This PR implements a graceful shutdown API for Deno.serve, allowing all
current connections to drain from the server before shutting down, while
preventing new connections from being started or new transactions on
existing connections from being created.
We split the cancellation handle into two parts: a listener handle, and
a connection handle. A graceful shutdown cancels the listener only,
while allowing the connections to drain. The connection handle aborts
all futures. If the listener handle is cancelled, we put the connections
into graceful shutdown mode, which disables keep-alive on http/1.1 and
uses http/2 mechanisms for http/2 connections.
In addition, we now guarantee that all connections are complete or
cancelled, and all resources are cleaned up when the server `finished`
promise resolves -- we use a Rust-side server refcount for this.
Performance impact: does not appear to affect basic serving performance
by more than 1% (~126k -> ~125k)
---------
Co-authored-by: Bartek Iwańczuk <biwanczuk@gmail.com>
When a TCP connection is force-closed (ie: browser refresh), the
underlying future we pass to Hyper is dropped which may cause us to try
to drop the body resource while the OpState lock is still held.
Preconditions for this bug to trigger:
- The body resource must have been taken
- The response must return a resource (which requires us to take the
OpState lock)
- The TCP connection must have been dropped before this
Fixes #20315 and #20298
Deno.serve's fast streaming implementation was not keeping the request
body resource ID alive. We were taking the `Rc<Resource>` from the
resource table during the response, so a hairpin duplex response that
fed back the request body would work.
However, if any JS code attempted to read from the request body (which
requires the resource ID to be valid), the response would fail with a
difficult-to-diagnose "EOF" error.
This was affecting more complex duplex uses of `Deno.fetch` (though as
far as I can tell was unreported).
Simple test:
```ts
const reader = request.body.getReader();
return new Response(
new ReadableStream({
async pull(controller) {
const { done, value } = await reader.read();
if (done) {
controller.close();
} else {
controller.enqueue(value);
}
},
}),
```
And then attempt to use the stream in duplex mode:
```ts
async function testDuplex(
reader: ReadableStreamDefaultReader<Uint8Array>,
writable: WritableStreamDefaultWriter<Uint8Array>,
) {
await writable.write(new Uint8Array([1]));
const chunk1 = await reader.read();
assert(!chunk1.done);
assertEquals(chunk1.value, new Uint8Array([1]));
await writable.write(new Uint8Array([2]));
const chunk2 = await reader.read();
assert(!chunk2.done);
assertEquals(chunk2.value, new Uint8Array([2]));
await writable.close();
const chunk3 = await reader.read();
assert(chunk3.done);
}
```
In older versions of Deno, this would just lock up. I believe after
23ff0e722e, it started throwing a more
explicit error:
```
httpServerStreamDuplexJavascript => ./cli/tests/unit/serve_test.ts:1339:6
error: TypeError: request or response body error: error reading a body from connection: Connection reset by peer (os error 54)
at async Object.pull (ext:deno_web/06_streams.js:810:27)
```
Benchmarking shows numbers are pretty close, however this is recommended
for the best possible thread-local performance and may improve in future
Rust compiler revisions.
Fixes for various `Attemped to access invalid request` bugs (#19058,
#15427, #17213).
We did not wait for both a drop event and a completion event before
removing items from the slab table. This ensures that we do so.
In addition, the slab methods are refactored out into `slab.rs` for
maintainability.
