Stefan Eissing [Thu, 22 May 2025 13:09:58 +0000 (15:09 +0200)]
cf: replace the method get_host with query
Connection filters had a method `get_host()` which had not really been
documented. Since then, the cf had the `query()` method added. Replace
the separate get_host with query.
Add `CF_QUERY_HOST_PORT` as query to connection filters to retrieve
which remote hostname and port the filter (or its sub-filter) is talking
to. The query is implemented by HTTP and SOCKS filters, all others pass
it through.
Add `Curl_conn_get_current_host()` to retrieve the remote host and port
for a connection. During connect, this will return the host the
connection is talking to right now. Before/After connect, this will
return `conn->host.name`.
Stefan Eissing [Wed, 18 Jun 2025 10:34:43 +0000 (12:34 +0200)]
multi: add dirty bitset
Add a bitset `dirty` to the multi handle. The presence of a transfer int
he "dirty" set means: this transfer has something to do ASAP.
"dirty" is set by multiplexing protocols like HTTP/2 and 3 when
encountering response data for another transfer than the current one.
"dirty" is set by protocols that want to be called.
Implementation:
* just an additional `uint_bset` in the multi handle
* `Curl_multi_mark_dirty()` to add a transfer to the dirty set.
* `multi_runsingle()` clears the dirty bit of the transfer at
start. Without new dirty marks, this empties the set after
al dirty transfers have been run.
* `multi_timeout()` immediately gives the current time and
timeout_ms == 0 when dirty transfers are present.
* multi_event: marks all transfers tracked for a socket as dirty.
Then marks all expired transfers as dirty. Then it runs
all dirty transfers.
With this mechanism:
* Most uses of `EXPIRE_RUN_NOW` are replaced by `Curl_multi_mark_dirty()`
* `Curl_multi_mark_dirty()` is cheaper than querying if a transfer is
already dirty or set for timeout. There is no need to check, just do it.
* `data->state.select_bits` is eliminated. We need no longer to
simulate a poll event to make a transfer run.
Viktor Szakats [Sat, 21 Jun 2025 09:03:54 +0000 (11:03 +0200)]
build: tidy up `Makefile.inc` use in lib and src
- cmake: use `CURL_RCFILES` instead of literal.
- cmake: use `LIB_RCFILES` instead of literal.
- cmake: fix comments.
- autotools: use `CURL_RCFILES` in `EXTRA_DIST`.
- autotools: use `LIB_RCFILES` in `EXTRA_DIST`.
- autotools: fix comments.
- autotools: fix indentation.
DoI [Tue, 10 Jun 2025 11:13:35 +0000 (23:13 +1200)]
curl: implement non-blocking STDIN read on Windows
Implements a seperate read thread for STDIN on Windows when curl is run
with -T/--upload-file .
This uses a similar technique to the nmap/ncat project, spawning a
seperate thread which creates a loop-back bound socket, sending STDIN
into this socket, and reading from the other end of said TCP socket in a
non-blocking way in the rest of curl.
Joel Depooter [Fri, 20 Jun 2025 22:44:02 +0000 (15:44 -0700)]
libssh: fix incorrect return value in myssh_in_AUTH_PKEY_INIT
In the unlikely case that no SSH auth methods are supported, the
previous code would return 0 from myssh_in_AUTH_PKEY_INIT. However,
following the code path, it seems like it should be returning SSH_ERROR,
as set in myssh_to_ERROR (through myssh_to_GSSAPI_AUTH,
myssh_to_KEY_AUTH and myssh_to_PASSWD_AUTH).
In actuality, this is unlikely to occur, as the similar code in
myssh_in_AUTHLIST would have already returned an error in this scenario.
However setting a return value and then ignoring it is a bit fishy and
should be documented if this is intended.
I believe this used to return an error, but was changed in the recent
re-factoring of this code.
Viktor Szakats [Sat, 21 Jun 2025 00:22:04 +0000 (02:22 +0200)]
tests: drop `BUNDLE_SRC` variable
Derive it from `$BUNDLE` instead. autotools seems to be already relying
on `$BUNDLE_SRC` being equal to `$BUNDLE.c`. (I haven't realized this
before aaebb45f58b3f62876a68c17c71ac37d98f1b3bb.)
Also drop redundant `nodist_<target>_SOURCE` lines in tunits and units.
Viktor Szakats [Sat, 21 Jun 2025 08:00:22 +0000 (10:00 +0200)]
cmake: omit clang-tidy on internal libs curlu and curltool
Skip clang-tidy while compiling curlu and curltool internal libraries.
To save about 1 minute per run. These libraries compile the lib and src
sources a second time, with the `UNITTESTS` macro enabled, which makes
tiny difference, for internal use. I figure it's not worth the extra CI
(and local) time because finding extra issues in these passes is
unlikely, and if found, not critical.
autotools also doesn't check curlu and curltool with clang-tidy.
Fabrício Canedo [Tue, 17 Jun 2025 18:51:57 +0000 (15:51 -0300)]
docs: fix broken link in CODE_REVIEW.md
In CODE_REVIEW.md file, the link that points to CONTRIBUTE was broken,
so I fixed this issue changing the link from only "CONTRIBUTE.md" to
"https://curl.se/dev/contribute.html".
Fabrício Canedo [Tue, 17 Jun 2025 17:46:23 +0000 (14:46 -0300)]
docs: fix broken link in INSTALL.md
In INSTALL.md file, the link that points to CURL-DISABLE was broken,
so I fixed this issue changing the link from only "CURL-DISABLE.md" to
"https://github.com/curl/curl/blob/master/docs/CURL-DISABLE.md".
Stefan Eissing [Mon, 16 Jun 2025 07:54:20 +0000 (09:54 +0200)]
docs: fix documentation of connect_only 2
Setting CURLOPT_CONNECT_ONLY with value 2 is only defined
for WebSocket and the effect on other protocols is undetermined.
That includes the HTTP urls.
Stefan Eissing [Tue, 17 Jun 2025 11:13:26 +0000 (13:13 +0200)]
xfer: manage pause bits
Concentrate the handling of KEEP_RECV_PAUSE and KEEP_SEND_PAUSE into
common transfer functions. Setting or clearing these bits requires
subsequent actions involving connection events and client reader/writer
notifications. Have it in one place.
Stefan Eissing [Mon, 19 May 2025 12:38:22 +0000 (14:38 +0200)]
connection: eliminate member `remote_addr`
Used to be a pointer set (and cleared) by the socket connection filters
to a struct in their contexts. Instead, add a filter query method to
obtain the pointer when needed.
Viktor Szakats [Tue, 17 Jun 2025 18:10:28 +0000 (20:10 +0200)]
cmake: build `stubgss` library for libtests to match autotools
Used by test 2056 and 2057, in a way that's Linux- & autotools-specific.
This patch builds it for all Unix, syncing cmake with autotools.
Adapt the two tests to find the library in CMake builds as well.
Tested OK on Linux. (CI does not test this. The corresponding jobs build
in debug mode, while the `LD_PRELOAD` feature is locked to non-debug.)
On macOS it didn't load without building everything for aarch64e arch:
"../bld/tests/libtest/libstubgss.dylib' (mach-o file, but is
an incompatible architecture (have 'arm64', need 'arm64e'))"
With that fixed it still did not load correctly and/or the tests did not
pass. So, for macOS these tests remain disabled.
Also:
- GHA/macos: build for aarch64e. (recognized by Apple clang as of this
patch. llvm and gcc fall back to aarch64.)
Background:
As part of my work on reproducible builds for openSUSE, I check that
software still gives identical build results in the future. The usual
offset is +16 years, because that is how long I expect some software
will be used in some places. This showed up failing tests in our package
build. See https://reproducible-builds.org/ for why this matters.
I tested that it passes on x86_64 in year 2041 and i586 in year 2037.
(but on i586, I got `TESTFAIL: These test cases failed: 31 46 61 1415`)
Viktor Szakats [Thu, 19 Jun 2025 12:41:34 +0000 (14:41 +0200)]
warnless: drop parts of the `read`/`write` preprocessor hack (Windows)
The `#undef` hack is no longer necessary after changing the redifitions
to not map back to the original symbols.
This makes it unnecessary to repeat the redefinitions after compiling
`warnless.c` itself (in unity mode).
Which in turns makes it unnecessary to include `warnless.h` again, to
trigger such redefinition.
This also means that `read`/`write` are now redefined on Windows from
the first inclusion of `warnless.h`.
Also:
- tests/server: drop a repeat `warnless.h` include, that is unnecessary
after this patch.
- tests/unit: drop repeat `warnless.h` include.
- tests/libtest: drop repeat `warnless.h` includes.
- tests/libtest: formatting.
Viktor Szakats [Thu, 19 Jun 2025 12:05:25 +0000 (14:05 +0200)]
lib: make `CURLX_SET_BINMODE()` and use it
Use it from libtests' `first.c` and thus also from units, and tunits.
Also:
- cmake: drop stray `curltool` lib dependency for units.
- units: stop depending on `src` headers.
- tests/server: drop depending on `src` headers.
(the remaining one listed in the comments, `tool_xattr.h`, was not
actually used from servers.)
- tests/server: drop duplicate curlx headers.
(Except `warnless.h`, which is tricky on Windows.)
Viktor Szakats [Thu, 19 Jun 2025 09:54:47 +0000 (11:54 +0200)]
cmake: use `target_link_options()` when available
To pass `-municode` to the linker. Before this patch we passed this via
`target_link_libraries()` which is designed to pass libraries. Keep
using it for old CMake versions, where no better alternative existed.
Also:
- also pass `-municode` as `PRIVATE` for old cmake versions.
(it should not make a difference because no target depends on the curl
tool, but this seem to be the modern, non-ambiguous syntax.)
- unfold a bunch of split lines for greppability of `add_library()` and
`add_executable()` commands.
- quote a string.
Viktor Szakats [Wed, 18 Jun 2025 08:33:15 +0000 (10:33 +0200)]
autotools: simplify configuration in tests, examples
- GHA/windows: make a mingw autotools build static only.
- GHA/windows: fix a CI script issue with the build above.
- src: fix to pass `LIBCURL_PC_LIBS_PRIVATE` instead of `LINKFLAGS`.
This makes the libs propagate to tunits, making the local hack there
unnecessary. `LINKFLAGS` had this single use in the repo, and it was
empty in local tests.
- tests: drop passing redundant `LIBCURL_PC_LDFLAGS_PRIVATE`.
- tests: drop redundant target name from config variables.
- examples, tests/client: drop `LIBDIR` temp variables with single uses.
- examples, tests: formatting to sync `Makefile.am` scripts with each
other.
Stefan Eissing [Mon, 16 Jun 2025 11:40:15 +0000 (13:40 +0200)]
multi: do no expire a blocked transfer
When checking to expire a transfer with input data pending, also assess
the blocked status and do not EXPIRE_RUN_NOW a transfer that is blocked
on READ/WRITE.
Viktor Szakats [Mon, 16 Jun 2025 14:04:22 +0000 (16:04 +0200)]
tests/http/clients: drop hack and use `curl_setup.h` again
Sync build properties with libtests.
This allows accessing macros from `curl_config.h`, for feature flags.
Smoothens out platform bumps, allowing to drop local replicas from
client sources. It enables using Windows wrappers, e.g. for `fopen()`.
Also fix client sources to use `curl_mfprintf()` where curl format
strings are used. (To avoid build failure with older mingw-w64, e.g.
6.4.0 in CI.)
Stefan Eissing [Mon, 16 Jun 2025 10:19:52 +0000 (12:19 +0200)]
multi: fix polling with pending input
When multi creates the pollset of a transfer, it checks now if
a connection (FIRST/SECONDARY) socket waits on POLLIN and has input data
pending in filters (relevant to OpenSSL's new read ahead). If so, it
triggers a timeout on the transfer via EXPIRE_RUN_NOW.
This fixes sporadic stalls in test 988 when running event based.
Viktor Szakats [Sun, 15 Jun 2025 12:48:46 +0000 (14:48 +0200)]
tests: bundle http clients, de-dupe, enable for MSVC
To make building the http client tests faster, with no duplication, by
using the build method that other test binaries already use.
The difference compared to other tests is that these don't use internal
libcurl headers or code. With the exception of `curl_config.h`, for
a feature macro.
Before this patch, these tests were built like examples.
Also:
- de-duplicate code and give unique names to colliding symbols.
- add local getopt implementation and enable all code for MSVC.
Adapted for curl via Public Domain source:
https://github.com/skeeto/getopt/blob/4e618ef782dc80b2cf0307ea74b68e6a62b025de/getopt.h Credits-to: Christopher Wellons
Thanks!
Viktor Szakats [Mon, 16 Jun 2025 08:16:17 +0000 (10:16 +0200)]
windows: fixup `fopen()` in `CURLDEBUG` builds
Introduce an immutable `CURL_FOPEN()` macro to store the `fopen()`
mapping on Windows. Then use that instead `(fopen)` from `memdebug.c`.
It makes CURLDEBUG builds use the correct `fopen` wrapper on Windows.
This macro is only defined on Windows, as of this patch.
Also:
- curl_setup.h: de-dupe, simplify Windows file I/O function overrides.
- curl_memory.h: fix to reset `fopen` to `curlx_win32_fopen()` on
Windows. Before this patch it reset it to stock `fopen()`.
Viktor Szakats [Mon, 16 Jun 2025 00:07:31 +0000 (02:07 +0200)]
memdebug: include in unity batch
Before this patch `memdebug.c` was compiled as a separate source in
unity builds. This was necessary because `memdebug.c` failed to compile
if `memdebug.h` was included before it, in `CURLDEBUG` mode. This patch
fixes this issue and allows to compile `memdebug.c` as part of the unity
source batch. This removes an exception and makes builds perform a notch
better.
- introduce `CURL_SCLOSE()` macro as an immutable synonym of `sclose()`.
- memdebug: replace `sclose()` reference with `CURL_SCLOSE()` to compile
as expected when `sclose()` is overridden by `memdebug.h`.
- memdebug: make it not break when including `memdebug.h` before it in
`CURLDEBUG` mode. Do this by calling low-level functions as
`(function)`.
- autotools, cmake: drop memdebug exception, include it like any other
source file. This is now possible because `memdebug.c` doesn't break
if `memdebug.h` was included before it, in `CURLDEBUG` builds.
- mk-unity: drop `--exclude` option. No longer used after this patch.
- drop `MEMDEBUG_NODEFINES` macro hack. No longer necessary.
Viktor Szakats [Sun, 15 Jun 2025 21:28:08 +0000 (23:28 +0200)]
tests/server: drop memdebug option
I added it just in case when removing enabled-by-default memdebug
from test servers. Apparently it broke after recent changes. It's
probably not a widely used feature and does not seem to be worth fixing.
It creates odd dependencies as the error message indicates:
```
[28/54] Building C object tests/server/CMakeFiles/servers.dir/__/__/lib/memdebug.c.obj
FAILED: tests/server/CMakeFiles/servers.dir/__/__/lib/memdebug.c.obj
[...]
lib/memdebug.c: In function 'curl_dbg_log':
lib/memdebug.c:465:12: error: implicit declaration of function 'mvsnprintf'; did you mean 'vsnprintf'? [-Wimplicit-function-declaration]
465 | nchars = mvsnprintf(buf, sizeof(buf), format, ap);
| ^~~~~~~~~~
| vsnprintf
lib/memdebug.c:465:12: warning: nested extern declaration of 'mvsnprintf' [-Wnested-externs]
```
This patch is dropping these build options:
- cmake: `ENABLE_SERVER_DEBUG`
- autotools: `--enable-server-debug` / `--disable-server-debug`
Viktor Szakats [Sun, 15 Jun 2025 19:54:20 +0000 (21:54 +0200)]
build: sync tests unity builds between cmake and autotools
Instead of relying on CMake's built-in unity feature, use `mk-unity.pl`,
as already done with autotools. It simplified the build, shortens logs
and makes debugging easier because of the fewer build variations.
It also allows testing / fixing with cmake and those automatically apply
to autotools builds too. cmake builds can be much-much faster, esp.
when working the builds themselves.
It also enables "unity" in old cmake versions. Basically every test
target is a single generated .c source.
Also:
- drop a `lib` unity workaround for libtests with autotools after fixing
the issue in libtests itself. It drops a few exceptions and makes
libcurl build faster (in autotools unity).
- fix another `lib` autotools unity issue and drop the workaround for it
from `mk-unity.pl`. `srcdir` was missing from the header path.
- simplify `mk-unity.pl` command-lines, drop exclusions.
Viktor Szakats [Wed, 11 Jun 2025 05:32:16 +0000 (07:32 +0200)]
build: drop `HAVE_SYS_SOCKET_H` and `HAVE_SYS_TIME_H` macros
Both are available with well-known conditions, under non-Windows, and
`curl/curl.h` already uses them. `sys/time.h` is also necessary for
mingw-w64 for `gettimeofday()`.
Stefan Eissing [Fri, 13 Jun 2025 07:43:40 +0000 (09:43 +0200)]
lib: avoid reusing unclean connection
When `curl_easy_cleanup()` is invoked while still being part
of a multi handle, the code will auto-remove it. But since the
connection was detached first, the code in
`curl_multi_remove_handle()` that invalidates dirty connections
did not bite.
Attach the connection *after* the possible remove from a multi
handle, so that connection reuse can be prevented.
Add test753 to reproduce and verify the fix. This required adding
the new debug env var CURL_FTP_PWD_STOP, to have a transfer return
from multi_perform() early with a connection that does not show
and pending input.
Reported-by: Brian Harris
Fixes https://github.com/curl/curl/issues/17578
Closes https://github.com/curl/curl/pull/17607
Theodore A. Roth [Thu, 12 Jun 2025 17:36:02 +0000 (11:36 -0600)]
openssl: Fix openssl engines
- Return CURLE_OK if the engine successfully loaded.
Prior to this change:
When loading an openssl engine, the result code is initialized to
CURLE_SSL_ENGINE_NOTFOUND, but is never set to CURLE_OK when the engine
was successfully loaded. This causes curl to error out, falsely stating
engine not found when it actually was.
Broken since f2ce6c46 (precedes 8.14.0) which added support for using
engines and providers at the same time.
Rod Widdowson [Wed, 21 May 2025 19:10:36 +0000 (20:10 +0100)]
schannel: allow partial chains for manual peer verification
- Align --cacert behaviour with OpenSSL and LibreSSL.
This changes the default behavior of Schannel manual certificate
verification, which is used when the user provides their own CA
certificates for verification, to accept partial chains. In other words,
the user may provide an intermediate certificate without having to
provide the root CA.
Win8/Server2012 widened the PKIX chain traversal API to allow
certificate traversal to terminate at an intermediate.
This behaviour (terminate at the fist matching intermediate) is the
default for LibreSSL and OpenSSL (with OpenSSL allowing control via
CURLSSLOPT_NO_PARTIALCHAIN).
This change uses the new API if it is available, and also allows the
behaviour to revert legacy if CURLSSLOPT_NO_PARTIALCHAIN is present.
Viktor Szakats [Wed, 11 Jun 2025 09:31:14 +0000 (11:31 +0200)]
tests: always make bundles, adapt build and tests
Make test bundles the default. Drop non-bundle build mode.
Also do all the optimizations and tidy-ups this allows, simpler builds,
less bundle exceptions, streamlined build mechanics.
Also rework the init/deinit macro magic for unit tests. The new method
allows using unique init/deinit function names, and calling them with
arguments. This is in turn makes it possible to reduce the use of global
variables.
Note this drop existing build options `-DCURL_TEST_BUNDLES=` from cmake
and `--enable-test-bundles` / `--disable-test-bundles` from autotools.
Also:
- rename test entry functions to have unique names: `test_<testname>`
This removes the last exception that was handled in the generator.
- fix `make dist` to not miss test sources with test bundles enabled.
- sync and merge `tests/mk-bundle.pl` into `scripts/mk-unity.pl`.
- mk-unity.pl: add `--embed` option and use it when `CURL_CLANG_TIDY=ON`
to ensure that `clang-tidy` does not miss external test C sources.
(because `clang-tidy` ignores code that's #included.)
- tests/unit: drop no-op setup/stop functions.
- tests: reduce symbol scopes, global macros, other fixes and tidy-ups.
- tool1621: fix to run, also fix it to pass.
- sockfilt: fix Windows compiler warning in certain unity include order,
by explicitly including `warnless.h`.
Stefan Eissing [Thu, 12 Jun 2025 06:45:20 +0000 (08:45 +0200)]
openssl: fix handling of buffered data
`SSL_pending()` only checks if the *current* TLS packet has more data.
There might be more data in SSL's buffers.
`SSL_has_pending()` only checks if there is data in buffers, but does
*not* check if there is a complete TLS packet that can be decoded.
If we only check the first, we will poll on socket events without having
processed all data and may stall. If we only check the second, we would
busy loop without SSL_read() ever giving something.
Add the flag `connssl->input_pending` that is set on incoming data in
the BIO receive. Clear the flag when encountering a CURLE_AGAIN on
the filters receive (via SSL_read()) or see an EOF.
Daniel McCarney [Wed, 11 Jun 2025 12:59:11 +0000 (08:59 -0400)]
tls: remove Curl_ssl false_start
The secure transport vTLS backend was the only Curl_ssl struct instance
that populated the false_start field. Since its removed, we can now
remove that field entirely. This was a protocol feature specific to TLS
1.2 that has been replaced by the more widely adopted TLS 1.3 early data
mechanisms.
Viktor Szakats [Wed, 11 Jun 2025 14:28:42 +0000 (16:28 +0200)]
libtests: stop building the sames source multiple times
After this patch there is no more double/multiple compile of the same
libtest source under a different libtest ID. Each libtest is compiled
once, and changing behavior at runtime based on test ID.
- drop recently added physical clones for two prevously multi-compiled
tests:
- merge lib587 into lib554 again, branch at runtime.
- merge lib645 into lib643 again, branch at runtime.
- replace existing dynamic branching to use `testnum` instead of
a manually rolled `testno` based on an extra command-line argument.
lib1571, lib1576.
- mk-bundle.pl: stop defining `LIB*` macros. No longer used.
- libtests: drop all `LIB*_C` guards.
- Make these tests branch at runtime, stop building copies:
- lib585, based on lib500
- lib565, based on lib510
- lib529, based on lib525
- lib527, lib532, based on lib526
- lib545, based on lib544
- lib548, based on lib547
- lib696, based on lib556
- lib584, based on lib589
- lib1539, based on lib1514
- lib1543, based on lib1518
- lib1917, based on lib1916
- lib1946, based on lib1940
- lib671, 672, 673, based on lib670
Stefan Eissing [Wed, 4 Jun 2025 09:12:28 +0000 (11:12 +0200)]
openssl: error on SSL_ERROR_SYSCALL
Convert the debug-only handling of SSL_ERROR_SYSCALL so that it is
enabled in all builds with openssl. This should not make a difference in
supported OpenSSL versions, but if whatever version or fork we link
against *does* return SSL_ERROR_SYSCALL, handle this as a fatal error.
Fixes #17471 Reported-by: Michael Kaufmann
Closes #17531
Stefan Eissing [Tue, 10 Jun 2025 15:14:00 +0000 (17:14 +0200)]
pytest test_07_70, weaken early data check
Since the amount of early data sent in the upload test_07_70 varies
much with how fast the server respondes and completes the handshake,
assert that we did sent at least *some* early data instead of relying
on a specific amount.
Stefan Eissing [Tue, 10 Jun 2025 08:11:40 +0000 (10:11 +0200)]
url: fix connection lifetime checks
The checks for a connection being "too long idle" or "too old" where
rounding down the elapsed time to seconds before comparing to the
configured max values. This caused connections to be reused for up to
999ms longer than intended.
Change the compares to scale the configured seconds up to ms, so
connection will properly be "too old" 1 ms after the coonfigured values.
Fixes sporadic failures of test1542 on platforms where "sleep(2)"
returnes before 2 full seconds on the internal clock where passed.
Reported-by: Christian Weisgerber
URL: https://curl.se/mail/lib-2025-06/0004.html
Closes #17571
projects / thirdparty / curl.git / log
