Coalesce concurrent remote metadata requests (#36184)

## Why

Concurrent callers can request metadata for the same remote path, resulting in duplicate `fs/getMetadata` RPCs.

## What changed

- Share an unsandboxed metadata request with concurrent callers for the same path while the RPC is in flight.
- Remove completed and failed requests so later callers fetch fresh metadata, and clear in-flight entries after filesystem mutations, including failed mutations.
- Keep requests for different paths separate and bypass sharing for sandboxed requests.

## Testing

Add WebSocket-backed tests covering shared results and errors, retry after cancellation or failure, mutation invalidation, and path and sandbox isolation.

GitOrigin-RevId: f91e51b94db3fb984597aec8d302b8715778121c
This commit is contained in:
Charlie Marsh
2026-07-30 16:35:20 +00:00
committed by copyberry
parent 0dcad0c972
commit 483559cc75
2 changed files with 505 additions and 13 deletions
+67 -13
View File
@@ -1,7 +1,12 @@
use std::collections::HashMap;
use std::sync::Arc;
use base64::Engine as _;
use base64::engine::general_purpose::STANDARD;
use codex_utils_path_uri::PathUri;
use tokio::io;
use tokio::sync::Mutex;
use tokio::sync::OnceCell;
use tracing::trace;
use crate::CopyOptions;
@@ -35,14 +40,20 @@ const NOT_FOUND_ERROR_CODE: i64 = -32004;
#[path = "remote_file_stream.rs"]
mod file_stream;
type InFlightMetadataRequest = OnceCell<Result<FileMetadata, Arc<io::Error>>>;
pub(crate) struct RemoteFileSystem {
client: LazyRemoteExecServerClient,
metadata_requests: Mutex<HashMap<PathUri, Arc<InFlightMetadataRequest>>>,
}
impl RemoteFileSystem {
pub(crate) fn new(client: LazyRemoteExecServerClient) -> Self {
trace!("remote fs new");
Self { client }
Self {
client,
metadata_requests: Mutex::new(HashMap::new()),
}
}
async fn canonicalize(
@@ -108,14 +119,15 @@ impl RemoteFileSystem {
) -> FileSystemResult<()> {
trace!("remote fs write_file");
let client = self.client.get().await.map_err(map_remote_error)?;
client
let result = client
.fs_write_file(FsWriteFileParams {
path: path.clone(),
data_base64: STANDARD.encode(contents),
sandbox: remote_sandbox_context(sandbox),
})
.await
.map_err(map_remote_error)?;
.await;
self.metadata_requests.lock().await.clear();
result.map_err(map_remote_error)?;
Ok(())
}
@@ -127,21 +139,61 @@ impl RemoteFileSystem {
) -> FileSystemResult<()> {
trace!("remote fs create_directory");
let client = self.client.get().await.map_err(map_remote_error)?;
client
let result = client
.fs_create_directory(FsCreateDirectoryParams {
path: path.clone(),
recursive: Some(options.recursive),
sandbox: remote_sandbox_context(sandbox),
})
.await
.map_err(map_remote_error)?;
.await;
self.metadata_requests.lock().await.clear();
result.map_err(map_remote_error)?;
Ok(())
}
/// Shares identical unsandboxed metadata requests only while their RPC remains in flight.
async fn get_metadata(
&self,
path: &PathUri,
sandbox: Option<&FileSystemSandboxContext>,
) -> FileSystemResult<FileMetadata> {
if sandbox.is_some() {
return self.get_metadata_uncached(path, sandbox).await;
}
let request = {
let mut requests = self.metadata_requests.lock().await;
requests.retain(|_, in_flight| Arc::strong_count(in_flight) > 1);
Arc::clone(requests.entry(path.clone()).or_default())
};
let result = match request
.get_or_init(|| async {
self.get_metadata_uncached(path, /*sandbox*/ None)
.await
.map_err(Arc::new)
})
.await
{
Ok(metadata) => Ok(metadata.clone()),
Err(error) => Err(io::Error::new(error.kind(), error.to_string())),
};
let mut requests = self.metadata_requests.lock().await;
if requests
.get(path)
.is_some_and(|in_flight| Arc::ptr_eq(in_flight, &request))
{
requests.remove(path);
}
result
}
/// Sends a fresh metadata request without sharing another caller's sandbox or result.
async fn get_metadata_uncached(
&self,
path: &PathUri,
sandbox: Option<&FileSystemSandboxContext>,
) -> FileSystemResult<FileMetadata> {
trace!("remote fs get_metadata");
let client = self.client.get().await.map_err(map_remote_error)?;
@@ -226,15 +278,16 @@ impl RemoteFileSystem {
) -> FileSystemResult<()> {
trace!("remote fs remove");
let client = self.client.get().await.map_err(map_remote_error)?;
client
let result = client
.fs_remove(FsRemoveParams {
path: path.clone(),
recursive: Some(options.recursive),
force: Some(options.force),
sandbox: remote_sandbox_context(sandbox),
})
.await
.map_err(map_remote_error)?;
.await;
self.metadata_requests.lock().await.clear();
result.map_err(map_remote_error)?;
Ok(())
}
@@ -247,15 +300,16 @@ impl RemoteFileSystem {
) -> FileSystemResult<()> {
trace!("remote fs copy");
let client = self.client.get().await.map_err(map_remote_error)?;
client
let result = client
.fs_copy(FsCopyParams {
source_path: source_path.clone(),
destination_path: destination_path.clone(),
recursive: options.recursive,
sandbox: remote_sandbox_context(sandbox),
})
.await
.map_err(map_remote_error)?;
.await;
self.metadata_requests.lock().await.clear();
result.map_err(map_remote_error)?;
Ok(())
}
}
@@ -1,5 +1,7 @@
#![allow(clippy::expect_used)]
use codex_exec_server_protocol::JSONRPCError;
use codex_exec_server_protocol::JSONRPCErrorError;
use codex_exec_server_protocol::JSONRPCMessage;
use codex_exec_server_protocol::JSONRPCResponse;
use codex_http_client::HttpClientFactory;
@@ -27,7 +29,14 @@ use tokio_tungstenite::tungstenite::Message;
use super::*;
use crate::client_api::DEFAULT_REMOTE_EXEC_SERVER_CONNECT_TIMEOUT;
use crate::client_api::ExecServerTransportParams;
use crate::protocol::FS_COPY_METHOD;
use crate::protocol::FS_CREATE_DIRECTORY_METHOD;
use crate::protocol::FS_GET_METADATA_METHOD;
use crate::protocol::FS_READ_FILE_METHOD;
use crate::protocol::FS_REMOVE_METHOD;
use crate::protocol::FS_WRITE_FILE_METHOD;
use crate::protocol::FsGetMetadataParams;
use crate::protocol::FsGetMetadataResponse;
use crate::protocol::FsReadFileParams;
use crate::protocol::FsReadFileResponse;
use crate::protocol::INITIALIZE_METHOD;
@@ -86,6 +95,335 @@ async fn remote_file_system_sends_path_and_sandbox_cwd_uris_without_native_conve
server.await.expect("recording server should succeed");
}
#[tokio::test]
async fn concurrent_remote_metadata_requests_share_only_in_flight_results() {
let (abandoned_request_tx, abandoned_request_rx) = oneshot::channel();
let (websocket_url, captured_params, server) = record_metadata_params(vec![
MetadataResponse::Immediate(Ok(metadata_response(/*size*/ 42))),
MetadataResponse::Abandoned(abandoned_request_tx),
MetadataResponse::Immediate(Ok(metadata_response(/*size*/ 43))),
])
.await;
let file_system = std::sync::Arc::new(RemoteFileSystem::new(LazyRemoteExecServerClient::new(
ExecServerTransportParams::websocket_url(
websocket_url,
DEFAULT_REMOTE_EXEC_SERVER_CONNECT_TIMEOUT,
),
HttpClientFactory::new(OutboundProxyPolicy::ReqwestDefault),
)));
let path = PathUri::parse("file:///workspace/project/AGENTS.md").expect("valid path URI");
let (first, second) = tokio::join!(
file_system.get_metadata(&path, /*sandbox*/ None),
file_system.get_metadata(&path, /*sandbox*/ None),
);
let expected = FileMetadata {
is_directory: false,
is_file: true,
is_symlink: false,
size: 42,
created_at_ms: 10,
modified_at_ms: 20,
};
assert_eq!(first.expect("first metadata request"), expected);
assert_eq!(second.expect("second metadata request"), expected);
assert!(file_system.metadata_requests.lock().await.is_empty());
let initializing_file_system = std::sync::Arc::clone(&file_system);
let initializing_path = path.clone();
let initializer = tokio::spawn(async move {
initializing_file_system
.get_metadata(&initializing_path, /*sandbox*/ None)
.await
});
abandoned_request_rx
.await
.expect("server should receive the abandoned metadata request");
let follower = file_system.get_metadata(&path, /*sandbox*/ None);
tokio::pin!(follower);
assert!(futures::poll!(follower.as_mut()).is_pending());
assert_eq!(
file_system
.metadata_requests
.lock()
.await
.get(&path)
.map(std::sync::Arc::strong_count),
Some(3)
);
initializer.abort();
assert!(
initializer
.await
.expect_err("metadata initializer should be aborted")
.is_cancelled()
);
let mut refreshed = expected;
refreshed.size = 43;
assert_eq!(
follower
.await
.expect("waiting metadata request should retry the abandoned initializer"),
refreshed
);
assert!(file_system.metadata_requests.lock().await.is_empty());
assert_eq!(
captured_params.await.expect("captured metadata requests"),
vec![
FsGetMetadataParams {
path: path.clone(),
sandbox: None,
};
3
]
);
server.await.expect("metadata server should succeed");
}
#[tokio::test]
async fn concurrent_remote_metadata_errors_are_shared_but_retried() {
let (websocket_url, captured_params, server) = record_metadata_params(vec![
MetadataResponse::Immediate(Err(JSONRPCErrorError {
code: NOT_FOUND_ERROR_CODE,
data: None,
message: "metadata not found".to_string(),
})),
MetadataResponse::Immediate(Ok(metadata_response(/*size*/ 42))),
])
.await;
let file_system = RemoteFileSystem::new(LazyRemoteExecServerClient::new(
ExecServerTransportParams::websocket_url(
websocket_url,
DEFAULT_REMOTE_EXEC_SERVER_CONNECT_TIMEOUT,
),
HttpClientFactory::new(OutboundProxyPolicy::ReqwestDefault),
));
let path = PathUri::parse("file:///workspace/project/AGENTS.md").expect("valid path URI");
let (first, second) = tokio::join!(
file_system.get_metadata(&path, /*sandbox*/ None),
file_system.get_metadata(&path, /*sandbox*/ None),
);
assert_eq!(
first.expect_err("first metadata error").kind(),
std::io::ErrorKind::NotFound
);
assert_eq!(
second.expect_err("shared metadata error").kind(),
std::io::ErrorKind::NotFound
);
assert_eq!(
file_system
.get_metadata(&path, /*sandbox*/ None)
.await
.expect("failed metadata request should be retried")
.size,
42
);
assert_eq!(
captured_params.await.expect("captured metadata requests"),
vec![
FsGetMetadataParams {
path: path.clone(),
sandbox: None,
},
FsGetMetadataParams {
path,
sandbox: None,
},
]
);
server.await.expect("metadata server should succeed");
}
#[tokio::test]
async fn remote_metadata_starts_fresh_after_intervening_filesystem_mutation() {
let path = PathUri::parse("file:///workspace/project/AGENTS.md").expect("valid path URI");
let source_path =
PathUri::parse("file:///workspace/project/source.md").expect("valid source path URI");
for (mutation, mutation_response) in [
(MetadataMutation::Write, Ok(())),
(MetadataMutation::CreateDirectory, Ok(())),
(MetadataMutation::Remove, Ok(())),
(MetadataMutation::Copy, Ok(())),
(
MetadataMutation::Copy,
Err(JSONRPCErrorError {
code: INVALID_REQUEST_ERROR_CODE,
data: None,
message: "mutation partially failed".to_string(),
}),
),
] {
let mutation_fails = mutation_response.is_err();
let (websocket_url, captured_params, server) = record_metadata_params(vec![
MetadataResponse::Immediate(Ok(metadata_response(/*size*/ 42))),
MetadataResponse::Mutation(mutation_response),
MetadataResponse::Immediate(Ok(metadata_response(/*size*/ 43))),
])
.await;
let file_system = RemoteFileSystem::new(LazyRemoteExecServerClient::new(
ExecServerTransportParams::websocket_url(
websocket_url,
DEFAULT_REMOTE_EXEC_SERVER_CONNECT_TIMEOUT,
),
HttpClientFactory::new(OutboundProxyPolicy::ReqwestDefault),
));
file_system
.client
.get()
.await
.expect("remote filesystem client should connect");
let stale_request = file_system.get_metadata(&path, /*sandbox*/ None);
tokio::pin!(stale_request);
assert!(futures::poll!(stale_request.as_mut()).is_pending());
let result = match mutation {
MetadataMutation::Write => {
file_system
.write_file(&path, b"updated".to_vec(), /*sandbox*/ None)
.await
}
MetadataMutation::CreateDirectory => {
file_system
.create_directory(
&path,
CreateDirectoryOptions { recursive: true },
/*sandbox*/ None,
)
.await
}
MetadataMutation::Remove => {
file_system
.remove(
&path,
RemoveOptions {
recursive: true,
force: true,
},
/*sandbox*/ None,
)
.await
}
MetadataMutation::Copy => {
file_system
.copy(
&source_path,
&path,
CopyOptions { recursive: true },
/*sandbox*/ None,
)
.await
}
};
if mutation_fails {
assert_eq!(
result
.expect_err("remote filesystem mutation should fail")
.kind(),
std::io::ErrorKind::InvalidInput
);
} else {
result.expect("remote filesystem mutation should succeed");
}
let (refreshed, stale) = tokio::join!(
file_system.get_metadata(&path, /*sandbox*/ None),
stale_request.as_mut(),
);
assert_eq!(stale.expect("original metadata request").size, 42);
assert_eq!(refreshed.expect("post-mutation metadata request").size, 43);
assert_eq!(
captured_params.await.expect("captured metadata requests"),
vec![
FsGetMetadataParams {
path: path.clone(),
sandbox: None,
};
2
]
);
server.await.expect("metadata server should succeed");
}
}
#[tokio::test]
async fn remote_metadata_requests_do_not_cross_path_or_sandbox_boundaries() {
let metadata = metadata_response(/*size*/ 42);
let (websocket_url, captured_params, server) = record_metadata_params(vec![
MetadataResponse::Immediate(Ok(metadata.clone())),
MetadataResponse::Immediate(Ok(metadata.clone())),
MetadataResponse::Immediate(Ok(metadata.clone())),
MetadataResponse::Immediate(Ok(metadata)),
])
.await;
let file_system = RemoteFileSystem::new(LazyRemoteExecServerClient::new(
ExecServerTransportParams::websocket_url(
websocket_url,
DEFAULT_REMOTE_EXEC_SERVER_CONNECT_TIMEOUT,
),
HttpClientFactory::new(OutboundProxyPolicy::ReqwestDefault),
));
let first_path = PathUri::parse("file:///workspace/project/AGENTS.md").expect("valid path URI");
let second_path = PathUri::parse("file:///workspace/project/SKILL.md").expect("valid path URI");
let sandbox = FileSystemSandboxContext::from_permission_profile_with_cwd(
PermissionProfile::from_runtime_permissions(
&FileSystemSandboxPolicy::restricted(vec![FileSystemSandboxEntry {
path: FileSystemPath::Special {
value: FileSystemSpecialPath::project_roots(/*subpath*/ None),
},
access: FileSystemAccessMode::Write,
missing_path_behavior: None,
}]),
NetworkSandboxPolicy::Restricted,
),
non_native_cwd(),
);
let (first, second) = tokio::join!(
file_system.get_metadata(&first_path, /*sandbox*/ None),
file_system.get_metadata(&second_path, /*sandbox*/ None),
);
first.expect("metadata for first path");
second.expect("metadata for second path");
let (first, second) = tokio::join!(
file_system.get_metadata(&first_path, Some(&sandbox)),
file_system.get_metadata(&first_path, Some(&sandbox)),
);
first.expect("first sandboxed metadata request");
second.expect("second sandboxed metadata request");
let captured_params = captured_params.await.expect("captured metadata requests");
assert_eq!(
captured_params
.iter()
.filter(|params| params.path == first_path && params.sandbox.is_none())
.count(),
1
);
assert_eq!(
captured_params
.iter()
.filter(|params| params.path == second_path && params.sandbox.is_none())
.count(),
1
);
assert_eq!(
captured_params
.iter()
.filter(|params| params.path == first_path && params.sandbox.as_ref() == Some(&sandbox))
.count(),
2
);
server.await.expect("metadata server should succeed");
}
async fn record_read_file_params(
expected_requests: usize,
) -> (
@@ -137,6 +475,106 @@ async fn record_read_file_params(
(websocket_url, captured_params_rx, server)
}
enum MetadataResponse {
Immediate(Result<FsGetMetadataResponse, JSONRPCErrorError>),
Abandoned(oneshot::Sender<()>),
Mutation(Result<(), JSONRPCErrorError>),
}
fn metadata_response(size: u64) -> FsGetMetadataResponse {
FsGetMetadataResponse {
is_directory: false,
is_file: true,
is_symlink: false,
size,
created_at_ms: 10,
modified_at_ms: 20,
}
}
#[derive(Clone, Copy)]
enum MetadataMutation {
Write,
CreateDirectory,
Remove,
Copy,
}
async fn record_metadata_params(
responses: Vec<MetadataResponse>,
) -> (
String,
oneshot::Receiver<Vec<FsGetMetadataParams>>,
tokio::task::JoinHandle<()>,
) {
let listener = TcpListener::bind("127.0.0.1:0")
.await
.expect("listener should bind");
let websocket_url = format!("ws://{}", listener.local_addr().expect("listener address"));
let (captured_params_tx, captured_params_rx) = oneshot::channel();
let server = tokio::spawn(async move {
let (stream, _) = listener.accept().await.expect("listener should accept");
let mut websocket = accept_async(stream)
.await
.expect("websocket handshake should succeed");
complete_websocket_initialize(&mut websocket).await;
let mut captured_params = Vec::with_capacity(responses.len());
for response in responses {
let request = match read_jsonrpc_websocket(&mut websocket).await {
JSONRPCMessage::Request(request) => request,
other => panic!("expected filesystem request, got {other:?}"),
};
if matches!(response, MetadataResponse::Mutation(_)) {
assert!(matches!(
request.method.as_str(),
FS_WRITE_FILE_METHOD
| FS_CREATE_DIRECTORY_METHOD
| FS_REMOVE_METHOD
| FS_COPY_METHOD
));
} else {
assert_eq!(request.method, FS_GET_METADATA_METHOD);
let params: FsGetMetadataParams = serde_json::from_value(
request.params.expect("fs/getMetadata params should exist"),
)
.expect("fs/getMetadata params should deserialize");
captured_params.push(params);
}
let response = match response {
MetadataResponse::Immediate(Ok(metadata)) => {
JSONRPCMessage::Response(JSONRPCResponse {
id: request.id,
result: serde_json::to_value(metadata)
.expect("fs/getMetadata response should serialize"),
})
}
MetadataResponse::Immediate(Err(error))
| MetadataResponse::Mutation(Err(error)) => JSONRPCMessage::Error(JSONRPCError {
error,
id: request.id,
}),
MetadataResponse::Mutation(Ok(())) => JSONRPCMessage::Response(JSONRPCResponse {
id: request.id,
result: serde_json::json!({}),
}),
MetadataResponse::Abandoned(received_tx) => {
received_tx
.send(())
.expect("abandoned request observer should stay open");
continue;
}
};
write_jsonrpc_websocket(&mut websocket, response).await;
}
captured_params_tx
.send(captured_params)
.expect("captured params receiver should stay open");
});
(websocket_url, captured_params_rx, server)
}
fn non_native_cwd() -> PathUri {
#[cfg(unix)]
let uri = "file://server/share/checkout";