Traefik: respondingTimeouts.readTimeout is not applied to HTTP/3, leaving slow-body uploads unbounded

Summary

There is a medium severity vulnerability in Traefik's HTTP/3 entry points: the respondingTimeouts settings were not applied to the HTTP/3 request path. readTimeout in particular is on by default at 60s and is documented as bounding the time to read the entire request including its body, but it is enforced as a deadline on the TCP connection, which cannot reach a QUIC stream, and Traefik's HTTP/3 server was constructed with no timeout of any kind. An unauthenticated client that trickles a request body therefore holds a request open for as long as it chooses, and with it one upstream connection per request, at negligible cost to itself. Backends with bounded connection pools are the practical pressure point.

The HTTP/3 path lost these timeouts in v2.8.2, when a quic-go API change removed the embedded http.Server that had carried them; every release from v2.8.2 onward is affected, and releases before v2.8.2 are not. Traefik v2.8.2 through v2.10.x and v3.0 through v3.6 are affected and are no longer maintained: they will not receive a patch on their own line, and the remedy for their users is to upgrade to v2.11.56 or v3.7.12.

Patches

• https://github.com/traefik/traefik/releases/tag/v2.11.56
• https://github.com/traefik/traefik/releases/tag/v3.7.12

For more information

If you have any questions or comments about this advisory, please open an issue.

<details> <summary>Original Description</summary>

Summary

entryPoints.<name>.transport.respondingTimeouts.readTimeout is documented as:

"Set the timeouts for incoming requests to the Traefik instance. This is the maximum duration for reading the entire request, including the body." — Default: 60s

It is on by default and it works over HTTP/1.1 and HTTP/2. It has no effect on HTTP/3.

The consequence is not that a hardening option was left unset. It is that every Traefik deployment with http3 enabled carries a 60-second bound that the operator has every reason to believe is in force, and which is silently absent on that protocol. A single client trickling one body byte every few seconds holds a request open indefinitely, and with it one upstream connection per request.

readTimeout is applied as a deadline on the TCP connection. HTTP/3 does not have one, and Traefik's HTTP/3 server is constructed with no timeout of any kind.

Steps to reproduce

No containers, VMs or cloud services — the official release binary, curl, openssl, and a 68-line Python standard-library backend. Everything is attached.

bash reproduce.sh                 # readTimeout 5s, ~40 seconds
MODE=default bash reproduce.sh    # the stock 60s default, ~4 minutes

By hand:

1. Static config (conf/traefik.yml, complete and unredacted). Note there is no respondingTimeouts block at all — this is the documented 60s default:

global:
  checkNewVersion: false
  sendAnonymousUsage: false

log:
  level: DEBUG

entryPoints:
  websecure:
    address: ":8443"
    http3:
      advertisedPort: 8443

providers:
  file:
    filename: conf/dynamic.yml

api:
  dashboard: false

2. Dynamic config (conf/dynamic.yml):

http:
  routers:
    backend-router:
      rule: "PathPrefix(`/`)"
      service: backend-svc
      entryPoints: [websecure]
      tls: {}
  services:
    backend-svc:
      loadBalancer:
        servers:
          - url: "http://127.0.0.1:8080"
tls:
  certificates:
    - certFile: cert.pem
      keyFile: key.pem

3. A self-signed cert, so no CA install and no sudo:

openssl req -x509 -newkey rsa:2048 -keyout key.pem -out cert.pem -days 30 -nodes \
    -subj "/CN=localhost" -addext "subjectAltName=DNS:localhost,IP:127.0.0.1"

4. A backend that reads the request body before responding, as a real HTTP/1.1 server does (backend.py, standard library only). This matters: a backend that answers the request headers alone lets Traefik release the upstream connection immediately, which hides the behaviour entirely.

python3 backend.py 8080 &
./traefik --configFile=conf/traefik.yml

5. The same slow upload over each protocol. The hold must exceed the timeout under test, so 92 seconds against the 60-second default:

{ for i in $(seq 1 23); do printf 'x'; sleep 4; done; } | \
    curl -v -k -T - --http1.1     https://localhost:8443/

{ for i in $(seq 1 23); do printf 'x'; sleep 4; done; } | \
    curl -v -k -T - --http3-only  https://localhost:8443/

Result

Traefik v3.7.10 (langres, go1.26.5), official traefik_v3.7.10_linux_amd64.tar.gz, sha256 01811bb12d44f17280550f425f5e3128d6c325f2665c09e67a651ca535f490ce.

Upstream connection lifetime measured at the backend — the client's view only shows the absence of a server action, which proves nothing on its own:

The HTTP/1.1 column is the control and it is the point of the exercise: the timeout is demonstrably live on this exact binary, releasing the upstream at its configured value. The HTTP/3 request, in the same run against the same instance with the same setting, ran to completion with the upstream pinned throughout. Traefik returned 499 on the aborted HTTP/1.1 arm and took no action at all on HTTP/3 — no RST_STREAM, no H3_REQUEST_INCOMPLETE, no connection close.

The setting is not being silently discarded: Traefik's own DEBUG log prints the loaded static configuration including "respondingTimeouts":{"idleTimeout":"3m0s","readTimeout":"5s"}. Full curl -v output, backend logs and Traefik DEBUG logs for every arm are in evidence/.

Cause

readTimeout is a TCP connection deadline — pkg/server/server_entrypoint_tcp.go:273:

if e.transportConfiguration.RespondingTimeouts.ReadTimeout > 0 {
    err := writeCloser.SetReadDeadline(time.Now().Add(time.Duration(e.transportConfiguration.RespondingTimeouts.ReadTimeout)))

A deadline on a TCP connection cannot reach a QUIC stream.

And the HTTP/3 server is given no timeout of any kind — pkg/server/server_entrypoint_tcp_http3.go:65:

h3.Server = &http3.Server{
    Addr:      config.GetAddress(),
    Port:      config.HTTP3.AdvertisedPort,
    Handler:   httpsServer.Server.(*http.Server).Handler,
    TLSConfig: &tls.Config{GetConfigForClient: h3.getTLSConfigForClient},
    QUICConfig: &quic.Config{
        Allow0RTT: false,
    },
    ConnContext: func(ctx context.Context, c *quic.Conn) context.Context {

It reuses the HTTPS server's handler and inherits none of its timeouts. There is therefore no duration control on the HTTP/3 request path at all — not readTimeout, not idleTimeout, nothing.

Note that this cannot be fixed by passing a field through: quic-go's http3.Server exposes no request-read deadline. The remedy has to be enforced around the request body inside the handler, or upstream in quic-go.

Suggested remedy

In preference order:

1. Enforce readTimeout on the HTTP/3 path in the handler. Traefik already passes the HTTPS server's handler to http3.Server, so when RespondingTimeouts.ReadTimeout > 0 it can wrap r.Body for HTTP/3 requests in a reader that enforces the deadline.
2. Add a request-read deadline upstream in quic-go's http3.Server and pass it through.
3. At minimum, document it. See below — the current documentation does not tell an operator this.

A sketch of option 1

Offered as a description of the shape, not as a patch. I have not built or tested this against Traefik, and I am not going to present untested code as though I had. If a working, tested patch would be useful, say so and I will prepare one properly and verify it against the reproducer above.

Where the HTTP/3 handler is wired up — server_entrypoint_tcp_http3.go, around the http3.Server construction — the handler passed in could be wrapped so that, when ReadTimeout is configured, an HTTP/3 request body carries the same deadline the TCP path gets from SetReadDeadline:

// Roughly: for HTTP/3 requests only, and only when the timeout is set.
if readTimeout > 0 && r.ProtoMajor == 3 && r.Body != nil {
    r.Body = deadlineBody(r.Body, readTimeout)
}

The part worth knowing, because it is what makes the approach work rather than merely look tidy: closing the request body cancels the QUIC stream read. In quic-go, http3's body Close() calls str.CancelRead(...), which unblocks a Read that is already parked waiting on the client. So a timer that closes the body on expiry bounds both a client that trickles and a client that simply stops sending — the latter being the case a wrapper that only checks the clock on each returning Read would miss entirely.

Returning os.ErrDeadlineExceeded from the wrapped Read keeps the failure classified as a timeout rather than a client abort, which matters for whatever status code and logging you decide is right.

Two design questions I would not want to answer on your behalf: whether HTTP/3 should reuse respondingTimeouts.readTimeout or get its own setting, and whether enforcement belongs in the handler wrapper or somewhere closer to the entrypoint. Both are your call.

A documentation issue, separately

Two things in the docs are worth correcting regardless of how the code question is resolved.

1. The readTimeout description carries no protocol qualification. It says "the maximum duration for reading the entire request, including the body", which is exactly what an operator relies on. The entrypoint page does note that respondingTimeouts have "no effect for UDP entryPoints", but that does not cover this case: HTTP/3 here is served on an HTTP entrypoint with an http3: block, not on a Traefik UDP entrypoint, which is a separate feature for UDP routers. An operator who adds http3: to their existing HTTPS entrypoint has not created a UDP entrypoint and has no reason to read that caveat as applying to them.

2. SECURITY.md's supported-versions table is stale. It lists 3.6.x as supported and < 3.6.x as unsupported, while 3.7.10 is the current release. Anyone checking whether their version is in scope before reporting gets a confusing answer.

Impact

Each held request occupies one upstream connection for as long as the client chooses, at negligible cost to the client, and the same client can open many. Backends with bounded connection pools are the practical pressure point.

I have not measured a concurrency ceiling on Traefik itself, so I am not asserting one. If that number matters to your assessment, tell me and I will measure it.

LLM ("AI") use disclosure

Not required by your policy, but stated because it is true and you should be able to weigh it. I am a penetration tester, not a Go developer. The finding, the attack concept and the decision to measure the upstream leg rather than the client are mine. An LLM coding assistant (Claude) built the test harness, ran the matrix, and located the two source citations; I verified those by hand against the v3.7.10 tag and ran the reproducer myself. I am a human and I will be the one replying in this thread.

Disclosure

Bishop Fox operates a 90-day disclosure policy, starting the day this is submitted, with extensions where a fix is in progress. Tell me what you would prefer and I will work to it.

Environment

• Traefik v3.7.10, official traefik_v3.7.10_linux_amd64.tar.gz, sha256 01811bb12d44f17280550f425f5e3128d6c325f2665c09e67a651ca535f490ce
• Kali GNU/Linux (WSL2), kernel 6.6.114.1
• curl 8.19.0 with ngtcp2 1.21.0 / nghttp3 1.15.0### Summary Short summary of the problem. Make the impact and severity as clear as possible. For example: An unsafe deserialization vulnerability allows any unauthenticated user to execute arbitrary code on the server.

Details

Give all details on the vulnerability. Pointing to the incriminated source code is very helpful for the maintainer.

PoC

Complete instructions, including specific configuration details, to reproduce the vulnerability.

Impact

What kind of vulnerability is it? Who is impacted?

</details> ---