- explicitly install `libldap-dev` to not rely on test-specific packages
installing it implicitly, to have the same `curl -V` output for each
TLS backend build pair.
Follow-up to 0455d8772a#18509
- install `libev-dev` for tests. It's a runtime dependency for
the local build of `nghttpx`. Missing it made pytest skip 178 tests.
Also skewing the 'Gain' time. I estimate it to account for 3 minutes,
making the total gain ~20 minutes.
Follow-up to 0455d8772a#18509
(It may be a better solution to disable libev for the local nghttp2
build, to avoid this hidden dependency.)
- fix quiche jobs to use the local build of `libnghttp2`.
- stop installing the `clang` package for Windows-cross. `clang` and
`clang-tidy` tools are preinstalled on the Ubuntu 24.04 runner.
Closes#18519
In the last couple of months some jobs started taking a lot of time and
often timing out due to slow `apt install` from the Azure Ubuntu mirror.
The jobs affected were those that installed large packages:
GHA/http3-linux and the 3 cross-build jobs in GHA/windows.
This patch reduces the installed packaged to the minimum required
to complete the jobs. Saving a minute+ for each http3-linux job (a total
of 20+ minutes for the workflow.) Also saving bandwidth and reducing
the chance for long downloads or timeouts with slow Azure repos.
Details:
- http3: delete redundant packages from the `build-cache` job.
- http3: install gnutls dependencies for gnutls jobs only.
- http3: do not install test dependencies in jobs not running tests.
- http3: drop redundant packages from the curl jobs.
- Windows-cross: replace `mingw-w64` with `gcc-mingw-w64-x86-64-win32`
for the 3 Windows cross-build job. Dropping C++, 32-bit, and 64-bit
POSIX-threaded parts. Saving time and significant bandwidth for each
of the 3 jobs:
Download size: 277 MB -> 65 MB (installed: 1300 MB -> 400 MB)
- Windows-cross: restore previous job time limit of 15m (from 45m)
Follow-up to ff5140a25f#18163
Before:
https://github.com/curl/curl/actions/runs/17611514207 (http3)
https://github.com/curl/curl/actions/runs/17611514185/job/50034354923 (Windows cross)
After:
https://github.com/curl/curl/actions/runs/17628406362?pr=18509 (http3)
https://github.com/curl/curl/actions/runs/17627562551/job/50088055529?pr=18509 (Windows cross)
http3 job | Bef. | Aft. |
:------------------ | ------: | ------: |
Build caches (hot) | 10s | 12s |
AM awslc | 3m 0s | 1m 54s |
CM awslc | 4m 32s | 3m 4s |
AM boringssl | 3m 9s | 1m 48s |
CM boringssl | 3m 43s | 3m 2s |
AM gnutls | 3m 9s | 2m 18s |
CM gnutls | 4m 19s | 2m 55s |
AM libressl | 2m 14s | 1m 24s |
CM libressl | 5m 30s | 2m 57s |
AM openssl | 5m 16s | 4m 17s |
CM openssl | 1m 50s | 1m 47s |
AM openssl-quic | 2m 58s | 1m 7s |
CM openssl-quic | 4m 16s | 2m 43s |
AM quiche | 2m 54s | 1m 34s |
CM quiche | 5m 0s | 3m 15s |
AM quictls | 2m 34s | 1m 13s |
CM quictls | 4m 20s | 3m 17s |
AM wolfssl | 2m 48s | 1m 30s |
CM wolfssl | 4m 49s | 3m 22s |
Total: | 66m 21s | 43m 27s |
Gain: | | 22m 54s |
Out of curiousity, build times as seen in the http3 build-cache job:
- TLS backends:
- openssl: 2m25s
- libressl: 27s
- aws-lc: 41s
- boringssl: 1m8s
- quictls: 1m46s
- gnutls: 6m30s
- wolfssl: 51s
- quiche + boringssl: 1m9s
- ng* libs (not yet optimized for build speed):
- nghttp3: 13s
- ngtcp2: 52s (with 6 backends, 3 runs)
- ngtcp2: 19s (boringssl)
- nghttp2: 21s
Ref: https://github.com/curl/curl/actions/runs/17626120054/job/50083344805
A similar effort in curl-for-win, affecting 2 GHA/curl-for-win Windows
jobs (though they use the default Debian repo, with no issues):
- with llvm/clang:
Download size: 648 MB -> 430 MB (installed: 3344 MB -> 2333 MB)
- with gcc:
Download size: 550 MB -> 328 MB (installed: 2815 MB -> 1804 MB)
Ref: e19665d948
Ref: 6b14c3946a
Bug: https://github.com/curl/curl/pull/18502#issuecomment-3270259744Closes#18509
To fix non-`CURLDEBUG` builds on 32-bit AIX, where `fopen` is a system
macro.
Ref: #18502
Ref: 793a375ce3
Follow-up to 3bb5e58c10#17827
Reported-by: Andrew Kirillov
Fixes#18510Closes#18514
When passing an address to curl_easy_getinfo to retrieve a value and the
address is NULL, return CURLE_BAD_FUNCTION_ARGUMENT instead of
CURLE_UNKNOWN_OPTION.
Closes#18512
Check the easy handles magic in calls to curl_easy_getinfo().
In Curl_close() clear the magic after DNS shutdown since we'd
like to see tracing for this.
When clearing the magic, also clear the verbose flag so we
no longer call DEBUGFUNCTION on such a handle.
Closes#18511
Some OSes (Linux, macOS, more?) will generate an EMSGSIZE socket error
on the next recv all after receiving an ICMP Packet Too Big on an
unconnected UDP socket.
These can be safely ignored as QUIC's DPLPMTUD uses MTU probes that do
not rely on receiving ICMP packets.
Closes#18505
To avoid overriding the system symbol `accept`, which is a macro on some
systems (AIX), and thus can't be called via the `(function)` PP trick.
It's also problematic to reset such macro to its original value.
Follow-up to 3bb5e58c10#17827
Reported-by: Andrew Kirillov
Fixes#18500Closes#18501Closes#18502
It's a `-Weverything` warning that appeared in llvm/clang 21.
```
lib/openldap.c:1297:19: warning: duplicate declaration of 'ldapsb_tls' is invalid in C++ [-Wtentative-definition-compat]
1297 | static Sockbuf_IO ldapsb_tls =
| ^
lib/openldap.c:499:19: note: previous declaration is here
499 | static Sockbuf_IO ldapsb_tls;
| ^
```
Reported-by: correctmost on github
Fixes#18470
Cherry-picked from #18477Closes#18485
The auto PONG frames were inserted into the connection at the time
a PING had been decoded, irregardless if an upstream frame was just
in the middle of being assembled.
Add PONG frames only to the buffer if there is no frame currently
assemebled and, if it is, set the control frame aside. This control
frame is then added on the first opportunity of a "clean" send buffer.
There is only a single control frame set aside at a time. This means
a double PING will, when the PONG cannot be sent right away, only
send the last PONG.
I imagine this is fine. We want to prevent the endless buffering of
PONG frames on a connection where the server sends but does no receives.
Reported-by: Calvin Ruocco
Fixes#16706Closes#18479
Sending websocket data did not clear the "upload_done" flag of
the initial HTTP Upgrade request, leading to KEEP_SEND never be
cleared. This caused the socket to be polled for INOUT after all
the websocket data had been sent. A busy loop.
Closes#18476
Added in 2011, but has seen little use in the code. The necessary
compiler feature is missing in some compilers (e.g. MSVC), thus in most
places the portable `(void)` cast is used in addition.
Also:
- vtls/rustls: silence unused argument warning with `(void)`.
Necessary for MSVC, for example.
Ref: ee4ed46128Closes#18455
Tweaks around handling of --limit-rate:
* tracing: trace outstanding timeouts by name
* multi: do not mark transfer as dirty that have
an EXPIRE_TOOFAST set
* multi: have one static function to asses speed limits
* multi: when setting EXPIRE_TOOFAST remove the transfers
from the dirty set
* progress: rename vars and comment on how speed limit
timeouts are calculated, for clarity
* transfer: when speed limiting, exit the receive loop
after a quarter of the limit has been received, not
on the first chunk received.
* cf-ip-happy.c: clear EXPIRE_HAPPY_EYEBALLS on connect
* scorecard: add --limit-rate parameter to test with
speed limits in effect
Closes#18454
Only actually needed servers should be listed and none is then implied
if no servers are listed.
Outputs a warning if "none" is still set as a server.
Closes#18466
A previous refactor of cookie logic changed Curl_cookie_getlist to no
longer return a list of copied cookies, but instead return a linked list
pointing to existing cookies. The returned linked list is accessed
outside of the scope of the cookie share lock in http_cookies, which
leads to issues if the shared cookie list is modified at the same time.
This is the relevant commit: be39ed1Closes#18457
