Adding a test that tries to exercise the race condition for the websocket path. The sequence is as follows:
1. The test starts by writing the payload to the mock server
2. Then, it closes the write side of the connection.
3. It waits for 3 seconds to make the race window explicit: the bug in cloudflared would have closed the connection and we would not be able to read.
4. Read the message.
With the fix applied, we should be able to read the full message.
This PR addresses an issue where cloudflared prematurely closes the origin connection before the upstream-to-downstream goroutine finishes reading, causing intermittent connection drops when a client immediately closes the write-side of a connection.
When a client finishes writing data, it immediately closes its side of the connection. Under the current implementation in cloudflared's downstream-to-upstream goroutine does not wait for the second stream to complete. It unblocks the moment the first stream writes to the channel. Once this happens, the pipe returns control to `proxyTCPStream`, which prematurely closes the origin connection. Consequently, when the upstream-to-downstream goroutine attempts to read the remaining data from the origin connection, the connection is already gone, leading to unexpected failures.
We started propagating the `TimeoutAfterFirstClose` configuration/parameter. This allows the proxy to wait for a designated period, giving the second stream sufficient time to finish processing and read all remaining data before `proxyTCPStream` tears down the origin connection.
systemctl list-units exits 0 whether or not any pattern matches, so the
filter was a no-op and every unit was always treated as installed.
Detect installed units by checking for the unit file on disk instead.
Generate a clean init script at /etc/init.d/cloudflared backed by
OpenRC's supervise-daemon supervisor, plus a companion
/etc/conf.d/cloudflared for operator overrides.
This MR contains the following updates:
| Package | Type | Update | Change |
|---|---|---|---|
| gcr.io/distroless/base-debian13 | final | pinDigest | → `ab7554b` |
---
### Configuration
📅 **Schedule**: (UTC)
- Branch creation
- At 12:00 AM through 04:59 AM and 10:00 PM through 11:59 PM, Monday through Friday (`* 0-4,22-23 * * 1-5`)
- Only on Sunday and Saturday (`* * * * 0,6`)
- Automerge
- At any time (no schedule defined)
🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied.
♻ **Rebasing**: Whenever MR is behind base branch, or you tick the rebase/retry checkbox.
🔕 **Ignore**: Close this MR and you won't be reminded about this update again.
---
- [ ] <!-- rebase-check -->If you want to rebase/retry this MR, check this box
---
This MR has been generated by [Mend Renovate](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0My4yMjQuMCIsInVwZGF0ZWRJblZlciI6IjQzLjIyNC4wIiwidGFyZ2V0QnJhbmNoIjoibWFzdGVyIiwibGFiZWxzIjpbXX0=-->
As it stands, cloudflared prechecks are not taking the `protocol` flag into consideration and is instead falling back to the default protocol, which is QUIC. Prechecks should report the protocol cloudflared will use, not the default protocol.
Bumps quic-go to v0.59.1 (chungthuang fork rebased from upstream v0.45 onto
v0.59.1). Upstream removed the `logging` package and replaced its
callback-based ConnectionTracer with the structured `qlog`/`qlogwriter` event
API, which required migrating cloudflared's QUIC metrics collection.
Migrations:
- quic/tracing.go: connTracer no longer fills a logging.ConnectionTracer
callback struct. It implements qlogwriter.Trace + qlogwriter.Recorder and
dispatches qlog events (PacketSent, PacketReceived, MetricsUpdated, ...) to
the collector through RecordEvent. NewClientTracer now returns a function
compatible with quic.Config.Tracer.
- quic/metrics.go: collector methods take qlog types (qlog.Frame,
qlog.PacketType, qlog.MetricsUpdated, ...) and plain int64 in place of the
removed logging.ByteCount/Frame/RTTStats/TransportParameters.
- quic/conversion.go: PacketType, PacketDropReason and PacketLossReason are
strings upstream rather than numeric iotas, so the converters become
pass-through allowlists. CongestionState is also a string;
congestionStateToFloat maps it back to the numeric gauge values cloudflared
exports.
- quic.Connection/quic.Stream became *quic.Conn/*quic.Stream; updated
ConnWithCloser, SafeStreamCloser and the connection package accordingly.
Tests and generated mocks (mocks/mock_quic_connection.go) were adapted to
the new pointer-based API.
Closes TUN-10557
Publishing internal image in cloudflared. This allows us to remove the dependency from cloudflare/plat/dockerfiles. In addition, our acceptance tests should now be able to use the latest image instead of relying on a fixed version for testing, which will allow us to detect potential failures earlier.
The bump of the QUIC library introduces a cyclic dependency between the connection and quic modules hence it is necessary to break this coupling.
Right now, the connection module depends on the quic module for the datagram v2/v3 and to which a QUIC connection (currently an interface) is passed.
As it is there is no issue however, under the hood, interface is a wrapper around an UDP connection and a QUIC connection meaning this type must be exposed to the quic module since the QUIC Connection will no longer be a interface but a struct.
Given the above, these changes introduce an interface, QUICConnection, with the surface used today in cloudflared and a struct, ConnWithCloser, that implements said interface within the quic module.
Closes TUN-10563
Pipes have a finite OS buffer (\~64KB Linux, \~4KB macOS, \~4KB Windows). Since nobody was reading stdout/stderr during the process lifetime, cloudflared would block once the buffer filled up. The post-terminate()/read() could only get whatever fit in the buffer, causing truncated logs.
There was also a race between terminate() and read(): the process might not have flushed its final output yet.
We're also deleting `test_default_only`. Since we changed `edge-ip-version` to auto, this test became redundant.
Fixing some bugs with DNS targets. Most importantly, these changes also fix some wrong assumptionsmade when trying to add support for the `--edge` flag:
1. Removes `StaticEdgeDNSResolver` in favor `resolveStaticEdge`. Since --edge does not imply resolving DNS, this fixes that assumption.
2. Adds EdgeAddrs, which allows us to skip DNS probes when set. This fixes the targets in the DNS rows.
3. Added a new `ResolvedTarget` struct, which joins addresses with the respective DNS results. This avoids the brittle logic we had before, where we assumed there were always two groups (one for each region) when running probes. So this not only makes the code more extensible in case we want to add more regions in the future but also adds support for multiple targets supplied via `--edge`.
4. Changes the existing nomenclature, going from calling things `region` to `target`. The term `region` works when resolving production regions (region1 and region2), but becomes misleading when we add the logic for `--edge`.
The end result of these changes is that we now see the correct addresses when you supply targets via `--edge`, while also making the code a bit clearer.
Instead of having the --precheck flag in cloudflared, we allow controlling prechecks via a DNS flag, so we can short-circuit this behavior in case anything goes wrong. Although we don't expect pre-checks to add that much traffic, we should still guarantee that we can stop pre-checks in case something goes wrong.
Introduce a new crypto/ package as the single source of truth for TLS
curve preferences used on every edge-facing connection, and adopt
X25519MLKEM768 as the primary post-quantum key exchange for both QUIC
and HTTP/2:
PQ Prefer (default): X25519MLKEM768, P256Kyber768Draft00, CurveP256
PQ Strict (--post-quantum): X25519MLKEM768, P256Kyber768Draft00
The curve list is identical under FIPS and non-FIPS builds, so
crypto.GetCurvePreferences takes only a features.PostQuantumMode and
returns a fresh slice on every call.
HTTP/2 now applies these curve preferences the same way QUIC does. The
previous PostQuantumStrict rejection in serveHTTP2 and the forced
QUIC-only selection in NewProtocolSelector are removed since both
transports support the same post-quantum curves; the needPQ parameter
is dropped from NewProtocolSelector accordingly.
Also fix a shared tls.Config race: both the QUIC and HTTP/2 paths now
Clone() the per-protocol entry from TunnelConfig.EdgeTLSConfigs before
mutating CurvePreferences instead of writing through the shared map
entry.
Legacy Kyber draft curve X25519Kyber768Draft00
and the unused removeDuplicates helper are removed along with the old
supervisor/pqtunnels.go / _test.go files.
AGENTS.md is updated with guidance on the new crypto/ package, the
cfdcrypto import alias, the tls.Config cloning rule, and the lint
workflow implications of .golangci.yaml's whole-files: true setting.
Final run method, which runs cloudlflared pre-checks for both the normal startup procedure, as well as cloudflared diag.
For cloudflared diag, this produces a new json output to which is added to the final zip file.
Also added in a new flag to prevent this from running all the time, at least for now until we are 100% sure this works as intended. We will later remove this flag, only leaving in `--no-prechecks`, so this runs by default for everyone using cloudflared.
Tested pre-checks locally with origintunneld. The results show all pre-checks succeeding. In this case, it ran with only 1 region, since locally we run it with `--edge origintunneld1:7844`.
{width=900 height=217}
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
This PR allows us to use edge overrides with pre-checks. I forgot to account for the --edge flag when developing pre-check probes in TUN-10388. This should allow us to wire this flag in.
DNSResolver will still be injected via RunDialers, and we will take care of initialization in cmd.go. This allows us to keep pre-checks testable and inject mock DNSResolvers when needed.
I thought about overriding `edgediscovery.ResolveEdge` and `edgeDiscovery.StaticEdge` instead of `allregion.EdgeDiscovery` and `allregionResolveAddrs`, respectively. But that would imply changing our existing probe logic to support `Regions` instead of `[][]EdgeAddr`, which would mean more work. Additionally (and perhaps most importantly), using `Regions` would also require us to create new functions to extract a list of edge addresses for our probe tests. I don't think this would go well with the current implementation of `Regions`, as I believe it's intent is to encapsulate the logic around managing addresses per regions. Adding these functions would mean breaking this encapsulation.
## What
Bumps go-boring from 1.26.0-1 to 1.26.2-1 and CI builder image from \`3501-fc698419a625\` to \`3595-779e088c0ec4\`.
go1.26.2 (released 2026-04-07) includes security fixes to the \`go\` command, the compiler, and the \`archive/tar\`, \`crypto/tls\`, \`crypto/x509\`, \`html/template\`, and \`os\` packages, as well as bug fixes to the \`net\`, \`net/http\`, and \`net/url\` packages.
### Security fixes (relevant)
- **crypto/tls**: multiple CVEs — cloudflared uses TLS extensively for tunnel connections
- **crypto/x509**: CVE-2026-32280 (excessive chain-building in \`Verify\`), CVE-2026-32281 (quadratic work in policy validation)
### Net bug fixes (not applicable)
- **net/url #78111**: \`url.Parse\` regression for MongoDB-style multi-host URLs — not used in cloudflared
- **net/http #78019**: race condition on Windows when using \`os.File\` as HTTP request body — cloudflared does not pass \`os.File\` as a request body
- **net #77885**: \`ReadMsgUDP\`/\`WriteMsgUDP\` WSAEFAULT on Windows with empty non-nil oob — quic-go uses \`basicConn\` on Windows (\`ReadFrom\`, not \`ReadMsgUDP\`)
## Jira
[TUN-10507](https://jira.cfdata.org/browse/TUN-10507)
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
Fixing missing TLS settings. While developing the pre-check probes, I forgot to add the certificate settings, which are essential for establishing a connection to origintunneld. I discovered this while testing cloudflared locally.
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
Fixing warnings in cloudflared log collector.
This attempts to fix errors like the ones shown below:
```
diagnostic/diagnostic.go:132:23: Error return value of `logHandle.Close` is not checked (errcheck)
defer logHandle.Close()
diagnostic/diagnostic.go:134:26: G303: File creation in shared tmp directory without using ioutil.Tempfile (gosec)
outputLogHandle, err := os.Create(filepath.Join(os.TempDir(), logFilename))
```
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
This introduces the main precheck logic. This will follow concurrency model, timeout handling, and retry logic established in the SPEC. We will follow the decision flow in the [connectivity pre-checks SPEC](https://wiki.cfdata.org/spaces/TUN/pages/1374967685/Connectivity+Pre-checks+for+cloudflared). You can find an attached image of what the decision flow should look like.
{width=900 height=235}
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
Replace the lock file mechanism with PID+start-time based stale
detection so that no cleanup is required on process death.
When both org and app token locks were held, the first signal handler
to call os.Exit() would kill the process before the second handler
could delete its lock file. The orphaned lock file then caused the
next invocation to wait ~128 seconds in an exponential backoff loop
before forcibly deleting it. The same issue occurred on SIGKILL, OOM,
or any non-signal death.
Lock files now contain the holder's PID and process start time as
JSON. On acquisition, if a lock file already exists, the recorded
process is checked for liveness via gopsutil. Stale locks are
reclaimed immediately with no backoff. Atomic O_CREATE|O_EXCL
prevents races between concurrent acquirers.
Also adds a companion .url file so processes waiting on an active
lock can print the auth URL for the user.
This PR implements all the dialers and resolvers needed to make pre-checks happen. So this task focuses on the following:
1. Implement the DNS probe: call DNSResolver.Resolve(region)
2. Implement the QUIC probe: call QUICDialer.DialQuic (handshake only, no stream opened) and record the result.
3. Implement the HTTP/2 probe: call TCPDialer.DialEdge (TCP + TLS handshake only, no frames sent) and record the result.
4. Implement the Management API probe: call ManagementDialer.DialContext to api.cloudflare.com:443 and record the result.
5. Export edgeDiscovery as EdgeDiscovery in edgediscovery/allregions/discovery.go so the pre-check can reuse the production DNS path.
This sets up the main components to implement the checker.
Check / check (1.22.x, macos-latest) (push) Has been cancelled
Check / check (1.22.x, ubuntu-latest) (push) Has been cancelled
Check / check (1.22.x, windows-latest) (push) Has been cancelled
Semgrep config / semgrep/ci (push) Has been cancelled
Switching `SuggestedProtocol` to a pointer, so we can pass nil whenever both QUIC and HTTP2 fail. We should not be suggesting anything when all our protocols have failed.
Miguel da Costa Martins Marcelino
João "Pisco" Fernandes
Renovate
Christopher Meng
Christopher Meng
lneto
Gonçalo Garcia
Evan Raw