This abstracts across the two HTTP/2 backends: nghttp2 and Hyper.
Add our own define for the "h2" ALPN protocol, so TLS backends can use
it without depending on a specific HTTP backend.
Closes#6959
Previously if a caller set CURLOPT_WRITEFUNCTION but did not set a
CURLOPT_HEADERDATA buffer, Hyper would still attempt to write headers to
the data->set.writeheader header buffer, even though it is null. This
led to NPE segfaults attempting to use libcurl+Hyper with Git, for
example.
Instead, process the client write for the status line using the same
logic we use to process the client write for the later HTTP headers,
which contains the appropriate guard logic. As a side benefit,
data->set.writeheader is now only read in one file instead of two.
Fixes#6619Fixesabetterinternet/crustls#49Fixeshyperium/hyper#2438Closes#6971
- Disable auto credentials by default. This is a breaking change
for clients that are using it, wittingly or not.
- New libcurl ssl option value CURLSSLOPT_AUTO_CLIENT_CERT tells libcurl
to automatically locate and use a client certificate for
authentication, when requested by the server.
- New curl tool options --ssl-auto-client-cert and
--proxy-ssl-auto-client-cert map to CURLSSLOPT_AUTO_CLIENT_CERT.
This option is only supported for Schannel (the native Windows SSL
library). Prior to this change Schannel would, with no notification to
the client, attempt to locate a client certificate and send it to the
server, when requested by the server. Since the server can request any
certificate that supports client authentication in the OS certificate
store it could be a privacy violation and unexpected.
Fixes https://github.com/curl/curl/issues/2262
Reported-by: Jeroen Ooms
Assisted-by: Wes Hinsley
Assisted-by: Rich FitzJohn
Ref: https://curl.se/mail/lib-2021-02/0066.html
Reported-by: Morten Minde Neergaard
Closes https://github.com/curl/curl/pull/6673
Input challenges and returned messages are now in binary.
Conversions from/to base64 are performed by callers (currently curl_sasl.c
and http_ntlm.c).
Closes#6654
A struct bufref holds a buffer pointer, a data size and a destructor.
When freed or its contents are changed, the previous buffer is implicitly
released by the associated destructor. The data size, although not used
internally, allows binary data support.
A unit test checks its handling methods: test 1661
Closes#6654
Reset FD_WRITE by sending zero bytes which is permissible
and will be treated by implementations as successful send.
Without this we won't be notified in case a socket is still
writable if we already received such a notification and did
not send any data afterwards on the socket. This would lead
to waiting forever on a writable socket being writable again.
Assisted-by: Tommy Odom
Reviewed-by: Jay Satiro
Reviewed-by: Marcel Raad
Tested-by: tmkk on github
Bug: #6146Closes#6245
1. Consolidate pre-checks into a single Curl_poll call:
This is an attempt to restructure the code in Curl_multi_wait
in such a way that less syscalls are made by removing individual
calls to Curl_socket_check via SOCKET_READABLE/SOCKET_WRITABLE.
2. Avoid resetting the WinSock event multiple times:
We finally call WSAResetEvent anyway, so specifying it as
an optional parameter to WSAEnumNetworkEvents is redundant.
3. Wakeup directly in case no sockets are being monitoring:
Fix the WinSock based implementation to skip extra waiting by
not sleeping in case no sockets are to be waited on and just
the WinSock event is being monitored for wakeup functionality.
Assisted-by: Tommy Odom
Reviewed-by: Jay Satiro
Reviewed-by: Marcel Raad
Bug: #6146Closes#6245
This reverts commit 2260e0ebe6,
also restoring previous follow up changes which were reverted.
Authored-by: rcombs on github
Authored-by: Marc Hörsken
Reviewed-by: Jay Satiro
Reviewed-by: Marcel Raad
Restores #5634
Reverts #6281
Part of #6245
This reverts commit 1cba36d216.
CMake provides properties that can be set on a target to rename the
output artifact without changing the name of a target.
Ref: #6899
Add parser for CURLOPT_SSL_CIPHER_LIST option for Secure Transport (ST)
back-end. Similar to NSS and GSKit back-ends, new code parses string
value and configures ST library to use those ciphers for communication.
Create cipher spec data structure and initialize the array of specs with
cipher number, name, alias, and 'weak' flag.
Mark triple-DES ciphers as 'weak', and exclude them from the default
ciphers list.
Closes#6464
By using #ifdef on the symbol names to work on anything that don't
provide them. SCO OpenServer 5.0.7, sys/socket.h does not define either
SHUT_RDWR, SHUT_RD, and SHUT_WR.
Reported-by: Kevin R. Bulgrien
Bug: https://curl.se/mail/lib-2021-04/0073.htmlCloses#6925
The ConnectionExists() function will note that the new transfer wants
less then h2 and that it can't multiplex it and therefor opt to open a
new connection instead.
Storing a stream error in the per-connection struct was an error that lead to
race conditions as subsequent stream handling could overwrite the error code
before it was used for the stream with the actual problem.
Closes#6910
This was this one condition where the stream could be closed due to an
error and the function would still wrongly just return 0 for it.
Reported-by: Gergely Nagy
Fixes#6862Closes#6910
When the host name in a URL is given as an IPv4 numerical address, the
address can be specified with dotted numericals in four different ways:
a32, a.b24, a.b.c16 or a.b.c.d and each part can be specified in
decimal, octal (0-prefixed) or hexadecimal (0x-prefixed).
Instead of passing on the name as-is and leaving the handling to the
underlying name functions, which made them not work with c-ares but work
with getaddrinfo, this change now makes the curl URL API itself detect
and "normalize" host names specified as IPv4 numericals.
The WHATWG URL Spec says this is an okay way to specify a host name in a
URL. RFC 3896 does not allow them, but curl didn't prevent them before
and it seems other RFC 3896-using tools have not either. Host names used
like this are widely supported by other tools as well due to the
handling being done by getaddrinfo and friends.
I decided to add the functionality into the URL API itself so that all
users of these functions get the benefits, when for example wanting to
compare two URLs. Also, it makes curl built to use c-ares now support
them as well and make curl builds more consistent.
The normalization makes HTTPS and virtual hosted HTTP work fine even
when curl gets the address specified using one of the "obscure" formats.
Test 1560 is extended to verify.
Fixes#6863Closes#6871
... by fixing macros to do-while constructs and moving out the calls to
"break" outside of the actual macro. It also fixes the problem where the
macro was used witin a loop and the break didn't do right.
Reported-by: Emil Engler
Fixes#6847Closes#6909
... previously they were supported if a TLS library would (unexpectedly)
still support them, but from this change they will be refused already in
curl_easy_setopt(). SSLv2 and SSLv3 have been known to be insecure for
many years now.
Closes#6773
The code in cr_recv was returning prematurely as soon as the socket
reported no more data to read. However, this could be leaving some
unread plaintext data in the rustls session from a previous call,
causing causing the transfer to hang if the socket never receives
further data.
We need to ensure that the session is fully drained of plaintext data
before returning CURLE_AGAIN to the caller.
Reviewed-by: Jacob Hoffman-Andrews
Closes#6894
Add test 676 to verify that setting CURLOPT_COOKIEFILE to NULL again clears
the cookiejar from memory.
Reported-by: Stefan Karpinski
Fixes#6889Closes#6891
According to Microsoft document MS-NLMP, current flags usage is not
accurate: flag NTLMFLAG_NEGOTIATE_NTLM2_KEY controls the use of
extended security in an NTLM authentication message and NTLM version 2
cannot be negotiated within the protocol.
The solution implemented here is: if the extended security flag is set,
prefer using NTLM version 2 (as a server featuring extended security
should also support version 2). If version 2 has been disabled at
compile time, use extended security.
Tests involving NTLM are adjusted to this new behavior.
Fixes#6813Closes#6849
In 2682e5f5, several instances of AC_HEADER_TIME were removed since
it is a deprecated autoconf macro. However, this was the macro that
defined TIME_WITH_SYS_TIME, which was used to indicate that <time.h>
can be included alongside <sys/time.h>. TIME_WITH_SYS_TIME is still
used in the configure test body and since it is no longer defined,
<time.h> is *not* included on systems that have <sys/time.h>.
In particular, at least on musl libc and glibc, <sys/time.h> does
not implicitly include <time.h> and does not declare clock_gettime,
gmtime_r, or localtime_r. This causes configure to fail to detect
those functions.
The AC_HEADER_TIME macro deprecation text says
> All current systems provide time.h; it need not be checked for.
> Not all systems provide sys/time.h, but those that do, all allow
> you to include it and time.h simultaneously.
So, to fix this issue, simply include <time.h> unconditionally when
testing for time-related functions and in libcurl, and don't bother
checking for it.
Closes#6859
This was previously defined by the obsolete AC_TYPE_SIGNAL macro,
which was removed in 2682e5f5. The deprecation text says
> Your code may safely assume C89 semantics that RETSIGTYPE is void.
So, remove it and just use void instead.
Closes#6861
This removes the last occurrence of PATH_MAX inside our libssh
implementation by calculating the path length from the string length of
the two components.
Closes#6829
instead of 13, before the server has told how many streams it
accepts. The server can always reject new streams anyway if we go above
what it accepts.
Ref: #6826Closes#6852
After 957bc1881e686f9714c4e6a01bf33535091f0e21, we no longer compute an
expected_size for directories. This has the upshot that when we compare
even an empty Range with the available size, we fail.
This brings back the previous behaviour, which was to succeed, but with
empty content. This also removes the "Accept-ranges: bytes" header,
which is nonsensical on directories.
Adds test 3016
Fixes#6845Closes#6846
for GnuTLS, BearSSL, mbedTLS, NSS, SChannnel, Secure Transport and
wolfSSL...
Regression since 88dd1a8a11 (shipped in 7.76.0)
Reported-by: Kenneth Davidson
Reported-by: romamik om github
Fixes#6825Closes#6827
- Define Curl_resolver_error function only when USE_CURL_ASYNC.
Prior to this change building curl without an asynchronous resolver
backend (c-ares or threaded) and without DoH (DNS-over-HTTPS, which is
also asynchronous but independent of resolver backend) would cause a
build error since Curl_resolver_error is called by and evaluates
variables only available in asynchronous builds.
Reported-by: Benbuck Nason
Fixes https://github.com/curl/curl/issues/6831
Closes https://github.com/curl/curl/pull/6832
By making sure never to send off more than the allowed number of bytes
per second the speed limit logic is given more room to actually work.
Reported-by: Fabian Keil
Bug: https://curl.se/mail/lib-2021-03/0042.htmlCloses#6797
Both were used for the same purposes and there was no logical separation
between them. Combined, this also saves 16 bytes in less holes in my
test build.
Closes#6798
Otherwise libcurl is likely to reuse the connection again in the next
attempt since the connection reuse logic doesn't take downgrades into
account.
Reported-by: Anthony Ramine
Fixes#6788Closes#6793
Otherwise, the transfer will be NULL in the trace function when the
early handshake details arrive and then curl won't show them.
Regresssion in 7.75.0
Reported-by: David Hu
Fixes#6783Closes#6792
Instead of clearing the callback argument in disconnect, set it to the
(new) transfer to make sure the correct data is passed to the callbacks.
Follow-up to e467ea3bd9
Assisted-by: Patrick Monnerat
Closes#6787
After the recent conn/data refactor in this source file, this function
was mistakenly still getting the old struct pointer which would lead to
crash on servers with keyboard-interactive auth enabled.
Follow-up to a304051620 (shipped in 7.75.0)
Reported-by: Christian Schmitz
Fixes#6691Closes#6782
To make sure the Host: header and the URL provide the same authority
portion when sent to the proxy, strip the default port number from the
URL if one was provided.
Reported-by: Michael Brown
Fixes#6769Closes#6778
If libssh2_knownhost_init() returns NULL, like in an OOM situation, the
ssh session was freed but the pointer wasn't cleared which made libcurl
later call libssh2 to cleanup using the stale pointer.
Fixes#6764Closes#6766
If we get a close_notify, treat that as EOF. If we get an EOF from the
TCP stream, treat that as an error (because we should have ended the
connection earlier, when we got a close_notify).
Closes#6763
- Document in DOH that some SSL settings are inherited but DOH hostname
and peer verification are not and are controlled separately.
- Document that CURLOPT_SSL_CTX_FUNCTION is inherited by DOH handles but
we're considering changing behavior to no longer inherit it. Request
feedback.
Closes https://github.com/curl/curl/pull/6688
When asked to resume a download, libcurl will convert that to HTTP logic
and if then the entire file is already transferred it will result in a
416 response from the HTTP server. With CURLOPT_FAILONERRROR set in that
scenario, it should *not* lead to an error return.
Updated test 1156, added test 1273
Reported-by: Jonathan Watt
Fixes#6740Closes#6753
The duration of a connect and the total transfer are calculated from two
different time-stamps. It can end up with the total timeout triggering
before the connect timeout expires and we should make sure to
acknowledge whichever timeout that is reached first.
This is especially notable when a transfer first sits in PENDING, as
that time is counted in the total time but the connect timeout is based
on the time since the handle changed to the CONNECT state.
The CONNECTTIMEOUT is per connect attempt. The TIMEOUT is for the entire
operation.
Fixes#6744Closes#6745
Reported-by: Andrei Bica
Assisted-by: Jay Satiro
Previously, rustls was using an on-stack array for TLS data. However,
crustls has an (unusual) requirement that buffers it deals with are
initialized before writing to them. By using calloc, we can ensure the
buffer is initialized once and then reuse it across calls.
Closes#6742
Align conditions for NTLM features between CMake and configure
builds by differentiating between USE_NTLM and USE_CURL_NTLM_CORE,
just like curl_setup.h does internally to detect support of:
- USE_NTLM: required for NTLM crypto authentication feature
- USE_CURL_NTLM_CORE: required for SMB protocol
Implement USE_WIN32_CRYPTO detection by checking for Crypt functions
in wincrypt.h which are not available in the Windows App environment.
Link advapi32 and crypt32 for Crypto API and Schannel SSL backend.
Fix condition of Schannel SSL backend in CMake build accordingly.
Reviewed-by: Marcel Raad
Closes#6277
Move the detection of the restricted Windows App environment
in curl_setup.h before the definition of USE_WIN32_CRYPTO
via included config-win32.h in case no build system is used.
Reviewed-by: Marcel Raad
Part of #6277
MAX_HSTS_SUBLEN and MAX_HSTS_SUBLENSTR were unused from the initial commit,
and mostly likely leftovers from early development. Remove as they're not
used for anything.
Closes#6741
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
If after a transfer is complete Curl_GetFTPResponse() returns an error,
curl would not free the ftp->pathalloc block.
Found by torture-testing test 576
Closes#6737
This requires the latest main branch of crustls, which provides
rustls_client_config_builder_dangerous_set_certificate_verifier and
rustls_client_config_builder_set_enable_sni.
This refactors the session setup into its own function, and adds a new
function cr_hostname_is_ip. Because crustls doesn't support verification
of IP addresses, special handling is needed: We disable SNI and set a
placeholder hostname (which never actually gets sent on the wire).
Closes#6719
Curl_cookie_init can be called with data being NULL, and this can in turn
be passed to Curl_cookie_add, meaning that both functions must be careful
to only use data where it's checked for being a NULL pointer. The libpsl
support code does however dereference data without checking, so if we are
indeed having an unset data pointer we cannot PSL check the cookiedomain.
This is currently not a reachable dereference, as the only caller with a
NULL data isn't passing a file to initialize cookies from, but since the
API has this contract let's ensure we hold it.
Closes#6731
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
- Change use of those options from CURLOPT_SSL_OPTIONS that are not
already evaluated via SSL_SET_OPTION in schannel and secure transport
to use that instead of data->set.ssl.optname.
Example:
Evaluate SSL_SET_OPTION(no_revoke) instead of data->set.ssl.no_revoke.
This change is because options set via CURLOPT_SSL_OPTIONS
(data->set.ssl.optname) are separate from those set for HTTPS proxy via
CURLOPT_PROXY_SSL_OPTIONS (data->set.proxy_ssl.optname). The
SSL_SET_OPTION macro determines whether the connection is for HTTPS
proxy and based on that which option to evaluate.
Since neither Schannel nor Secure Transport backends currently support
HTTPS proxy in libcurl, this change is for posterity and has no other
effect.
Closes https://github.com/curl/curl/pull/6690
unescaped is coming from Curl_urldecode and not a unicode conversion
function, so reclaiming its memory should be performed with a normal
call to free rather than curlx_unicodefree. In reality, this is the
same thing as curlx_unicodefree is implemented as a call to free but
that's not guaranteed to always hold. Using the curlx macro present
issues with memory debugging as well.
Closes#6671
Reviewed-by: Jay Satiro <raysatiro@yahoo.com>
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
openssl: use SSL_get_version to get connection protocol
Replace our bespoke get_ssl_version_txt in favor of SSL_get_version.
We can get rid of few lines of code, since SSL_get_version achieve
the exact same thing
Closes#6665
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Signed-off-by: Jean-Philippe Menil <jpmenil@gmail.com>
Commit e06fa7462a removed support for libgcrypt leaving only
support for nettle which has been the default crypto library in
GnuTLS for a long time. There were however a few conditionals on
USE_GNUTLS_NETTLE which cause compilation errors in the metalink
code (as it used the gcrypt fallback instead as a result). See the
below autobuild for an example of the error:
https://curl.se/dev/log.cgi?id=20210225123226-30704#prob1
This removes all uses of USE_GNUTLS_NETTLE and also removes the
gcrypt support from the metalink code while at it.
Closes#6656
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
Align header with project style of using named parameters in the
function prototypes to aid readability and self-documentation.
Closes#6653
Reviewed-by: Daniel Stenberg <daniel@haxx.se>
... since the state machine might go to RATELIMITING and then back to
PERFORMING doing once-per-transfer inits in that function is wrong and
it caused problems with receiving chunked HTTP and it set the
PRETRANSFER time much too often...
Regression from b68dc34af3 (shipped in 7.75.0)
Reported-by: Amaury Denoyelle
Fixes#6640Closes#6641
(Unless 32-bit `time_t` is selected manually via the `_USE_32BIT_TIME_T`
mingw macro.)
Previously, 64-bit `time_t` was enabled on VS2005 and newer only, and
32-bit `time_t` was used on all other Windows builds.
Assisted-by: Jay Satiro
Closes#6636
- Use atexit to register a dbg cleanup function that closes the logfile.
LeakSantizier (LSAN) calls _exit() instead of exit() when a leak is
detected on exit so the logfile must be closed explicitly or data could
be lost. Though _exit() does not call atexit handlers such as this,
LSAN's call to _exit() comes after the atexit handlers are called.
Prior to this change the logfile was not explicitly closed so it was
possible that if LSAN detected a leak and called _exit (which does
not flush or close files like exit) then the logfile could be missing
data. That could then cause curl's memanalyze to report false leaks
(eg a malloc was recorded to the logfile but the corresponding free was
discarded from the buffer instead of written to the logfile, then
memanalyze reports that as a leak).
Ref: https://github.com/google/sanitizers/issues/1374
Bug: https://github.com/curl/curl/pull/6591#issuecomment-780396541
Closes https://github.com/curl/curl/pull/6620
- Change the Windows char <-> UTF-8 conversion functions to return an
allocated copy of the passed in string instead of the original.
Prior to this change the curlx_convert_ functions would, as what I
assume was an optimization, not make a copy of the passed in string if
no conversion was required. No conversion is required in non-UNICODE
Windows builds since our tchar strings are type char and remain in
whatever the passed in encoding is, which is assumed to be UTF-8 but may
be other encoding.
In contrast the UNICODE Windows builds require conversion
(wchar <-> char) and do return a copy. That inconsistency could lead to
programming errors where the developer expects a copy, and does not
realize that won't happen in all cases.
Closes https://github.com/curl/curl/pull/6602
- add CURLINFO_REFERER libcurl option
- add --write-out '%{referer}' command-line option
- extend --xattr command-line option to fill user.xdg.referrer.url extended
attribute with the referrer (if there was any)
Closes#6591
... with the help of Curl_resolver_error() which now is moved from
asyn-thead.c and is provided globally for this purpose.
Follow-up to 35ca04ce1b
Makes test 1188 work for c-ares builds
Closes#6626
This caused a memory leak as the session id cache entry was still
erroneously stored with a NULL sessionid and that would later be treated
as not needed to get freed.
Reported-by: Gisle Vanem
Fixes#6616Closes#6617
While working on documenting the states it dawned on me that step one is
to use more descriptive names on the states. This also changes prefix on
the states to make them shorter in the source.
State names NOT ending with *ing are transitional ones.
Closes#6612
The Curl_easy pointer struct entry in connectdata is now gone. Just
before commit 215db086e0 landed on January 8, 2021 there were 919
references to conn->data.
Closes#6608
- Share the shared object from the user's easy handle with the DOH
handles.
Prior to this change if the user had set a shared object with shared
cached DNS (CURL_LOCK_DATA_DNS) for their easy handle then that wasn't
used by any associated DOH handles, since they used the multi's default
hostcache.
This change means all the handles now use the same hostcache, which is
either the shared hostcache from the user created shared object if it
exists or if not then the multi's default hostcache.
Reported-by: Manuj Bhatia
Fixes https://github.com/curl/curl/issues/6589
Closes https://github.com/curl/curl/pull/6607
... but instead use a private alternative that points to the "driving
transfer" from the connection. We set the "user data" associated with
the connection to be the connectdata struct, but when we drive transfers
the code still needs to know the pointer to the transfer. We can change
the user data to become the Curl_easy handle, but with older nghttp2
version we cannot dynamically update that pointer properly when
different transfers are used over the same connection.
Closes#6520
We still make the trace callback function get the connectdata struct
passed to it, since the callback is anchored on the connection.
Repeatedly updating the callback pointer to set 'data' with
SSL_CTX_set_msg_callback_arg() doesn't seem to work, probably because
there might already be messages in the queue with the old pointer.
This code therefore makes sure to set the "logger" handle before using
OpenSSL calls so that the right easy handle gets used for tracing.
Closes#6522
- New libcurl options CURLOPT_DOH_SSL_VERIFYHOST,
CURLOPT_DOH_SSL_VERIFYPEER and CURLOPT_DOH_SSL_VERIFYSTATUS do the
same as their respective counterparts.
- New curl tool options --doh-insecure and --doh-cert-status do the same
as their respective counterparts.
Prior to this change DOH SSL certificate verification settings for
verifyhost and verifypeer were supposed to be inherited respectively
from CURLOPT_SSL_VERIFYHOST and CURLOPT_SSL_VERIFYPEER, but due to a bug
were not. As a result DOH verification remained at the default, ie
enabled, and it was not possible to disable. This commit changes
behavior so that the DOH verification settings are independent and not
inherited.
Ref: https://github.com/curl/curl/pull/4579#issuecomment-554723676
Fixes https://github.com/curl/curl/issues/4578
Closes https://github.com/curl/curl/pull/6597
- Guard some Curl_async accesses with USE_CURL_ASYNC instead of
!CURLRES_SYNCH.
This is another follow-up to 8335c64 which moved the async struct from
the connectdata struct into the Curl_easy struct. A previous follow-up
6cd167a fixed building for sync resolver by guarding some async struct
accesses with !CURLRES_SYNCH. The problem is since DOH (DNS-over-HTTPS)
is available as an asynchronous secondary resolver the async struct may
be used even when libcurl is built for the sync resolver. That means
that CURLRES_SYNCH and USE_CURL_ASYNC may be defined at the same time.
Closes https://github.com/curl/curl/pull/6603
HTTP auth "accidentally" worked before this cleanup since the code would
always overwrite the connection credentials with the credentials from
the most recent transfer and since HTTP auth is typically done first
thing, this has not been an issue. It was still wrong and subject to
possible race conditions or future breakage if the sequence of functions
would change.
The data.set.str[] strings MUST remain unmodified exactly as set by the
user, and the credentials to use internally are instead set/updated in
state.aptr.*
Added test 675 to verify different credentials used in two requests done
over a reused HTTP connection, which previously behaved wrongly.
Fixes#6542Closes#6545
Rename it to 'httpwant' and make a cloned field in the state struct as
well for run-time updates.
Also: refuse non-supported HTTP versions. Verified with test 129.
Closes#6585
