trillium_client/response_body.rs
1use crate::{Error, Pool, pool::PoolEntry};
2use encoding_rs::Encoding;
3use futures_lite::{AsyncRead, AsyncReadExt, AsyncWriteExt};
4use std::{
5 fmt::{self, Debug, Formatter},
6 io, mem,
7 pin::{Pin, pin},
8 sync::{
9 Arc,
10 atomic::{AtomicBool, Ordering},
11 },
12 task::{Context, Poll, Wake, Waker, ready},
13 time::{Duration, Instant},
14};
15use trillium_http::{
16 Body, BodySource, Headers, HttpConfig, MutCow, ReceivedBody, ReceivedBodyState,
17};
18use trillium_server_common::{Runtime, Transport, url::Origin};
19
20/// A response body received from a server.
21///
22/// Most of the time this represents a body that will be read from the underlying transport, but it
23/// can also wrap an override body installed by middleware via [`ConnExt::set_response_body`]
24/// — e.g. cache hits, mocked responses, or circuit-breaker short-circuits. Reads, encoding
25/// handling, and `max_len` enforcement work transparently across both cases.
26///
27/// [`ConnExt::set_response_body`]: crate::ConnExt::set_response_body
28///
29/// ```rust
30/// use trillium_client::Client;
31/// use trillium_testing::{client_config, with_server};
32///
33/// with_server("hello from trillium", |url| async move {
34/// let client = Client::new(client_config());
35/// let mut conn = client.get(url).await?;
36/// let body = conn.response_body(); //<-
37/// assert_eq!(Some(19), body.content_length());
38/// assert_eq!("hello from trillium", body.read_string().await?);
39/// Ok(())
40/// });
41/// ```
42///
43/// ## Bounds checking
44///
45/// Every `ResponseBody` has a maximum length beyond which it will return an error, expressed as a
46/// u64. To override this on the specific `ResponseBody`, use [`ResponseBody::with_max_len`] or
47/// [`ResponseBody::set_max_len`]. The bound is enforced on override bodies as well as
48/// transport-backed ones, so a user-set memory cap holds even when middleware has replaced the
49/// body with externally-sourced bytes.
50pub struct ResponseBody<'a> {
51 inner: ResponseBodyInner<'a>,
52 /// Set on `'static` Received bodies built via
53 /// [`Conn::take_response_body`][crate::Conn::take_response_body]. `recycle` and `Drop`
54 /// consult it to decide whether to drain (keepalive) or close the underlying transport.
55 /// `None` for borrowed bodies (the conn handles their cleanup) and for Override bodies (no
56 /// transport is attached at this layer — `take_response_body` already evicted any leftover
57 /// transport before returning).
58 cleanup: Option<CleanupContext>,
59 /// Trailers harvested off the inner [`ReceivedBody`] when it reaches `End`. The
60 /// EOF-driven recycle in `poll_read` moves the `ReceivedBody` out before the caller can
61 /// observe its trailers, so they're captured here to outlive it and surfaced through
62 /// [`BodySource::trailers`].
63 trailers: Option<Headers>,
64}
65
66#[allow(clippy::large_enum_variant)]
67enum ResponseBodyInner<'a> {
68 Received(ReceivedBody<'a, Box<dyn Transport>>),
69 Override(OverrideBody<'a>),
70 Closing(Pin<Box<dyn Future<Output = ()> + Send + Sync + 'static>>),
71 Closed,
72}
73
74type H1Pool = Pool<Origin, Box<dyn Transport>>;
75
76/// Carries everything `Drop for ResponseBody` and [`ResponseBody::recycle`] need to release
77/// a transport without re-deriving from a [`crate::Conn`] (which is gone by then).
78///
79/// `pool_origin: Some` means "keepalive transport, pool is configured — insert here on
80/// completion." `None` means "close on completion" (non-keepalive *or* no pool). The same
81/// instance is cloned into the body's `on_completion` callback in
82/// [`Conn::take_received_body`][crate::Conn::take_received_body], so the user-driven
83/// drain/read paths and the Drop/recycle paths share one source of truth for what to do
84/// with the transport when the body finishes.
85#[derive(Clone)]
86pub(crate) struct CleanupContext {
87 pub(crate) runtime: Runtime,
88 pub(crate) h1_pool_origin: Option<(H1Pool, Origin)>,
89 /// Idle lifetime stamped onto a pooled h1 transport, counted from the moment it returns to
90 /// the pool. `None` disables expiry. Sourced from
91 /// [`Client::h1_idle_timeout`][crate::Client::h1_idle_timeout].
92 pub(crate) h1_idle_timeout: Option<Duration>,
93}
94
95/// The expiry [`Instant`] to stamp on an h1 [`PoolEntry`] returning to the pool now, given the
96/// configured idle timeout — i.e. `now + timeout`, or `None` when expiry is disabled.
97fn pool_expiry(h1_idle_timeout: Option<Duration>) -> Option<Instant> {
98 h1_idle_timeout.map(|d| Instant::now() + d)
99}
100
101impl CleanupContext {
102 /// Hand a freshly-released transport off to its destination — pool insert (sync) or
103 /// spawn close.
104 pub(crate) fn handoff(&self, mut transport: Box<dyn Transport>) {
105 match &self.h1_pool_origin {
106 Some((pool, origin)) => {
107 log::trace!("body transferred, returning to pool");
108 pool.insert(
109 origin.clone(),
110 PoolEntry::new(transport, pool_expiry(self.h1_idle_timeout)),
111 );
112 }
113 None => {
114 self.runtime.clone().spawn(async move {
115 log_close_result(transport.close().await);
116 });
117 }
118 }
119 }
120}
121
122pub(crate) struct OverrideBody<'a> {
123 body: MutCow<'a, Body>,
124 encoding: &'static Encoding,
125 max_len: u64,
126 initial_len: usize,
127 max_preallocate: usize,
128}
129
130impl AsyncRead for OverrideBody<'_> {
131 fn poll_read(
132 mut self: Pin<&mut Self>,
133 cx: &mut Context<'_>,
134 buf: &mut [u8],
135 ) -> Poll<io::Result<usize>> {
136 let remaining = self.max_len.saturating_sub(self.body.bytes_read());
137 if remaining == 0 && !buf.is_empty() {
138 return Poll::Ready(Err(io::Error::other(Error::ReceivedBodyTooLong(
139 self.max_len,
140 ))));
141 }
142 let cap = remaining.min(buf.len() as u64) as usize;
143 Pin::new(&mut *self.body).poll_read(cx, &mut buf[..cap])
144 }
145}
146
147impl<'a> OverrideBody<'a> {
148 pub(crate) fn new(
149 body: impl Into<MutCow<'a, Body>>,
150 encoding: &'static Encoding,
151 http_config: &HttpConfig,
152 ) -> Self {
153 Self {
154 body: body.into(),
155 encoding,
156 max_len: http_config.received_body_max_len(),
157 max_preallocate: http_config.received_body_max_preallocate(),
158 initial_len: http_config.received_body_initial_len(),
159 }
160 }
161}
162
163impl Debug for ResponseBody<'_> {
164 fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
165 match &self.inner {
166 ResponseBodyInner::Received(rb) => f.debug_tuple("ResponseBody").field(rb).finish(),
167 ResponseBodyInner::Override(o) => f
168 .debug_struct("ResponseBody (override)")
169 .field("body", &*o.body)
170 .field("encoding", &o.encoding.name())
171 .field("max_len", &o.max_len)
172 .finish(),
173 ResponseBodyInner::Closing(_) => f.debug_tuple("ResponseBody (closing)").finish(),
174 ResponseBodyInner::Closed => f.debug_tuple("ResponseBody (closed)").finish(),
175 }
176 }
177}
178
179impl AsyncRead for ResponseBody<'_> {
180 fn poll_read(
181 mut self: Pin<&mut Self>,
182 cx: &mut Context<'_>,
183 buf: &mut [u8],
184 ) -> Poll<io::Result<usize>> {
185 let mut bytes = 0;
186 loop {
187 match &mut self.inner {
188 ResponseBodyInner::Received(rb) => bytes = ready!(Pin::new(rb).poll_read(cx, buf))?,
189 ResponseBodyInner::Override(o) => bytes = ready!(Pin::new(o).poll_read(cx, buf))?,
190 ResponseBodyInner::Closing(fut) => {
191 ready!(fut.as_mut().poll(cx));
192 self.inner = ResponseBodyInner::Closed;
193 break;
194 }
195
196 ResponseBodyInner::Closed => break,
197 };
198
199 // Inline transport settlement — see take_received_body's `cleanup` param for
200 // why this isn't done via on_completion.
201 if bytes == 0
202 && let Some((mut rb, cleanup)) = self.prepare_for_recycle()
203 && rb.state() == ReceivedBodyState::End
204 && let Some(mut transport) = rb.take_transport()
205 {
206 self.trailers = Pin::new(&mut rb).trailers();
207 if let Some((pool, origin)) = cleanup.h1_pool_origin {
208 pool.insert(
209 origin,
210 PoolEntry::new(transport, pool_expiry(cleanup.h1_idle_timeout)),
211 );
212 } else {
213 self.inner = ResponseBodyInner::Closing(Box::pin(async move {
214 log_close_result(transport.close().await);
215 }));
216 }
217 } else {
218 break;
219 }
220 }
221
222 Poll::Ready(Ok(bytes))
223 }
224}
225
226impl ResponseBody<'_> {
227 fn take_inner(&mut self) -> ResponseBodyInner<'_> {
228 mem::replace(&mut self.inner, ResponseBodyInner::Closed)
229 }
230
231 fn max_preallocate(&self) -> usize {
232 match &self.inner {
233 ResponseBodyInner::Received(rb) => rb.max_preallocate(),
234 ResponseBodyInner::Override(override_body) => override_body.max_preallocate,
235 _ => 0,
236 }
237 }
238
239 fn max_len(&self) -> u64 {
240 match &self.inner {
241 ResponseBodyInner::Received(rb) => rb.max_len(),
242 ResponseBodyInner::Override(override_body) => override_body.max_len,
243 _ => 0,
244 }
245 }
246
247 fn initial_len(&self) -> usize {
248 match &self.inner {
249 ResponseBodyInner::Received(rb) => rb.initial_len(),
250 ResponseBodyInner::Override(override_body) => override_body.initial_len,
251 _ => 0,
252 }
253 }
254
255 fn encoding(&self) -> &'static Encoding {
256 match &self.inner {
257 ResponseBodyInner::Received(rb) => rb.encoding(),
258 ResponseBodyInner::Override(override_body) => override_body.encoding,
259 _ => encoding_rs::UTF_8,
260 }
261 }
262
263 /// Similar to [`ResponseBody::read_string`], but returns the raw bytes. This is useful for
264 /// bodies that are not text.
265 ///
266 /// You can use this in conjunction with `encoding` if you need different handling of malformed
267 /// character encoding than the lossy conversion provided by [`ResponseBody::read_string`].
268 ///
269 /// An empty or nonexistent body will yield an empty Vec, not an error.
270 ///
271 /// # Errors
272 ///
273 /// This will return an error if there is an IO error on the underlying transport such as a
274 /// disconnect.
275 ///
276 /// This will also return an error if the length exceeds the maximum length. To configure the
277 /// value on this specific request body, use [`ResponseBody::with_max_len`] or
278 /// [`ResponseBody::set_max_len`]
279 pub async fn read_bytes(mut self) -> Result<Vec<u8>, Error> {
280 let mut vec = if let Some(len) = self.content_length() {
281 if len > self.max_len() {
282 return Err(Error::ReceivedBodyTooLong(self.max_len()));
283 }
284
285 let len =
286 usize::try_from(len).map_err(|_| Error::ReceivedBodyTooLong(self.max_len()))?;
287
288 Vec::with_capacity(len.min(self.max_preallocate()))
289 } else {
290 Vec::with_capacity(self.initial_len())
291 };
292
293 self.read_to_end(&mut vec).await?;
294
295 Ok(vec)
296 }
297
298 /// Reads the entire body to a `String`.
299 ///
300 /// Uses the encoding determined by the content-type (mime) charset. If an encoding problem
301 /// is encountered, the returned `String` will contain utf8 replacement characters.
302 ///
303 /// Note that this can only be performed once per Conn, as the underlying data is not cached
304 /// anywhere. This is the only copy of the body contents.
305 ///
306 /// An empty or nonexistent body will yield an empty String, not an error
307 ///
308 /// # Errors
309 ///
310 /// This will return an error if there is an IO error on the
311 /// underlying transport such as a disconnect
312 ///
313 ///
314 /// This will also return an error if the length exceeds the maximum length. To configure the
315 /// value on this specific response body, use [`ResponseBody::with_max_len`] or
316 /// [`ResponseBody::set_max_len`].
317 pub async fn read_string(self) -> Result<String, Error> {
318 let encoding = self.encoding();
319 let bytes = self.read_bytes().await?;
320 let (s, _, _) = encoding.decode(&bytes);
321 Ok(s.to_string())
322 }
323
324 /// Set the maximum content length to read, returning self
325 ///
326 /// This protects against a memory-use denial-of-service attack wherein an untrusted peer sends
327 /// an unbounded request body. This is especially important when using
328 /// [`ResponseBody::read_string`] and [`ResponseBody::read_bytes`] instead of streaming with
329 /// `AsyncRead`.
330 ///
331 /// The default value can be found documented [in the trillium-http
332 /// crate](https://docs.trillium.rs/trillium_http/struct.HttpConfig#method.received_body_max_len)
333 #[must_use]
334 pub fn with_max_len(mut self, max_len: u64) -> Self {
335 self.set_max_len(max_len);
336 self
337 }
338
339 /// Set the maximum content length to read
340 ///
341 /// This protects against a memory-use denial-of-service attack wherein an untrusted peer sends
342 /// an unbounded request body. This is especially important when using
343 /// [`ResponseBody::read_string`] and [`ResponseBody::read_bytes`] instead of streaming with
344 /// `AsyncRead`.
345 ///
346 /// The default value can be found documented [in the trillium-http
347 /// crate](https://docs.trillium.rs/trillium_http/struct.HttpConfig#method.received_body_max_len)
348 pub fn set_max_len(&mut self, max_len: u64) -> &mut Self {
349 match &mut self.inner {
350 ResponseBodyInner::Received(rb) => {
351 rb.set_max_len(max_len);
352 }
353 ResponseBodyInner::Override(o) => {
354 o.max_len = max_len;
355 }
356 _ => {}
357 }
358 self
359 }
360
361 /// The trailers received after the response body, if any.
362 ///
363 /// Returns `None` until the body has been read to end-of-stream, and only on protocols
364 /// that delivered a trailer section (HTTP/1.1 chunked with trailers, HTTP/2, HTTP/3).
365 /// Reading the body via [`read_string`](Self::read_string)/[`read_bytes`](Self::read_bytes)
366 /// consumes it, so to observe trailers drive the body to completion through its
367 /// [`AsyncRead`](futures_lite::AsyncRead) interface and then call this.
368 pub fn trailers(&self) -> Option<&Headers> {
369 match &self.inner {
370 ResponseBodyInner::Received(rb) => rb.trailers_ref(),
371 // Captured off the inner ReceivedBody when it was recycled at end-of-stream.
372 _ => self.trailers.as_ref(),
373 }
374 }
375
376 /// The content-length of this body, if available.
377 ///
378 /// Usually derived from the content-length header. If the response uses
379 /// transfer-encoding chunked, this will be `None`.
380 pub fn content_length(&self) -> Option<u64> {
381 match &self.inner {
382 ResponseBodyInner::Received(rb) => rb.content_length(),
383 ResponseBodyInner::Override(o) => o.body.len(),
384 _ => None,
385 }
386 }
387
388 fn prepare_for_recycle(
389 &mut self,
390 ) -> Option<(
391 ReceivedBody<'static, Box<dyn Transport + 'static>>,
392 CleanupContext,
393 )> {
394 let cleanup = self.cleanup.take()?;
395
396 let ResponseBodyInner::Received(rb) = self.take_inner() else {
397 return None;
398 };
399
400 let rb = rb.try_into_owned()?;
401
402 Some((rb, cleanup))
403 }
404}
405
406async fn drain(rb: &mut ReceivedBody<'static, Box<dyn Transport + 'static>>) -> io::Result<u64> {
407 let copy_loops_per_yield = rb.copy_loops_per_yield();
408 trillium_http::copy(rb, futures_lite::io::sink(), copy_loops_per_yield).await
409}
410
411/// Best-effort synchronous drain of whatever body remainder is already buffered.
412///
413/// Small responses are often slurped in full during head parsing, so reaching
414/// `End` frequently requires no transport IO at all — just consuming buffered
415/// bytes. This drives [`drain`] with a wake-flag waker: cooperative yields
416/// re-poll immediately, and the first genuine IO `Pending` (or a drain error)
417/// stops the attempt, leaving the body wherever it got to for an async caller
418/// to finish. Never parks and never performs IO waits, so it is safe to call
419/// from `Drop` in any context.
420fn drain_buffered(rb: &mut ReceivedBody<'static, Box<dyn Transport + 'static>>) {
421 struct WakeFlag(AtomicBool);
422 impl Wake for WakeFlag {
423 fn wake(self: Arc<Self>) {
424 self.wake_by_ref();
425 }
426
427 fn wake_by_ref(self: &Arc<Self>) {
428 self.0.store(true, Ordering::Relaxed);
429 }
430 }
431
432 let flag = Arc::new(WakeFlag(AtomicBool::new(false)));
433 let waker = Waker::from(flag.clone());
434 let mut cx = Context::from_waker(&waker);
435 let mut drain = pin!(drain(rb));
436 loop {
437 match drain.as_mut().poll(&mut cx) {
438 Poll::Ready(_) => return,
439 Poll::Pending if flag.0.swap(false, Ordering::Relaxed) => {}
440 Poll::Pending => return,
441 }
442 }
443}
444
445/// Report the result of closing a transport we're discarding. `NotConnected` is the expected
446/// "already closed" signal from a finished multiplexed stream — notably h3/QUIC, whose `close`
447/// (unlike h2's) isn't idempotent and errors once the stream has finished — so it's absorbed at
448/// trace. Other errors are unexpected and warned.
449fn log_close_result(result: io::Result<()>) {
450 match result {
451 Ok(()) => {}
452 Err(e) if e.kind() == io::ErrorKind::NotConnected => {
453 log::trace!("transport already closed during recycle: {e}");
454 }
455 Err(e) => log::warn!("transport close failed during recycle: {e}"),
456 }
457}
458
459async fn recycle(
460 mut rb: ReceivedBody<'static, Box<dyn Transport + 'static>>,
461 h1_pool_origin: Option<(H1Pool, Origin)>,
462 h1_idle_timeout: Option<Duration>,
463) {
464 if let Some((pool, origin)) = h1_pool_origin {
465 match drain(&mut rb).await {
466 Ok(drained) => {
467 if rb.state() == ReceivedBodyState::End
468 && let Some(transport) = rb.take_transport()
469 {
470 log::trace!(
471 "drained {drained} bytes, returning transport to pool for {origin:?}"
472 );
473 pool.insert(
474 origin,
475 PoolEntry::new(transport, pool_expiry(h1_idle_timeout)),
476 );
477 return;
478 }
479 }
480 Err(e) => log::warn!("drain failed during recycle: {e}"),
481 }
482 }
483
484 if let Some(mut transport) = rb.take_transport() {
485 log_close_result(transport.close().await);
486 }
487}
488
489impl Drop for ResponseBody<'_> {
490 fn drop(&mut self) {
491 let Some((mut rb, cleanup)) = self.prepare_for_recycle() else {
492 return;
493 };
494
495 // Sync fast path for reclaiming an owned http/1.1 keepalive body: consume any
496 // remainder that's already buffered, and if that reaches End, pool the transport
497 // before drop returns — so a subsequent request on this client deterministically
498 // reuses it. Bodies with bytes still on the wire fall through to a spawned drain
499 // (Drop can't await IO), as does the no-pool path (`transport.close()` is async).
500 if let Some((pool, origin)) = cleanup.h1_pool_origin {
501 if rb.state() != ReceivedBodyState::End {
502 drain_buffered(&mut rb);
503 }
504
505 if rb.state() == ReceivedBodyState::End
506 && let Some(transport) = rb.take_transport()
507 {
508 pool.insert(
509 origin,
510 PoolEntry::new(transport, pool_expiry(cleanup.h1_idle_timeout)),
511 );
512 return;
513 }
514
515 cleanup
516 .runtime
517 .spawn(recycle(rb, Some((pool, origin)), cleanup.h1_idle_timeout));
518 } else {
519 cleanup
520 .runtime
521 .spawn(recycle(rb, None, cleanup.h1_idle_timeout));
522 }
523 }
524}
525
526impl BodySource for ResponseBody<'static> {
527 fn trailers(self: Pin<&mut Self>) -> Option<Headers> {
528 let this = self.get_mut();
529 match &mut this.inner {
530 ResponseBodyInner::Received(rb) => Pin::new(rb).trailers(),
531 ResponseBodyInner::Override(o) => o.body.trailers(),
532 // Recycled at EOF — trailers were captured off the ReceivedBody before it was
533 // moved out. See `ResponseBody::trailers`.
534 _ => this.trailers.take(),
535 }
536 }
537}
538
539impl<'a> From<ReceivedBody<'a, Box<dyn Transport>>> for ResponseBody<'a> {
540 fn from(received_body: ReceivedBody<'a, Box<dyn Transport>>) -> Self {
541 Self {
542 inner: ResponseBodyInner::Received(received_body),
543 cleanup: None,
544 trailers: None,
545 }
546 }
547}
548
549impl<'a> From<OverrideBody<'a>> for ResponseBody<'a> {
550 fn from(o: OverrideBody<'a>) -> Self {
551 Self {
552 inner: ResponseBodyInner::Override(o),
553 cleanup: None,
554 trailers: None,
555 }
556 }
557}
558
559impl ResponseBody<'static> {
560 pub(crate) fn received_owned(
561 body: ReceivedBody<'static, Box<dyn Transport>>,
562 cleanup: CleanupContext,
563 ) -> Self {
564 Self {
565 inner: ResponseBodyInner::Received(body),
566 cleanup: Some(cleanup),
567 trailers: None,
568 }
569 }
570
571 /// Drains and pools the underlying transport when worthwhile, closes it otherwise.
572 ///
573 /// Use this to release a keepalive transport synchronously before reissuing a request on
574 /// the same client — the redirect/retry handler pattern. For an h1.1 keepalive transport
575 /// this drives the body to EOF and returns the transport to the pool. For a non-keepalive
576 /// transport this calls `transport.close()` directly without draining (since draining
577 /// would just waste bytes on a connection we're about to close).
578 ///
579 /// For an Override body (cache hit, mocked response, tee), this is a no-op — the body's
580 /// own components handle their own teardown when dropped.
581 pub async fn recycle(mut self) {
582 let Some((rb, cleanup)) = self.prepare_for_recycle() else {
583 return;
584 };
585
586 recycle(rb, cleanup.h1_pool_origin, cleanup.h1_idle_timeout).await;
587 }
588}
589
590impl<'a> IntoFuture for ResponseBody<'a> {
591 type IntoFuture = Pin<Box<dyn Future<Output = Self::Output> + Send + 'a>>;
592 type Output = trillium_http::Result<String>;
593
594 fn into_future(self) -> Self::IntoFuture {
595 Box::pin(async move { self.read_string().await })
596 }
597}
598
599const _: fn() = || {
600 fn assert_send_sync<T: Send + Sync + ?Sized>() {}
601 assert_send_sync::<ResponseBody<'static>>();
602 assert_send_sync::<ResponseBody<'_>>();
603};