- Send no more than 150 cookies per request
- Cap the max length used for a cookie: header to 8K
- Cap the max number of received Set-Cookie: headers to 50
Bug: https://curl.se/docs/CVE-2022-32205.html
CVE-2022-32205
Reported-by: Harry Sintonen
Closes#9048
clang 14 warns about its use. It is being deprecated by the working
group for the programming language C: "The macro ATOMIC_VAR_INIT is
basically useless for the purpose for which it was designed"
Ref: https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2886.htm
Reported-by: Tatsuhiro Tsujikawa
Fixes#9041Closes#9042
Testing on macOS 12.4, sendmsg() fails with EINVAL when a msg_control
buffer is provided in sengmsg(), even though msg_controllen was set to
0.
Initialize msg.msg_controllen just as needed and also perform the size
assertion only when needed.
Closes#9039
connect_init() (lib/http_proxy.c) swaps out the protocol state while
working on the proxy connection, this is then restored by
Curl_connect_done() after the connection completes.
ftp_do_more() extracted the protocol state pointer to a local variable
at the start of the function then calls Curl_proxy_connect(). If the proxy
connection completes, Curl_proxy_connect() will call Curl_connect_done()
(via Curl_proxyCONNECT()), which restores data->req.p to point to the ftp
protocol state instead of the http proxy protocol state, but the local
variable in ftp_do_more still pointed to the old value.
Ultimately this meant that the state worked on by ftp_do_more() was the
http proxy state not the ftp state initialised by ftp_connect(), but
subsequent calls to any ftp_ function would use the original state.
For my use-case, the visible consequence was that ftp->downloadsize was
never set and so downloaded data was never returned to the application.
This commit updates the ftp protocol state pointer in ftp_do_more() after
Curl_proxy_connect() returns, ensuring that the correct state pointer is
used.
Fixes#8737Closes#9043
Prior to this change _mingw.h needed to be included in each unit before
evaluating __MINGW{32,64}_xxx_VERSION macros since it defines them. It
is included only in some mingw headers (eg stdio.h) and not others
(eg windows.h) so it's better to explicitly include it once.
Closes https://github.com/curl/curl/pull/9036
- Prevent CMake to auto-detect /dev/urandom when cross-building.
Before this patch, it would detect it in a cross-build scenario on *nix
hosts with this device present. This was a problem for example with
Windows builds, but it could affect any target system with this device
missing. This also syncs detection behaviour with autotools, which also
skips it for cross-builds.
- Also, make sure to never use the file RANDOM_FILE as entropy for libcurl's
fallback random number generator on Windows. Windows does not have the
concept of reading a random stream from a filename, nor any guaranteed
non-world-writable path on disk. With this, a manual misconfiguration or
an overeager auto-detection can no longer result in a user-controllable
seed source.
Reviewed-by: Daniel Stenberg
Closes#9038
Since this [1] commit in 2011, `_WIN32_WINNT` was set fixed to Windows
XP when the `-ipv6` option is selected. Maybe this was added to support
pre-XP Windows versions (?). These days libcurl builds fine for both XP
and post-XP versions with IPv6 support enabled. The relevance of pre-XP
version is also low by now. Other build methods also do not impose such
limitation for a similar configuration. So, drop this hard-wired
`_WIN32_WINNT` limit from `Makefile.m32`, thus building for the default
Windows version set by the compiler. This is Vista for recent MinGW
versions.
Old behaviour can be restored by setting this envvar:
export CURL_CFLAG_EXTRAS=-D_WIN32_WINNT=0x0501
[1] 98a61d8e2eCloses#9035
Append to the upload buffer when only small amount remains in buffer
rather than performing a separate tiny send to empty buffer.
Avoid degenerative upload behavior which might cause curl to send mostly
1-byte DATA frames after exhausing the h2 send window size
Related discussion: https://github.com/nghttp2/nghttp2/issues/1722
Signed-off-by: Glenn Strauss <gstrauss@gluelogic.com>
Closes#8965
As per the documentation :
> Setting a part to a NULL pointer will effectively remove that
> part's contents from the CURLU handle.
But currently clearing CURLUPART_URL does nothing and returns
CURLUE_OK. This change will clear all parts of the URL at once.
Closes#9028
.. and update some docs to explain curl_global_* is now thread-safe.
Follow-up to 23af112 which made curl_global_init/cleanup thread-safe.
Closes https://github.com/curl/curl/pull/9016
Referring to Daniel's article [1], making the init function thread-safe
was the last bit to make libcurl thread-safe as a whole. So the name of
the feature may as well be the more concise 'threadsafe', also telling
the story that libcurl is now fully thread-safe, not just its init
function. Chances are high that libcurl wants to remain so in the
future, so there is little likelihood of ever needing any other distinct
`threadsafe-<name>` feature flags.
For consistency we also shorten `CURL_VERSION_THREADSAFE_INIT` to
`CURL_VERSION_THREADSAFE`, update its description and reference libcurl's
thread safety documentation.
[1]: https://daniel.haxx.se/blog/2022/06/08/making-libcurl-init-more-thread-safe/
Reviewed-by: Daniel Stenberg
Reviewed-by: Jay Satiro
Closes#8989
Add licensing and copyright information for all files in this repository. This
either happens in the file itself as a comment header or in the file
`.reuse/dep5`.
This commit also adds a Github workflow to check pull requests and adapts
copyright.pl to the changes.
Closes#8869
Adds two new error codes: CURLE_UNRECOVERABLE_POLL and
CURLM_UNRECOVERABLE_POLL one each for the easy and the multi interfaces.
Reported-by: Harry Sintonen
Fixes#8921Closes#8961
This flag can be used to make sure that curl_global_init() is
thread-safe.
This can be useful for libraries that can't control what other
dependencies are doing with Curl.
Closes#8680
The callback set by CURLOPT_SSH_HOSTKEYFUNCTION is called to check
wether or not the connection should continue.
The host key is passed in argument with a custom handle for the
application.
It overrides CURLOPT_SSH_KNOWNHOSTS
Closes#7959
RFC 7616 (and 2617) requires values to be "unquoted" before used for
digest calculations. The only place where unquoting can be done
correctly is header parsing function (realm="DOMAIN\\host" and
realm=DOMAN\\host are different realms).
This commit adds unquoting (de-escaping) of all values during header
parsing and quoting of the values during header forming. This approach
should be most straightforward and easy to read/maintain as all values
are processed in the same way as required by RFC.
Closes#8912
Enable Linux GSO in ngtcp2 QUIC. In order to recover from the
EAGAIN/EWOULDBLOCK by sendmsg with multiple packets in one GSO write,
packet buffer is now held by struct quicsocket. GSO write might fail in
runtime depending on NIC. Disable GSO if sendmsg returns EIO.
Closes#8909
As that would indicate an illegal header. The fuzzer reached the assert
in unfold_value() proving that this case can happen.
Follow-up to c9b60f0053Closes#8939
The .netrc parser now accepts strings within double-quotes in order to
deal with for example passwords containing white space - which
previously was not possible.
A password that starts with a double-quote also ends with one, and
double-quotes themselves are escaped with backslashes, like \". It also
supports \n, \r and \t for newline, carriage return and tabs
respectively.
If the password does not start with a double quote, it will end at first
white space and no escaping is performed.
WARNING: this change is not entirely backwards compatible. If anyone
previously used a double-quote as the first letter of their password,
the parser will now get it differently compared to before. This is
highly unfortunate but hard to avoid.
Reported-by: ImpatientHippo on GitHub
Fixes#8908Closes#8937
