Connection filter had a `get_select_socks()` method, inspired by the
various `getsocks` functions involved during the lifetime of a
transfer. These, depending on transfer state (CONNECT/DO/DONE/ etc.),
return sockets to monitor and flag if this shall be done for POLLIN
and/or POLLOUT.
Due to this design, sockets and flags could only be added, not
removed. This led to problems in filters like HTTP/2 where flow control
prohibits the sending of data until the peer increases the flow
window. The general transfer loop wants to write, adds POLLOUT, the
socket is writeable but no data can be written.
This leads to cpu busy loops. To prevent that, HTTP/2 did set the
`SEND_HOLD` flag of such a blocked transfer, so the transfer loop cedes
further attempts. This works if only one such filter is involved. If a
HTTP/2 transfer goes through a HTTP/2 proxy, two filters are
setting/clearing this flag and may step on each other's toes.
Connection filters `get_select_socks()` is replaced by
`adjust_pollset()`. They get passed a `struct easy_pollset` that keeps
up to `MAX_SOCKSPEREASYHANDLE` sockets and their `POLLIN|POLLOUT`
flags. This struct is initialized in `multi_getsock()` by calling the
various `getsocks()` implementations based on transfer state, as before.
After protocol handlers/transfer loop have set the sockets and flags
they want, the `easy_pollset` is *always* passed to the filters. Filters
"higher" in the chain are called first, starting at the first
not-yet-connection one. Each filter may add sockets and/or change
flags. When all flags are removed, the socket itself is removed from the
pollset.
Example:
* transfer wants to send, adds POLLOUT
* http/2 filter has a flow control block, removes POLLOUT and adds
POLLIN (it is waiting on a WINDOW_UPDATE from the server)
* TLS filter is connected and changes nothing
* h2-proxy filter also has a flow control block on its tunnel stream,
removes POLLOUT and adds POLLIN also.
* socket filter is connected and changes nothing
* The resulting pollset is then mixed together with all other transfers
and their pollsets, just as before.
Use of `SEND_HOLD` is no longer necessary in the filters.
All filters are adapted for the changed method. The handling in
`multi.c` has been adjusted, but its state handling the the protocol
handlers' `getsocks` method are untouched.
The most affected filters are http/2, ngtcp2, quiche and h2-proxy. TLS
filters needed to be adjusted for the connecting handshake read/write
handling.
No noticeable difference in performance was detected in local scorecard
runs.
Closes#11833
Put the instructions to run tests right at the top of tests/README.md.
Give instructions to read the runtests.1 man page for information
about flags. Delete redundant copy of the flags documentation in the
README.
Add a mention in README.md of the important parallelism flag, to make
test runs go much faster.
Move documentation of output line format into the runtests.1 man page,
and update it with missing flags.
Fix the order of two flags in the man page.
Closes#12193
The goal of this patch is to avoid unnecessary feature detection work
when doing Windows builds with CMake. Do this by pre-filling well-known
detection results for Windows and specifically for mingw-w64 and MSVC
compilers. Also limit feature checks to platforms where the results are
actually used. Drop a few redundant ones. And some tidying up.
- pre-fill remaining detection values in Windows CMake builds.
Based on actual detection results observed in CI runs, preceding
similar work over libssh2 and matching up values with
`lib/config-win32.h`.
This brings down CMake configuration time from 58 to 14 seconds on the
same local machine.
On AppVeyor CI this translates to:
- 128 seconds -> 50 seconds VS2022 MSVC with OpenSSL (per CMake job):
https://ci.appveyor.com/project/curlorg/curl/builds/48208419/job/4gw66ecrjpy7necb#L296https://ci.appveyor.com/project/curlorg/curl/builds/48217440/job/8m4fwrr2fe249uo8#L186
- 62 seconds -> 16 seconds VS2017 MINGW (per CMake job):
https://ci.appveyor.com/project/curlorg/curl/builds/48208419/job/s1y8q5ivlcs7ub29?fullLog=true#L290https://ci.appveyor.com/project/curlorg/curl/builds/48217440/job/pchpxyjsyc9kl13a?fullLog=true#L194
The formula is about 1-3 seconds delay for each detection. Almost all
of these trigger a full compile-link cycle behind the scenes, slow
even today, both cross and native, mingw-w64 and apparently MSVC too.
Enabling .map files or other custom build features slows it down
further. (Similar is expected for autotools configure.)
- stop detecting `idn2.h` if idn2 was deselected.
autotools does this.
- stop detecting `idn2.h` if idn2 was not found.
This deviates from autotools. Source code requires both header and
lib, so this is still correct, but faster.
- limit `ADDRESS_FAMILY` detection to Windows.
- normalize `HAVE_WIN32_WINNT` value to lowercase `0x0a12` format.
- pre-fill `HAVE_WIN32_WINNT`-dependent detection results.
Saving 4 (slow) feature-detections in most builds: `getaddrinfo`,
`freeaddrinfo`, `inet_ntop`, `inet_pton`
- fix pre-filled `HAVE_SYS_TIME_H`, `HAVE_SYS_PARAM_H`,
`HAVE_GETTIMEOFDAY` for mingw-w64.
Luckily this do not change build results, as `WIN32` took
priority over `HAVE_GETTIMEOFDAY` with the current source
code.
- limit `HAVE_CLOCK_GETTIME_MONOTONIC_RAW` and
`HAVE_CLOCK_GETTIME_MONOTONIC` detections to non-Windows.
We're not using these in the source code for Windows.
- reduce compiler warning noise in CMake internal logs:
- fix to include `winsock2.h` before `windows.h`.
Apply it to autotools test snippets too.
- delete previous `-D_WINSOCKAPI_=` hack that aimed to fix the above.
- cleanup `CMake/CurlTests.c` to emit less warnings.
- delete redundant `HAVE_MACRO_SIGSETJMP` feature check.
It was the same check as `HAVE_SIGSETJMP`.
- delete 'experimental' marking from `CURL_USE_OPENSSL`.
- show CMake version via `CMakeLists.txt`.
Credit to the `zlib-ng` project for the idea:
61e181c8ae/CMakeLists.txt (L7)
- make `CMake/CurlTests.c` pass `checksrc`.
- `CMake/WindowsCache.cmake` tidy-ups.
- replace `WIN32` guard with `_WIN32` in `CMake/CurlTests.c`.
Closes#12044
Currently the verbose output does not include which algorithms are used
for the signature and key exchange when using OpenSSL. Including the
algorithms used will enable better debugging when working on using new
algorithm implementations. Know what algorithms are used has become more
important with the fast growing research into new quantum-safe
algorithms.
This implementation includes a build time check for the OpenSSL version
to use a new function that will be included in OpenSSL 3.2 that was
introduced in openssl/openssl@6866824
Based-on-patch-by: Martin Schmatz <mrt@zurich.ibm.com>
Closes#12030
Getting nghttp2's error message helps users understand what's going
on. For example when the connection is brought down due a forbidden
header is used - as that header is then not displayed by curl itself.
Example:
curl: (92) Invalid HTTP header field was received: frame type: 1,
stream: 1, name: [upgrade], value: [h2,h2c]
Ref: #12172Closes#12179
... and make the code require both symbol and declaration.
This is because for Android, the symbol is always present in the lib at
build-time even when not actually available in run-time.
Assisted-by: Viktor Szakats
Reported-by: 12932 on github
Fixes#12086Closes#12158
An orphan call to `CheckQuicSupportInOpenSSL()` remained after a recent
update when checking QUIC for quiche. Move back QUIC detection to
a function and fixup callers to use that. Also make sure that quiche
gets QUIC from BoringSSL, because it doesn't support other forks at this
time.
Regression from dee310d54261f9a8416e87d50bccfe2cbe404949 #11555
Reported-by: Casey Bodley <cbodley@redhat.com>
Fixes#12160Closes#12162
Fixes a minor memory leak on LDAP connection reuse.
Doing the allocation already in *setup_connection() is wrong since that
connect struct might get discarded early when an existing connection is
reused instead.
Closes#12166
Remove the CURL_CA_FALLBACK logic. That build option was added to allow
primarily OpenSSL to use the default paths for loading the CA certs. For
GnuTLS it was instead made to load the "system certs", which is
different and not desirable.
The native CA store loading is now asked for with this option.
Follow-up to 7b55279d1d856
Co-authored-by: Jay Satiro
Closes#12137
- fix HTTP header parsing to report incomplete
lines it buffers as consumed!
- re-implement the RTP parser for interleave RTP
messages for robustness. It is now keeping its
state at the connection
- RTSP protocol handler "readwrite" implementation
now tracks if the response is before/in/after
header parsing or "in" a bod by calling
"Curl_http_readwrite_headers()" itself. This
allows it to know when non-RTP bytes are "junk"
or HEADER or BODY.
- tested with #12035 and various small receive
sizes where current master fails
Closes#12052
- fold the code to convert dynhds to the nghttp2 structs
into a dynhds internal method
- saves code duplication
- pacifies compiler analyzers
Closes#12097
Python precheck/postcheck alternatives were included but commented out.
Since these are not used and perl is guaranteed to be available to run
the perl versions anyway, the Python ones are removed.
To avoid the state machine to start over and redownload all the files
*again*.
Reported-by: lkordos on github
Regression from 843b3baa3e3cb228 (shipped in 8.1.0)
Bisect-by: Dan Fandrich
Fixes#11775Closes#12156
The checkcmd() and checktestcmd() functions would not have worked on
Windows due to hard-coding the UNIX PATH separator character and not
adding .exe file extension. This meant that tools like stunnel, valgrind
and nghttpx would not have been found and used on Windows, and
inspection of previous test runs show none of those being found in pure
Windows CI builds.
With this fixed, they can be used to detect the handle64.exe program
before attempting to use it. When handle64.exe was called
unconditionally without it existing, it caused perl to abort the test
run with the error
The running command stopped because the preference variable
"ErrorActionPreference" or common parameter is set to Stop:
sh: handle64.exe: command not found
Closes#12115
- Add 'threadsafe' to the feature list shown during build if POSIX
threads are being used.
This is a follow-up to 5adb6000 which added support for building a
thread-safe libcurl with older versions of gcc where atomic is not
available but pthread is.
Reported-by: Dan Fandrich
Co-authored-by: Dan Fandrich
Fixes https://github.com/curl/curl/issues/12125
Closes https://github.com/curl/curl/pull/12127
The idea of `check_library_exists_concat()` is that it detects an
optional component and adds it to the list of libs that we also use in
subsequent component checks. This caused problems when detecting
components with unnecessary dependencies that were not yet built.
CMake offers the `CMAKE_REQUIRED_LIBRARIES` variable to set libs used
for component checks, which we already use in most cases. That left 4
uses of `check_library_exists_concat()`. Only one of these actually
needed the 'concat' feature (ldap/lber).
Delete this function and replace it with standard
`check_library_exists()` and manual management of our `CURL_LIBS`
list we use when linking build targets. And special logic to handle the
ldap/lber case.
(We have a similar function for headers: `check_include_file_concat()`.
It works, but problematic for performance reasons and because it hides
the actual headers required in `check_symbol_exists()` calls.)
Ref: #11537#11558Fixes#11285Fixes#11648Closes#12070
- Pass missing parameter for 'lpNumberOfCharsWritten' to WriteConsoleW()
function.
Apparently this parameter was *not* optional on older Windows versions.
Issue observed on Windows XP SP2. Issue not observed on Windows 7 SP1.
So at some point between those two Microsoft changed the behavior.
Prior to this change, on those versions if parameter is NULL then the
function call fails with error ERROR_INVALID_ACCESS.
Regression since af3f4e41.
Ref: https://github.com/MicrosoftDocs/Console-Docs/issues/299
Fixes https://github.com/curl/curl/issues/12131
Closes https://github.com/curl/curl/pull/12130
- Add additional checking for missing and too-short SOCKS5 handshake
messages.
Prior to this change the SOCKS5 test server did not check that all parts
of the handshake were received successfully. If those parts were missing
or too short then the server would access uninitialized memory.
This issue was discovered in CI job 'memory-sanitizer' test results.
Test 2055 was failing due to the SOCKS5 test server not running. It was
not running because either it crashed or memory sanitizer aborted it
during Test 728. Test 728 connects to the SOCKS5 test server on a
redirect but does not send any data on purpose. The test server was not
prepared for that.
Reported-by: Dan Fandrich
Fixes https://github.com/curl/curl/issues/12117
Closes https://github.com/curl/curl/pull/12118
Currently, curl allows users to specify absurd request rates that might
be higher than the number of milliseconds in the unit (ex: curl --rate
3600050/h http://localhost:8080 does not error out despite there being
only 3600000ms in a hour).
This change adds a conditional check before the millisecond calculation
making sure that the number is not higher than the numerator (the unit)
If the number is higher, curl errors out with PARAM_NUMBER_TOO_LARGE
Closes#12116
- Remove DEBUGASSERT that an internal handle must not have user
private_data set before calling the user's debug callback.
This is a follow-up to 0dc40b2a. The user can distinguish their easy
handle from an internal easy handle by setting CURLOPT_PRIVATE on their
easy handle. I had wrongly assumed that meant the user couldn't then
set CURLOPT_PRIVATE on an internal handle as well.
Bug: https://github.com/curl/curl/pull/12060#issuecomment-1754594697
Reported-by: Daniel Stenberg
Closes https://github.com/curl/curl/pull/12104
