fix(getobject): distinguish downstream output closure (#5220)

fix(getobject): preserve internal broken pipe failures

Co-authored-by: heihutu <heihutu@gmail.com>
This commit is contained in:
houseme
2026-07-25 16:22:20 +08:00
committed by GitHub
parent ffd1b94e1f
commit 9ddb30139d
5 changed files with 235 additions and 38 deletions
+60 -5
View File
@@ -14,8 +14,8 @@
use crate::disk::error::DiskError;
use crate::error::StorageError;
use std::io;
use std::time::Instant;
use std::{error::Error as StdError, fmt, io};
pub(crate) const GET_OBJECT_PATH_CODEC_STREAMING: &str = "codec_streaming";
pub(crate) const GET_OBJECT_PATH_CODEC_STREAMING_LEGACY_ENGINE: &str = "codec_streaming_legacy_engine";
@@ -207,9 +207,42 @@ pub(crate) fn classify_disk_error(err: &DiskError) -> GetObjectFailureReason {
}
}
pub(crate) fn classify_io_error(err: &io::Error) -> GetObjectFailureReason {
#[derive(Debug)]
struct GetObjectDownstreamClosed {
source: io::Error,
}
impl fmt::Display for GetObjectDownstreamClosed {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(f, "GetObject output closed: {}", self.source)
}
}
impl StdError for GetObjectDownstreamClosed {
fn source(&self) -> Option<&(dyn StdError + 'static)> {
Some(&self.source)
}
}
pub(crate) fn mark_get_object_downstream_closed(err: io::Error) -> io::Error {
match err.kind() {
io::ErrorKind::BrokenPipe | io::ErrorKind::ConnectionReset | io::ErrorKind::ConnectionAborted => {
io::Error::new(err.kind(), GetObjectDownstreamClosed { source: err })
}
_ => err,
}
}
pub(crate) fn classify_io_error(err: &io::Error) -> GetObjectFailureReason {
let mut source = err.get_ref().map(|source| source as &(dyn StdError + 'static));
while let Some(error) = source {
if error.is::<GetObjectDownstreamClosed>() {
return GetObjectFailureReason::DownstreamClosed;
}
source = error.source();
}
match err.kind() {
io::ErrorKind::BrokenPipe | io::ErrorKind::ConnectionReset => GetObjectFailureReason::DownstreamClosed,
io::ErrorKind::TimedOut => GetObjectFailureReason::Timeout,
io::ErrorKind::UnexpectedEof => GetObjectFailureReason::ShortRead,
io::ErrorKind::InvalidInput | io::ErrorKind::InvalidData => GetObjectFailureReason::RangeOrLengthInvalid,
@@ -260,6 +293,16 @@ mod tests {
classify_storage_error(&StorageError::InvalidRangeSpec("bad range".to_string())),
GetObjectFailureReason::RangeOrLengthInvalid
);
let internal_broken_pipe = StorageError::Io(io::Error::from(io::ErrorKind::BrokenPipe));
assert_eq!(classify_storage_error(&internal_broken_pipe), GetObjectFailureReason::Io);
let downstream_closed = StorageError::Io(mark_get_object_downstream_closed(io::Error::from(io::ErrorKind::BrokenPipe)));
assert_eq!(
classify_storage_error(&downstream_closed),
GetObjectFailureReason::DownstreamClosed,
"the legacy duplex producer must suppress only explicitly tagged output closes"
);
}
#[test]
@@ -277,8 +320,8 @@ mod tests {
#[test]
fn classifies_io_errors_for_get_pipeline() {
let cases = [
(io::ErrorKind::BrokenPipe, GetObjectFailureReason::DownstreamClosed),
(io::ErrorKind::ConnectionReset, GetObjectFailureReason::DownstreamClosed),
(io::ErrorKind::BrokenPipe, GetObjectFailureReason::Io),
(io::ErrorKind::ConnectionReset, GetObjectFailureReason::Io),
(io::ErrorKind::TimedOut, GetObjectFailureReason::Timeout),
(io::ErrorKind::UnexpectedEof, GetObjectFailureReason::ShortRead),
(io::ErrorKind::InvalidInput, GetObjectFailureReason::RangeOrLengthInvalid),
@@ -292,6 +335,18 @@ mod tests {
}
}
#[test]
fn classifies_marked_output_close_as_downstream_closed() {
for kind in [
io::ErrorKind::BrokenPipe,
io::ErrorKind::ConnectionReset,
io::ErrorKind::ConnectionAborted,
] {
let err = mark_get_object_downstream_closed(io::Error::from(kind));
assert_eq!(classify_io_error(&err), GetObjectFailureReason::DownstreamClosed, "kind={kind:?}");
}
}
#[test]
fn keeps_metric_labels_stable() {
assert_eq!(GetObjectFailureReason::ReadQuorum.as_str(), "read_quorum");
+3
View File
@@ -4446,6 +4446,7 @@ mod tests {
use crate::layout::endpoints::SetupType;
use crate::object_api::BLOCK_SIZE_V2;
use crate::object_api::ObjectInfo;
use crate::set_disk::read::GetObjectOutputKind;
use crate::storage_api_contracts::{
heal::HealOperations as _, lifecycle::TransitionedObject, list::ListOperations as _, multipart::CompletePart,
namespace::NamespaceLocking as _, object::ObjectIO as _, object::ObjectOperations as _,
@@ -8400,6 +8401,7 @@ mod tests {
range_offset,
range_length as i64,
&mut writer,
GetObjectOutputKind::HttpResponse,
fi,
files,
&disks,
@@ -8511,6 +8513,7 @@ mod tests {
0,
total_size as i64,
&mut writer,
GetObjectOutputKind::HttpResponse,
fi,
files,
&disks,
+42 -18
View File
@@ -21,6 +21,7 @@
use super::super::*;
use super::bitrot_self_verify::{BitrotSelfVerifyTarget, drop_failed_writer_disks, verify_written_bitrot_shards};
use crate::set_disk::read::GetObjectOutputKind;
use crate::bucket::lifecycle::{
tier_delete_journal::{persist_tier_delete_journal_entry, remove_tier_delete_journal_entry},
@@ -31,6 +32,7 @@ use crate::bucket::lifecycle::{
save_transition_transaction_record,
},
};
use crate::diagnostics::get::GetObjectFailureReason;
use crate::disk::OldCurrentSize;
use crate::object_api::{GetObjectBodySource, get_object_body_cache_hook_suppressed};
use crate::services::tier::tier::{TierConfigMgr, TierOperationLease};
@@ -522,6 +524,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
0,
object_info.size,
&mut output,
GetObjectOutputKind::Internal,
fi,
files,
&disks,
@@ -650,6 +653,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
offset,
length,
&mut writer,
GetObjectOutputKind::HttpResponse,
fi,
files,
&disks,
@@ -664,24 +668,43 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
.await
{
let reason = classify_storage_error(&e);
record_get_object_pipeline_failure(GET_STAGE_EMIT, reason);
error!(
event = EVENT_SET_DISK_WRITE,
component = LOG_COMPONENT_ECSTORE,
subsystem = LOG_SUBSYSTEM_SET_DISK,
bucket,
object,
pool_index,
set_index,
offset,
requested_length = length,
skip_verify_bitrot = skip_verify,
state = "read_pipeline_failed",
stage = GET_STAGE_EMIT,
reason = reason.as_str(),
error = ?e,
"Set disk object read pipeline failed"
);
if reason == GetObjectFailureReason::DownstreamClosed {
debug!(
event = EVENT_SET_DISK_WRITE,
component = LOG_COMPONENT_ECSTORE,
subsystem = LOG_SUBSYSTEM_SET_DISK,
bucket,
object,
pool_index,
set_index,
offset,
requested_length = length,
state = "downstream_closed",
stage = GET_STAGE_EMIT,
reason = reason.as_str(),
error = ?e,
"Set disk object read pipeline stopped after downstream closed"
);
} else {
record_get_object_pipeline_failure(GET_STAGE_EMIT, reason);
error!(
event = EVENT_SET_DISK_WRITE,
component = LOG_COMPONENT_ECSTORE,
subsystem = LOG_SUBSYSTEM_SET_DISK,
bucket,
object,
pool_index,
set_index,
offset,
requested_length = length,
skip_verify_bitrot = skip_verify,
state = "read_pipeline_failed",
stage = GET_STAGE_EMIT,
reason = reason.as_str(),
error = ?e,
"Set disk object read pipeline failed"
);
}
};
});
@@ -3356,6 +3379,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
0,
cloned_fi.size,
&mut writer,
GetObjectOutputKind::Internal,
cloned_fi,
meta_arr,
&online_disks,
+114 -10
View File
@@ -34,8 +34,8 @@ use crate::diagnostics::get::{
GET_STAGE_METADATA_RESOLVE, GET_STAGE_RANGE, GET_STAGE_READER_SETUP, GET_STAGE_READER_SETUP_DROP_PENDING,
GET_STAGE_READER_SETUP_SCHEDULE, GET_STAGE_READER_SETUP_WAIT_QUORUM, GET_STAGE_READER_TASK_BITROT_READER_INIT,
GET_STAGE_READER_TASK_FILE_OPEN, GET_STAGE_READER_TASK_READER_CONSTRUCTION, GetObjectFailureReason, classify_disk_error,
get_stage_timer_if_enabled, record_get_object_pipeline_failure, record_get_object_pipeline_failure_for_path,
record_get_stage_duration_if_enabled,
get_stage_timer_if_enabled, mark_get_object_downstream_closed, record_get_object_pipeline_failure,
record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
};
use crate::erasure::coding::BitrotReader;
use crate::io_support::bitrot::{
@@ -53,12 +53,60 @@ use std::{
task::{Context, Poll},
time::{Duration, Instant},
};
use tokio::io::{AsyncRead, ReadBuf};
use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
use tokio::sync::RwLock;
use tokio::task::JoinSet;
use super::core::io_primitives::*;
#[derive(Clone, Copy)]
pub(super) enum GetObjectOutputKind {
HttpResponse,
Internal,
}
struct GetObjectOutputWriter<'a, W> {
inner: &'a mut W,
output_kind: GetObjectOutputKind,
}
impl<'a, W> GetObjectOutputWriter<'a, W> {
fn new(inner: &'a mut W, output_kind: GetObjectOutputKind) -> Self {
Self { inner, output_kind }
}
}
fn map_get_object_output_error(output_kind: GetObjectOutputKind, err: std::io::Error) -> std::io::Error {
if matches!(output_kind, GetObjectOutputKind::HttpResponse) {
mark_get_object_downstream_closed(err)
} else {
err
}
}
impl<W: AsyncWrite + Unpin> AsyncWrite for GetObjectOutputWriter<'_, W> {
fn poll_write(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &[u8]) -> Poll<std::io::Result<usize>> {
let output_kind = self.output_kind;
Pin::new(&mut *self.inner)
.poll_write(cx, buf)
.map(|result| result.map_err(|err| map_get_object_output_error(output_kind, err)))
}
fn poll_flush(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
let output_kind = self.output_kind;
Pin::new(&mut *self.inner)
.poll_flush(cx)
.map(|result| result.map_err(|err| map_get_object_output_error(output_kind, err)))
}
fn poll_shutdown(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
let output_kind = self.output_kind;
Pin::new(&mut *self.inner)
.poll_shutdown(cx)
.map(|result| result.map_err(|err| map_get_object_output_error(output_kind, err)))
}
}
impl SetDisks {
async fn get_object_metadata_cache_bypass_reason(
&self,
@@ -586,6 +634,7 @@ impl SetDisks {
offset: usize,
length: i64,
writer: &mut W,
output_kind: GetObjectOutputKind,
fi: FileInfo,
files: Vec<FileInfo>,
disks: &[Option<DiskStore>],
@@ -969,9 +1018,10 @@ impl SetDisks {
let unattempted_data_shards = !reader_setup.data_shards_attempted(erasure.data_shards);
let readers = reader_setup.readers;
let deferred_stripe_handles = reader_setup.deferred_stripe_handles;
let mut output_writer = GetObjectOutputWriter::new(writer, output_kind);
let (written, err) = erasure
.decode_with_stripe_handles(
writer,
&mut output_writer,
readers,
part_offset,
part_length,
@@ -989,7 +1039,7 @@ impl SetDisks {
metrics_size_bucket,
decode_elapsed.as_secs_f64(),
);
if decode_elapsed >= SLOW_OBJECT_READ_LOG_THRESHOLD || err.is_some() {
if decode_elapsed >= SLOW_OBJECT_READ_LOG_THRESHOLD && err.is_none() {
warn!(
event = EVENT_SET_DISK_READ,
component = LOG_COMPONENT_ECSTORE,
@@ -1005,9 +1055,8 @@ impl SetDisks {
available_shards,
missing_shards,
elapsed_ms = decode_elapsed.as_millis(),
error = ?err,
state = if err.is_some() { "decode_failed_or_slow" } else { "decode_slow" },
"Set disk object decode stage is slow or failed"
state = "decode_slow",
"Set disk object decode stage is slow"
);
}
debug!(
@@ -1051,7 +1100,7 @@ impl SetDisks {
if has_err {
let reason = classify_disk_error(&de_err);
if reason == GetObjectFailureReason::DownstreamClosed {
warn!(
debug!(
bucket,
object,
part_index = current_part,
@@ -1969,6 +2018,7 @@ mod metadata_cache_tests {
2,
1,
&mut output,
GetObjectOutputKind::Internal,
valid_test_fileinfo(object),
Vec::new(),
&[],
@@ -1992,6 +2042,7 @@ mod metadata_cache_tests {
usize::MAX,
1,
&mut output,
GetObjectOutputKind::Internal,
overflow,
Vec::new(),
&[],
@@ -2013,6 +2064,7 @@ mod metadata_cache_tests {
1,
1,
&mut output,
GetObjectOutputKind::Internal,
valid_test_fileinfo(object),
Vec::new(),
&[],
@@ -2036,6 +2088,7 @@ mod metadata_cache_tests {
0,
1,
&mut output,
GetObjectOutputKind::Internal,
invalid_erasure,
Vec::new(),
&[],
@@ -2079,6 +2132,7 @@ mod metadata_cache_tests {
0,
1,
&mut output,
GetObjectOutputKind::Internal,
fi,
Vec::new(),
&[],
@@ -2920,11 +2974,60 @@ mod tests {
Arc,
atomic::{AtomicUsize, Ordering},
};
use tokio::io::AsyncReadExt;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
const CODEC_STREAMING_TEST_BUCKET: &str = "bucket";
const CODEC_STREAMING_TEST_OBJECT: &str = "object";
struct ClosedOutputWriter;
impl AsyncWrite for ClosedOutputWriter {
fn poll_write(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &[u8]) -> Poll<std::io::Result<usize>> {
Poll::Ready(Err(std::io::Error::from(ErrorKind::BrokenPipe)))
}
fn poll_flush(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
Poll::Ready(Ok(()))
}
fn poll_shutdown(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
Poll::Ready(Ok(()))
}
}
#[tokio::test]
async fn output_writer_marks_closed_duplex_reader_as_downstream_close() {
let (reader, mut inner) = tokio::io::duplex(64);
drop(reader);
let mut writer = GetObjectOutputWriter::new(&mut inner, GetObjectOutputKind::HttpResponse);
let err = writer
.write_all(b"range payload")
.await
.expect_err("closed duplex reader must fail the output write");
assert_eq!(
classify_disk_error(&DiskError::from(err)),
GetObjectFailureReason::DownstreamClosed,
"only the output adapter may classify a broken pipe as a downstream close"
);
}
#[tokio::test]
async fn output_writer_preserves_remote_broken_pipe_as_io_failure() {
let mut inner = ClosedOutputWriter;
let mut writer = GetObjectOutputWriter::new(&mut inner, GetObjectOutputKind::Internal);
let err = writer
.write_all(b"transition payload")
.await
.expect_err("closed remote writer must fail the output write");
assert_eq!(
classify_disk_error(&DiskError::from(err)),
GetObjectFailureReason::Io,
"remote transition writer failures must not be mistaken for HTTP client cancellation"
);
}
async fn local_test_disks(count: usize, bucket: &str) -> (Vec<tempfile::TempDir>, Vec<Option<crate::disk::DiskStore>>) {
let mut dirs = Vec::with_capacity(count);
let mut disks = Vec::with_capacity(count);
@@ -4019,6 +4122,7 @@ mod tests {
0,
part_data.len() as i64,
&mut output,
GetObjectOutputKind::Internal,
fi,
files,
&disks,
+16 -5
View File
@@ -1500,11 +1500,7 @@ impl<R> GetObjectStreamingReader<R> {
{
return "read_quorum";
}
if error_msg.contains("connection reset")
|| error_msg.contains("broken pipe")
|| error_msg.contains("downstream")
|| error_msg.contains("remote closed")
{
if error_msg.contains("getobject output closed:") {
return "downstream_closed";
}
}
@@ -10608,6 +10604,21 @@ mod tests {
assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
}
#[test]
fn get_object_streaming_reader_distinguishes_internal_and_output_broken_pipes() {
let internal = std::io::Error::from(std::io::ErrorKind::BrokenPipe);
assert_eq!(GetObjectStreamingReader::<std::io::Cursor<Vec<u8>>>::classify_read_error(&internal), "io");
let output_closed = std::io::Error::new(
std::io::ErrorKind::BrokenPipe,
std::io::Error::other("GetObject output closed: broken pipe"),
);
assert_eq!(
GetObjectStreamingReader::<std::io::Cursor<Vec<u8>>>::classify_read_error(&output_closed),
"downstream_closed"
);
}
#[tokio::test]
async fn put_object_body_read_timeout_guard_aborts_on_stall() {
// Inner stream never yields and never reports EOF (a proxy that forwarded