fix bitrot

This commit is contained in:
weisd
2025-06-18 18:22:07 +08:00
parent 85421657ad
commit 5b1b527851
12 changed files with 177 additions and 129 deletions
+1 -1
View File
@@ -28,7 +28,7 @@ pub async fn create_bitrot_reader(
checksum_algo: HashAlgorithm,
) -> disk::error::Result<Option<BitrotReader<Box<dyn AsyncRead + Send + Sync + Unpin>>>> {
// Calculate the total length to read, including the checksum overhead
let length = offset.div_ceil(shard_size) * checksum_algo.size() + length;
let length = length.div_ceil(shard_size) * checksum_algo.size() + length;
if let Some(data) = inline_data {
// Use inline data
+3 -5
View File
@@ -1399,8 +1399,6 @@ impl DiskAPI for LocalDisk {
#[tracing::instrument(level = "debug", skip(self))]
async fn create_file(&self, origvolume: &str, volume: &str, path: &str, _file_size: i64) -> Result<FileWriter> {
// warn!("disk create_file: origvolume: {}, volume: {}, path: {}", origvolume, volume, path);
if !origvolume.is_empty() {
let origvolume_dir = self.get_bucket_path(origvolume)?;
if !skip_access_checks(origvolume) {
@@ -1431,8 +1429,6 @@ impl DiskAPI for LocalDisk {
#[tracing::instrument(level = "debug", skip(self))]
// async fn append_file(&self, volume: &str, path: &str, mut r: DuplexStream) -> Result<File> {
async fn append_file(&self, volume: &str, path: &str) -> Result<FileWriter> {
warn!("disk append_file: volume: {}, path: {}", volume, path);
let volume_dir = self.get_bucket_path(volume)?;
if !skip_access_checks(volume) {
access(&volume_dir)
@@ -1497,7 +1493,9 @@ impl DiskAPI for LocalDisk {
return Err(DiskError::FileCorrupt);
}
f.seek(SeekFrom::Start(offset as u64)).await?;
if offset > 0 {
f.seek(SeekFrom::Start(offset as u64)).await?;
}
Ok(Box::new(f))
}
+15 -2
View File
@@ -608,7 +608,14 @@ impl DiskAPI for RemoteDisk {
#[tracing::instrument(level = "debug", skip(self))]
async fn read_file_stream(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<FileReader> {
info!("read_file_stream {}/{}/{}", self.endpoint.to_string(), volume, path);
// warn!(
// "disk remote read_file_stream {}/{}/{} offset={} length={}",
// self.endpoint.to_string(),
// volume,
// path,
// offset,
// length
// );
let url = format!(
"{}/rustfs/rpc/read_file_stream?disk={}&volume={}&path={}&offset={}&length={}",
self.endpoint.grid_host(),
@@ -641,7 +648,13 @@ impl DiskAPI for RemoteDisk {
#[tracing::instrument(level = "debug", skip(self))]
async fn create_file(&self, _origvolume: &str, volume: &str, path: &str, file_size: i64) -> Result<FileWriter> {
info!("create_file {}/{}/{}", self.endpoint.to_string(), volume, path);
// warn!(
// "disk remote create_file {}/{}/{} file_size={}",
// self.endpoint.to_string(),
// volume,
// path,
// file_size
// );
let url = format!(
"{}/rustfs/rpc/put_file_stream?disk={}&volume={}&path={}&append={}&size={}",
+44 -37
View File
@@ -1,7 +1,9 @@
use bytes::Bytes;
use pin_project_lite::pin_project;
use rustfs_utils::{HashAlgorithm, read_full, write_all};
use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite};
use rustfs_utils::HashAlgorithm;
use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
use tracing::error;
use uuid::Uuid;
pin_project! {
/// BitrotReader reads (hash+data) blocks from an async reader and verifies hash integrity.
@@ -12,10 +14,11 @@ pin_project! {
shard_size: usize,
buf: Vec<u8>,
hash_buf: Vec<u8>,
hash_read: usize,
data_buf: Vec<u8>,
data_read: usize,
hash_checked: bool,
// hash_read: usize,
// data_buf: Vec<u8>,
// data_read: usize,
// hash_checked: bool,
id: Uuid,
}
}
@@ -32,10 +35,11 @@ where
shard_size,
buf: Vec::new(),
hash_buf: vec![0u8; hash_size],
hash_read: 0,
data_buf: Vec::new(),
data_read: 0,
hash_checked: false,
// hash_read: 0,
// data_buf: Vec::new(),
// data_read: 0,
// hash_checked: false,
id: Uuid::new_v4(),
}
}
@@ -51,30 +55,31 @@ where
let hash_size = self.hash_algo.size();
// Read hash
let mut hash_buf = vec![0u8; hash_size];
if hash_size > 0 {
self.inner.read_exact(&mut hash_buf).await?;
self.inner.read_exact(&mut self.hash_buf).await.map_err(|e| {
error!("bitrot reader read hash error: {}", e);
e
})?;
}
let data_len = read_full(&mut self.inner, out).await?;
// // Read data
// let mut data_len = 0;
// while data_len < out.len() {
// let n = self.inner.read(&mut out[data_len..]).await?;
// if n == 0 {
// break;
// }
// data_len += n;
// // Only read up to one shard_size block
// if data_len >= self.shard_size {
// break;
// }
// }
// Read data
let mut data_len = 0;
while data_len < out.len() {
let n = self.inner.read(&mut out[data_len..]).await.map_err(|e| {
error!("bitrot reader read data error: {}", e);
e
})?;
if n == 0 {
break;
}
data_len += n;
}
if hash_size > 0 {
let actual_hash = self.hash_algo.hash_encode(&out[..data_len]);
if actual_hash.as_ref() != hash_buf.as_slice() {
if actual_hash.as_ref() != self.hash_buf.as_slice() {
error!("bitrot reader hash mismatch, id={} data_len={}, out_len={}", self.id, data_len, out.len());
return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
}
}
@@ -145,22 +150,20 @@ where
self.buf.extend_from_slice(buf);
// Write hash+data in one call
let mut n = write_all(&mut self.inner, &self.buf).await?;
self.inner.write_all(&self.buf).await?;
if n < hash_algo.size() {
return Err(std::io::Error::new(
std::io::ErrorKind::WriteZero,
"short write: not enough bytes written",
));
}
self.inner.flush().await?;
n -= hash_algo.size();
let n = self.buf.len();
self.buf.clear();
Ok(n)
}
pub async fn shutdown(&mut self) -> std::io::Result<()> {
self.inner.shutdown().await
}
}
pub fn bitrot_shard_file_size(size: usize, shard_size: usize, algo: HashAlgorithm) -> usize {
@@ -330,6 +333,10 @@ impl BitrotWriterWrapper {
self.bitrot_writer.write(buf).await
}
pub async fn shutdown(&mut self) -> std::io::Result<()> {
self.bitrot_writer.shutdown().await
}
/// Extract the inline buffer data, consuming the wrapper
pub fn into_inline_data(self) -> Option<Vec<u8>> {
match self.writer_type {
+21 -23
View File
@@ -67,36 +67,34 @@ where
}
// 使用并发读取所有分片
let mut read_futs = Vec::with_capacity(self.readers.len());
let read_futs: Vec<_> = self
.readers
.iter_mut()
.enumerate()
.map(|(i, opt_reader)| {
if let Some(reader) = opt_reader.as_mut() {
for (i, opt_reader) in self.readers.iter_mut().enumerate() {
let future = if let Some(reader) = opt_reader.as_mut() {
Box::pin(async move {
let mut buf = vec![0u8; shard_size];
// 需要move i, buf
Some(async move {
match reader.read(&mut buf).await {
Ok(n) => {
buf.truncate(n);
(i, Ok(buf))
}
Err(e) => (i, Err(Error::from(e))),
match reader.read(&mut buf).await {
Ok(n) => {
buf.truncate(n);
(i, Ok(buf))
}
})
} else {
None
}
})
.collect();
Err(e) => (i, Err(Error::from(e))),
}
}) as std::pin::Pin<Box<dyn std::future::Future<Output = (usize, Result<Vec<u8>, Error>)> + Send>>
} else {
// reader是None时返回FileNotFound错误
Box::pin(async move { (i, Err(Error::FileNotFound)) })
as std::pin::Pin<Box<dyn std::future::Future<Output = (usize, Result<Vec<u8>, Error>)> + Send>>
};
read_futs.push(future);
}
// 过滤掉Nonejoin_all
let mut results = join_all(read_futs.into_iter().flatten()).await;
let results = join_all(read_futs).await;
let mut shards: Vec<Option<Vec<u8>>> = vec![None; self.readers.len()];
let mut errs = vec![None; self.readers.len()];
for (i, shard) in results.drain(..) {
for (i, shard) in results.into_iter() {
match shard {
Ok(data) => {
if !data.is_empty() {
+8 -1
View File
@@ -97,6 +97,13 @@ impl<'a> MultiWriter<'a> {
.join(", ")
)))
}
pub async fn _shutdown(&mut self) -> std::io::Result<()> {
for writer in self.writers.iter_mut().flatten() {
writer.shutdown().await?;
}
Ok(())
}
}
impl Erasure {
@@ -147,7 +154,7 @@ impl Erasure {
}
let (reader, total) = task.await??;
// writers.shutdown().await?;
Ok((reader, total))
}
}
+16 -2
View File
@@ -555,6 +555,13 @@ mod tests {
use super::*;
#[test]
fn test_shard_file_size_cases2() {
let erasure = Erasure::new(12, 4, 1024 * 1024);
assert_eq!(erasure.shard_file_size(1572864), 131074);
}
#[test]
fn test_shard_file_size_cases() {
let erasure = Erasure::new(4, 2, 8);
@@ -577,6 +584,8 @@ mod tests {
assert_eq!(erasure.shard_file_size(1248739), 312186); // 1248739/8=156092, last=3, 3 div_ceil 4=1, 156092*2+1=312185
assert_eq!(erasure.shard_file_size(43), 12); // 43/8=5, last=3, 3 div_ceil 4=1, 5*2+1=11
assert_eq!(erasure.shard_file_size(1572864), 393216); // 43/8=5, last=3, 3 div_ceil 4=1, 5*2+1=11
}
#[test]
@@ -677,9 +686,14 @@ mod tests {
#[test]
fn test_shard_file_offset() {
let erasure = Erasure::new(4, 2, 8);
let offset = erasure.shard_file_offset(0, 16, 32);
let erasure = Erasure::new(8, 8, 1024 * 1024);
let offset = erasure.shard_file_offset(0, 86, 86);
println!("offset={}", offset);
assert!(offset > 0);
let total_length = erasure.shard_file_size(86);
println!("total_length={}", total_length);
assert!(total_length > 0);
}
#[tokio::test]
+2 -2
View File
@@ -243,7 +243,7 @@ impl ECStore {
return Err(err);
}
error!("rebalanceMeta: not found, rebalance not started");
warn!("rebalanceMeta: not found, rebalance not started");
}
}
@@ -501,7 +501,7 @@ impl ECStore {
if let Some(meta) = rebalance_meta.as_mut() {
meta.cancel = Some(tx)
} else {
error!("start_rebalance: rebalance_meta is None exit");
warn!("start_rebalance: rebalance_meta is None exit");
return;
}
-2
View File
@@ -1859,8 +1859,6 @@ impl SetDisks {
let (last_part_index, _) = fi.to_part_offset(end_offset)?;
// let erasure = Erasure::new(fi.erasure.data_blocks, fi.erasure.parity_blocks, fi.erasure.block_size);
let erasure = erasure_coding::Erasure::new(fi.erasure.data_blocks, fi.erasure.parity_blocks, fi.erasure.block_size);
let mut total_readed = 0;