audron/garage
1
0
Fork 0
forked from Deuxfleurs/garage
Code Issues Pull requests Projects Releases Wiki Activity

Merge branch 'main' into improve-deps

Browse source
This commit is contained in:
Alex 2022-09-07 18:13:27 +02:00
commit 8adc654713
Signed by untrusted user: lx
GPG key ID: 0E496D15096376BE
9 changed files with 725 additions and 528 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.drone.yml
View file

@ -269,6 +269,6 @@ trigger:
--- ---
kind: signature kind: signature
hmac: fa1f98f327abf88486c0c54984287285a4b951efa3776af9dd33b4d782b50815 hmac: 362639b4c9541ad9bd06ff7f72b5235b2b0216bcb16eececd25285b6fe94ba6f
... ...

1
src/block/lib.rs
View file

@ -3,6 +3,7 @@ extern crate tracing;
pub mod manager; pub mod manager;
pub mod repair; pub mod repair;
pub mod resync;
mod block; mod block;
mod metrics; mod metrics;

567
src/block/manager.rs
View file

@ -1,32 +1,19 @@
use std::convert::TryInto;
use std::path::PathBuf; use std::path::PathBuf;
use std::sync::Arc; use std::sync::Arc;
use std::time::Duration; use std::time::Duration;
use arc_swap::ArcSwapOption;
use async_trait::async_trait; use async_trait::async_trait;
use serde::{Deserialize, Serialize}; use serde::{Deserialize, Serialize};
use futures::future::*;
use tokio::fs; use tokio::fs;
use tokio::io::{AsyncReadExt, AsyncWriteExt}; use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::select; use tokio::sync::{mpsc, Mutex};
use tokio::sync::{mpsc, watch, Mutex, Notify};
use opentelemetry::{
trace::{FutureExt as OtelFutureExt, TraceContextExt, Tracer},
Context, KeyValue,
};
use garage_db as db; use garage_db as db;
use garage_db::counted_tree_hack::CountedTree;
use garage_util::background::*;
use garage_util::data::*; use garage_util::data::*;
use garage_util::error::*; use garage_util::error::*;
use garage_util::metrics::RecordDuration; use garage_util::metrics::RecordDuration;
use garage_util::time::*;
use garage_util::tranquilizer::Tranquilizer;
use garage_rpc::system::System; use garage_rpc::system::System;
use garage_rpc::*; use garage_rpc::*;
@ -37,23 +24,13 @@ use crate::block::*;
use crate::metrics::*; use crate::metrics::*;
use crate::rc::*; use crate::rc::*;
use crate::repair::*; use crate::repair::*;
use crate::resync::*;
/// Size under which data will be stored inlined in database instead of as files /// Size under which data will be stored inlined in database instead of as files
pub const INLINE_THRESHOLD: usize = 3072; pub const INLINE_THRESHOLD: usize = 3072;
// Timeout for RPCs that read and write blocks to remote nodes // Timeout for RPCs that read and write blocks to remote nodes
const BLOCK_RW_TIMEOUT: Duration = Duration::from_secs(30); pub(crate) const BLOCK_RW_TIMEOUT: Duration = Duration::from_secs(30);
// Timeout for RPCs that ask other nodes whether they need a copy
// of a given block before we delete it locally
const NEED_BLOCK_QUERY_TIMEOUT: Duration = Duration::from_secs(5);
// The delay between the time where a resync operation fails
// and the time when it is retried, with exponential backoff
// (multiplied by 2, 4, 8, 16, etc. for every consecutive failure).
const RESYNC_RETRY_DELAY: Duration = Duration::from_secs(60);
// The minimum retry delay is 60 seconds = 1 minute
// The maximum retry delay is 60 seconds * 2^6 = 60 seconds << 6 = 64 minutes (~1 hour)
const RESYNC_RETRY_DELAY_MAX_BACKOFF_POWER: u64 = 6;
// The delay between the moment when the reference counter // The delay between the moment when the reference counter
// drops to zero, and the moment where we allow ourselves // drops to zero, and the moment where we allow ourselves
@ -90,22 +67,18 @@ pub struct BlockManager {
pub data_dir: PathBuf, pub data_dir: PathBuf,
compression_level: Option<i32>, compression_level: Option<i32>,
background_tranquility: u32,
mutation_lock: Mutex<BlockManagerLocked>, mutation_lock: Mutex<BlockManagerLocked>,
pub(crate) rc: BlockRc, pub(crate) rc: BlockRc,
pub resync: BlockResyncManager,
resync_queue: CountedTree,
resync_notify: Notify,
resync_errors: CountedTree,
pub(crate) system: Arc<System>, pub(crate) system: Arc<System>,
endpoint: Arc<Endpoint<BlockRpc, Self>>, pub(crate) endpoint: Arc<Endpoint<BlockRpc, Self>>,
metrics: BlockManagerMetrics, pub(crate) metrics: BlockManagerMetrics,
tx_scrub_command: ArcSwapOption<mpsc::Sender<ScrubWorkerCommand>>, tx_scrub_command: mpsc::Sender<ScrubWorkerCommand>,
} }
// This custom struct contains functions that must only be ran // This custom struct contains functions that must only be ran
@ -113,18 +86,11 @@ pub struct BlockManager {
// it INSIDE a Mutex. // it INSIDE a Mutex.
struct BlockManagerLocked(); struct BlockManagerLocked();
enum ResyncIterResult {
BusyDidSomething,
BusyDidNothing,
IdleFor(Duration),
}
impl BlockManager { impl BlockManager {
pub fn new( pub fn new(
db: &db::Db, db: &db::Db,
data_dir: PathBuf, data_dir: PathBuf,
compression_level: Option<i32>, compression_level: Option<i32>,
background_tranquility: u32,
replication: TableShardedReplication, replication: TableShardedReplication,
system: Arc<System>, system: Arc<System>,
) -> Arc<Self> { ) -> Arc<Self> {
@ -133,17 +99,7 @@ impl BlockManager {
.expect("Unable to open block_local_rc tree"); .expect("Unable to open block_local_rc tree");
let rc = BlockRc::new(rc); let rc = BlockRc::new(rc);
let resync_queue = db let resync = BlockResyncManager::new(db, &system);
.open_tree("block_local_resync_queue")
.expect("Unable to open block_local_resync_queue tree");
let resync_queue =
CountedTree::new(resync_queue).expect("Could not count block_local_resync_queue");
let resync_errors = db
.open_tree("block_local_resync_errors")
.expect("Unable to open block_local_resync_errors tree");
let resync_errors =
CountedTree::new(resync_errors).expect("Could not count block_local_resync_errors");
let endpoint = system let endpoint = system
.netapp .netapp
@ -151,32 +107,39 @@ impl BlockManager {
let manager_locked = BlockManagerLocked(); let manager_locked = BlockManagerLocked();
let metrics = BlockManagerMetrics::new(resync_queue.clone(), resync_errors.clone()); let metrics = BlockManagerMetrics::new(resync.queue.clone(), resync.errors.clone());
let (scrub_tx, scrub_rx) = mpsc::channel(1);
let block_manager = Arc::new(Self { let block_manager = Arc::new(Self {
replication, replication,
data_dir, data_dir,
compression_level, compression_level,
background_tranquility,
mutation_lock: Mutex::new(manager_locked), mutation_lock: Mutex::new(manager_locked),
rc, rc,
resync_queue, resync,
resync_notify: Notify::new(),
resync_errors,
system, system,
endpoint, endpoint,
metrics, metrics,
tx_scrub_command: ArcSwapOption::new(None), tx_scrub_command: scrub_tx,
}); });
block_manager.endpoint.set_handler(block_manager.clone()); block_manager.endpoint.set_handler(block_manager.clone());
block_manager.clone().spawn_background_workers(); // Spawn a bunch of resync workers
for index in 0..MAX_RESYNC_WORKERS {
let worker = ResyncWorker::new(index, block_manager.clone());
block_manager.system.background.spawn_worker(worker);
}
// Spawn scrub worker
let scrub_worker = ScrubWorker::new(block_manager.clone(), scrub_rx);
block_manager.system.background.spawn_worker(scrub_worker);
block_manager block_manager
} }
/// Ask nodes that might have a (possibly compressed) block for it /// Ask nodes that might have a (possibly compressed) block for it
async fn rpc_get_raw_block(&self, hash: &Hash) -> Result<DataBlock, Error> { pub(crate) async fn rpc_get_raw_block(&self, hash: &Hash) -> Result<DataBlock, Error> {
let who = self.replication.read_nodes(hash); let who = self.replication.read_nodes(hash);
let resps = self let resps = self
.system .system
@ -228,20 +191,6 @@ impl BlockManager {
Ok(()) Ok(())
} }
/// Get lenght of resync queue
pub fn resync_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.resync_queue.len())
}
/// Get number of blocks that have an error
pub fn resync_errors_len(&self) -> Result<usize, Error> {
// (see resync_queue_len comment)
Ok(self.resync_errors.len())
}
/// Get number of items in the refcount table /// Get number of items in the refcount table
pub fn rc_len(&self) -> Result<usize, Error> { pub fn rc_len(&self) -> Result<usize, Error> {
Ok(self.rc.rc.len()?) Ok(self.rc.rc.len()?)
@ -249,13 +198,7 @@ impl BlockManager {
/// Send command to start/stop/manager scrub worker /// Send command to start/stop/manager scrub worker
pub async fn send_scrub_command(&self, cmd: ScrubWorkerCommand) { pub async fn send_scrub_command(&self, cmd: ScrubWorkerCommand) {
let _ = self let _ = self.tx_scrub_command.send(cmd).await;
.tx_scrub_command
.load()
.as_ref()
.unwrap()
.send(cmd)
.await;
} }
//// ----- Managing the reference counter ---- //// ----- Managing the reference counter ----
@ -276,7 +219,7 @@ impl BlockManager {
// we will fecth it from someone. // we will fecth it from someone.
let this = self.clone(); let this = self.clone();
tokio::spawn(async move { tokio::spawn(async move {
if let Err(e) = this.put_to_resync(&hash, 2 * BLOCK_RW_TIMEOUT) { if let Err(e) = this.resync.put_to_resync(&hash, 2 * BLOCK_RW_TIMEOUT) {
error!("Block {:?} could not be put in resync queue: {}.", hash, e); error!("Block {:?} could not be put in resync queue: {}.", hash, e);
} }
}); });
@ -298,7 +241,9 @@ impl BlockManager {
// after that delay has passed. // after that delay has passed.
let this = self.clone(); let this = self.clone();
tokio::spawn(async move { tokio::spawn(async move {
if let Err(e) = this.put_to_resync(&hash, BLOCK_GC_DELAY + Duration::from_secs(10)) if let Err(e) = this
.resync
.put_to_resync(&hash, BLOCK_GC_DELAY + Duration::from_secs(10))
{ {
error!("Block {:?} could not be put in resync queue: {}.", hash, e); error!("Block {:?} could not be put in resync queue: {}.", hash, e);
} }
@ -310,7 +255,11 @@ impl BlockManager {
// ---- Reading and writing blocks locally ---- // ---- Reading and writing blocks locally ----
/// Write a block to disk /// Write a block to disk
async fn write_block(&self, hash: &Hash, data: &DataBlock) -> Result<BlockRpc, Error> { pub(crate) async fn write_block(
&self,
hash: &Hash,
data: &DataBlock,
) -> Result<BlockRpc, Error> {
let write_size = data.inner_buffer().len() as u64; let write_size = data.inner_buffer().len() as u64;
let res = self let res = self
@ -346,7 +295,7 @@ impl BlockManager {
Ok(c) => c, Ok(c) => c,
Err(e) => { Err(e) => {
// Not found but maybe we should have had it ?? // Not found but maybe we should have had it ??
self.put_to_resync(hash, 2 * BLOCK_RW_TIMEOUT)?; self.resync.put_to_resync(hash, 2 * BLOCK_RW_TIMEOUT)?;
return Err(Into::into(e)); return Err(Into::into(e));
} }
}; };
@ -373,24 +322,37 @@ impl BlockManager {
.await .await
.move_block_to_corrupted(hash, self) .move_block_to_corrupted(hash, self)
.await?; .await?;
self.put_to_resync(hash, Duration::from_millis(0))?; self.resync.put_to_resync(hash, Duration::from_millis(0))?;
return Err(Error::CorruptData(*hash)); return Err(Error::CorruptData(*hash));
} }
Ok(data) Ok(data)
} }
/// Check if this node should have a block, but don't actually have it /// Check if this node has a block and whether it needs it
async fn need_block(&self, hash: &Hash) -> Result<bool, Error> { pub(crate) async fn check_block_status(&self, hash: &Hash) -> Result<BlockStatus, Error> {
let BlockStatus { exists, needed } = self self.mutation_lock
.mutation_lock
.lock() .lock()
.await .await
.check_block_status(hash, self) .check_block_status(hash, self)
.await?; .await
}
/// Check if this node should have a block, but don't actually have it
async fn need_block(&self, hash: &Hash) -> Result<bool, Error> {
let BlockStatus { exists, needed } = self.check_block_status(hash).await?;
Ok(needed.is_nonzero() && !exists) Ok(needed.is_nonzero() && !exists)
} }
/// Delete block if it is not needed anymore
pub(crate) async fn delete_if_unneeded(&self, hash: &Hash) -> Result<(), Error> {
self.mutation_lock
.lock()
.await
.delete_if_unneeded(hash, self)
.await
}
/// Utility: gives the path of the directory in which a block should be found /// Utility: gives the path of the directory in which a block should be found
fn block_dir(&self, hash: &Hash) -> PathBuf { fn block_dir(&self, hash: &Hash) -> PathBuf {
let mut path = self.data_dir.clone(); let mut path = self.data_dir.clone();
@ -417,305 +379,6 @@ impl BlockManager {
path.set_extension(""); path.set_extension("");
fs::metadata(&path).await.map(|_| false).map_err(Into::into) fs::metadata(&path).await.map(|_| false).map_err(Into::into)
} }
// ---- Resync loop ----
// This part manages a queue of blocks that need to be
// "resynchronized", i.e. that need to have a check that
// they are at present if we need them, or that they are
// 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
// about the status of blocks that need to be resynchronized:
//
// - resync_queue: a tree that is ordered first by a timestamp
// (in milliseconds since Unix epoch) that is the time at which
// the resync must be done, and second by block hash.
// The key in this tree is just:
// concat(timestamp (8 bytes), hash (32 bytes))
// The value is the same 32-byte hash.
//
// - resync_errors: a tree that indicates for each block
// if the last resync resulted in an error, and if so,
// the following two informations (see the ErrorCounter struct):
// - how many consecutive resync errors for this block?
// - when was the last try?
// These two informations are used to implement an
// exponential backoff retry strategy.
// The key in this tree is the 32-byte hash of the block,
// and the value is the encoded ErrorCounter value.
//
// We need to have these two trees, because the resync queue
// is not just a queue of items to process, but a set of items
// that are waiting a specific delay until we can process them
// (the delay being necessary both internally for the exponential
// backoff strategy, and exposed as a parameter when adding items
// to the queue, e.g. to wait until the GC delay has passed).
// This is why we need one tree ordered by time, and one
// ordered by identifier of item to be processed (block hash).
//
// When the worker wants to process an item it takes from
// resync_queue, it checks in resync_errors that if there is an
// exponential back-off delay to await, it has passed before we
// process the item. If not, the item in the queue is skipped
// (but added back for later processing after the time of the
// delay).
//
// An alternative that would have seemed natural is to
// only add items to resync_queue with a processing time that is
// after the delay, but there are several issues with this:
// - This requires to synchronize updates to resync_queue and
// resync_errors (with the current model, there is only one thread,
// the worker thread, that accesses resync_errors,
// so no need to synchronize) by putting them both in a lock.
// This would mean that block_incref might need to take a lock
// before doing its thing, meaning it has much more chances of
// not completing successfully if something bad happens to Garage.
// Currently Garage is not able to recover from block_incref that
// doesn't complete successfully, because it is necessary to ensure
// the consistency between the state of the block manager and
// information in the BlockRef table.
// - If a resync fails, we put that block in the resync_errors table,
// and also add it back to resync_queue to be processed after
// the exponential back-off delay,
// but maybe the block is already scheduled to be resynced again
// at another time that is before the exponential back-off delay,
// and we have no way to check that easily. This means that
// in all cases, we need to check the resync_errors table
// in the resync loop at the time when a block is popped from
// the resync_queue.
// Overall, the current design is therefore simpler and more robust
// because it tolerates inconsistencies between the resync_queue
// and resync_errors table (items being scheduled in resync_queue
// for times that are earlier than the exponential back-off delay
// is a natural condition that is handled properly).
fn spawn_background_workers(self: Arc<Self>) {
// Launch a background workers for background resync loop processing
let background = self.system.background.clone();
let worker = ResyncWorker::new(self.clone());
tokio::spawn(async move {
tokio::time::sleep(Duration::from_secs(10)).await;
background.spawn_worker(worker);
});
// Launch a background worker for data store scrubs
let (scrub_tx, scrub_rx) = mpsc::channel(1);
self.tx_scrub_command.store(Some(Arc::new(scrub_tx)));
let scrub_worker = ScrubWorker::new(self.clone(), scrub_rx);
self.system.background.spawn_worker(scrub_worker);
}
pub(crate) fn put_to_resync(&self, hash: &Hash, delay: Duration) -> db::Result<()> {
let when = now_msec() + delay.as_millis() as u64;
self.put_to_resync_at(hash, when)
}
fn put_to_resync_at(&self, hash: &Hash, when: u64) -> db::Result<()> {
trace!("Put resync_queue: {} {:?}", when, hash);
let mut key = u64::to_be_bytes(when).to_vec();
key.extend(hash.as_ref());
self.resync_queue.insert(key, hash.as_ref())?;
self.resync_notify.notify_waiters();
Ok(())
}
async fn resync_iter(&self) -> Result<ResyncIterResult, db::Error> {
if let Some((time_bytes, hash_bytes)) = self.resync_queue.first()? {
let time_msec = u64::from_be_bytes(time_bytes[0..8].try_into().unwrap());
let now = now_msec();
if now >= time_msec {
let hash = Hash::try_from(&hash_bytes[..]).unwrap();
if let Some(ec) = self.resync_errors.get(hash.as_slice())? {
let ec = ErrorCounter::decode(&ec);
if now < ec.next_try() {
// if next retry after an error is not yet,
// don't do resync and return early, but still
// make sure the item is still in queue at expected time
self.put_to_resync_at(&hash, ec.next_try())?;
// ec.next_try() > now >= time_msec, so this remove
// is not removing the one we added just above
// (we want to do the remove after the insert to ensure
// that the item is not lost if we crash in-between)
self.resync_queue.remove(time_bytes)?;
return Ok(ResyncIterResult::BusyDidNothing);
}
}
let tracer = opentelemetry::global::tracer("garage");
let trace_id = gen_uuid();
let span = tracer
.span_builder("Resync block")
.with_trace_id(
opentelemetry::trace::TraceId::from_hex(&hex::encode(
&trace_id.as_slice()[..16],
))
.unwrap(),
)
.with_attributes(vec![KeyValue::new("block", format!("{:?}", hash))])
.start(&tracer);
let res = self
.resync_block(&hash)
.with_context(Context::current_with_span(span))
.bound_record_duration(&self.metrics.resync_duration)
.await;
self.metrics.resync_counter.add(1);
if let Err(e) = &res {
self.metrics.resync_error_counter.add(1);
warn!("Error when resyncing {:?}: {}", hash, e);
let err_counter = match self.resync_errors.get(hash.as_slice())? {
Some(ec) => ErrorCounter::decode(&ec).add1(now + 1),
None => ErrorCounter::new(now + 1),
};
self.resync_errors
.insert(hash.as_slice(), err_counter.encode())?;
self.put_to_resync_at(&hash, err_counter.next_try())?;
// err_counter.next_try() >= now + 1 > now,
// the entry we remove from the queue is not
// the entry we inserted with put_to_resync_at
self.resync_queue.remove(time_bytes)?;
} else {
self.resync_errors.remove(hash.as_slice())?;
self.resync_queue.remove(time_bytes)?;
}
Ok(ResyncIterResult::BusyDidSomething)
} else {
Ok(ResyncIterResult::IdleFor(Duration::from_millis(
time_msec - now,
)))
}
} else {
// Here we wait either for a notification that an item has been
// added to the queue, or for a constant delay of 10 secs to expire.
// The delay avoids a race condition where the notification happens
// between the time we checked the queue and the first poll
// to resync_notify.notified(): if that happens, we'll just loop
// back 10 seconds later, which is fine.
Ok(ResyncIterResult::IdleFor(Duration::from_secs(10)))
}
}
async fn resync_block(&self, hash: &Hash) -> Result<(), Error> {
let BlockStatus { exists, needed } = self
.mutation_lock
.lock()
.await
.check_block_status(hash, self)
.await?;
if exists != needed.is_needed() || exists != needed.is_nonzero() {
debug!(
"Resync block {:?}: exists {}, nonzero rc {}, deletable {}",
hash,
exists,
needed.is_nonzero(),
needed.is_deletable(),
);
}
if exists && needed.is_deletable() {
info!("Resync block {:?}: offloading and deleting", hash);
let mut who = self.replication.write_nodes(hash);
if who.len() < self.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()));
}
who.retain(|id| *id != self.system.id);
let msg = Arc::new(BlockRpc::NeedBlockQuery(*hash));
let who_needs_fut = who.iter().map(|to| {
self.system.rpc.call_arc(
&self.endpoint,
*to,
msg.clone(),
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_timeout(NEED_BLOCK_QUERY_TIMEOUT),
)
});
let who_needs_resps = join_all(who_needs_fut).await;
let mut need_nodes = vec![];
for (node, needed) in who.iter().zip(who_needs_resps.into_iter()) {
match needed.err_context("NeedBlockQuery RPC")? {
BlockRpc::NeedBlockReply(needed) => {
if needed {
need_nodes.push(*node);
}
}
m => {
return Err(Error::unexpected_rpc_message(m));
}
}
}
if !need_nodes.is_empty() {
trace!(
"Block {:?} needed by {} nodes, sending",
hash,
need_nodes.len()
);
for node in need_nodes.iter() {
self.metrics
.resync_send_counter
.add(1, &[KeyValue::new("to", format!("{:?}", node))]);
}
let put_block_message = self.read_block(hash).await?;
self.system
.rpc
.try_call_many(
&self.endpoint,
&need_nodes[..],
put_block_message,
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_quorum(need_nodes.len())
.with_timeout(BLOCK_RW_TIMEOUT),
)
.await
.err_context("PutBlock RPC")?;
}
info!(
"Deleting unneeded block {:?}, offload finished ({} / {})",
hash,
need_nodes.len(),
who.len()
);
self.mutation_lock
.lock()
.await
.delete_if_unneeded(hash, self)
.await?;
self.rc.clear_deleted_block_rc(hash)?;
}
if needed.is_nonzero() && !exists {
info!(
"Resync block {:?}: fetching absent but needed block (refcount > 0)",
hash
);
let block_data = self.rpc_get_raw_block(hash).await?;
self.metrics.resync_recv_counter.add(1);
self.write_block(hash, &block_data).await?;
}
Ok(())
}
} }
#[async_trait] #[async_trait]
@ -734,80 +397,9 @@ impl EndpointHandler<BlockRpc> for BlockManager {
} }
} }
struct ResyncWorker { pub(crate) struct BlockStatus {
manager: Arc<BlockManager>, pub(crate) exists: bool,
tranquilizer: Tranquilizer, pub(crate) needed: RcEntry,
next_delay: Duration,
}
impl ResyncWorker {
fn new(manager: Arc<BlockManager>) -> Self {
Self {
manager,
tranquilizer: Tranquilizer::new(30),
next_delay: Duration::from_secs(10),
}
}
}
#[async_trait]
impl Worker for ResyncWorker {
fn name(&self) -> String {
"Block resync worker".into()
}
fn info(&self) -> Option<String> {
let mut ret = vec![];
let qlen = self.manager.resync_queue_len().unwrap_or(0);
let elen = self.manager.resync_errors_len().unwrap_or(0);
if qlen > 0 {
ret.push(format!("{} blocks in queue", qlen));
}
if elen > 0 {
ret.push(format!("{} blocks in error state", elen));
}
if !ret.is_empty() {
Some(ret.join(", "))
} else {
None
}
}
async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
self.tranquilizer.reset();
match self.manager.resync_iter().await {
Ok(ResyncIterResult::BusyDidSomething) => Ok(self
.tranquilizer
.tranquilize_worker(self.manager.background_tranquility)),
Ok(ResyncIterResult::BusyDidNothing) => Ok(WorkerState::Busy),
Ok(ResyncIterResult::IdleFor(delay)) => {
self.next_delay = delay;
Ok(WorkerState::Idle)
}
Err(e) => {
// The errors that we have here are only Sled 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,
// 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
Err(e.into())
}
}
}
async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
select! {
_ = tokio::time::sleep(self.next_delay) => (),
_ = self.manager.resync_notify.notified() => (),
};
WorkerState::Busy
}
}
struct BlockStatus {
exists: bool,
needed: RcEntry,
} }
impl BlockManagerLocked { impl BlockManagerLocked {
@ -912,50 +504,3 @@ impl BlockManagerLocked {
Ok(()) Ok(())
} }
} }
/// Counts the number of errors when resyncing a block,
/// and the time of the last try.
/// Used to implement exponential backoff.
#[derive(Clone, Copy, Debug)]
struct ErrorCounter {
errors: u64,
last_try: u64,
}
impl ErrorCounter {
fn new(now: u64) -> Self {
Self {
errors: 1,
last_try: now,
}
}
fn decode(data: &[u8]) -> Self {
Self {
errors: u64::from_be_bytes(data[0..8].try_into().unwrap()),
last_try: u64::from_be_bytes(data[8..16].try_into().unwrap()),
}
}
fn encode(&self) -> Vec<u8> {
[
u64::to_be_bytes(self.errors),
u64::to_be_bytes(self.last_try),
]
.concat()
}
fn add1(self, now: u64) -> Self {
Self {
errors: self.errors + 1,
last_try: now,
}
}
fn delay_msec(&self) -> u64 {
(RESYNC_RETRY_DELAY.as_millis() as u64)
<< std::cmp::min(self.errors - 1, RESYNC_RETRY_DELAY_MAX_BACKOFF_POWER)
}
fn next_try(&self) -> u64 {
self.last_try + self.delay_msec()
}
}

34
src/block/repair.rs
View file

@ -19,7 +19,17 @@ use garage_util::tranquilizer::Tranquilizer;
use crate::manager::*; use crate::manager::*;
const SCRUB_INTERVAL: Duration = Duration::from_secs(3600 * 24 * 30); // full scrub every 30 days // Full scrub every 30 days
const SCRUB_INTERVAL: Duration = Duration::from_secs(3600 * 24 * 30);
// Scrub tranquility is initially set to 4, but can be changed in the CLI
// and the updated version is persisted over Garage restarts
const INITIAL_SCRUB_TRANQUILITY: u32 = 4;
// ---- ---- ----
// FIRST KIND OF REPAIR: FINDING MISSING BLOCKS/USELESS BLOCKS
// This is a one-shot repair operation that can be launched,
// checks everything, and then exits.
// ---- ---- ----
pub struct RepairWorker { pub struct RepairWorker {
manager: Arc<BlockManager>, manager: Arc<BlockManager>,
@ -102,7 +112,9 @@ impl Worker for RepairWorker {
} }
for hash in batch_of_hashes.into_iter() { for hash in batch_of_hashes.into_iter() {
self.manager.put_to_resync(&hash, Duration::from_secs(0))?; self.manager
.resync
.put_to_resync(&hash, Duration::from_secs(0))?;
self.next_start = Some(hash) self.next_start = Some(hash)
} }
@ -114,7 +126,9 @@ impl Worker for RepairWorker {
// This allows us to find blocks we are storing but don't actually need, // This allows us to find blocks we are storing but don't actually need,
// so that we can offload them if necessary and then delete them locally. // so that we can offload them if necessary and then delete them locally.
if let Some(hash) = bi.next().await? { if let Some(hash) = bi.next().await? {
self.manager.put_to_resync(&hash, Duration::from_secs(0))?; self.manager
.resync
.put_to_resync(&hash, Duration::from_secs(0))?;
Ok(WorkerState::Busy) Ok(WorkerState::Busy)
} else { } else {
Ok(WorkerState::Done) Ok(WorkerState::Done)
@ -128,7 +142,13 @@ impl Worker for RepairWorker {
} }
} }
// ---- // ---- ---- ----
// SECOND KIND OF REPAIR: SCRUBBING THE DATASTORE
// This is significantly more complex than the process above,
// as it is a continuously-running task that triggers automatically
// every SCRUB_INTERVAL, but can also be triggered manually
// and whose parameter (esp. speed) can be controlled at runtime.
// ---- ---- ----
pub struct ScrubWorker { pub struct ScrubWorker {
manager: Arc<BlockManager>, manager: Arc<BlockManager>,
@ -176,7 +196,7 @@ impl ScrubWorker {
Ok(v) => v, Ok(v) => v,
Err(_) => ScrubWorkerPersisted { Err(_) => ScrubWorkerPersisted {
time_last_complete_scrub: 0, time_last_complete_scrub: 0,
tranquility: 4, tranquility: INITIAL_SCRUB_TRANQUILITY,
corruptions_detected: 0, corruptions_detected: 0,
}, },
}; };
@ -343,7 +363,9 @@ impl Worker for ScrubWorker {
} }
} }
// ---- // ---- ---- ----
// UTILITY FOR ENUMERATING THE BLOCK STORE
// ---- ---- ----
struct BlockStoreIterator { struct BlockStoreIterator {
path: Vec<ReadingDir>, path: Vec<ReadingDir>,

590
src/block/resync.rs Normal file
View file

@ -0,0 +1,590 @@
use std::collections::HashSet;
use std::convert::TryInto;
use std::sync::{Arc, Mutex};
use std::time::Duration;
use arc_swap::ArcSwap;
use async_trait::async_trait;
use serde::{Deserialize, Serialize};
use futures::future::*;
use tokio::select;
use tokio::sync::{watch, Notify};
use opentelemetry::{
trace::{FutureExt as OtelFutureExt, TraceContextExt, Tracer},
Context, KeyValue,
};
use garage_db as db;
use garage_db::counted_tree_hack::CountedTree;
use garage_util::background::*;
use garage_util::data::*;
use garage_util::error::*;
use garage_util::metrics::RecordDuration;
use garage_util::persister::Persister;
use garage_util::time::*;
use garage_util::tranquilizer::Tranquilizer;
use garage_rpc::system::System;
use garage_rpc::*;
use garage_table::replication::TableReplication;
use crate::manager::*;
// Timeout for RPCs that ask other nodes whether they need a copy
// of a given block before we delete it locally
pub(crate) const NEED_BLOCK_QUERY_TIMEOUT: Duration = Duration::from_secs(5);
// The delay between the time where a resync operation fails
// and the time when it is retried, with exponential backoff
// (multiplied by 2, 4, 8, 16, etc. for every consecutive failure).
pub(crate) const RESYNC_RETRY_DELAY: Duration = Duration::from_secs(60);
// The minimum retry delay is 60 seconds = 1 minute
// The maximum retry delay is 60 seconds * 2^6 = 60 seconds << 6 = 64 minutes (~1 hour)
pub(crate) const RESYNC_RETRY_DELAY_MAX_BACKOFF_POWER: u64 = 6;
// No more than 4 resync workers can be running in the system
pub(crate) const MAX_RESYNC_WORKERS: usize = 4;
// Resync tranquility is initially set to 2, but can be changed in the CLI
// and the updated version is persisted over Garage restarts
const INITIAL_RESYNC_TRANQUILITY: u32 = 2;
pub struct BlockResyncManager {
pub(crate) queue: CountedTree,
pub(crate) notify: Notify,
pub(crate) errors: CountedTree,
busy_set: BusySet,
persister: Persister<ResyncPersistedConfig>,
persisted: ArcSwap<ResyncPersistedConfig>,
}
#[derive(Serialize, Deserialize, Clone, Copy)]
struct ResyncPersistedConfig {
n_workers: usize,
tranquility: u32,
}
enum ResyncIterResult {
BusyDidSomething,
BusyDidNothing,
IdleFor(Duration),
}
type BusySet = Arc<Mutex<HashSet<Vec<u8>>>>;
struct BusyBlock {
time_bytes: Vec<u8>,
hash_bytes: Vec<u8>,
busy_set: BusySet,
}
impl BlockResyncManager {
pub(crate) fn new(db: &db::Db, system: &System) -> Self {
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 = Persister::new(&system.metadata_dir, "resync_cfg");
let persisted = match persister.load() {
Ok(v) => v,
Err(_) => ResyncPersistedConfig {
n_workers: 1,
tranquility: INITIAL_RESYNC_TRANQUILITY,
},
};
Self {
queue,
notify: Notify::new(),
errors,
busy_set: Arc::new(Mutex::new(HashSet::new())),
persister,
persisted: ArcSwap::new(Arc::new(persisted)),
}
}
/// 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())
}
/// Get number of blocks that have an error
pub fn errors_len(&self) -> Result<usize, Error> {
// (see queue_len comment)
Ok(self.errors.len())
}
// ---- Resync loop ----
// This part manages a queue of blocks that need to be
// "resynchronized", i.e. that need to have a check that
// they are at present if we need them, or that they are
// 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
// about the status of blocks that need to be resynchronized:
//
// - resync.queue: a tree that is ordered first by a timestamp
// (in milliseconds since Unix epoch) that is the time at which
// the resync must be done, and second by block hash.
// The key in this tree is just:
// concat(timestamp (8 bytes), hash (32 bytes))
// The value is the same 32-byte hash.
//
// - resync.errors: a tree that indicates for each block
// if the last resync resulted in an error, and if so,
// the following two informations (see the ErrorCounter struct):
// - how many consecutive resync errors for this block?
// - when was the last try?
// These two informations are used to implement an
// exponential backoff retry strategy.
// The key in this tree is the 32-byte hash of the block,
// and the value is the encoded ErrorCounter value.
//
// We need to have these two trees, because the resync queue
// is not just a queue of items to process, but a set of items
// that are waiting a specific delay until we can process them
// (the delay being necessary both internally for the exponential
// backoff strategy, and exposed as a parameter when adding items
// to the queue, e.g. to wait until the GC delay has passed).
// This is why we need one tree ordered by time, and one
// ordered by identifier of item to be processed (block hash).
//
// When the worker wants to process an item it takes from
// resync.queue, it checks in resync.errors that if there is an
// exponential back-off delay to await, it has passed before we
// process the item. If not, the item in the queue is skipped
// (but added back for later processing after the time of the
// delay).
//
// An alternative that would have seemed natural is to
// only add items to resync.queue with a processing time that is
// after the delay, but there are several issues with this:
// - This requires to synchronize updates to resync.queue and
// resync.errors (with the current model, there is only one thread,
// the worker thread, that accesses resync.errors,
// so no need to synchronize) by putting them both in a lock.
// This would mean that block_incref might need to take a lock
// before doing its thing, meaning it has much more chances of
// not completing successfully if something bad happens to Garage.
// Currently Garage is not able to recover from block_incref that
// doesn't complete successfully, because it is necessary to ensure
// the consistency between the state of the block manager and
// information in the BlockRef table.
// - If a resync fails, we put that block in the resync.errors table,
// and also add it back to resync.queue to be processed after
// the exponential back-off delay,
// but maybe the block is already scheduled to be resynced again
// at another time that is before the exponential back-off delay,
// and we have no way to check that easily. This means that
// in all cases, we need to check the resync.errors table
// in the resync loop at the time when a block is popped from
// the resync.queue.
// Overall, the current design is therefore simpler and more robust
// because it tolerates inconsistencies between the resync.queue
// and resync.errors table (items being scheduled in resync.queue
// for times that are earlier than the exponential back-off delay
// is a natural condition that is handled properly).
pub(crate) fn put_to_resync(&self, hash: &Hash, delay: Duration) -> db::Result<()> {
let when = now_msec() + delay.as_millis() as u64;
self.put_to_resync_at(hash, when)
}
pub(crate) fn put_to_resync_at(&self, hash: &Hash, when: u64) -> db::Result<()> {
trace!("Put resync_queue: {} {:?}", when, hash);
let mut key = u64::to_be_bytes(when).to_vec();
key.extend(hash.as_ref());
self.queue.insert(key, hash.as_ref())?;
self.notify.notify_waiters();
Ok(())
}
async fn resync_iter(&self, manager: &BlockManager) -> Result<ResyncIterResult, db::Error> {
if let Some(block) = self.get_block_to_resync()? {
let time_msec = u64::from_be_bytes(block.time_bytes[0..8].try_into().unwrap());
let now = now_msec();
if now >= time_msec {
let hash = Hash::try_from(&block.hash_bytes[..]).unwrap();
if let Some(ec) = self.errors.get(hash.as_slice())? {
let ec = ErrorCounter::decode(&ec);
if now < ec.next_try() {
// if next retry after an error is not yet,
// don't do resync and return early, but still
// make sure the item is still in queue at expected time
self.put_to_resync_at(&hash, ec.next_try())?;
// ec.next_try() > now >= time_msec, so this remove
// is not removing the one we added just above
// (we want to do the remove after the insert to ensure
// that the item is not lost if we crash in-between)
self.queue.remove(&block.time_bytes)?;
return Ok(ResyncIterResult::BusyDidNothing);
}
}
let tracer = opentelemetry::global::tracer("garage");
let trace_id = gen_uuid();
let span = tracer
.span_builder("Resync block")
.with_trace_id(
opentelemetry::trace::TraceId::from_hex(&hex::encode(
&trace_id.as_slice()[..16],
))
.unwrap(),
)
.with_attributes(vec![KeyValue::new("block", format!("{:?}", hash))])
.start(&tracer);
let res = self
.resync_block(manager, &hash)
.with_context(Context::current_with_span(span))
.bound_record_duration(&manager.metrics.resync_duration)
.await;
manager.metrics.resync_counter.add(1);
if let Err(e) = &res {
manager.metrics.resync_error_counter.add(1);
warn!("Error when resyncing {:?}: {}", hash, e);
let err_counter = match self.errors.get(hash.as_slice())? {
Some(ec) => ErrorCounter::decode(&ec).add1(now + 1),
None => ErrorCounter::new(now + 1),
};
self.errors.insert(hash.as_slice(), err_counter.encode())?;
self.put_to_resync_at(&hash, err_counter.next_try())?;
// err_counter.next_try() >= now + 1 > now,
// the entry we remove from the queue is not
// the entry we inserted with put_to_resync_at
self.queue.remove(&block.time_bytes)?;
} else {
self.errors.remove(hash.as_slice())?;
self.queue.remove(&block.time_bytes)?;
}
Ok(ResyncIterResult::BusyDidSomething)
} else {
Ok(ResyncIterResult::IdleFor(Duration::from_millis(
time_msec - now,
)))
}
} else {
// Here we wait either for a notification that an item has been
// added to the queue, or for a constant delay of 10 secs to expire.
// The delay avoids a race condition where the notification happens
// between the time we checked the queue and the first poll
// to resync_notify.notified(): if that happens, we'll just loop
// back 10 seconds later, which is fine.
Ok(ResyncIterResult::IdleFor(Duration::from_secs(10)))
}
}
fn get_block_to_resync(&self) -> Result<Option<BusyBlock>, db::Error> {
let mut busy = self.busy_set.lock().unwrap();
for it in self.queue.iter()? {
let (time_bytes, hash_bytes) = it?;
if !busy.contains(&time_bytes) {
busy.insert(time_bytes.clone());
return Ok(Some(BusyBlock {
time_bytes,
hash_bytes,
busy_set: self.busy_set.clone(),
}));
}
}
Ok(None)
}
async fn resync_block(&self, manager: &BlockManager, hash: &Hash) -> Result<(), Error> {
let BlockStatus { exists, needed } = manager.check_block_status(hash).await?;
if exists != needed.is_needed() || exists != needed.is_nonzero() {
debug!(
"Resync block {:?}: exists {}, nonzero rc {}, deletable {}",
hash,
exists,
needed.is_nonzero(),
needed.is_deletable(),
);
}
if exists && needed.is_deletable() {
info!("Resync block {:?}: offloading and deleting", hash);
let mut who = manager.replication.write_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()));
}
who.retain(|id| *id != manager.system.id);
let msg = Arc::new(BlockRpc::NeedBlockQuery(*hash));
let who_needs_fut = who.iter().map(|to| {
manager.system.rpc.call_arc(
&manager.endpoint,
*to,
msg.clone(),
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_timeout(NEED_BLOCK_QUERY_TIMEOUT),
)
});
let who_needs_resps = join_all(who_needs_fut).await;
let mut need_nodes = vec![];
for (node, needed) in who.iter().zip(who_needs_resps.into_iter()) {
match needed.err_context("NeedBlockQuery RPC")? {
BlockRpc::NeedBlockReply(needed) => {
if needed {
need_nodes.push(*node);
}
}
m => {
return Err(Error::unexpected_rpc_message(m));
}
}
}
if !need_nodes.is_empty() {
trace!(
"Block {:?} needed by {} nodes, sending",
hash,
need_nodes.len()
);
for node in need_nodes.iter() {
manager
.metrics
.resync_send_counter
.add(1, &[KeyValue::new("to", format!("{:?}", node))]);
}
let put_block_message = manager.read_block(hash).await?;
manager
.system
.rpc
.try_call_many(
&manager.endpoint,
&need_nodes[..],
put_block_message,
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_quorum(need_nodes.len())
.with_timeout(BLOCK_RW_TIMEOUT),
)
.await
.err_context("PutBlock RPC")?;
}
info!(
"Deleting unneeded block {:?}, offload finished ({} / {})",
hash,
need_nodes.len(),
who.len()
);
manager.delete_if_unneeded(hash).await?;
manager.rc.clear_deleted_block_rc(hash)?;
}
if needed.is_nonzero() && !exists {
info!(
"Resync block {:?}: fetching absent but needed block (refcount > 0)",
hash
);
let block_data = manager.rpc_get_raw_block(hash).await?;
manager.metrics.resync_recv_counter.add(1);
manager.write_block(hash, &block_data).await?;
}
Ok(())
}
async fn update_persisted(
&self,
update: impl Fn(&mut ResyncPersistedConfig),
) -> Result<(), Error> {
let mut cfg: ResyncPersistedConfig = *self.persisted.load().as_ref();
update(&mut cfg);
self.persister.save_async(&cfg).await?;
self.persisted.store(Arc::new(cfg));
self.notify.notify_waiters();
Ok(())
}
pub async fn set_n_workers(&self, n_workers: usize) -> Result<(), Error> {
if !(1..=MAX_RESYNC_WORKERS).contains(&n_workers) {
return Err(Error::Message(format!(
"Invalid number of resync workers, must be between 1 and {}",
MAX_RESYNC_WORKERS
)));
}
self.update_persisted(|cfg| cfg.n_workers = n_workers).await
}
pub async fn set_tranquility(&self, tranquility: u32) -> Result<(), Error> {
self.update_persisted(|cfg| cfg.tranquility = tranquility)
.await
}
}
impl Drop for BusyBlock {
fn drop(&mut self) {
let mut busy = self.busy_set.lock().unwrap();
busy.remove(&self.time_bytes);
}
}
pub(crate) struct ResyncWorker {
index: usize,
manager: Arc<BlockManager>,
tranquilizer: Tranquilizer,
next_delay: Duration,
}
impl ResyncWorker {
pub(crate) fn new(index: usize, manager: Arc<BlockManager>) -> Self {
Self {
index,
manager,
tranquilizer: Tranquilizer::new(30),
next_delay: Duration::from_secs(10),
}
}
}
#[async_trait]
impl Worker for ResyncWorker {
fn name(&self) -> String {
format!("Block resync worker #{}", self.index + 1)
}
fn info(&self) -> Option<String> {
let persisted = self.manager.resync.persisted.load();
if self.index >= persisted.n_workers {
return Some("(unused)".into());
}
let mut ret = vec![];
ret.push(format!("tranquility = {}", persisted.tranquility));
let qlen = self.manager.resync.queue_len().unwrap_or(0);
if qlen > 0 {
ret.push(format!("{} blocks in queue", qlen));
}
let elen = self.manager.resync.errors_len().unwrap_or(0);
if elen > 0 {
ret.push(format!("{} blocks in error state", elen));
}
Some(ret.join(", "))
}
async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
if self.index >= self.manager.resync.persisted.load().n_workers {
return Ok(WorkerState::Idle);
}
self.tranquilizer.reset();
match self.manager.resync.resync_iter(&self.manager).await {
Ok(ResyncIterResult::BusyDidSomething) => Ok(self
.tranquilizer
.tranquilize_worker(self.manager.resync.persisted.load().tranquility)),
Ok(ResyncIterResult::BusyDidNothing) => Ok(WorkerState::Busy),
Ok(ResyncIterResult::IdleFor(delay)) => {
self.next_delay = delay;
Ok(WorkerState::Idle)
}
Err(e) => {
// The errors that we have here are only Sled 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,
// 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
Err(e.into())
}
}
}
async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
while self.index >= self.manager.resync.persisted.load().n_workers {
self.manager.resync.notify.notified().await
}
select! {
_ = tokio::time::sleep(self.next_delay) => (),
_ = self.manager.resync.notify.notified() => (),
};
WorkerState::Busy
}
}
/// Counts the number of errors when resyncing a block,
/// and the time of the last try.
/// Used to implement exponential backoff.
#[derive(Clone, Copy, Debug)]
struct ErrorCounter {
errors: u64,
last_try: u64,
}
impl ErrorCounter {
fn new(now: u64) -> Self {
Self {
errors: 1,
last_try: now,
}
}
fn decode(data: &[u8]) -> Self {
Self {
errors: u64::from_be_bytes(data[0..8].try_into().unwrap()),
last_try: u64::from_be_bytes(data[8..16].try_into().unwrap()),
}
}
fn encode(&self) -> Vec<u8> {
[
u64::to_be_bytes(self.errors),
u64::to_be_bytes(self.last_try),
]
.concat()
}
fn add1(self, now: u64) -> Self {
Self {
errors: self.errors + 1,
last_try: now,
}
}
fn delay_msec(&self) -> u64 {
(RESYNC_RETRY_DELAY.as_millis() as u64)
<< std::cmp::min(self.errors - 1, RESYNC_RETRY_DELAY_MAX_BACKOFF_POWER)
}
fn next_try(&self) -> u64 {
self.last_try + self.delay_msec()
}
}

32
src/garage/admin.rs
View file

@ -15,6 +15,8 @@ use garage_table::*;
use garage_rpc::*; use garage_rpc::*;
use garage_block::repair::ScrubWorkerCommand;
use garage_model::bucket_alias_table::*; use garage_model::bucket_alias_table::*;
use garage_model::bucket_table::*; use garage_model::bucket_table::*;
use garage_model::garage::Garage; use garage_model::garage::Garage;
@ -782,13 +784,13 @@ impl AdminRpcHandler {
writeln!( writeln!(
&mut ret, &mut ret,
" resync queue length: {}", " resync queue length: {}",
self.garage.block_manager.resync_queue_len()? self.garage.block_manager.resync.queue_len()?
) )
.unwrap(); .unwrap();
writeln!( writeln!(
&mut ret, &mut ret,
" blocks with resync errors: {}", " blocks with resync errors: {}",
self.garage.block_manager.resync_errors_len()? self.garage.block_manager.resync.errors_len()?
) )
.unwrap(); .unwrap();
@ -839,6 +841,32 @@ impl AdminRpcHandler {
let workers = self.garage.background.get_worker_info(); let workers = self.garage.background.get_worker_info();
Ok(AdminRpc::WorkerList(workers, opt)) Ok(AdminRpc::WorkerList(workers, opt))
} }
WorkerCmd::Set { opt } => match opt {
WorkerSetCmd::ScrubTranquility { tranquility } => {
let scrub_command = ScrubWorkerCommand::SetTranquility(tranquility);
self.garage
.block_manager
.send_scrub_command(scrub_command)
.await;
Ok(AdminRpc::Ok("Scrub tranquility updated".into()))
}
WorkerSetCmd::ResyncNWorkers { n_workers } => {
self.garage
.block_manager
.resync
.set_n_workers(n_workers)
.await?;
Ok(AdminRpc::Ok("Number of resync workers updated".into()))
}
WorkerSetCmd::ResyncTranquility { tranquility } => {
self.garage
.block_manager
.resync
.set_tranquility(tranquility)
.await?;
Ok(AdminRpc::Ok("Resync tranquility updated".into()))
}
},
} }
} }
} }

19
src/garage/cli/structs.rs
View file

@ -501,6 +501,12 @@ pub enum WorkerCmd {
#[structopt(flatten)] #[structopt(flatten)]
opt: WorkerListOpt, opt: WorkerListOpt,
}, },
/// Set worker parameter
#[structopt(name = "set", version = version::garage())]
Set {
#[structopt(subcommand)]
opt: WorkerSetCmd,
},
} }
#[derive(Serialize, Deserialize, StructOpt, Debug, Eq, PartialEq, Clone, Copy)] #[derive(Serialize, Deserialize, StructOpt, Debug, Eq, PartialEq, Clone, Copy)]
@ -512,3 +518,16 @@ pub struct WorkerListOpt {
#[structopt(short = "e", long = "errors")] #[structopt(short = "e", long = "errors")]
pub errors: bool, pub errors: bool,
} }
#[derive(Serialize, Deserialize, StructOpt, Debug, Eq, PartialEq, Clone)]
pub enum WorkerSetCmd {
/// Set tranquility of scrub operations
#[structopt(name = "scrub-tranquility", version = version::garage())]
ScrubTranquility { tranquility: u32 },
/// Set number of concurrent block resync workers
#[structopt(name = "resync-n-workers", version = version::garage())]
ResyncNWorkers { n_workers: usize },
/// Set tranquility of block resync operations
#[structopt(name = "resync-tranquility", version = version::garage())]
ResyncTranquility { tranquility: u32 },
}

1
src/model/garage.rs
View file

@ -185,7 +185,6 @@ impl Garage {
&db, &db,
config.data_dir.clone(), config.data_dir.clone(),
config.compression_level, config.compression_level,
config.block_manager_background_tranquility,
data_rep_param, data_rep_param,
system.clone(), system.clone(),
); );

7
src/util/config.rs
View file

@ -23,10 +23,6 @@ pub struct Config {
#[serde(default = "default_block_size")] #[serde(default = "default_block_size")]
pub block_size: usize, pub block_size: usize,
/// Size of data blocks to save to disk
#[serde(default = "default_block_manager_background_tranquility")]
pub block_manager_background_tranquility: u32,
/// Replication mode. Supported values: /// Replication mode. Supported values:
/// - none, 1 -> no replication /// - none, 1 -> no replication
/// - 2 -> 2-way replication /// - 2 -> 2-way replication
@ -145,9 +141,6 @@ fn default_sled_flush_every_ms() -> u64 {
fn default_block_size() -> usize { fn default_block_size() -> usize {
1048576 1048576
} }
fn default_block_manager_background_tranquility() -> u32 {
2
}
/// Read and parse configuration /// Read and parse configuration
pub fn read_config(config_file: PathBuf) -> Result<Config, Error> { pub fn read_config(config_file: PathBuf) -> Result<Config, Error> {