quorum() { return Err(Error::Message("Not trying to offload block because we don't have a quorum of nodes to write to".to_string())); } @@ -385,7 +385,7 @@ impl BlockResyncManager { let who_needs_resps = manager .system - .rpc + .rpc_helper() .call_many( &manager.endpoint, &who, @@ -431,10 +431,10 @@ impl BlockResyncManager { .with_stream_from_buffer(bytes); manager .system - .rpc + .rpc_helper() .try_call_many( &manager.endpoint, - &need_nodes[..], + &need_nodes, put_block_message, RequestStrategy::with_priority(PRIO_BACKGROUND | PRIO_SECONDARY) .with_quorum(need_nodes.len()), @@ -462,7 +462,15 @@ impl BlockResyncManager { let block_data = manager .rpc_get_raw_block(hash, PRIO_BACKGROUND | PRIO_SECONDARY, None) - .await?; + .await; + if matches!(block_data, Err(Error::MissingBlock(_))) { + warn!( + "Could not fetch needed block {:?}, no node returned valid data. Checking that refcount is correct.", + hash + ); + manager.rc.recalculate_rc(hash)?; + } + let block_data = block_data?; manager.metrics.resync_recv_counter.add(1); @@ -543,9 +551,9 @@ impl Worker for ResyncWorker { Ok(WorkerState::Idle) } Err(e) => { - // The errors that we have here are only Sled errors + // The errors that we have here are only db errors // We don't really know how to handle them so just ¯\_(ツ)_/¯ - // (there is kind of an assumption that Sled won't error on us, + // (there is kind of an assumption that the db won't error on us, // if it does there is not much we can do -- TODO should we just panic?) // Here we just give the error to the worker manager, // it will print it to the logs and increment a counter |