garage/src/model/block.rs

507 lines
13 KiB
Rust
Raw Normal View History

2020-04-09 21:45:07 +00:00
use std::path::PathBuf;
use std::sync::Arc;
use std::time::Duration;
2020-04-09 21:45:07 +00:00
use arc_swap::ArcSwapOption;
use futures::future::*;
use futures::select;
use futures::stream::*;
2020-04-18 17:21:34 +00:00
use serde::{Deserialize, Serialize};
2020-04-09 21:45:07 +00:00
use tokio::fs;
use tokio::prelude::*;
use tokio::sync::{watch, Mutex, Notify};
2020-04-09 21:45:07 +00:00
2020-04-24 10:10:01 +00:00
use garage_util::data;
use garage_util::data::*;
use garage_util::error::Error;
2020-04-24 10:10:01 +00:00
use garage_rpc::membership::System;
use garage_rpc::rpc_client::*;
use garage_rpc::rpc_server::*;
2020-04-23 17:05:46 +00:00
2021-03-11 15:54:15 +00:00
use garage_table::replication::{sharded::TableShardedReplication, TableReplication};
2020-04-23 17:05:46 +00:00
2020-04-24 10:10:01 +00:00
use crate::block_ref_table::*;
2020-04-23 17:05:46 +00:00
2020-04-24 10:10:01 +00:00
use crate::garage::Garage;
2020-04-11 21:00:26 +00:00
pub const INLINE_THRESHOLD: usize = 3072;
const BLOCK_RW_TIMEOUT: Duration = Duration::from_secs(42);
const NEED_BLOCK_QUERY_TIMEOUT: Duration = Duration::from_secs(5);
const RESYNC_RETRY_TIMEOUT: Duration = Duration::from_secs(10);
2020-04-18 17:21:34 +00:00
#[derive(Debug, Serialize, Deserialize)]
pub enum Message {
Ok,
GetBlock(Hash),
PutBlock(PutBlockMessage),
NeedBlockQuery(Hash),
NeedBlockReply(bool),
}
#[derive(Debug, Serialize, Deserialize)]
pub struct PutBlockMessage {
pub hash: Hash,
#[serde(with = "serde_bytes")]
pub data: Vec<u8>,
}
2020-04-18 17:21:34 +00:00
impl RpcMessage for Message {}
2020-04-11 21:00:26 +00:00
pub struct BlockManager {
2020-04-23 17:05:46 +00:00
pub replication: TableShardedReplication,
2020-04-11 21:00:26 +00:00
pub data_dir: PathBuf,
pub data_dir_lock: Mutex<()>,
2020-04-11 21:00:26 +00:00
pub rc: sled::Tree,
pub resync_queue: sled::Tree,
pub resync_notify: Notify,
pub system: Arc<System>,
2020-04-18 17:21:34 +00:00
rpc_client: Arc<RpcClient<Message>>,
pub garage: ArcSwapOption<Garage>,
2020-04-09 21:45:07 +00:00
}
2020-04-11 21:00:26 +00:00
impl BlockManager {
2020-04-18 17:21:34 +00:00
pub fn new(
db: &sled::Db,
data_dir: PathBuf,
2020-04-23 17:05:46 +00:00
replication: TableShardedReplication,
2020-04-18 17:21:34 +00:00
system: Arc<System>,
rpc_server: &mut RpcServer,
) -> Arc<Self> {
let rc = db
.open_tree("block_local_rc")
2020-04-11 21:00:26 +00:00
.expect("Unable to open block_local_rc tree");
rc.set_merge_operator(rc_merge);
let resync_queue = db
.open_tree("block_local_resync_queue")
.expect("Unable to open block_local_resync_queue tree");
2020-04-18 17:21:34 +00:00
let rpc_path = "block_manager";
let rpc_client = system.rpc_client::<Message>(rpc_path);
let block_manager = Arc::new(Self {
2020-04-23 17:05:46 +00:00
replication,
data_dir,
data_dir_lock: Mutex::new(()),
2020-04-11 21:00:26 +00:00
rc,
resync_queue,
resync_notify: Notify::new(),
system,
2020-04-18 17:21:34 +00:00
rpc_client,
garage: ArcSwapOption::from(None),
2020-04-18 17:21:34 +00:00
});
block_manager
.clone()
.register_handler(rpc_server, rpc_path.into());
block_manager
}
fn register_handler(self: Arc<Self>, rpc_server: &mut RpcServer, path: String) {
let self2 = self.clone();
2020-04-18 17:21:34 +00:00
rpc_server.add_handler::<Message, _, _>(path, move |msg, _addr| {
let self2 = self2.clone();
async move { self2.handle(&msg).await }
2020-04-18 17:21:34 +00:00
});
let self2 = self.clone();
self.rpc_client
.set_local_handler(self.system.id, move |msg| {
let self2 = self2.clone();
async move { self2.handle(&msg).await }
});
}
async fn handle(self: Arc<Self>, msg: &Message) -> Result<Message, Error> {
match msg {
Message::PutBlock(m) => self.write_block(&m.hash, &m.data).await,
Message::GetBlock(h) => self.read_block(h).await,
Message::NeedBlockQuery(h) => self.need_block(h).await.map(Message::NeedBlockReply),
2020-11-08 14:04:30 +00:00
_ => Err(Error::BadRPC(format!("Unexpected RPC message"))),
}
}
pub fn spawn_background_worker(self: Arc<Self>) {
// Launch 2 simultaneous workers for background resync loop preprocessing
for i in 0..2u64 {
let bm2 = self.clone();
let background = self.system.background.clone();
tokio::spawn(async move {
tokio::time::delay_for(Duration::from_secs(10 * (i + 1))).await;
background.spawn_worker(format!("block resync worker {}", i), move |must_exit| {
bm2.resync_loop(must_exit)
});
});
}
2020-04-11 21:00:26 +00:00
}
pub async fn write_block(&self, hash: &Hash, data: &[u8]) -> Result<Message, Error> {
let _lock = self.data_dir_lock.lock().await;
2020-04-11 21:00:26 +00:00
let mut path = self.block_dir(hash);
fs::create_dir_all(&path).await?;
2020-04-09 21:45:07 +00:00
2020-04-11 21:00:26 +00:00
path.push(hex::encode(hash));
if fs::metadata(&path).await.is_ok() {
return Ok(Message::Ok);
}
2020-04-09 21:45:07 +00:00
2020-04-11 21:00:26 +00:00
let mut f = fs::File::create(path).await?;
f.write_all(data).await?;
drop(f);
Ok(Message::Ok)
2020-04-09 21:45:07 +00:00
}
2020-04-11 21:00:26 +00:00
pub async fn read_block(&self, hash: &Hash) -> Result<Message, Error> {
2020-04-17 17:20:17 +00:00
let path = self.block_path(hash);
2020-04-09 21:45:07 +00:00
2020-04-17 16:38:11 +00:00
let mut f = match fs::File::open(&path).await {
Ok(f) => f,
Err(e) => {
// Not found but maybe we should have had it ??
2020-04-17 19:08:43 +00:00
self.put_to_resync(hash, 0)?;
2020-04-17 16:38:11 +00:00
return Err(Into::into(e));
}
};
2020-04-11 21:00:26 +00:00
let mut data = vec![];
f.read_to_end(&mut data).await?;
drop(f);
if data::blake2sum(&data[..]) != *hash {
let _lock = self.data_dir_lock.lock().await;
2020-04-21 12:54:55 +00:00
warn!("Block {:?} is corrupted. Deleting and resyncing.", hash);
fs::remove_file(path).await?;
self.put_to_resync(&hash, 0)?;
return Err(Error::CorruptData(*hash));
}
2020-04-11 21:00:26 +00:00
Ok(Message::PutBlock(PutBlockMessage { hash: *hash, data }))
2020-04-11 21:00:26 +00:00
}
2020-04-09 21:45:07 +00:00
pub async fn need_block(&self, hash: &Hash) -> Result<bool, Error> {
let needed = self
.rc
.get(hash.as_ref())?
.map(|x| u64_from_bytes(x.as_ref()) > 0)
.unwrap_or(false);
if needed {
2020-04-17 17:20:17 +00:00
let path = self.block_path(hash);
let exists = fs::metadata(&path).await.is_ok();
Ok(!exists)
} else {
Ok(false)
}
}
2020-04-11 21:00:26 +00:00
fn block_dir(&self, hash: &Hash) -> PathBuf {
let mut path = self.data_dir.clone();
path.push(hex::encode(&hash.as_slice()[0..1]));
path.push(hex::encode(&hash.as_slice()[1..2]));
path
}
2020-04-17 17:20:17 +00:00
fn block_path(&self, hash: &Hash) -> PathBuf {
let mut path = self.block_dir(hash);
path.push(hex::encode(hash.as_ref()));
path
}
2020-04-09 21:45:07 +00:00
2020-04-11 21:00:26 +00:00
pub fn block_incref(&self, hash: &Hash) -> Result<(), Error> {
2020-04-17 19:14:06 +00:00
let old_rc = self.rc.get(&hash)?;
self.rc.merge(&hash, vec![1])?;
if old_rc.map(|x| u64_from_bytes(&x[..]) == 0).unwrap_or(true) {
self.put_to_resync(&hash, BLOCK_RW_TIMEOUT.as_millis() as u64)?;
}
2020-04-11 21:00:26 +00:00
Ok(())
}
2020-04-09 21:45:07 +00:00
pub fn block_decref(&self, hash: &Hash) -> Result<(), Error> {
let new_rc = self.rc.merge(&hash, vec![0])?;
2020-04-17 19:14:06 +00:00
if new_rc.map(|x| u64_from_bytes(&x[..]) == 0).unwrap_or(true) {
self.put_to_resync(&hash, 0)?;
}
Ok(())
}
fn put_to_resync(&self, hash: &Hash, delay_millis: u64) -> Result<(), Error> {
let when = now_msec() + delay_millis;
2020-04-21 12:54:55 +00:00
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();
Ok(())
}
async fn resync_loop(
self: Arc<Self>,
mut must_exit: watch::Receiver<bool>,
) -> Result<(), Error> {
let mut n_failures = 0usize;
while !*must_exit.borrow() {
if let Some((time_bytes, hash_bytes)) = self.resync_queue.pop_min()? {
let time_msec = u64_from_bytes(&time_bytes[0..8]);
let now = now_msec();
if now >= time_msec {
let hash = Hash::try_from(&hash_bytes[..]).unwrap();
if let Err(e) = self.resync_iter(&hash).await {
2020-04-21 12:54:55 +00:00
warn!("Failed to resync block {:?}, retrying later: {}", hash, e);
self.put_to_resync(&hash, RESYNC_RETRY_TIMEOUT.as_millis() as u64)?;
n_failures += 1;
if n_failures >= 10 {
warn!("Too many resync failures, throttling.");
tokio::time::delay_for(Duration::from_secs(1)).await;
}
} else {
n_failures = 0;
}
} else {
self.resync_queue.insert(time_bytes, hash_bytes)?;
let delay = tokio::time::delay_for(Duration::from_millis(time_msec - now));
select! {
_ = delay.fuse() => (),
_ = self.resync_notify.notified().fuse() => (),
_ = must_exit.recv().fuse() => (),
}
}
} else {
select! {
_ = self.resync_notify.notified().fuse() => (),
_ = must_exit.recv().fuse() => (),
}
2020-04-11 21:00:26 +00:00
}
}
Ok(())
2020-04-11 21:00:26 +00:00
}
async fn resync_iter(&self, hash: &Hash) -> Result<(), Error> {
2020-04-17 17:20:17 +00:00
let path = self.block_path(hash);
let exists = fs::metadata(&path).await.is_ok();
let needed = self
.rc
.get(hash.as_ref())?
.map(|x| u64_from_bytes(x.as_ref()) > 0)
.unwrap_or(false);
2020-04-21 12:54:55 +00:00
if exists != needed {
info!(
"Resync block {:?}: exists {}, needed {}",
hash, exists, needed
);
}
if exists && !needed {
2021-02-24 10:58:03 +00:00
trace!("Offloading block {:?}", hash);
let ring = self.system.ring.borrow().clone();
let mut who = self.replication.replication_nodes(&hash, &ring);
if who.len() < self.replication.write_quorum(&self.system) {
return Err(Error::Message(format!("Not trying to offload block because we don't have a quorum of nodes to write to")));
}
2021-02-24 10:58:03 +00:00
who.retain(|id| *id != self.system.id);
let msg = Arc::new(Message::NeedBlockQuery(*hash));
let who_needs_fut = who.iter().map(|to| {
self.rpc_client
.call_arc(*to, msg.clone(), 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? {
Message::NeedBlockReply(needed) => {
if needed {
need_nodes.push(*node);
}
}
2021-02-24 10:58:03 +00:00
_ => {
return Err(Error::Message(format!(
"Unexpected response to NeedBlockQuery RPC"
)));
}
}
2021-02-24 10:58:03 +00:00
}
if need_nodes.len() > 0 {
trace!(
"Block {:?} needed by {} nodes, sending",
hash,
need_nodes.len()
);
let put_block_message = self.read_block(hash).await?;
2021-03-12 20:52:19 +00:00
self.rpc_client
.try_call_many(
&need_nodes[..],
put_block_message,
2021-03-12 20:52:19 +00:00
RequestStrategy::with_quorum(need_nodes.len())
.with_timeout(BLOCK_RW_TIMEOUT),
)
.await?;
}
trace!(
"Deleting block {:?}, offload finished ({} / {})",
hash,
need_nodes.len(),
who.len()
);
2021-02-24 10:58:03 +00:00
fs::remove_file(path).await?;
self.resync_queue.remove(&hash)?;
}
if needed && !exists {
// TODO find a way to not do this if they are sending it to us
// Let's suppose this isn't an issue for now with the BLOCK_RW_TIMEOUT delay
// between the RC being incremented and this part being called.
2020-04-18 17:21:34 +00:00
let block_data = self.rpc_get_block(&hash).await?;
self.write_block(hash, &block_data[..]).await?;
}
Ok(())
}
2020-04-18 17:21:34 +00:00
pub async fn rpc_get_block(&self, hash: &Hash) -> Result<Vec<u8>, Error> {
2020-04-23 17:05:46 +00:00
let who = self.replication.read_nodes(&hash, &self.system);
let resps = self
.rpc_client
.try_call_many(
&who[..],
Message::GetBlock(*hash),
RequestStrategy::with_quorum(1)
.with_timeout(BLOCK_RW_TIMEOUT)
.interrupt_after_quorum(true),
)
.await?;
for resp in resps {
if let Message::PutBlock(msg) = resp {
return Ok(msg.data);
2020-04-18 17:21:34 +00:00
}
}
Err(Error::Message(format!(
"Unable to read block {:?}: no valid blocks returned",
hash
)))
}
pub async fn rpc_put_block(&self, hash: Hash, data: Vec<u8>) -> Result<(), Error> {
2020-04-23 17:05:46 +00:00
let who = self.replication.write_nodes(&hash, &self.system);
2020-04-18 17:21:34 +00:00
self.rpc_client
.try_call_many(
&who[..],
Message::PutBlock(PutBlockMessage { hash, data }),
RequestStrategy::with_quorum(self.replication.write_quorum(&self.system))
.with_timeout(BLOCK_RW_TIMEOUT),
2020-04-18 17:21:34 +00:00
)
.await?;
Ok(())
}
2020-04-19 20:36:36 +00:00
pub async fn repair_data_store(&self, must_exit: &watch::Receiver<bool>) -> Result<(), Error> {
2020-04-19 20:36:36 +00:00
// 1. Repair blocks from RC table
let garage = self.garage.load_full().unwrap();
let mut last_hash = None;
let mut i = 0usize;
2021-03-11 15:54:15 +00:00
for entry in garage.block_ref_table.data.store.iter() {
2020-04-19 20:36:36 +00:00
let (_k, v_bytes) = entry?;
let block_ref = rmp_serde::decode::from_read_ref::<_, BlockRef>(v_bytes.as_ref())?;
if Some(&block_ref.block) == last_hash.as_ref() {
continue;
}
if !block_ref.deleted.get() {
last_hash = Some(block_ref.block);
2020-04-19 20:36:36 +00:00
self.put_to_resync(&block_ref.block, 0)?;
}
i += 1;
if i & 0xFF == 0 && *must_exit.borrow() {
return Ok(());
}
}
// 2. Repair blocks actually on disk
self.repair_aux_read_dir_rec(&self.data_dir, must_exit)
.await?;
2020-04-19 20:36:36 +00:00
2021-02-24 10:58:03 +00:00
Ok(())
}
fn repair_aux_read_dir_rec<'a>(
&'a self,
path: &'a PathBuf,
must_exit: &'a watch::Receiver<bool>,
) -> BoxFuture<'a, Result<(), Error>> {
2021-02-25 09:53:33 +00:00
// Lists all blocks on disk and adds them to the resync queue.
// 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.
2021-02-24 10:58:03 +00:00
async move {
let mut ls_data_dir = fs::read_dir(path).await?;
while let Some(data_dir_ent) = ls_data_dir.next().await {
let data_dir_ent = data_dir_ent?;
let name = data_dir_ent.file_name();
let name = match name.into_string() {
2020-04-19 20:36:36 +00:00
Ok(x) => x,
Err(_) => continue,
};
2021-02-24 10:58:03 +00:00
let ent_type = data_dir_ent.file_type().await?;
if name.len() == 2 && hex::decode(&name).is_ok() && ent_type.is_dir() {
self.repair_aux_read_dir_rec(&data_dir_ent.path(), must_exit)
.await?;
2021-02-24 10:58:03 +00:00
} else if name.len() == 64 {
let hash_bytes = match hex::decode(&name) {
Ok(h) => h,
Err(_) => continue,
};
let mut hash = [0u8; 32];
hash.copy_from_slice(&hash_bytes[..]);
self.put_to_resync(&hash.into(), 0)?;
2020-04-19 20:36:36 +00:00
}
if *must_exit.borrow() {
2021-02-24 10:58:03 +00:00
break;
2020-04-19 20:36:36 +00:00
}
}
2021-02-24 10:58:03 +00:00
Ok(())
}
.boxed()
2020-04-19 20:36:36 +00:00
}
}
fn u64_from_bytes(bytes: &[u8]) -> u64 {
assert!(bytes.len() == 8);
let mut x8 = [0u8; 8];
x8.copy_from_slice(bytes);
u64::from_be_bytes(x8)
2020-04-11 21:00:26 +00:00
}
fn rc_merge(_key: &[u8], old: Option<&[u8]>, new: &[u8]) -> Option<Vec<u8>> {
let old = old.map(u64_from_bytes).unwrap_or(0);
2020-04-11 21:00:26 +00:00
assert!(new.len() == 1);
let new = match new[0] {
0 => {
if old > 0 {
old - 1
} else {
0
}
}
1 => old + 1,
_ => unreachable!(),
};
if new == 0 {
None
} else {
Some(u64::to_be_bytes(new).to_vec())
}
2020-04-09 21:45:07 +00:00
}