The idea of linking dependencies found to `libcurl.pc` turns out not
to work in practice in some cases.
Specifically: gss, ldap, mbedtls, libmsh3, rustls
A `.pc` may not work or be missing for a couple of reasons:
- not all build methods generate it: mbedTLS, Rustls
- generated file is broken: msh3
Ref: https://github.com/nibanks/msh3/pull/225
- installed package flavour isn't shipping with one:
FreeBSD GSS, OmniOS LDAP, macOS LDAP
The effect of such issues shall be subtle in theory, because
`libcurl.pc` normally lists these dependencies in the `Requires.private`
section meant for static linking. But, e.g. `pkg-config --exists`
requires these to be present, and builds sometimes use this check
regardless of build type. This bug is not present in `pkgconf`; it only
checks for them when `--static` is also passed.
Fix these by adding affected `.pc` references to `libcurl.pc` only when
we detected the dependency via `pkg-config`.
There are a few side-effects of this solution:
- references are never added for dependencies where curl doesn't
implement `pkg-config` detection. These are:
- autotools: ldap, mbedtls, msh3
- cmake: ldap (pending #15273)
- generated `libcurl.pc` depends on the build-time environment.
- generated `libcurl.pc` depends on curl build tool (cmake, autotools).
- generated `libcurl.pc` depends on curl build implementation details.
Make an exception for GNU GSS, where I blindly guess that `gss.pc` is
always available, as no issues were reported.
Other, not mentioned, dependencies continue to be added regardless
of the detection method.
Reported-by: Harmen Stoppels, Thomas, Daniel Engberg, Andy Fiddaman
Fixes#15469Fixes#15507Fixes#15535
Fixes https://github.com/curl/curl/pull/15163#issuecomment-2473358444Closes#15573
`libcurl.pc` `Libs.private` (since 8.11.0, and in `Libs` before 7.20.0)
and `curl-config` `--static-libs` (since 7.17.1, and in `Libs` between
7.7.2-7.25.0). This included all flags inherited from the environment,
in addition to those coming from dependency detections.
To avoid spilling all linker flags inherited from the environment to
the libcurl config files, this patch omits them all, except `-L`, `-F`,
`--library-path=` and `-framework` options, which are still passed.
The rationale for the exceptions is that `LIBS` is passed as-is, and
`LDFLAGS`, `LIBS` are the canonical way to pass custom libs options
to a build. `LIBS` may not work without a matching custom libpath.
This brings autotools behaviour closer to cmake, and `curl-config`
closer to `libcurl.pc`.
Follow-up to 9f56bb608e#14681
Follow-up to 4c8adc8fee
Reported-by: Peter Marko
Fixes#15533Closes#15550
Different variable names were used in flag save and restore operations,
which could cause existing `CPPFLAGS` and `LDFLAGS` be accidentally lost
when detecting Rustls.
Follow-up to 647e86a3ef#13179Closes#15546
Build wolfSSL master with
./configure --prefix=/path --enable-ip-alt-name --enable-quic
--enable-earlydata --enable-psk --enable-opensslcoexist
and configure curl with openssl + wolfssl. Normal tests run.
pytest session resumption fails, as wolfssl does not handle the
new_session callback without opensslextra right now.
Closes#15481
The gethostbyname(3) family was removed in POSIX-1.2008 in favor of
getaddrinfo(3) introduced in POSIX-1.2001. Modern POSIX systems such as
Sortix does not have gethostbyname nor the related definitions and
structures.
curl already only uses getaddrinfo(3) if available and thread safe,
although there is mild breakage if the related gethostbyname definitions
are missing.
This change attempts to fix that breakage:
Remove an unnecessary configure error if gethostbyname is missing since
getaddrinfo is enough as a fallback.
Rewrite Curl_ip2addr to not use struct hostent as it no longer is
standardized and create the struct Curl_addrinfo directly.
Only define the Curl_he2ai function on non-getaddrinfo systems where it
is going to be used with struct hoestent.
Revoke the fallback logic for when it's unknown whether getaddrinfo is
thread safe. It doesn't appear to make any sense since h_errno is
unrelated to getaddrinfo. The logic prevents new POSIX.1-2024 systems
from passing the thread safety test since h_errno does not exist anymore
and POSIX already requires getaddrinfo to be thread safe. There's
already a denylist in place for operating systems with known buggy
implementations.
Closes#15475
Added in 37eba37019 (2009-06-17) to help
detecting socket functions.
But, this `socket.h` isn't used in the source code since
90dd1fc664#8288 (2022-01-16).
Closes#15173
Before this patch they were detected via manual methods, then with
`AC_CHECK_FUNCS()`.
Delete the manual checks and keep the latter.
Also delete `CURL_INCLUDES_POLL()` which is no longer used after
the above.
Closes#15170
poll() on macOS 10.12 was deemed broken in 2016 when we discovered that
it misbehaves when provided with no sockets to wait for. The
HAVE_POLL_FINE is used to mark a poll() implementation that behaves
correctly: it *should* still wait the timeout time.
curl has therefore opted to use select() on Apple operating systems ever
since. To avoid the risk that this or other breakage cause problems.
However, using select() internally is also bad because it suffers from
problems when using file descriptors beyond 1024.
This change makes poll() used if it is present, but if there is no
sockets to wait for it avoids using poll() and instead falls back to
select() - but without any sockets to wait for there is no 1024 problem.
This removes all previous special-handling involving HAVE_POLL_FINE.
ref: https://daniel.haxx.se/blog/2016/10/11/poll-on-mac-10-12-is-broken/Closes#15096
Since it seems the _POSIX_C_SOURCE "trick" does not work there, the
check does not find poll().
Fixes#15013
Reported-by: vvb2060 on github
Closes#15037
- lib557: suppress `-Wformat-overflow` warning in source.
Fixes:
```
lib557.c: In function ‘test_float_formatting’:
lib557.c:1408:37: error: ‘%*f’ directive output of 2147483648 bytes exceeds ‘INT_MAX’ [-Werror=format-overflow=]
1408 | curl_msnprintf(buf, sizeof(buf), "%*f", INT_MIN, 9.1);
| ^~~
lib557.c:1408:3: note: ‘curl_msnprintf’ output 2147483649 bytes
1408 | curl_msnprintf(buf, sizeof(buf), "%*f", INT_MIN, 9.1);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
```
Ref: https://app.circleci.com/pipelines/github/curl/curl/10226/workflows/87642ee9-cda6-4916-8206-c82aac5f595e/jobs/107669?invite=true#step-106-40996_46
The root cause of why this option gets enabled remains undiscovered.
Reported-by: Daniel Stenberg
Fixes#15008
Follow-up to 71cf0d1fca#14772
- build: drop `-Wno-format-overflow` from picky warning list.
These options only get used with picky warnings enabled.
Follow-up to 145f87b9e8#14598
- unit1652: suppress in source (and not rely on picky warnings anymore.)
Closes#15012
Before this patch, only these triplets were considered Apple:
`<cpu>-apple-darwin`
After this patch, these are also considered Apple:
`<cpu>-apple-(ios*|tvos*|visionos*|watchos*|<ETC>)`
`$host_os` (the last third of the triplet) still has a valid use
to differentiate between OS flavours, though for now this isn't
used, aligning with CMake.
Closes#14728
- cmake: drop `configure.os`.
This also includes OS version, but thus far it's not important enough
to include it.
- autotools: drop redundant, autotools-only `{target|host}.vendor`.
(it's part of the triplet in `{target|host}`.)
- swap order to `*.cpu` -> `*.os` to match triplet-order.
- cmake: drop redundant `target`.
It's manually filled and only in a (so far) few CI jobs. Let's revisit
when this becomes useful.
- move `buildinfo.txt` to build root.
- dist: add `buildinfo.txt` to `DISTCLEANFILES`.
- autotools: detect human readable compiler version.
- autotools: replace `XXYY` `compiler.version` with "X.Y"-style.
(also to match cmake.)
- autotools: use distinct `compiler_id` for Apple clang: `APPLECLANG`.
To match cmake and also because the the "X.Y"-style version number
is the Apple version, while `XXYY` was a value roughly translated to
mainline llvm/clang version.
- show buildinfo at the end of the configure stage, when run in CI, or
when `CURL_BUILDINFO` or `CURL_CI` env is set.
Follow-up to 1fdea16846#14802
Assisted-by: Dan Fandrich
Ref: https://github.com/curl/curl/pull/14802#issuecomment-2334942991Closes#14822
It could previously be set with configure/cmake and used in rare cases
for reading randomness: with ancient mbedTLS or rustls without
arc4random.
We now get randomness in this order:
1. The TLS library's way to provide random
2. On Windows: Curl_win32_random
3. if arc4random exists, use that
4. weak non-crytographically strong pseudo-random
Closes#14749
For cross-builds rely on `_POSIX_C_SOURCE` to decide if `poll()` is
supported, rather than just assuming it isn't.
This may still miss to detect `poll()` support, as seen for example with
Linux MUSL cross-builds.
Also:
- GHA/curl-for-win: enable RISC-V 64 cross-target for Linux MUSL.
(to test this case with cmake, with a false-negative.)
The first RISC-V 64 build in curl's CI.
- GHA/curl-for-win: add arm64/intel64 job for Linux glibc.
(to test this case with cmake, and succeed.)
- cmake: delete unnecessary `#include <sys/time.h>` from non-cross-build
`poll()` detection snippet.
Follow-up tp cc8b813765#14718Fixes#14714Closes#14734
Drop Apple-specific detection logic for `poll()`. This detection snippet
has been disabled for Apple in both configure and cmake, for `poll()`
being broken on Apple since 10.12 Sierra (2016).
Also replace `exit(1);` with `return 1;` in configure, to make the
snippets match.
Added in 9297ca49f5#1057 (2016-10-11).
Disabled for:
configure/darwin in a34c7ce754 (2016-10-18)
cmake/macOS in 825911be58#7619
cmake/iOS in d14831233d#8244
cmake/all Apple in a86254b393#12515Closes#14718
POSIX sed doesn't support extended regular expressions, so convert a
call to the basic format. This caused a problem on AIX. Also, use the
detected sed binary name instead of hard-coding one.
-Wformat-overflow is not a warning that we want enabled as it does not
help us. It can only bring us false positives since it warns on bad uses
of sprintf and vsprintf ("that might overflow the destination buffer").
Two functions we explicitly ban in curl code.
The only way this flag triggers warnings in curl code is false positives
for functions we have marked with the CURL_PRINTF() macro.
Further: it seems -Wformat-trunaction option might in turn also enable
-Wformat-overflow, so if this second option is used, we need to
explicitly set -Wno-format-overflow - not just skip setting
-Wformat-overflow.
Reported-by: Viktor Szakats
Fixes#14168Closes#14598
- sync cmake macOS `OS` value manually with the autotools one.
- stop exporting/subst-ing `HAVE_OPENSSL_QUIC` from autotools.
The variable was only used internally.
- exclude a dependency detection symbol.
- allow to run when the workflow itself was updated.
- simplify cmake command.
- fix indentation.
Closes#14546
Use these words and casing more consistently across text, comments and
one curl tool output:
AIX, ALPN, ANSI, BSD, Cygwin, Darwin, FreeBSD, GitHub, HP-UX, Linux,
macOS, MS-DOS, MSYS, MinGW, NTLM, POSIX, Solaris, UNIX, Unix, Unicode,
WINE, WebDAV, Win32, winbind, WinIDN, Windows, Windows CE, Winsock.
Mostly OS names and a few more.
Also a couple of other minor text fixups.
Closes#14360
`nettle` is a direct dependency of curl, when building with GnuTLS.
Add a new `Find` module to detect it.
Also:
- GHA/macos: drop `nettle` hack no longer necessary.
- add `nettle` to `libcurl.pc`.
- also add `nettle` to `libcurl.pc` in autotools builds.
Follow-up to 781242ffa4#11967Closes#14285
Before this patch, `SystemConfiguration` detection failed due to this
error when compiling the detection snippet:
```
/Applications/Xcode_15.3.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/TargetConditionals.h:140:50: error: missing binary operator before token "("
140 | #if !defined(__has_extension) || !__has_extension(define_target_os_macros)
| ^
```
Ref: https://github.com/curl/curl/actions/runs/9821817534/job/27117929218#step:6:1079
It occured with gcc-11 when combined with macOS SDK 14.4 and 14.5
(default SDKs in Xcode 15.3 and 15.4 respectively). It did not happen
with earlier releases.
Despite the failure in `./configure`, `lib/macos.c` compiled with
Apple's `TargetConditionals.h` just fine.
Turns out that including the `sys/types.h` header before the SDK
header fixes the error and makes the detection snippet compile.
Cherry-picked from #14097Closes#14130
- cmake: enable Apple-specific `-Werror=partial-availability` to match
autotools.
- autotools: enable `-pedantic-errors` with llvm/clang to match gcc and
CMake.
- autotools: enable `-Werror-implicit-function-declaration` for
llvm/clang to match gcc.
- cmake: enable `-Werror-implicit-function-declaration` to match
autotools.
- move `-Wpointer-bool-conversion` from autotools to the local file
(`sectransp.c`) it was meant to apply. This way it applies to all
build methods.
- autotoos: show `CURL_CFLAG_EXTRAS` in the `./configure` summary.
(it may contain `-Werror` and/or `-pedentic-errors`.)
Cherry-picked from #14097Closes#14128
- sort features case-insensitively to match `curl -V` and cmake.
`sort -f` is POSIX, but check if it's available anyway.
- make protocols lowercase to match `curl -V` and cmake.
- replace two outlier `$()` with backticks.
Closes#14117
Also:
- remove stray `ECH` and `HTTPSRR` from cmake protocol list.
- stop excluding `Debug` and `TrackMemory` in `test1013.pl`.
- configure: delete `CURL_CHECK_CURLDEBUG` check.
Ref: 065047dc62
This check was effectively doing nothing, except disabling
`--enable-curldebug` in `curl-config` for
Cygwin/MSYS/cegcc/OS2/AIX targets with c-ares enabled.
Closes#14096
- when checking for QUIC support in OpenSSL, also check
for it being at least 3.3.0
- remove workarounds for features buggy or missing in 3.2
Closes#14026
- no longer warns for mbedtls
- warns for each item on individual lines
- no longer shows irrelevant TLS libraries when multiple are selected
- removes ech repetition
Closes#13941
- cmake: populate for dependencies.
- autotools: populate for dependencies.
(including mbedtls, though the script does not detect
mbedtls through pkgconfig. mbedtls 3.6.0 now supports it.)
Skip dealing with gssapi in this patch.
Fixes#864Closes#13911
MIPSPro and the predecessor compiler which was part of the IDO (IRIS
Development Option) were only ever shipped on the SGI IRIX operating
system (with MIPSPro on 6.0+ which was released in 1994). Limit the
autoconf check to IRIX when probing for these compilers to save some
cycles on other platforms.
Closes: #13611
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
It's necessary to set this var to "yes" _after_ AC_DEFINE and AC_SUBST
in order for a later `test` to pass so that `check_for_ca_bundle=1` ends
up being set. This is in turn required for the default CA certificate
bundle to be set when building w/ rustls & pkg-config.
Reported-by: Matt Jolly
Fixes#13248Closes#13251
