This patch adds two major proxy capabilities to curl (ngtcp2 QUIC):
- HTTP/3 Proxy CONNECT: Tunnel HTTP/1.1 or HTTP/2 traffic through an
HTTPS proxy that speaks HTTP/3 (QUIC) using the standard CONNECT
method over an HTTP/3 connection.
- MASQUE CONNECT-UDP: Tunnel HTTP/3 (QUIC) traffic through an HTTP
proxy (speaking HTTP/1.1, HTTP/2, or HTTP/3) using the extended
CONNECT method with the CONNECT-UDP protocol (RFC9297 & RFC9298).
Public API additions:
- `CURLPROXY_HTTPS3`: new proxy type constant for HTTP/3 proxy
- `--proxy-http3`: new CLI flag to negotiate HTTP/3 with HTTPS proxy
The implementation adds two new filters:
- `H3-PROXY` - enables negotiating HTTP/3 (QUIC) to the proxy and
running CONNECT/CONNECT-UDP through that proxy transport.
- `CAPSULE` - dedicated filter inserted between QUIC transport and
HTTP-PROXY to handle datagram capsule encapsulation/decapsulation.
Here is how the curl filter chaining looks in different scenarios:
- HTTP/3 Proxy CONNECT (tunneling TCP protocols over QUIC proxy):
conn -> HTTP/1.1 or HTTP/2 -> SSL -> HTTP-PROXY ->
H3-PROXY -> HAPPY-EYEBALLS -> UDP
- MASQUE CONNECT-UDP (tunneling QUIC over any proxy):
conn -> HTTP/3 -> CAPSULE -> HTTP-PROXY -> H3-PROXY ->
HAPPY-EYEBALLS -> UDP
conn -> HTTP/3 -> CAPSULE -> HTTP-PROXY -> H1-PROXY or H2-PROXY ->
SSL -> HAPPY-EYEBALLS -> TCP
- Both features currently require the ngtcp2 QUIC backend.
- Both features are experimental (disabled by default). Enable with
`--enable-proxy-http3`(autotools) or `-DUSE_PROXY_HTTP3=ON`(CMake).
Tests:
- tests/unit/unit3400.c: Unit tests for capsule protocol encode/decode
- tests/http/test_60_h3_proxy.py: Comprehensive pytest integration suite
- tests/http/testenv/h2o.py: Managing h2o instances with HTTP/1.1, HTTP/2,
and HTTP/3 (QUIC) listeners, proxy.connect and proxy.connect-udp enabled.
References:
RFC 9297 - HTTP Datagrams and the Capsule Protocol
RFC 9298 - Proxying UDP in HTTP
RFC 9000 §16 — Variable-Length Integer Encoding
Signed-off-by: Aritra Basu <aritrbas+gh@cisco.com>
Closes#21153
- fix macOS locale tests to clear existing variables.
(Without this, the system-defined `LC_ALL` takes precedence, and
the custom envs in CI are ignored.)
- trigger test 1981 issue by setting `LC_TIME` to non-English, on macOS.
(On Linux it'd require explicitly installing a non-English locale, I
skipped this for simplicity.)
```
[...]
-Time: 01/Aug/2025 08:31:43.037103 +0000 UTC[CR][LF]
+Time: 01/ao%c3%bb/2025 08:31:43.037103 +0000 UTC[CR][LF]
[...]
FAIL 1981: '%time output with --write-out' HTTP, HTTP GET
```
Follow-up to 90a7732d46#21749
Follow-up to 1cc8a5235f#17988
Follow-up to c221c0ee59#17938Closes#21753
Detect latest tarball version via the https://curl.se/downloads.html
page, download the signing key from a public keyserver then verify
source download signatures.
To ensure that public downloads are intact.
Closes#21759
Also switch back to ftp.openbsd.org download server. More often than not
the GitHub release entry is missing the download artifacts at the time
of detecting a new version, breaking automatic bumps. We cache the
download so it does not bang the origin server with many requests.
Follow-up to 800b0bec18#19082Closes#21742Closes#21754
Since all API features we need for TLSv1.3 earlydata support do exist
only from version 3.7.2 onwards, make that the minimal version required.
Fixes#21750
Reported-by: Johannes Schlatow
Closes#21751
The mev_sh_entry object might be removed if curl_easy_pause() is called
from within the socket callback.
Introduced a 'magic' struct field to to 'mev_sh_entry' to make it easier
to programmatically detect/assert if the pointer is bad - in debug
builds.
Reported-by: Joshua Rogers
Closes#21748
Host key type mismatch from libssh was incorrectly reported as missing,
causing key callbacks to accept instead of reject.
Reported by: Joshua Rogers (Aisle Research)
Closes#21724
The old CMake bug about exporting -lpthread instead of Threads::Threads
no longer matches current master. As of 2d546d239e
("cmake: use Threads::Threads imported target for POSIX Threads"),
the build now uses Threads::Threads and the generated CMake package
config resolves the dependency explicitly, so this KNOWN_BUGS entry is stale.
Closes#21734
The HTTP/2 feature is deprecated, few servers implement it and our
implementation is complicated by its state management. Make the two
CURLOPT_* involved a nop and deprecate them.
Closes#21723
The entry is about GnuTLS not sending the client cert when it doesn't
match the `DN` the server requested. OpenSSL does the opposite.
The issue was already fixed by #4958 and removed from KNOWN_BUGS,
but it was added back to the list by #16677, seemingly by mistake.
The issue is still fixed for GnuTLS >= 3.5.0.
As curl only supports GnuTLS >= 3.6.5, remove the bug entry from
KNOWN_BUGS.md
Fixes#21720Closes#21722
CI time cost is 1s.
It may replace existing pedantic check, if this level isn't bringing
false-positives or annoyance. Officially it's not meant for CI, but curl
has been passing this in the last couple of months when checked locally.
Closes#21718
- use `UNITTEST_BEGIN_SIMPLE`/`UNITTEST_END_SIMPLE` where missing.
- drop redundant `(void)arg;` where using `UNITTEST_BEGIN*`.
- unit1636: drop redundant block after `UNITTEST_BEGIN*`.
- unit1609: fix typo in comment.
- unit1627: merge to `if`s.
Closes#21715
- GHA/windows: bump dl-mingw job from gcc 15 to 16.
- multi: silence warning while building libcurlu:
```
In function 'multi_ischanged',
inlined from 'multi_socket.isra' at D:/a/curl/curl/lib/multi.c:3282:6:
D:/a/curl/curl/lib/multi.c:1710:17: error: null pointer dereference [-Werror=null-dereference]
1710 | bool retval = (bool)multi->recheckstate;
| ^~~~~~~~~~~~~~~~~~~~~~~~~
D:/a/curl/curl/lib/multi.c:1712:25: error: null pointer dereference [-Werror=null-dereference]
1712 | multi->recheckstate = FALSE;
| ^
```
Ref: https://github.com/curl/curl/actions/runs/26217071531/job/77142119137?pr=21707
- multi: silence another `-Wnull-dereference`, popping up in libcurl
with gcc 13 after the previous silencing:
```
In function 'Curl_multi_xfers_running',
inlined from 'multi_socket.isra' at ../../lib/multi.c:3292:28:
../../lib/multi.c:4132:15: error: null pointer dereference [-Werror=null-dereference]
4132 | return multi->xfers_alive;
| ~~~~~^~~~~~~~~~~~~
```
Ref: https://github.com/curl/curl/actions/runs/26218822231/job/77148186045
- multi: also add `DEBUGASSERT(multi)` to the two updated functions.
Closes#21707
If so much data is being sent over the internal IPC pipe that the pipe
buffer fills and the syscall blocks, the program will hang. Add an
assert to ensure that this limit is never reached. The buffer size is
going to be different on different platforms, so choose 1KB which is
likely to be a reasonable lower bound on just about any system.
Currently, the maximum amount ever written is <100 bytes, so this should
provide plenty of headroom.
Spotted by Codex Security
Closes#21688
- Fix revoke_best_effort reading wrong ssl config.
Prior to this change the revoke_best_effort setting for the proxy was
wrongly ignored in favor of the same setting for the destination host.
In other words, CURLSSLOPT_REVOKE_BEST_EFFORT set via
CURLOPT_PROXY_SSL_OPTIONS did not apply to the proxy and
CURLSSLOPT_REVOKE_BEST_EFFORT set via CURLOPT_SSL_OPTIONS wrongly
applied to the proxy.
Closes https://github.com/curl/curl/pull/21683
When a transfer gets paused after the connection has been established,
any data sitting in the kernel socket buffers will no longer get read.
Prevent the sockets form being added to the pollsets, because they
will trigger POLLIN endlessly and cause a busy poll loop. Same in event
based processing.
Reported-by: Harry Sintonen
Fixes https://github.com/curl/curl/issues/21671
Closes https://github.com/curl/curl/pull/21675
- Clear credentials on AcquireCredentialsHandle failure so it is not
used on a subsequent call.
SSPI initialization may evaluate the credentials pointer to determine
whether or not a prior call to AcquireCredentialsHandle was successful,
therefore we must clear it on a failed call.
Closes https://github.com/curl/curl/pull/21642
Example:
```
Wed, 20 May 2026 09:51:48 GMT Updating Avalon repository catalogue...
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT repository Avalon has no meta file, using default settings
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT pkg: An error occurred while fetching package: No error
Wed, 20 May 2026 09:51:48 GMT Unable to update repository Avalon
Wed, 20 May 2026 09:51:48 GMT Error updating repositories!
Wed, 20 May 2026 09:51:48 GMT Error: Process completed with exit code 3.
```
As tested over at libssh2, retrying the install command also does not
help, only repeats the same failure.
Also: fix whitespace in matrix.
Follow-up to b158d1c9f7#21681Closes#21694
- bump cross-platform-actions to v1.1.0.
Ref: https://github.com/cross-platform-actions/action/releases/tag/v1.1.0
- merge BSD jobs into a single matrix.
- split BSD jobs into build steps as used for other platforms.
A new feature of cross-platform-actions v1.1.0.
- sync BSD build steps with other platforms.
- add DragonFlyBSD and MidnightBSD to the BSD matrix.
New features of cross-platform-actions v1.1.0.
MidnightBSD uses GnuTLS to add variation, also the preinstalled
OpenSSL is too old (v1.1.1w) for curl.
Stick with autotools for DragonFlyBSD; I could not figure out how
to install cmake.
Refs:
https://en.wikipedia.org/wiki/DragonFly_BSDhttps://en.wikipedia.org/wiki/MidnightBSD
- bump Intel FreeBSD jobs from v14.3 to v15.0.
- fix to show `gcc` in the NetBSD job name.
All these saved 50 lines of YAML. The two new jobs take 2m15s each. The
bump to FreeBSD 15 needs and extra minute in total.
Note, the DragonFlyBSD job seems to have reliability issues. If it
remains an issue, I'll comment it out or delete it in a future commit.
Closes#21681
"Lines 244-245 overwrite global variables `$runnerr` and `$runnerw` that
were already assigned in the child process (lines 205-206). In the
parent process context, these assignments appear incorrect and could
cause issues if `runner_init` is called multiple times. The parent
should only store references in the controller hashes."
It could never cause an actual issue, but clarifies the intent of the
code.
Spotted and fixed by GitHub Code Quality
Cherry-picked from #21646Closes#21672
