* perf: bypass header key normalization for canonical literals
http.Header's Get, Set, Add and Del run textproto.CanonicalMIMEHeaderKey
on every call. When the key is a canonical literal the result is the key
itself, so the scan is pure overhead. Access the map directly instead.
Only hot, allocation-free paths are converted, and only where the key is
a canonical literal. Call sites taking a key from config, a template or
the request itself still go through the normalizing methods, since they
need it for correctness.
allHeaderValues and lastHeaderValue dropped their http.CanonicalHeaderKey
call for the same reason: all callers pass canonical literals.
benchstat, Apple M5 Pro, go1.27.1:
PrepareRequest-15 659.4n ± 2% -> 457.7n ± 2% -30.58% (p=0.000 n=12)
AcceptedEncodings-15 330.8n ± 5% -> 309.4n ± 7% -6.44% (p=0.019 n=10)
ResponseWriterWriteHeader/event-stream-15 1007.5n ± 3% -> 732.5n ± 1% -27.30%
ResponseWriterWriteHeader/informational-15 60.0n ± 3% -> 48.0n ± 2% -20.01%
ResponseWriterWriteHeader/not-modified-15 214.3n ± 6% -> 175.3n ± 4% -18.20%
ResponseWriterWriteHeader/connect-15 60.2n ± 3% -> 51.9n ± 4% -13.85%
ResponseWriterWriteHeader/html-15 206.5n ± 4% -> 197.9n ± 2% -4.19%
geomean -14.38%
Allocations are unchanged.
* tests: verify http.Header map keys are canonical
Indexing http.Header directly skips the canonicalization Get/Set/Add/Del
perform, which also puts the key outside what CI checks. canonicalheader
only inspects method calls. staticcheck's SA1008 covers reads, but stops
descending as soon as an assignment writes to a header, so it misses the
write, the delete builtin and the read-modify-write append. A key that is
not canonical in those positions is a silent lookup miss, not a compile
error.
Walk the loaded packages and verify every string constant used as a key in
an http.Header map operation matches http.CanonicalHeaderKey. A helper that
takes the key from its caller would hide the literal again, so a parameter
named with a canonical prefix opts its arguments into the same check.
The WebSocket headers Caddy deliberately writes with RFC 6455 casing are
allowed explicitly; they are only ever written out, never looked up against
a parsed header map.
Covers 64 operations, runs in under a second, and promotes golang.org/x/tools
from an indirect to a direct dependency at the version already in the module
graph.
* headercheck: resolve aliases of http.Header
* headercheck: check http.Header composite literal keys
* headercheck: note overlap with canonicalheader
* headercheck: drop parameter-name check
Helpers now take the indexed values, so every key sits in an index expression the analyzer and the linter already check.
* headercheck: scope RFC 6455 exceptions to their call sites
A key-based allow-list also accepted reads such as h["Sec-WebSocket-Key"], which miss because parsed headers are canonical. Honor //nolint:canonicalheader on the line instead, which the linter will use too.
* encode: read every Accept-Encoding field line
AcceptedEncodings read only the first field line, so a client that split
the list across several lines had the rest ignored. RFC 9110 section 5.3
treats repeated list-field lines as one comma-joined list. Iterate over
the items of every line instead of joining them, and skip empty items,
which used to yield an empty encoding. file_server's precompressed
selection uses the same function and is fixed with it.
Fixes https://github.com/caddyserver/caddy/issues/8272
Co-authored-by: Kağan Küçük <kaankck@gmail.com>
* encode: read every Cache-Control and Content-Encoding field line
Only the first field line was read, but RFC 9110 section 5.3 makes
multiple lines of a list-based field equivalent to one comma-joined
value:
- Cache-Control: no-transform on a later line was ignored, so the
request or response was compressed anyway.
- Content-Encoding: an empty first line followed by a real coding let
the handler encode the body a second time.
* encode: parse Cache-Control no-transform per RFC 9111
The substring match missed No-Transform, though RFC 9111 section 5.2 compares directive names case-insensitively, and fired on text that is not the directive: longer tokens such as x-no-transform-ext, quoted arguments such as community="no-transform", and semicolon-separated values, which are not valid Cache-Control syntax. Split the field on commas outside quoted strings and match only a well-formed no-transform element.
* encode: keep accepted encoding preferences on the stack
AcceptedEncodings counted commas to size the preference slice, and the variable capacity put it on the heap. A constant capacity covers typical Accept-Encoding headers and lets the compiler keep the slice on the stack; append still moves it to the heap for longer ones. BenchmarkAcceptedEncodings: 151.8ns -> 120.5ns (-21%), 192 -> 64 B/op, 2 -> 1 allocs/op (n=8).
* encode: build the encoded ETag without fmt
fmt.Sprintf boxed both arguments and parsed the format on every encoded response with a strong ETag. Concatenation: 47ns -> 16.5ns, 56 -> 24 B, 3 -> 1 allocs.
* reverseproxy: access response Server and Via headers directly
finalizeResponse runs on every proxied response; skip the key normalization as prepareRequest already does on the request side.
* encode: access headers directly in Write, WriteHeader and Close
These run once per response; convert the remaining literal-key Get, Set and Add calls like the rest of the handler.
* encode: match Cache-Control no-transform with httpguts
httpguts.HeaderValuesContainsToken already compares list elements case-insensitively across field lines, so the hand-written list parser is unnecessary. A no-transform inside a comma-separated quoted argument now also disables encoding, which only errs toward not transforming.
* encode: inline the no-transform check
isEncodeAllowed only wrapped one httpguts call. Its unit test now drives the response directive through ServeHTTP instead.
---------
Co-authored-by: Kağan Küçük <kaankck@gmail.com>
* reverseproxy: strip hop-by-hop headers when copying response headers
copy_response_headers copied the upstream header map before
finalizeResponse could strip hop-by-hop fields. Clone and strip
first, so include and exclude cannot forward them, and leave the
upstream response unchanged for later handlers in the route.
Co-authored-by: Grok 4.7 <noreply@x.ai>
* reverseproxy: close backend body and hijacked socket on upgrade setup failures
When an upstream response indicates an HTTP upgrade (101 Switching Protocols),
handleUpgradeResponse takes ownership of res.Body and subsequently hijacks
the client connection.
However, if any setup check fails prior to starting handleUpgradeTunnel
(such as mismatched upgrade headers, non-writable body assertion failure,
hijack unsupported, or handshake buffer flush errors), the function returned
early without closing res.Body or the hijacked client socket.
Because calling Hijack() permanently disables net/http's automatic connection
cleanup, and raw TCP sockets are not closed by the garbage collector, these
early failure exits leaked file descriptors into CLOSE_WAIT/ESTABLISHED.
Fix this by:
1. Adding a deferred cleanup guard that closes res.Body and any hijacked
front connection unless the tunnel successfully takes ownership.
2. Wrapping connections with onceCloseConn to ensure idempotent closure
while preserving optional CloseWrite (TCP half-close) support.
3. Adding TestUpgradeResponseCleanupOnFailure to verify cleanup.
* reverseproxy: test post-hijack cleanup and remove once-close wrappers
Cover handshake flush and buffered backend-write failures after successful hijacking, checking that both connections close. Keep setup-failure cleanup while using the original connections directly.
Code and tests were AI-assisted using an OpenAI coding assistant in Pi.
* reverseproxy: emit OpenTelemetry client spans for upstream requests
When the request carries a valid span context (set by the tracing
handler), each upstream round trip, including retries, is recorded as a
client span and that span's context is propagated to the upstream.
Requests that are not traced use the transport directly.
The transport is wrapped at the call site rather than replacing
h.Transport, so transport interface checks and health checks are
unaffected. The tracer provider comes from the span in the request
context, so no global provider is needed.
Fixes#7223
* reverseproxy: assert client span URL, upstream address and status
Use a provisioned HTTPTransport so the test fails if roundTrip stops
setting the scheme before the traced transport records url.full.
* reverseproxy: add license header to tracing test
When a forward proxy is configured (network_proxy, forward_proxy_url or
HTTP_PROXY), net/http wraps dial errors in a *net.OpError with Op
"proxyconnect". tryAgain used a type assertion to detect DialError, so it
missed the wrapped error and treated it as a failure after the request
was sent. Non-GET requests were then not retried, even though no
connection to the proxy was ever established.
Use errors.As so a DialError is recognized anywhere in the error chain.
Errors that do not contain a DialError are unaffected, so requests that
may have reached the proxy or upstream are still not retried by default.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fastcgi: derive SCRIPT_FILENAME from the normalized request path
Path matchers evaluate a normalized request path — MatchPath applies
normalizeWindowsPath and then CleanPath before comparing — but the FastCGI
transport split r.URL.Path verbatim. Caddy does not rewrite r.URL.Path itself,
so the two disagreed whenever the path contained dot segments after the split
token: /admin.php/../../other.php was matched as /other.php while
SCRIPT_FILENAME resolved to <root>/admin.php, and on Windows the same held for
/admin.php\..\other.php. A route gated by a path matcher could therefore execute
a script that matcher never evaluated.
php_fastcgi generates a rewrite route that normally normalizes the path first,
but that route can fail to match, so the transport establishes the invariant
itself rather than inheriting it. The reason it can fail to match is addressed
in the following commit; this change does not depend on it.
caddyhttp.NormalizedPath is added so a consumer acting on a path a matcher has
already judged derives it exactly as the matcher did, rather than reassembling
the two steps and risking drift.
Scope — two divergences of the same class are NOT addressed here:
* SCRIPT_NAME remains a prefix of the request path, so a matcher naming a
script exactly (path /admin.php) is still reachable via PATH_INFO at
/admin.php/anything. Matchers protecting a script need a trailing wildcard
regardless; that is inherent to CGI, not to this bug.
* On Windows, NTFS 8.3 short names are not rejected here. fileserver's
ServeHTTP rejects them (and ':' for ADS) precisely because matchers cannot
see through them, so GET /ADMIN~1.PHP is judged as /admin~1.php by a
path matcher while NTFS resolves SCRIPT_FILENAME to admin.php. Closing that
means sharing fileserver's hasWindowsShortName across packages, which is a
larger change than this fix; it is reported alongside this one and a patch
is available on request.
Side effects: PATH_INFO and DOCUMENT_URI are now derived from the normalized
path, so repeated slashes are merged and "." segments removed (/index.php//a//b
yields PATH_INFO=/a/b rather than //a//b), and on Windows per-component trailing
dots and spaces are folded. An application routing on a raw PATH_INFO will see
this. An empty URL.Path stays empty rather than becoming ".". REQUEST_URI is
unchanged and still carries the original target.
Tests: a buildEnv regression over literal and percent-encoded dot segments,
covering that legitimate PATH_INFO is preserved and an empty path stays empty; a
Windows-only counterpart whose cases each fail without the Windows
normalization; and an end-to-end test of a basic_auth-protected php_fastcgi
route against a FastCGI responder reporting the SCRIPT_FILENAME it received. The
end-to-end test uses `index off` so no rewrite route exists, ensuring it
exercises the transport rather than the rewrite.
* fileserver: make first_exist_fallback fall back when the file is absent
The try_policy first_exist_fallback is documented to match unconditionally,
falling back to the last try_files entry when no earlier candidate exists. It
did not: candidates are built with fs.Glob, which returns only paths that exist,
so an absent final entry left the list empty and the fallback branch — nested
inside the loop over that list — was unreachable.
php_fastcgi hard-codes this policy so its generated rewrite route always runs;
when the fallback failed, that rewrite was silently skipped.
Keep the pattern itself as a candidate when nothing globbed. This also aligns
Linux with Windows, where globResults is populated without an existence check
and the fallback already fired.
Behaviour change worth calling out: on a php_fastcgi site whose root has no
index.php (or no file matching the configured index), the rewrite now always
runs, so a request for a path that exists neither on the filesystem nor as a PHP
script is proxied to the backend as <root>/index.php instead of falling through
to file_server and returning 404. That is what the documented fallback promises,
but for a site of standalone .php scripts plus static assets it is a visible
change: unmatched requests now reach PHP-FPM.
The other try policies still test existence explicitly, so a candidate that does
not exist is skipped — but note this applies to ANY pattern that expands to
nothing, not only a malformed glob: a file whose literal name contains glob
metacharacters can now match under first_exist, largest_size and smallest_size
where previously globbing found it only via expansion.
* fileserver: synthesize the literal candidate only for the fallback entry
Address review: the previous commit appended the pattern itself as a
candidate inside makeCandidates whenever globbing found nothing. Since
makeCandidates is shared by every try_policy, first_exist, largest_size,
smallest_size and most_recently_modified could then stat the literal
pattern, so a file whose name contains glob metacharacters (for example
"lit[1].txt") matched where it did not before.
Restore makeCandidates to its original behaviour and build the literal
candidate only for the final entry of first_exist_fallback, the one
entry that is documented to match unconditionally. Placeholders in that
entry are expanded without glob escaping, since nothing is globbed, so
a request for /a[b].txt falls back to /a[b].txt rather than /a\[b].txt.
The path is still cleaned and joined with SanitizedPathJoin, so it
cannot escape the root.
This supersedes the last paragraph of the previous commit message: the
other policies now behave exactly as they did before that commit. The
php_fastcgi behaviour change described there still applies.
Add a regression test that creates a file literally named "lit[1].txt"
and checks that only the fallback entry of first_exist_fallback matches
it by its literal name, including the split_path remainder.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fastcgi: add license headers to new test files
Address review: phpfastcgi_test.go and fastcgi_windows_test.go were
added without the Apache license header used across the codebase. On
the Windows test, the header goes above the build constraint, matching
replacer_fuzz.go.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fastcgi: avoid hardcoded Unix paths in dot-segment tests
Address review: the new tests compared SCRIPT_FILENAME against strings
with forward slashes ("/srv/other.php", a "/admin.php" suffix), but the
transport makes the root absolute and joins it with filepath.Join, so
on Windows the value is C:\srv\other.php and the assertions would fail.
Use a temporary directory as the root and build the expected paths with
filepath instead.
Also lower the bcrypt cost of the integration test's basic_auth hash to
the minimum. At cost 14 the authenticated request exceeded the 5s test
client timeout under -race.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A CGI script that redirects by printing only a Location header got a
200 response with that header, which clients do not follow. RFC 3875
section 6.2.3 says the server sends 302 Found for this response, and
net/http/cgi, nginx and Apache do so.
With request_buffers or response_buffers set to unlimited, the error
from reading the body was discarded and whatever had been read was
treated as the complete body. A request cut short by request_body
max_size was forwarded truncated with a 200, and a truncated chunked
upstream response was sent to the client as complete.
Handle a read error the same way the limited case already does.
* caddy: pass ListenQUIC options as a struct
Collapse the TLS config, packet-conn wrappers, and Allow0RTT arguments
into ListenQUICOptions so the signature matches Listen's ctx/portOffset/
ListenConfig prefix. Zero-value opts keep the historical default of
allowing 0-RTT. The API remains experimental.
* caddyhttp: advertise WebTransport in HTTP/3 SETTINGS
Add github.com/quic-go/webtransport-go dep (built on quic-go, same
maintainers) and call webtransport.ConfigureHTTP3Server on the
http3.Server. This advertises WebTransport enablement in SETTINGS,
enables HTTP/3 DATAGRAMs, and stashes the *quic.Conn in each request's
context — a prerequisite for a later WebTransport-aware handler or
reverse-proxy transport to call webtransport.Server.Upgrade. Also
enable QUIC stream reset partial delivery, required by webtransport-go.
No user-visible behavior change: clients that don't speak WebTransport
ignore the extra SETTINGS, and no handler yet calls Upgrade.
Extract the http3.Server construction into buildHTTP3Server so the
SETTINGS assertions can be unit-tested without a live UDP listener.
* caddyhttp: add UnwrapResponseWriterAs helper
Go's type assertion `x.(T)` does not follow Unwrap() http.ResponseWriter
chains. Caddy wraps the writer multiple times (logging recorder,
intercept, encode, etc.), so code that needs interfaces implemented only
by the raw writer owned by the HTTP server — for example the
http3.Settingser/HTTPStreamer interfaces that webtransport.Server.Upgrade
type-asserts — cannot see through those wrappers.
UnwrapResponseWriterAs walks the Unwrap() chain and returns the first
writer that satisfies the requested interface (or the zero value if none
do). Mirrors the traversal http.ResponseController performs internally.
Used by upcoming WebTransport handler and reverse-proxy transport.
* caddyhttp: add terminating WebTransport handler
Introduces http.handlers.webtransport, an EXPERIMENTAL handler that
terminates an incoming WebTransport session on top of Caddy's HTTP/3
server and echoes bytes on each bidirectional stream. Primary use case
is as a test upstream for the forthcoming WebTransport reverse-proxy
transport; it also serves as the minimal proof that the server-side
WebTransport wiring works end-to-end.
Plumbing changes:
* caddyhttp.Server gains a *webtransport.Server field alongside
h3server. It's built in buildWebTransportServer(), wrapping the
existing http3.Server. Exposed via WebTransportServer() any on the
Server, so the caddyhttp public API does not name the upstream
webtransport-go type (per AGENTS.md).
* serveHTTP3 now runs a custom accept loop (serveH3AcceptLoop) that
dispatches each accepted QUIC connection to
webtransport.Server.ServeQUICConn instead of
http3.Server.ServeListener. The WebTransport server transparently
forwards non-WT streams to the underlying http3 request handler
(cost: one varint peek per stream), so behavior for non-WT clients
is unchanged.
* ListenQUIC enables EnableDatagrams and
EnableStreamResetPartialDelivery on the QUIC listener config.
These are capability bits negotiated during the QUIC handshake and
are prerequisites for any WebTransport session; they do not force
usage so non-WT H3 traffic is unaffected.
* Stop path closes wtServer after h3server.Shutdown to clean up any
remaining WebTransport session state.
The handler uses caddyhttp.UnwrapResponseWriterAs to reach the naked
http3.Settingser/HTTPStreamer writer through Caddy's wrapping chain
before calling webtransport.Server.Upgrade.
Includes unit tests for request-shape detection plus an integration
test (caddytest/integration/webtransport_test.go) that spins up a
Caddy HTTP/3 server with the handler, dials it with a real
webtransport.Dialer, and asserts end-to-end bidirectional-stream
echo.
* reverseproxy: add WebTransport upstream dialer helper
dialUpstreamWebTransport is a thin wrapper around webtransport.Dialer.Dial
that sets the QUIC config flags WebTransport requires (EnableDatagrams,
EnableStreamResetPartialDelivery) and forwards request headers on the
Extended CONNECT. Intended as an internal building block for the
upcoming WebTransport reverse-proxy transport; not yet wired into
ServeHTTP.
Unit-tested against an in-process webtransport.Server with a freshly
minted self-signed certificate. Covers: successful dial, header
forwarding, and connection-refused against an unbound port.
* reverseproxy: add WebTransport session pump
runWebTransportPump bridges two WebTransport sessions so every
bidirectional stream, unidirectional stream, and datagram opened on one
side is mirrored on the other. Uses six goroutines (bidi both ways, uni
both ways, datagrams both ways) and blocks until both sessions end.
Close propagation: when either session ends, the peer is closed via
CloseWithError. The code/message are read from the closing session's
stored close state (by probing AcceptStream with a short timeout),
since Receive{Datagram,UniStream} return the underlying stream error
rather than the SessionError and can win the propagation race. Close
propagation is best-effort for client-initiated close through a
Dialer-dedicated QUIC conn: webtransport-go tears down the QUIC
connection immediately after CloseWithError, so the upstream may
observe a QUIC ApplicationError before the WT_CLOSE_SESSION capsule is
parsed. The pump still closes the peer session; only the specific
error code may not survive.
Not yet wired into ServeHTTP.
Tests: topology of client -> frontend -> upstream where frontend runs
the pump. Exercises bidi both ways, uni client-to-upstream, datagram
round-trip, CloseWithError propagation both ways, and a basic
goroutine-leak check.
* reverseproxy: wire WebTransport pump into ServeHTTP
Extends the http reverse-proxy transport with a webtransport boolean
that opts the upstream into WebTransport passthrough. Must be combined
with versions: ["3"]; WebTransport rides on HTTP/3 exclusively.
When enabled, Handler.ServeHTTP detects Extended CONNECT with
:protocol=webtransport early — before any of the normal round-trip
machinery — and branches to serveWebTransport, which:
1. Pulls the *webtransport.Server off caddyhttp.Server (via
WebTransportServer()) and errors out cleanly if HTTP/3 isn't
enabled on the frontend.
2. Picks a single upstream through the configured load-balancer.
No retries: a failed dial closes the client session and returns.
3. Walks the response-writer Unwrap() chain to reach the raw http3
writer and calls webtransport.Server.Upgrade to terminate the
incoming session.
4. Uses dialUpstreamWebTransport to open a session to the selected
upstream, forwarding request headers on the Extended CONNECT.
5. Runs runWebTransportPump between the two sessions and blocks
until both close.
The transport's wtTLSConfig is built at Provision time from the
existing TLS config (same path h3Transport already uses) and reused
for every session.
Tests: adds TestWebTransport_ReverseProxyEndToEnd which spins up a
single Caddy instance with two HTTP/3 servers — one proxy on :9443,
one terminating echo upstream on :9444 — and drives a real
webtransport.Dialer through the proxy to assert end-to-end
bidirectional-stream echo.
* reverseproxy: add webtransport Caddyfile subdirective
Adds a `webtransport` subdirective to the `transport http {}` block of
reverse_proxy that sets the new WebTransport bool on the transport.
Takes no arguments; exclusivity with versions 3 is enforced at
Provision time so parse order doesn't matter.
Example:
reverse_proxy https://backend:9443 {
transport http {
versions 3
webtransport
tls_insecure_skip_verify
}
}
Includes a Caddyfile-to-JSON adapt test round-tripping the new
subdirective.
* reverseproxy: apply standard request preparation to WebTransport CONNECT
The WebTransport proxy path previously bypassed the request-preparation
pipeline that normal reverse-proxy traffic runs through. Reuse it so
`header_up`, `X-Forwarded-For`/`Host`/`Proto`, `Via`, `Rewrite`, the
`{http.reverse_proxy.upstream.*}` placeholders, dynamic upstreams,
`countFailure`, and the `{http.reverse_proxy.duration{_ms}}` timing
placeholder all behave the same as on the regular path.
Retries, `handle_response`, and response-header ops are intentionally
not run here — a WebTransport session has no HTTP response body to
post-process and is not idempotent. Integration test exercises the
header-forwarding contract end-to-end through a standalone (non-Caddy)
WebTransport upstream so the forwarded Extended CONNECT can be
inspected.
* reverseproxy: dial WebTransport upstream before upgrading client
Reorder serveWebTransport so the upstream is dialed first. If the
upstream is unreachable or refuses the CONNECT, a proper 5xx is returned
from the handler — the client's Dial() surfaces the real status instead
of a successful upgrade followed by an opaque session close.
Also apply `h.Headers.Response` (gated by `Require`, if configured)
against the upstream response status/headers; the ops run on the
client-visible response headers, which webtransport.Server.Upgrade
flushes with the 200 OK. If the client-side upgrade fails after the
upstream dial succeeded, close the upstream session cleanly.
Integration test drives a dial to an unbound loopback port and asserts
the client sees a 5xx status instead of a bare session close.
* reverseproxy: track WebTransport sessions in upstream in-flight counters
Bracket the pump's lifetime with Host.countRequest(±1) and
incInFlightRequest/decInFlightRequest so WT sessions participate in the
same accounting as the normal proxy path:
- MaxRequests gating (Upstream.Full) now blocks WT sessions past the
cap, instead of silently failing open.
- LeastConn / FirstAvailable selection sees WT load, instead of seeing
busy upstreams as idle.
- Admin /reverse_proxy/upstreams reports WT sessions under num_requests.
Integration test holds an upstream session open via a standalone WT
server, polls the admin API to assert num_requests increments during
the session and drops back to 0 after close.
* reverseproxy: extract shared upstream-selection helpers
The WebTransport proxy path in serveWebTransport duplicated the
dynamic-upstream-fallback block and the {http.reverse_proxy.upstream.*}
replacer-variable block from proxyLoopIteration. Francis flagged this
as a maintenance burden in review of #7669.
Extract two helpers:
* resolveUpstreams(r) returns the candidate upstream set — dynamic
when configured (with provisioning + fallback-on-error), static
otherwise. Caller runs the LB selection policy, since the two call
sites diverge on how selection failure is reported (retry loop vs.
fast 502 for long-lived WT sessions).
* setUpstreamReplacerVars(repl, up, di) publishes the seven
placeholders describing the selected upstream.
Both are used by proxyLoopIteration and serveWebTransport with
identical semantics to the inlined code they replace. No behavior
change for either path.
* reverseproxy: inline WebTransport protocol const and writer interface
Francis pointed out in review of #7669 that importing the whole
modules/caddyhttp/webtransport package solely to pull in one constant
and one interface wasn't worthwhile.
Move both into webtransport_transport.go as unexported identifiers
(webtransportProtocol, webtransportWriter). This removes reverseproxy's
dependency on the caddywt package and clears the way for moving the
echo handler itself out of the production module tree.
No behavior change.
* caddyhttp: move WebTransport echo handler to integration tests
Francis pointed out in review of #7669 that the echo handler — which
exists solely as a test upstream for the WebTransport reverse-proxy
tests — should not be a full-fledged module registered in every Caddy
binary. Mirroring the mockdns_test.go pattern, move it into a _test.go
file under caddytest/integration/.
The module ID http.handlers.webtransport is now registered only when
the integration test binary is built, which is when
caddytest/integration/webtransport_test.go references it by ID string
in its JSON configs. Production Caddy builds no longer include it.
Changes:
* New file: caddytest/integration/webtransport_echo_test.go —
contains the WebTransportEcho handler, its types and interface
guards, the isWebTransportEchoUpgrade helper, and the unit tests
that used to live in the deleted package's handler_test.go.
* Deleted: modules/caddyhttp/webtransport/ (handler.go + handler_test.go).
* Removed the blank import from modules/caddyhttp/standard/imports.go.
The Protocol const and Writer interface that this package used to
export were inlined into reverseproxy's own files in a preceding
commit, so nothing else depends on the deleted package.
* caddyhttp, reverseproxy: gate WebTransport behind enable_webtransport server flag
steadytao raised an architectural concern in review of #7669: the PR
put experimental WebTransport handling directly into Caddy's core
HTTP/3 accept path, so every HTTP/3 deployment paid for the feature
whether or not they used it.
Collapse the enablement surface to a single server-level opt-in that
matches Caddy's existing precedent for protocol-level features
(`protocols`, `allow_0rtt`, `enable_full_duplex`), and detect the
request shape at the handler the same way `reverse_proxy` detects a
WebSocket upgrade today — no per-handler config flag.
Core HTTP/3 path changes (modules/caddyhttp/server.go):
* New `EnableWebTransport bool` field on Server, marked EXPERIMENTAL.
* buildHTTP3Server now only calls webtransport.ConfigureHTTP3Server
and sets EnableStreamResetPartialDelivery when the flag is true.
When false, the constructed http3.Server is bit-for-bit identical
to the pre-WebTransport implementation.
* wtServer is constructed only when the flag is true.
* serveH3AcceptLoop falls back to http3.Server.ServeListener when
the flag is false — no varint peek, no per-connection dispatch.
Caddyfile wiring (caddyconfig/httpcaddyfile/serveroptions.go):
* New `enable_webtransport` global server option, modeled on
`enable_full_duplex`.
Reverse-proxy simplifications (modules/caddyhttp/reverseproxy/):
* Removed HTTPTransport.WebTransport field and its Provision-time
exclusivity check (no longer needed; H3 is validated separately).
* Removed the `webtransport` Caddyfile subdirective under
`transport http { }` — this neutralizes the prior commit that
introduced it.
* Removed Handler.webtransportEnabled cache. ServeHTTP now branches
on isWebTransportExtendedConnect(r) alone, matching how the
WebSocket upgrade branch works.
* serveWebTransport gains fail-fast guards with clear errors when
the parent server has enable_webtransport=false or when the
handler's transport does not include HTTP/3.
Tests:
* Existing TestServer_BuildHTTP3ServerEnablesWebTransport now sets
EnableWebTransport=true explicitly; new
TestServer_BuildHTTP3ServerWithoutWebTransport locks in the
regression guard that flag-off produces the pre-PR http3.Server.
* Integration tests updated: enable_webtransport: true added to
every H3 server block; "webtransport": true dropped from the
reverse_proxy transport JSON (auto-detected now).
* Caddyfile adapt test for the deleted `webtransport` subdirective
is removed; `enable_webtransport` is added to the existing
global_server_options_single adapt test alongside its peers.
No runtime behavior change when enable_webtransport is false. Diff
against master on the core HTTP/3 path is effectively zero in that
configuration.
* caddyhttp: micro-benchmark HTTP/3 server construction with/without WebTransport
Adds a hermetic pair of benchmarks on buildHTTP3Server to provide
quantitative evidence for the claim that deployments with
enable_webtransport=false pay no cost for the WebTransport feature.
Results on Apple M4, go1.25, -count=5:
BenchmarkBuildHTTP3Server_WebTransportOff ~70 ns/op 392 B/op 3 allocs/op
BenchmarkBuildHTTP3Server_WebTransportOn ~144 ns/op 600 B/op 6 allocs/op
The Off path is about half the cost on every dimension, confirming
that the work skipped when the flag is false is the
webtransport.ConfigureHTTP3Server call plus EnableStreamResetPartialDelivery.
Absolute cost is a one-time per-server setup so either branch is
negligible in practice, but the asymmetry locks in a regression guard:
a future refactor that accidentally re-enables the WT configuration
unconditionally would show up as a jump in the Off numbers.
This benchmark does not exercise the per-stream dispatch cost inside
webtransport.Server.ServeQUICConn — that would require a full QUIC
setup to measure in isolation and is follow-up work.
* reverseproxy: collapse WebTransport handler into main proxy loop
The WT path duplicated upstream resolution, LB selection, header ops,
replacer vars, and in-flight counters. Route WT through the shared
ServeHTTP -> proxyLoopIteration -> reverseProxy flow and swap RoundTrip
for a small webTransportHijack that only does WT-specific work (writer
unwrap, upstream dial, client upgrade, pump).
Rename roundtripSucceededError -> terminalError. The existing name
described when it was emitted (after a successful round-trip); the
new name describes its contract with the retry loop (stop looping,
propagate error unchanged). The WebTransport upgrade case is a second
natural caller for that same signal.
Comes with two behavior improvements that fall out of the collapse:
- WT upstream dial failures now surface as DialError, so the loop
can fail over across upstreams like normal proxies (today: 502).
- Passive health checks apply to WT dials (dial-failure countFailure
and UnhealthyLatency on dial duration) via the shared path.
Addresses reviewer feedback that the duplicated setup was a
maintenance risk.
* reverseproxy: inline upstream-resolution helpers back into the proxy loop
The earlier #7669 revision extracted resolveUpstreams and
setUpstreamReplacerVars to share between the WebTransport path and the
normal proxy path. A later commit collapsed the WebTransport path into
the main proxy loop, so the helpers became single-callers — they no
longer earn their abstraction.
Inline both back into proxyLoopIteration. The dynamic-upstream block
absorbs the CachingUpstreamSource cache-reset logic added by #7662,
which was preserved through the rebase.
No behavior change.
* caddyhttp: clarify WebTransportServer doc covers both disabled paths
WebTransportServer returns nil for two distinct configurations:
EnableWebTransport=false on the server, and HTTP/3 not in use. The
prior doc only mentioned the latter. Spell out both, per review.
* reverseproxy: drop nil-logger guard in runWebTransportPump
Production and test callers both pass a non-nil *zap.Logger. The
defensive nil check silently swapped in a no-op logger, which can mask
a real bug at the call site. Document the precondition in the doc
comment instead, per review.
* reverseproxy: use sync.WaitGroup.Go in WebTransport pump
Replace the wg.Add(6) + six "go func() { defer wg.Done(); fn() }()"
calls with sync.WaitGroup.Go (Go 1.25+, the project's minimum
version). Same number of goroutines; less boilerplate.
* caddyhttp: drop defensive checks in UnwrapResponseWriterAs
Per review, the nil-pointer guard and the next == w self-reference
guard defended against shapes that no in-tree ResponseWriter wrapper
produces. Drop both, simplifying the loop. Also remove the
self-reference test that exercised the now-removed guard.
* reverseproxy: use r.URL.RequestURI() to build the WebTransport upstream URL
Replace buildWebTransportUpstreamURL with an inline call to
r.URL.RequestURI() at the single call site. RequestURI handles the
empty-path → "/" fixup and query suffixing already, and uses
EscapedPath so encoded path segments (e.g. %2F) are preserved on the
upstream — fixing a small correctness gap where the previous helper
read r.URL.Path (decoded).
Aligns with how Caddy's normal proxy path lets stdlib render the
outbound request line via URL.RequestURI semantics.
* reverseproxy: use wg.Go in spliceBidi
* reverseproxy: expand SNI placeholders on the WebTransport upstream dial
The normal HTTP/3 path expands placeholders in tls_server_name at dial
time via a custom h3Transport.Dial hook (#7737). The WebTransport path
dials through its own webtransport.Dialer and bypasses that hook, so a
placeholder like {http.request.host} reached the QUIC ClientHello
literally. Expand it in webTransportHijack before dialing, cloning the
shared TLS config so concurrent sessions are unaffected.
Adds an integration test asserting the upstream observes the resolved
SNI rather than the literal "{...}" string.
* reverseproxy: bump webtransport-go to v0.11.0 (WebTransport draft-15)
webtransport-go v0.11.0 implements draft-15, which renames the Extended
CONNECT :protocol token from "webtransport" to "webtransport-h3" (the
server keeps accepting the legacy token for older clients). Detect both
tokens in the reverse-proxy upgrade check and in the test echo handler.
Pulls quic-go v0.59.1 -> v0.60.0, required by webtransport-go v0.11.0;
its only breaking change is requiring Go 1.25 (already met), with no API
changes affecting Caddy core. Also picks up the v0.11.0 fix for a
stream-close race on concurrent session close, relevant to the long-lived
pump.
* caddyhttp: log WebTransport QUIC connection serve errors
* caddy: gate QUIC WebTransport capabilities behind an explicit ListenQUIC parameter
* caddyhttp: fall back to plain HTTP/3 when a reused QUIC listener lacks WebTransport
* caddyhttp: enable WebTransport via a webtransport config object
Replace the enable_webtransport bool with a presence-gated
WebTransportConfig object so later options can be added without
renaming the key. JSON is "webtransport": {}, Caddyfile is
webtransport (optional empty block).
* reverseproxy: use origReq for the client WebTransport upgrade
Upgrade is the client-facing session: CheckOrigin and the CONNECT
shape must come from the downstream request. The upstream-directed
clone remains used for the dial URL, header_up, and SNI expansion.
* reverseproxy: relay WebTransport application protocol negotiation
Parse WT-Available-Protocols from the downstream request, set it on
the upstream Dialer, and copy WT-Protocol from the upstream 200 onto
the client response. The shared webtransport.Server does not select
a protocol of its own, so MOQ/MediaMTX clients receive the upstream's
choice instead of WT_ALPN_ERROR.
* caddytest: skip origin check on standalone WebTransport backends
header_up Host rewrites used by proxy tests would 400 the upstream
Upgrade via CheckOrigin. Test backends are not browsers.
* caddytest: cover MediaMTX Host-rewrite + moqt-19 through the WT proxy
Both of elee's #7669 reports fire together: header_up rewrites Host
(Upgrade must use origReq) and the client offers moqt-19 (the proxy
must relay WT-Available-Protocols / WT-Protocol). A bidi echo after
negotiation proves the session is usable, not just headers.
* caddyhttp: break self-referential Unwrap loops
A wrapper whose Unwrap returns itself would spin forever. Restore the
cycle-break mholt applied on #7669, without bringing back the nil-pointer
guard Weidi asked to drop.
* caddyhttp: close WebTransport sessions before HTTP/3 shutdown
ServeQUICConn waits for active WT sessions, and with the default
eternal grace period HTTP/3 Shutdown would hang until every client
disconnected. Close the WebTransport server first so stop/reload
can drain.
* caddyhttp: serve plain HTTP/3 when a reused listener lacks WebTransport
A pooled QUIC listener created without DATAGRAMs cannot upgrade WT
sessions. Falling back on the WT-configured http3.Server still
advertised SETTINGS the socket cannot honor. Build a separate plain
HTTP/3 server for that listener and assert it omits WT SETTINGS.
* caddyhttp: visit Unwrap writers by pointer identity
Comparing writers with == panics on uncomparable types and only
catches a one-node cycle. Track pointer identity, bound the walk,
and add tests for an uncomparable self-unwrap and an A→B→A cycle.
RecordHijackedStatus lets Upgrade write through the naked HTTP/3
writer while the access-log recorder still observes the 200.
* reverseproxy: honor prepared Host, header_down, and WT bookkeeping
Dialer.Dial always sets Host from the URL, so header_up Host never
reached the upstream. Dial the real address via DialAddr and put
the prepared Host on the CONNECT.
Parse WT-Available-Protocols from the prepared request so header_up
applies to negotiation. Copy filtered upstream response headers
before header_down, record the hijacked 200 for access logs, and
run the same latency / circuit-breaker / passive-health path as
a normal round-trip.
* caddytest: cover WT reload, access-log status, Host, and header ops
Reload with an eternal grace period must close an in-flight session.
Access logs must record 200 after Upgrade. header_up Host and
WT-Available-Protocols must reach the upstream; header_down and
non-hop upstream headers must reach the client.
* reverseproxy: build WebTransport against quic-go v0.63
webtransport-go v0.13.0 renames Dialer to Transport. v0.11.1 does not
compile against quic-go v0.63.
* reverseproxy: close the upstream body when a WebTransport dial fails
Dial can return a response with an error when the peer rejects the
CONNECT. Close that body before the error is returned.
* caddyhttp: close WebTransport sessions without closing the HTTP/3 server
webtransport.Server.Close also closes the shared HTTP/3 server and aborts
every QUIC connection before the grace period applies. Close the registered
sessions, then let HTTP/3 shutdown drain ordinary requests.
* chore: modernise
Signed-off-by: Zen Dodd <mail@steadytao.com>
* reverseproxy: preserve WT-Available-Protocols across upstream retries
The proxy loop reuses the request and only recopies headers when header
ops are configured. Deleting the offer on that map hid it from the next
upstream.
* caddyhttp: close WebTransport sessions when the grace period ends
Eternal grace leaves an active session until the peer ends it. A
configured grace period closes it at the deadline, while the socket
is still up, so the close capsule can be delivered.
* reverseproxy: use terminalError when refusing an Incremental response
The proxy-loop collapse renamed roundtripSucceededError to terminalError.
Master's Incremental refusal still returned the old type, so the rebased
package did not compile.
* caddyhttp: don't let idle HTTP/3 conns pin WebTransport shutdown
- Count in-flight requests on ServeQUICConn connections and close idle ones when shutdown starts
- Delay that close by a second so a just-finished response can be read
- Retain plain HTTP/3 fallbacks for listeners without datagrams and shut them down with the same grace period
* reverseproxy: wait for WebTransport stream workers
- Join bidi and uni splice workers after the accept loops return
- Cancel the opposite copy on a hard splice error
---------
Signed-off-by: Zen Dodd <mail@steadytao.com>
Co-authored-by: tomholford <tomholford@users.noreply.github.com>
Co-authored-by: Zen Dodd <mail@steadytao.com>
* reverseproxy: extract DNS SRV+A discovery into a reusable package
Move the SRV and A/AAAA resolution + caching logic out of reverseproxy
into a new, transport-neutral package (dynamicupstreams) that returns
neutral targets instead of *reverseproxy.Upstream. reverseproxy's
SRVUpstreams and AUpstreams now delegate to it and build their own
upstreams from the targets; behavior is unchanged.
This lets other proxies (e.g. third-party layer4 proxies) reuse the same
DNS discovery + caching instead of copying it, reducing duplication and
maintenance burden. reverseproxy/upstreams.go shrinks substantially.
RFC for the de-duplication discussed in caddyserver/caddy-l4#429.
* test(dynamicupstreams): cover ResetAllSRV, filtered records, cache bound, concurrent refresh
Brings the package to 96.8% statement coverage. Adds:
- TestResetAllSRV: full-cache drop
- TestSRVFilteredRecords: LookupSRV partial-error semantics (usable
records returned, error downgraded to a warning)
- TestSRVCacheBound / TestACacheBound: eviction keeps the cache bounded
at maxCacheEntries
- TestSRVConcurrentRefreshDeduplicates / TestAConcurrentRefreshDeduplicates:
two goroutines missing the read-lock fast path trigger only one lookup
(verified under -race)
The grace-period and filtered-records logs now use a level-enabled logger
so the log-write branches execute. The only uncovered statements are the
two double-checked-locking re-check returns, which are reachable solely in
a narrow race window (the same untested idiom as reverseproxy's SRV/A).
* reverseproxy: move DNS SRV+A discovery into modules/internal/network
Per review feedback, the shared discovery code does not belong in a
top-level package: it is internal helper code, not part of Caddy's
runtime surface. Move it from the top-level dynamicupstreams package into
the existing modules/internal/network package and update reverse_proxy's
upstreams.go to consume it from there.
Being under modules/internal means third-party modules (e.g. layer4)
cannot import it yet; they keep their own copy until layer4 is part of
Caddy's standard distribution, at which point they can switch over.
* test(dynamicupstreams): correct the concurrent-refresh test's claim
The two concurrent-refresh tests documented themselves as covering the
write-lock re-check branch ("the second re-checks under the write lock"),
but they do not: while the first goroutine holds the write lock the second
cannot acquire the read lock either, so by the time it acquires srvMu the
cache is already populated and it is served by the read-lock fast path. The
50ms sleep was presented as letting the second goroutine "queue on the write
lock" but proved nothing.
Make the comments and failure messages honest: these tests assert
deduplication (two concurrent callers for one uncached key share a single
lookup) and, under -race, race-safety on the shared cache -- not a specific
locking branch. Ordering is enforced by the channels, so the sleep is removed.
No production code change.
* reverseproxy: preserve DNS discovery logging and grace behavior
Keep refresh and discovery logs in the DNS lookup path, retaining the original SRV field types. Preserve error suppression when an SRV lookup fails with grace enabled before any result has been cached.
Add regression tests for refresh-only logging and cold-cache grace handling.
AI-assisted-by: OpenAI Codex (GPT-6)
---------
Co-authored-by: Francis Lavoie <lavofr@gmail.com>
* reverseproxy: Only retry responses for expressions that use response data
Since #7569 every lb_retry_match set with an expression is evaluated against
upstream responses. An expression that only tests the request, such as
{http.request.method} == "POST", then matched every successful response:
the request was sent again to the next upstreams and the client got a 502
once retries ran out. Only treat expressions referencing {rp.*} placeholders
as response conditions.
* reverseproxy: Only count placeholders set from the response as response data
* reverseproxy: Read response placeholders from the checked expression
The expression matcher records the placeholder names its checked CEL
expression reads, so retry_match finds direct ph() calls and ignores
escaped placeholder text in string literals. A placeholder name only
known at evaluation time counts as response data.
* reverseproxy: Match response placeholder names exactly
Only the header placeholders are matched by prefix now, and both
upstream latency placeholders are listed. MatchExpression.Placeholders
returns a copy, so callers cannot change the provisioned matcher.
* reverseproxy: Keep own headers after 1xx responses
Since #6656 the 1xx handler clears the whole header map after writing the
informational response, which also drops headers set before proxying (for
example by the header handler). Restore those after clearing, as #5091 did.
* reverseproxy: Use maps.Copy to restore own headers
* reverseproxy: Check the 103 reaches the client in the 1xx headers test
Wrap backend dial failures as reverseproxy.DialError so non-GET requests can be retried when no connection was made.
AI-assisted implementation and test; reviewed by zemamba.
Signed-off-by: zemamba <zemamba@gmail.com>
* reverseproxy: Optionally detach stream (websockets) from config lifecycle
* lint
* Improved logging facilities
* Adjustments from Weidi's review
* record bytes read and written for response writers unless detached
* simplify streaming handling
* correctly close detached streams
* make handleUpgradeTunnel a standalone func
* fix tests
* fix tests
* change the log level if hijacking without writing a status code first
* only clean up connections when stopped
* Move type and const down to the bottom
* Rename to tunnelTracker, reflow some comments
* Add note about capturing h
* Rename to stream_detached
* More comments
* reverseproxy: Preserve half-close behavior after rebase
Keep master half-close propagation and its unit coverage in the refactored tunnel lifecycle. Adapt response finalization tests to the selected upstream argument.
AI-assisted-by: Codex (GPT-6)
* reverseproxy: Fix detached stream cleanup and recorder chains
Snapshot the tracker registry before closing connections, and unregister empty trackers outside their mutex. Mark trackers stopped at cleanup entry so upstream removal closes detached streams even with stream_close_delay, while attached streams retain their delay.
Detach every supported response writer in the middleware chain so inner recorders do not retain state or count detached traffic. Add regressions for all four failures and an access-log/encode/intercept reload integration test.
AI-assisted-by: Codex (GPT-6)
---------
Co-authored-by: WeidiDeng <weidi_deng@icloud.com>
* Move websocket header normalization later
A regression since #6621, these headers can be rewritten by cloneHeaders
and by .Add from user and transport ops. Moving this closer to where
it's round-tripped feels like the nicest solution.
I've put the harness I used to reproduce the issue at
https://codeberg.org/nfreya/caddy-websocket-headers-repro
* Integration test for normalized websocket headers
---------
Co-authored-by: Zen Dodd <mail@steadytao.com>
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
A request whose body has no length of its own — Transfer-Encoding: chunked, or
HTTP/2 and HTTP/3 requests sent without the header — cannot be forwarded over
FastCGI until it has been buffered in full, because CGI/1.1 requires
CONTENT_LENGTH and php-fpm hangs when it is absent or wrong and the body is not
empty. When request_buffers is too small to hold the whole body, the length
stays unknown and the request is refused with 411.
The refusal said none of that. RoundTrip passes r.ContentLength to Post, which
is -1 for such a request, so the operator got a bare "411 Length Required" with
either no error at all or strconv's "invalid syntax" on a value they never
wrote. On #7386 that sent two people looking for the fault in their backend.
The 411 now carries the reason and the remedy, and the unknown-length case is
told apart from a genuinely malformed value. ParseUint becomes ParseInt plus an
explicit negative check, which is strictly tighter: values above MaxInt64 used
to be accepted and are unreachable anyway, since CONTENT_LENGTH is always
FormatInt of an int64 by the time Do sees it.
No status code changes. Every request that was refused before is still refused;
raising request_buffers is still what makes a large chunked body work.
Tests: the reachable path through Post, the unusable CONTENT_LENGTH values Do
itself guards against, an integration test driving a chunked body through a
fastcgi reverse_proxy over a unix socket at each side of the buffer boundary
(including request_buffers -1, which succeeds at any size and is the remedy),
and a boundary test recording that a body exactly the size of the buffer counts
as partial.
That last one is deliberately not changed here. Telling "exactly the limit"
apart from "more to come" costs a read that a paused stream may never answer,
and bufferedBody also backs response_buffers: measured against a body that
delivers exactly the limit and stops, the current code hands the buffered
prefix on immediately, while peeking one byte first delivers nothing at all.
Co-authored-by: Aditya <205600203+Rohilalala@users.noreply.github.com>
Co-authored-by: Zen Dodd <mail@steadytao.com>
* reverseproxy: build TLS client config once so HTTP/3 keeps tls_trust_pool
reuse the config built for the HTTP/1.1+2 transport and give the HTTP/3
transport a clone of it to honour the tls_trust_pool
Signed-off-by: Timo Stark <tippexs91@googlemail.com>
Co-authored-by: Faiyaz Rahman <faiyazrahman03@gmail.com>
* Adding issue reference
---------
Signed-off-by: Timo Stark <tippexs91@googlemail.com>
Co-authored-by: Faiyaz Rahman <faiyazrahman03@gmail.com>
* reverseproxy: propagate TCP half-close on upgraded streams
handleUpgradeResponse starts one copy goroutine per direction but returns as
soon as the first one reports a result, and returning runs two unconditional
deferred closes (the client connection and the backend connection). A client
that ends its send direction with a TCP half-close therefore causes the whole
tunnel to be torn down: the backend connection is reset mid-conversation and
whatever it was about to send is never delivered.
Treat a clean EOF as the end of one direction rather than the end of the
tunnel. When a copy finishes without error, propagate the half-close to the
destination with CloseWrite where the connection supports it, and keep waiting
for the other direction so pending bytes can still drain. Real copy errors, the
stream timeout and request cancellation/shutdown still tear down immediately.
If the destination does not support CloseWrite, the peer cannot be told that
the direction ended, so the previous behavior is kept and the tunnel is torn
down rather than held open until the stream timeout.
This is the same class that was fixed upstream in net/http/httputil
(https://go.dev/issue/35892); Caddy has its own upgraded-stream copier, so the
standard library fix does not apply here.
Fixes#8026
* reverseproxy: follow net/http/httputil shape for the half-close
Replace the copyResult struct and the closeWriter/closeWrite helpers with
upstream's plain error channel and errCopyDone sentinel, propagating
CloseWrite inside the copier, as requested in review.
Behaviour is unchanged. The wait is repeated twice inside the existing
select because handleUpgradeResponse also selects on the stream timeout
and sizes the channel at 2 so both goroutines can exit when it fires
(#7418), which net/http/httputil has no equivalent of.
* reverseproxy: explain the repeated wait next to it
Move the reasoning for why the wait is repeated inside the select, rather
than written as upstream's two-line form, into a comment beside the loop
so a future reader does not have to find the review discussion.
* reverseproxy: add an upgraded-stream half-close integration test
Drive a real upgraded stream through the handler and check that closing one
direction is propagated to the other side instead of tearing the tunnel down.
Mirrors net/http/httputil's TestReverseProxyWebSocketHalfTCP, which covers the
same class upstream (https://go.dev/issue/35892): a backend that hijacks and
hands back a *net.TCPConn, the handler in front of it, and a client that dials
it directly, then the four close-read / close-write combinations.
Against the unfixed handler the two half-close cases fail - the surviving
direction never delivers its pending bytes - while the two close-read cases
pass, so the test isolates exactly the reported behaviour.
* reverseproxy: give the half-close test reads a deadline
Without one, a regression blocks in ReadFull until the package timeout
instead of failing the test, which is the flakiness an end-to-end test of
this kind is usually accused of. Ten seconds is far above any real latency
here and turns a hang into a clear failure.
* reverseproxy: hijack the test backend via http.NewResponseController
Matches how the rest of the tree reaches Hijack (responsewriter.go,
caddyauth.go, streaming.go itself) and keeps working if the writer is
wrapped, where a direct http.Hijacker assertion would not.
The active health check body is user-supplied and commonly JSON, but it
was expanded with ReplaceAll, which blanks any `{...}` the replacer does
not recognize. Only globals are available here, so a JSON
health_request_body was sent as an empty body.
This is not limited to bodies without placeholders. A body that does use
one is mangled too: `{"token":"{env.SOME_VAR}"}` is sent as `"}`, since
the opening `{"token":"` is consumed as an unrecognized placeholder
before the real one is reached.
Use ReplaceKnown, matching the header values a few lines below, so
unrecognized braces are left intact and real placeholders still expand.
Only the body is changed. The Host header on the same path keeps
ReplaceAll deliberately: Go's HTTP server blanks a Host containing
braces, so leaving `{...}` intact there would send a Host the upstream
discards, where blanking it lets Go fall back to the URL host.
* Implement HTTPoxy mitigation in FastCGI
Added HTTPoxy mitigation to prevent trusting client-supplied Proxy header for HTTP_PROXY environment variable.
* Implement test for HTTPoxy vulnerability protection
Add test to ensure HTTPoxy vulnerability is mitigated by dropping client-supplied Proxy headers.
* gofmt: format code
* revert unrelated gofmt change to replacer_test.go
* reverseproxy: isolate active health-check state per distinct check config
Multiple reverse_proxy handlers configured with different active health
checks (health_uri, health_headers, ...) against the same upstream dial
address currently share a single Host in the global pool, so one
handler's failing probes mark the address unhealthy for every other
handler. Key the pool by dial address plus a stable fingerprint of the
active health-check config, so distinct checks get independent health
state.
The fingerprint is strictly internal to pool identity: the Prometheus
upstreams_healthy label and the /reverse_proxy/upstreams admin endpoint
continue to report the plain dial address, unchanged.
Dynamic upstreams are intentionally out of scope here: they resolve
through a separate per-lookup path (dynamicHosts) and collapsing there
has different lifetime semantics; noted for a follow-up.
Fixes#7870
* reverseproxy: use strings.Cut in hostKeyAddress
Satisfies the modernize linter; behaviour is unchanged, since Cut returns
the whole string when the separator is absent.
* reverseproxy: expose the health-check fingerprint as a public discriminator
Health state is now kept per (dial address, active health check config),
but both user-visible surfaces still reported address alone:
- caddy_reverse_proxy_upstreams_healthy was labeled only by upstream, so
every handler sharing an address wrote the same series concurrently and
the reported value was whichever updater ran last. The metric gains a
health_check label carrying the config fingerprint ("" when no active
checks), so each health target owns its series; aggregate across checks
with sum/min by (upstream).
- /reverse_proxy/upstreams reported one entry per pool key but with only
the plain address, so consumers indexing by address silently discarded
all but one entry. Entries now carry health_check (omitted when empty),
and the endpoint documents that (address, health_check) is the entry's
identity — one entry per health target, deliberately not aggregated,
since any aggregation here would be lossy and undocumented.
Tests: two handlers on one address with different checks must produce two
metric series reflecting their own state (fails if the fingerprint is
dropped from the label), and two admin entries distinguished by non-empty
fingerprints.
* reverseproxy: narrow the fix to per-Upstream active health counters
Move the consecutive active pass/fail counters from Host onto Upstream,
alongside the active unhealthy state that already lives there, instead of
re-keying the global host pool.
Host is keyed by dial address alone, but an active health check is
configured per handler, so two handlers dialing the same address with
different health_uri or health_headers share those counters and can push
each other over their own thresholds. Upstream is already per-handler and
already carries the active unhealthy flag, so the counters belong next to
it and the pool keeps its plain dial-address keys.
This drops the host key fingerprint and its exposure in the metric label
and the admin upstreams endpoint; the metric series identity is left for
separate consideration.
---------
Co-authored-by: SillyZir <269283839+SillyZir@users.noreply.github.com>
Co-authored-by: Zen Dodd <mail@steadytao.com>
Caddy's underscore header filter (GHSA-f59h-q822-g45g) only checked
for `_`. PHP folds `.` to `_` when registering $_SERVER keys the same
way CGI/FastCGI folds `-` to `_`, so a dotted alias (e.g. Remote.User)
survived the filter and collided with the legitimate hyphenated
header once it reached a PHP/FastCGI backend, bypassing forward_auth
copy_headers the same way the underscore alias did.
Extends the filter to drop `.` symmetrically, adds an
`expected_dot_headers` allowlist mirroring `expected_underscore_headers`,
and handles header names containing both separators (only an exact
allowlist entry is honored there, since a prefix glob's free-form
suffix can't be vetted for an embedded second separator).
Root cause identified by @iliaal in the FrankenPHP advisory
GHSA-49wc-4hcv-v58q.
Preallocate the sources slice in MultiUpstreams.Provision and the
netAddrs slice in UpstreamResolver.ParseAddresses now that their final
lengths are known, avoiding the intermediate append regrowth
allocations. For a 4-address resolver this reduces ParseAddresses from
11 to 9 allocs/op and 528 to 384 B/op; the benefit scales with the
number of sources/addresses.
| before | after |
| sec/op | sec/op vs base |
UpstreamResolverParseAddresses-8 | 2.289µ |1.946µ ~ (p=0.165) noisy |
B/op: 528 → 384 -27.27% (p=0.000)
allocs/op: 11 → 9 -18.18% (p=0.000)
Adds BenchmarkUpstreamResolverParseAddresses as per policy.
Build the Via request and response headers with strconv-based
concatenation instead of fmt.Sprintf, avoiding fmt's reflection and
formatting overhead on every proxied request/response. Same single
allocation, ~40% faster for that construction in isolation.
| before (fmt) | after (strconv) |
| sec/op | sec/op vs base |
PrepareRequest-8 | 6.136µ ± 11% | 4.491µ ± 33% -26.80% (p=0.007 n=10)|
B/op: 1008 -> 1008 (unchanged)
allocs/op: 12 -> 12 (unchanged)
No behavior change. Adds BenchmarkPrepareRequest, as per policy, exercising the
request preparation path.
* feat(fastcgi): populate SERVER_ADDR by default
Populate the SERVER_ADDR FastCGI environment variable using the local socket address of the incoming connection from the request context. This improves out-of-the-box compatibility for PHP applications that rely on $_SERVER['SERVER_ADDR'] (such as legacy frameworks and custom telemetry packages) when migrating from Nginx or Apache.
Co-developed-by: Gemini AI <renich+gemini@woralelandia.com>
Signed-off-by: Rénich Bon Ćirić <renich@woralelandia.com>
* fix(fastcgi): validate SERVER_ADDR IP format and add test coverage
Validate that SERVER_ADDR is only populated when the local address parses as a valid IP address, preventing Unix socket paths from being assigned if SplitHostPort fails. Add test cases covering IPv4, IPv6, missing context, Unix sockets, and explicit overrides.
Co-developed-by: Gemini AI <renich+gemini@woralelandia.com>
Signed-off-by: Rénich Bon Ćirić <renich@woralelandia.com>
The random_choose selection policy is meant to implement
power-of-d-choices: sample d available upstreams uniformly, then pick
the least-loaded of that sample. The sampling loop, however, was not a
correct reservoir sample (Algorithm R): it never filled the first k
reservoir slots unconditionally, instead writing every candidate to a
random slot j = rand(i+1), which can evict an earlier candidate while
leaving another slot nil. It also derived j from the upstream's index
in the pool rather than from the number of available upstreams seen,
skewing the sample whenever unavailable upstreams precede available
ones.
As a result the reservoir frequently held fewer than min(k, available)
upstreams, so the least-load comparison often never happened. Most
notably, with two upstreams and 'random_choose 2' (the canonical
power-of-two-choices setup), half of all requests were routed to the
more-loaded upstream even when it was saturated and the other idle --
identical behavior to plain 'random'. The nil slots this leaves behind
were also the cause of the panic reported in #3810, which was patched
by skipping nils in leastRequests rather than by fixing the sampling.
Replace the loop with a standard Algorithm R reservoir sample over the
available upstreams: fill the first k slots, then replace a random slot
with probability k/n. Every available upstream is now sampled uniformly
and the reservoir always holds min(k, available) candidates.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* reverseproxy: validate on weighted_round_robin policy
Validate that weighted_round_robin has a non-zero total weight.
This prevents configurations such as:
weighted_round_robin 0 0
from being accepted and causing a divide-by-zero panic during request handling.
* test: validation test on zero weight upstreams.
* test: provision called instead of totalweight setting
* reverseproxy: validate on negative upstream weights
* test: regression test on weighted_round_robin selection policy
