Fixes #21121 and #19498
Migrates fully to rustls_tokio_stream. We no longer need to maintain our
own TlsStream implementation to properly support duplex.
This should fix a number of errors with TLS and websockets, HTTP and
"other" places where it's failing.
Use HttpRecord as response body so requests can be tracked all the way
to response body completion.
This allows Request properties to be accessed while the response body is
streaming.
Graceful shutdown now awaits a future instead of async spinning waiting
for requests to finish.
On the minimal benchmark this refactor improves performance an
additional 2% over pooling alone for a net 3% increase over the previous
deno main branch.
Builds upon https://github.com/denoland/deno/pull/20809 and
https://github.com/denoland/deno/pull/20770.
---------
Co-authored-by: Matt Mastracci <matthew@mastracci.com>
Reuse existing existing allocations for HttpRecord and response
HeaderMap where possible.
At request end used allocations are returned to the pool and the pool
and the pool sized to 1/8th the current number of inflight requests.
For http1 hyper will reuse the response HeaderMap for the following
request on the connection.
Builds upon https://github.com/denoland/deno/pull/20770
---------
Co-authored-by: Matt Mastracci <matthew@mastracci.com>
Makes the JavaScript Request use a v8:External opaque pointer to
directly refer to the Rust HttpRecord.
The HttpRecord is now reference counted. To avoid leaks the strong count
is checked at request completion.
Performance seems unchanged on the minimal benchmark. 118614 req/s this
branch vs 118564 req/s on main, but variance between runs on my laptop
is pretty high.
---------
Co-authored-by: Matt Mastracci <matthew@mastracci.com>
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
<!--
Before submitting a PR, please read https://deno.com/manual/contributing
1. Give the PR a descriptive title.
Examples of good title:
- fix(std/http): Fix race condition in server
- docs(console): Update docstrings
- feat(doc): Handle nested reexports
Examples of bad title:
- fix #7123
- update docs
- fix bugs
2. Ensure there is a related issue and it is referenced in the PR text.
3. Ensure there are tests that cover the changes.
4. Ensure `cargo test` passes.
5. Ensure `./tools/format.js` passes without changing files.
6. Ensure `./tools/lint.js` passes.
7. Open as a draft PR if your work is still in progress. The CI won't
run
all steps, but you can add '[ci]' to a commit message to force it to.
8. If you would like to run the benchmarks on the CI, add the 'ci-bench'
label.
-->
As the title.
---------
Co-authored-by: Matt Mastracci <matthew@mastracci.com>
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)
```
Includes a lightly-modified version of hyper-util's `TokioIo` utility.
Hyper changes:
v1.0.0-rc.4 (2023-07-10)
Bug Fixes
http1:
http1 server graceful shutdown fix (#3261)
([f4b51300](f4b513009d))
send error on Incoming body when connection errors (#3256)
([52f19259](52f192593f),
closes https://github.com/hyperium/hyper/issues/3253)
properly end chunked bodies when it was known to be empty (#3254)
([fec64cf0](fec64cf0ab),
closes https://github.com/hyperium/hyper/issues/3252)
Features
client: Make clients able to use non-Send executor (#3184)
([d977f209](d977f209bc),
closes https://github.com/hyperium/hyper/issues/3017)
rt:
replace IO traits with hyper::rt ones (#3230)
([f9f65b7a](f9f65b7aa6),
closes https://github.com/hyperium/hyper/issues/3110)
add downcast on Sleep trait (#3125)
([d92d3917](d92d3917d9),
closes https://github.com/hyperium/hyper/issues/3027)
service: change Service::call to take &self (#3223)
([d894439e](d894439e00),
closes https://github.com/hyperium/hyper/issues/3040)
Breaking Changes
Any IO transport type provided must not implement hyper::rt::{Read,
Write} instead of tokio::io traits. You can grab a helper type from
hyper-util to wrap Tokio types, or implement the traits yourself, if
it's a custom type.
([f9f65b7a](f9f65b7aa6))
client::conn::http2 types now use another generic for an Executor. Code
that names Connection needs to include the additional generic parameter.
([d977f209](d977f209bc))
The Service::call function no longer takes a mutable reference to self.
The FnMut trait bound on the service::util::service_fn function and the
trait bound on the impl for the ServiceFn struct were changed from FnMut
to Fn.
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.
This is a new op system that will eventually replace `#[op]`.
Features
- More maintainable, generally less-coupled code
- More modern Rust proc-macro libraries
- Enforces correct `fast` labelling for fast ops, allowing for visual
scanning of fast ops
- Explicit marking of `#[string]`, `#[serde]` and `#[smi]` parameters.
This first version of op2 supports integer and Option<integer>
parameters only, and allows us to start working on converting ops and
adding features.
This switches syscall used in HTTP and WS server from "writev"
to "sendto".
"DENO_USE_WRITEV=1" can be used to enable using "writev" syscall.
Doing this for easier testing of various setups.
This PR attempts to resolve the first item on the list from
https://github.com/denoland/deno/issues/19330 which is about using a
flat list of interleaved key/value pairs, instead of a nested array of
tuples.
I can tackle some more if you can provide a quick example of using raw
v8 arrays, cc @mmastrac
Under heavy load, we often have requests queued up that don't need an
async call to retrieve. We can use a fast path sync op to drain this set
of ready requests, and then fall back to the async op once we run out of
work.
This is a .5-1% bump in req/s on an M2 mac. About 90% of the handlers go
through this sync phase (based on a simple instrumentation that is not
included in this PR) and skip the async machinery entirely.
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.
Merges `op_http_upgrade_next` and `op_ws_server_create`, significantly
simplifying websocket construction in ext/http (next), and removing one
JS -> Rust call. Also WS server now doesn't bypass
`HttpPropertyExtractor`.
Improve abstractions around listeners to support listener + connection
network stream combinations not previously possible (for example a
listener exposed as a Tcp, creating Unix network streams).
Partially supersedes #19016.
This migrates `spawn` and `spawn_blocking` to `deno_core`, and removes
the requirement for `spawn` tasks to be `Send` given our single-threaded
executor.
While we don't need to technically do anything w/`spawn_blocking`, this
allows us to have a single `JoinHandle` type that works for both cases,
and allows us to more easily experiment with alternative
`spawn_blocking` implementations that do not require tokio (ie: rayon).
Async ops (+~35%):
Before:
```
time 1310 ms rate 763358
time 1267 ms rate 789265
time 1259 ms rate 794281
time 1266 ms rate 789889
```
After:
```
time 956 ms rate 1046025
time 954 ms rate 1048218
time 924 ms rate 1082251
time 920 ms rate 1086956
```
HTTP serve (+~4.4%):
Before:
```
Running 10s test @ http://localhost:4500
2 threads and 10 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 68.78us 19.77us 1.43ms 86.84%
Req/Sec 68.78k 5.00k 73.84k 91.58%
1381833 requests in 10.10s, 167.36MB read
Requests/sec: 136823.29
Transfer/sec: 16.57MB
```
After:
```
Running 10s test @ http://localhost:4500
2 threads and 10 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 63.12us 17.43us 1.11ms 85.13%
Req/Sec 71.82k 3.71k 77.02k 79.21%
1443195 requests in 10.10s, 174.79MB read
Requests/sec: 142921.99
Transfer/sec: 17.31MB
```
Suggested-By: alice@ryhl.io
Co-authored-by: Bartek Iwańczuk <biwanczuk@gmail.com>
`Content-Encoding: gzip` support for `Deno.serve`. This doesn't support
Brotli (`br`) yet, however it should not be difficult to add. Heuristics
for compression are modelled after those in `Deno.serveHttp`.
Tests are provided to ensure that the gzip compression is correct. We
chunk a number of different streams (zeros, hard-to-compress data,
already-gzipped data) in a number of different ways (regular, random,
large/small, small/large).
Fixes #16699 and #18960 by ensuring that we release our HTTP
`spawn_local` tasks when the HTTP resource is dropped.
Because our cancel handle was being projected from the resource via
`RcMap`, the resource was never `Drop`ped. By splitting the handle out
into its own `Rc`, we can avoid keeping the resource alive and let it
drop to cancel everything.
This implements HTTP/2 prior-knowledge connections, allowing clients to
request HTTP/2 over plaintext or TLS-without-ALPN connections. If a
client requests a specific protocol via ALPN (`h2` or `http/1.1`),
however, the protocol is forced and must be used.
