path: root/docs/KNOWN_BUGS

                                  _   _ ____  _
                              ___| | | |  _ \| |
                             / __| | | | |_) | |
                            | (__| |_| |  _ <| |___
                             \___|\___/|_| \_\_____|

                                  Known Bugs

These are problems and bugs known to exist at the time of this release. Feel
free to join in and help us correct one or more of these. Also be sure to
check the changelog of the current development status, as one or more of these
problems may have been fixed or changed somewhat since this was written.

 1. HTTP
 1.2 Multiple methods in a single WWW-Authenticate: header
 1.3 STARTTRANSFER time is wrong for HTTP POSTs
 1.4 multipart formposts file name encoding
 1.5 Expect-100 meets 417
 1.6 Unnecessary close when 401 received waiting for 100
 1.7 Deflate error after all content was received
 1.8 DoH is not used for all name resolves when enabled
 1.11 CURLOPT_SEEKFUNCTION not called with CURLFORM_STREAM

 2. TLS
 2.1 CURLINFO_SSL_VERIFYRESULT has limited support
 2.2 DER in keychain
 2.3 Unable to use PKCS12 certificate with Secure Transport
 2.4 Secure Transport will not import PKCS#12 client certificates without a password
 2.5 Client cert handling with Issuer DN differs between backends
 2.6 CURL_GLOBAL_SSL
 2.7 Client cert (MTLS) issues with Schannel
 2.8 Schannel disable CURLOPT_SSL_VERIFYPEER and verify hostname
 2.9 TLS session cache does not work with TFO
 2.10 Store TLS context per transfer instead of per connection
 2.11 Schannel TLS 1.2 handshake bug in old Windows versions
 2.12 FTPS with Schannel times out file list operation
 2.14 Secure Transport disabling hostname validation also disables SNI
 2.15 Renegotiate from server may cause hang for OpenSSL backend

 3. Email protocols
 3.1 IMAP SEARCH ALL truncated response
 3.2 No disconnect command
 3.3 POP3 expects "CRLF.CRLF" eob for some single-line responses
 3.4 AUTH PLAIN for SMTP is not working on all servers

 4. Command line
 4.1 -J and -O with %-encoded file names
 4.2 -J with -C - fails
 4.3 --retry and transfer timeouts

 5. Build and portability issues
 5.1 OS400 port requires deprecated IBM library
 5.2 curl-config --libs contains private details
 5.3 curl compiled on OSX 10.13 failed to run on OSX 10.10
 5.4 Build with statically built dependency
 5.5 cannot handle Unicode arguments in non-Unicode builds on Windows
 5.6 make distclean loops forever
 5.7 Visual Studio project gaps
 5.8 configure finding libs in wrong directory
 5.9 Utilize Requires.private directives in libcurl.pc
 5.10 curl hangs on SMB upload over stdin
 5.11 configure --with-gssapi with Heimdal is ignored on macOS
 5.12 flaky Windows CI builds

 6. Authentication
 6.1 NTLM authentication and unicode
 6.2 MIT Kerberos for Windows build
 6.3 NTLM in system context uses wrong name
 6.4 Negotiate and Kerberos V5 need a fake user name
 6.5 NTLM does not support password with § character
 6.6 libcurl can fail to try alternatives with --proxy-any
 6.7 Do not clear digest for single realm
 6.8 RTSP authentication breaks without redirect support
 6.9 SHA-256 digest not supported in Windows SSPI builds
 6.10 curl never completes Negotiate over HTTP
 6.11 Negotiate on Windows fails
 6.12 cannot use Secure Transport with Crypto Token Kit

 7. FTP
 7.1 FTP without or slow 220 response
 7.2 FTP with CONNECT and slow server
 7.3 FTP with NOBODY and FAILONERROR
 7.4 FTP with ACCT
 7.5 ASCII FTP
 7.6 FTP with NULs in URL parts
 7.7 FTP and empty path parts in the URL
 7.8 Premature transfer end but healthy control channel
 7.9 Passive transfer tries only one IP address
 7.10 FTPS needs session reuse
 7.11 FTPS upload data loss with TLS 1.3

 8. TELNET
 8.1 TELNET and time limitations do not work
 8.2 Microsoft telnet server

 9. SFTP and SCP
 9.1 SFTP does not do CURLOPT_POSTQUOTE correct
 9.2 wolfssh: publickey auth does not work
 9.3 Remote recursive folder creation with SFTP

 10. SOCKS
 10.3 FTPS over SOCKS
 10.4 active FTP over a SOCKS

 11. Internals
 11.1 Curl leaks .onion hostnames in DNS
 11.2 error buffer not set if connection to multiple addresses fails
 11.3 Disconnects do not do verbose
 11.4 HTTP test server 'connection-monitor' problems
 11.5 Connection information when using TCP Fast Open
 11.7 signal-based resolver timeouts
 11.8 DoH leaks memory after followlocation
 11.9 DoH does not inherit all transfer options
 11.10 Blocking socket operations in non-blocking API
 11.11 A shared connection cache is not thread-safe
 11.12 'no_proxy' string-matches IPv6 numerical addresses
 11.13 wakeup socket disconnect causes havoc
 11.14 Multi perform hangs waiting for threaded resolver
 11.15 CURLOPT_OPENSOCKETPAIRFUNCTION is missing
 11.16 libcurl uses renames instead of locking for atomic operations

 12. LDAP
 12.1 OpenLDAP hangs after returning results
 12.2 LDAP on Windows does authentication wrong?
 12.3 LDAP on Windows does not work
 12.4 LDAPS with NSS is slow

 13. TCP/IP
 13.1 --interface for ipv6 binds to unusable IP address
 13.2 Trying local ports fails on Windows

 14. DICT
 14.1 DICT responses show the underlying protocol

 15. CMake
 15.1 use correct SONAME
 15.2 support build with GnuTLS
 15.3 unusable tool_hugehelp.c with MinGW
 15.4 build docs/curl.1
 15.5 build on Linux links libcurl to libdl
 15.6 uses -lpthread instead of Threads::Threads
 15.7 generated .pc file contains strange entries
 15.8 libcurl.pc uses absolute library paths
 15.9 cert paths autodetected when cross-compiling
 15.10 libpsl is not supported
 15.11 ExternalProject_Add does not set CURL_CA_PATH
 15.12 cannot enable LDAPS on Windows
 15.13 CMake build with MIT Kerberos does not work

 16. Applications

 17. HTTP/2
 17.1 Excessive HTTP/2 packets with TCP_NODELAY
 17.2 HTTP/2 frames while in the connection pool kill reuse
 17.3 ENHANCE_YOUR_CALM causes infinite retries
 17.4 Connection failures with parallel HTTP/2
 17.5 HTTP/2 connections through HTTPS proxy frequently stall

 18. HTTP/3
 18.1 If the HTTP/3 server closes connection during upload curl hangs
 18.2 Transfer closed with n bytes remaining to read
 18.4 timeout when reusing a http3 connection
 18.9 connection migration does not work

==============================================================================

1. HTTP

1.2 Multiple methods in a single WWW-Authenticate: header

 The HTTP responses headers WWW-Authenticate: can provide information about
 multiple authentication methods as multiple headers or as several methods
 within a single header. The latter way, several methods in the same physical
 line, is not supported by libcurl's parser. (For no good reason.)

1.3 STARTTRANSFER time is wrong for HTTP POSTs

 Wrong STARTTRANSFER timer accounting for POST requests Timer works fine with
 GET requests, but while using POST the time for CURLINFO_STARTTRANSFER_TIME
 is wrong. While using POST CURLINFO_STARTTRANSFER_TIME minus
 CURLINFO_PRETRANSFER_TIME is near to zero every time.

 https://github.com/curl/curl/issues/218
 https://curl.se/bug/view.cgi?id=1213

1.4 multipart formposts file name encoding

 When creating multipart formposts. The file name part can be encoded with
 something beyond ascii but currently libcurl will only pass in the verbatim
 string the app provides. There are several browsers that already do this
 encoding. The key seems to be the updated draft to RFC2231:
 https://datatracker.ietf.org/doc/html/draft-reschke-rfc2231-in-http-02

1.5 Expect-100 meets 417

 If an upload using Expect: 100-continue receives an HTTP 417 response, it
 ought to be automatically resent without the Expect:. A workaround is for
 the client application to redo the transfer after disabling Expect:.
 https://curl.se/mail/archive-2008-02/0043.html

1.6 Unnecessary close when 401 received waiting for 100

 libcurl closes the connection if an HTTP 401 reply is received while it is
 waiting for the 100-continue response.
 https://curl.se/mail/lib-2008-08/0462.html

1.7 Deflate error after all content was received

 There's a situation where we can get an error in an HTTP response that is
 compressed, when that error is detected after all the actual body contents
 have been received and delivered to the application. This is tricky, but is
 ultimately a broken server.

 See https://github.com/curl/curl/issues/2719

1.8 DoH is not used for all name resolves when enabled

 Even if DoH is specified to be used, there are some name resolves that are
 done without it. This should be fixed. When the internal function
 `Curl_resolver_wait_resolv()` is called, it does not use DoH to complete the
 resolve as it otherwise should.

 See https://github.com/curl/curl/pull/3857 and
 https://github.com/curl/curl/pull/3850

1.11 CURLOPT_SEEKFUNCTION not called with CURLFORM_STREAM

 When using libcurl to POST form data using a FILE* with the CURLFORM_STREAM
 option of curl_formadd(). I notice that if the connection drops at just the
 right time, the POST is reattempted without the data from the file. It seems
 like the file stream position is not getting reset to the beginning of the
 file. I found the CURLOPT_SEEKFUNCTION option and set that with a function
 that performs an fseek() on the FILE*. However, setting that did not seem to
 fix the issue or even get called. See https://github.com/curl/curl/issues/768


2. TLS

2.1 CURLINFO_SSL_VERIFYRESULT has limited support

 CURLINFO_SSL_VERIFYRESULT is only implemented for the OpenSSL, NSS and
 GnuTLS backends, so relying on this information in a generic app is flaky.

2.2 DER in keychain

 Curl does not recognize certificates in DER format in keychain, but it works
 with PEM.  https://curl.se/bug/view.cgi?id=1065

2.3 Unable to use PKCS12 certificate with Secure Transport

 See https://github.com/curl/curl/issues/5403

2.4 Secure Transport will not import PKCS#12 client certificates without a password

 libcurl calls SecPKCS12Import with the PKCS#12 client certificate, but that
 function rejects certificates that do not have a password.
 https://github.com/curl/curl/issues/1308

2.5 Client cert handling with Issuer DN differs between backends

 When the specified client certificate does not match any of the
 server-specified DNs, the OpenSSL and GnuTLS backends behave differently.
 The github discussion may contain a solution.

 See https://github.com/curl/curl/issues/1411

2.6 CURL_GLOBAL_SSL

 Since libcurl 7.57.0, the flag CURL_GLOBAL_SSL is a no-op. The change was
 merged in https://github.com/curl/curl/commit/d661b0afb571a

 It was removed since it was

 A) never clear for applications on how to deal with init in the light of
    different SSL backends (the option was added back in the days when life
    was simpler)

 B) multissl introduced dynamic switching between SSL backends which
    emphasized (A) even more

 C) libcurl uses some TLS backend functionality even for non-TLS functions (to
    get "good" random) so applications trying to avoid the init for
    performance reasons would do wrong anyway

 D) not documented carefully so all this mostly just happened to work
    for some users

 However, in spite of the problems with the feature, there were some users who
 apparently depended on this feature and who now claim libcurl is broken for
 them. The fix for this situation is not obvious as a downright revert of the
 patch is totally ruled out due to those reasons above.

 https://github.com/curl/curl/issues/2276

2.7 Client cert (MTLS) issues with Schannel

 See https://github.com/curl/curl/issues/3145

2.8 Schannel disable CURLOPT_SSL_VERIFYPEER and verify hostname

 This seems to be a limitation in the underlying Schannel API.

 https://github.com/curl/curl/issues/3284

2.9 TLS session cache does not work with TFO

 See https://github.com/curl/curl/issues/4301

2.10 Store TLS context per transfer instead of per connection

 The GnuTLS `backend->cred` and the OpenSSL `backend->ctx` data and their
 proxy versions (and possibly other TLS backends), could be better moved to be
 stored in the Curl_easy handle instead of in per connection so that a single
 transfer that makes multiple connections can reuse the context and reduce
 memory consumption.

 https://github.com/curl/curl/issues/5102

2.11 Schannel TLS 1.2 handshake bug in old Windows versions

 In old versions of Windows such as 7 and 8.1 the Schannel TLS 1.2 handshake
 implementation likely has a bug that can rarely cause the key exchange to
 fail, resulting in error SEC_E_BUFFER_TOO_SMALL or SEC_E_MESSAGE_ALTERED.

 https://github.com/curl/curl/issues/5488

2.12 FTPS with Schannel times out file list operation

 "Instead of the command completing, it just sits there until the timeout
 expires." - the same command line seems to work with other TLS backends and
 other operating systems. See https://github.com/curl/curl/issues/5284.

2.14 Secure Transport disabling hostname validation also disables SNI

 SNI is the hostname that is sent by the TLS library to the server as part of
 the TLS handshake. Secure Transport does not send SNI when hostname validation
 is disabled. Servers that host multiple websites may not know which
 certificate to serve without SNI or which backend server to connect to. The
 server may serve the certificate of a default server or abort.

 If a server aborts a handshake then curl shows error "SSL peer handshake
 failed, the server most likely requires a client certificate to connect".
 In this case the error may also have been caused by lack of SNI.

 https://github.com/curl/curl/issues/6347

2.15 Renegotiate from server may cause hang for OpenSSL backend

 A race condition has been observed when, immediately after the initial
 handshake, curl has sent an HTTP request to the server and at the same time
 the server has sent a TLS hello request (renegotiate) to curl. Both are
 waiting for the other to respond. OpenSSL is supposed to send a handshake
 response but does not.

 https://github.com/curl/curl/issues/6785
 https://github.com/openssl/openssl/issues/14722

3. Email protocols

3.1 IMAP SEARCH ALL truncated response

 IMAP "SEARCH ALL" truncates output on large boxes. "A quick search of the
 code reveals that pingpong.c contains some truncation code, at line 408, when
 it deems the server response to be too large truncating it to 40 characters"
 https://curl.se/bug/view.cgi?id=1366

3.2 No disconnect command

 The disconnect commands (LOGOUT and QUIT) may not be sent by IMAP, POP3 and
 SMTP if a failure occurs during the authentication phase of a connection.

3.3 POP3 expects "CRLF.CRLF" eob for some single-line responses

 You have to tell libcurl not to expect a body, when dealing with one line
 response commands. Please see the POP3 examples and test cases which show
 this for the NOOP and DELE commands. https://curl.se/bug/?i=740

3.4 AUTH PLAIN for SMTP is not working on all servers

 Specifying "--login-options AUTH=PLAIN" on the command line does not seem to
 work correctly.

 See https://github.com/curl/curl/issues/4080

4. Command line

4.1 -J and -O with %-encoded file names

 -J/--remote-header-name does not decode %-encoded file names. RFC6266 details
 how it should be done. The can of worm is basically that we have no charset
 handling in curl and ascii >=128 is a challenge for us. Not to mention that
 decoding also means that we need to check for nastiness that is attempted,
 like "../" sequences and the like. Probably everything to the left of any
 embedded slashes should be cut off.
 https://curl.se/bug/view.cgi?id=1294

 -O also does not decode %-encoded names, and while it has even less
 information about the charset involved the process is similar to the -J case.

 Note that we will not add decoding to -O without the user asking for it with
 some other means as well, since -O has always been documented to use the name
 exactly as specified in the URL.

4.2 -J with -C - fails

 When using -J (with -O), automatically resumed downloading together with "-C
 -" fails. Without -J the same command line works. This happens because the
 resume logic is worked out before the target file name (and thus its
 pre-transfer size) has been figured out.
 https://curl.se/bug/view.cgi?id=1169

4.3 --retry and transfer timeouts

 If using --retry and the transfer timeouts (possibly due to using -m or
 -y/-Y) the next attempt does not resume the transfer properly from what was
 downloaded in the previous attempt but will truncate and restart at the
 original position where it was at before the previous failed attempt. See
 https://curl.se/mail/lib-2008-01/0080.html and Mandriva bug report
 https://qa.mandriva.com/show_bug.cgi?id=22565

5. Build and portability issues

5.1 OS400 port requires deprecated IBM library

 curl for OS400 requires QADRT to build, which provides ASCII wrappers for
 libc/POSIX functions in the ILE, but IBM no longer supports or even offers
 this library to download.

 See https://github.com/curl/curl/issues/5176

5.2 curl-config --libs contains private details

 "curl-config --libs" will include details set in LDFLAGS when configure is
 run that might be needed only for building libcurl. Further, curl-config
 --cflags suffers from the same effects with CFLAGS/CPPFLAGS.

5.3 curl compiled on OSX 10.13 failed to run on OSX 10.10

 See https://github.com/curl/curl/issues/2905

5.4 Build with statically built dependency

 The build scripts in curl (autotools, cmake and others) are primarily done to
 work with shared/dynamic third party dependencies. When linking with shared
 libraries, the dependency "chain" is handled automatically by the library
 loader - on all modern systems.

 If you instead link with a static library, we need to provide all the
 dependency libraries already at the link command line.

 Figuring out all the dependency libraries for a given library is hard, as it
 might also involve figuring out the dependencies of the dependencies and they
 may vary between platforms and even change between versions.

 When using static dependencies, the build scripts will mostly assume that
 you, the user, will provide all the necessary additional dependency libraries
 as additional arguments in the build. With configure, by setting LIBS/LDFLAGS
 on the command line.

 We welcome help to improve curl's ability to link with static libraries, but
 it is likely a task that we can never fully support.

5.5 cannot handle Unicode arguments in non-Unicode builds on Windows

 If a URL or filename cannot be encoded using the user's current codepage then
 it can only be encoded properly in the Unicode character set. Windows uses
 UTF-16 encoding for Unicode and stores it in wide characters, however curl
 and libcurl are not equipped for that at the moment except when built with
 _UNICODE and UNICODE defined. And, except for Cygwin, Windows cannot use UTF-8
 as a locale.

  https://curl.se/bug/?i=345
  https://curl.se/bug/?i=731
  https://curl.se/bug/?i=3747

5.6 make distclean loops forever

 Due to an issue (probably) in automake, "make distclean" can end up in a
 never-ending loop.

 See https://github.com/curl/curl/issues/7716

5.7 Visual Studio project gaps

 The Visual Studio projects lack some features that the autoconf and nmake
 builds offer, such as the following:

  - support for zlib and nghttp2
  - use of static runtime libraries
  - add the test suite components

 In addition to this the following could be implemented:

  - support for other development IDEs
  - add PATH environment variables for third-party DLLs

5.8 configure finding libs in wrong directory

 When the configure script checks for third-party libraries, it adds those
 directories to the LDFLAGS variable and then tries linking to see if it
 works. When successful, the found directory is kept in the LDFLAGS variable
 when the script continues to execute and do more tests and possibly check for
 more libraries.

 This can make subsequent checks for libraries wrongly detect another
 installation in a directory that was previously added to LDFLAGS by another
 library check.

 A possibly better way to do these checks would be to keep the pristine LDFLAGS
 even after successful checks and instead add those verified paths to a
 separate variable that only after all library checks have been performed gets
 appended to LDFLAGS.

5.9 Utilize Requires.private directives in libcurl.pc

 https://github.com/curl/curl/issues/864

5.10 curl hangs on SMB upload over stdin

 See https://github.com/curl/curl/issues/7896

5.11 configure --with-gssapi with Heimdal is ignored on macOS

 ... unless you also pass --with-gssapi-libs

 https://github.com/curl/curl/issues/3841

5.12 flaky Windows CI builds

 We run many CI builds for each commit and PR on github, and especially a
 number of the Windows builds are flaky. This means that we rarely get all CI
 builds go green and complete without errors. This is unfortunate as it makes
 us sometimes miss actual build problems and it is surprising to newcomers to
 the project who (rightfully) do not expect this.

 See https://github.com/curl/curl/issues/6972

6. Authentication

6.1 NTLM authentication and unicode

 NTLM authentication involving unicode user name or password only works
 properly if built with UNICODE defined together with the Schannel
 backend. The original problem was mentioned in:
 https://curl.se/mail/lib-2009-10/0024.html
 https://curl.se/bug/view.cgi?id=896

 The Schannel version verified to work as mentioned in
 https://curl.se/mail/lib-2012-07/0073.html

6.2 MIT Kerberos for Windows build

 libcurl fails to build with MIT Kerberos for Windows (KfW) due to KfW's
 library header files exporting symbols/macros that should be kept private to
 the KfW library. See ticket #5601 at https://krbdev.mit.edu/rt/

6.3 NTLM in system context uses wrong name

 NTLM authentication using SSPI (on Windows) when (lib)curl is running in
 "system context" will make it use wrong(?) user name - at least when compared
 to what winhttp does. See https://curl.se/bug/view.cgi?id=535

6.4 Negotiate and Kerberos V5 need a fake user name

 In order to get Negotiate (SPNEGO) authentication to work in HTTP or Kerberos
 V5 in the email protocols, you need to  provide a (fake) user name (this
 concerns both curl and the lib) because the code wrongly only considers
 authentication if there's a user name provided by setting
 conn->bits.user_passwd in url.c  https://curl.se/bug/view.cgi?id=440 How?
 https://curl.se/mail/lib-2004-08/0182.html A possible solution is to
 either modify this variable to be set or introduce a variable such as
 new conn->bits.want_authentication which is set when any of the authentication
 options are set.

6.5 NTLM does not support password with § character

 https://github.com/curl/curl/issues/2120

6.6 libcurl can fail to try alternatives with --proxy-any

 When connecting via a proxy using --proxy-any, a failure to establish an
 authentication will cause libcurl to abort trying other options if the
 failed method has a higher preference than the alternatives. As an example,
 --proxy-any against a proxy which advertise Negotiate and NTLM, but which
 fails to set up Kerberos authentication will not proceed to try authentication
 using NTLM.

 https://github.com/curl/curl/issues/876

6.7 Do not clear digest for single realm

 https://github.com/curl/curl/issues/3267

6.8 RTSP authentication breaks without redirect support

 RTSP authentication broke in 7.66.0. A work-around is to enable RTSP in
 CURLOPT_REDIR_PROTOCOLS. Authentication should however not be considered an
 actual redirect so a "proper" fix needs to be different and not require users
 to allow redirects to RTSP to work.

 See https://github.com/curl/curl/pull/4750

6.9 SHA-256 digest not supported in Windows SSPI builds

 Windows builds of curl that have SSPI enabled use the native Windows API calls
 to create authentication strings. The call to InitializeSecurityContext fails
 with SEC_E_QOP_NOT_SUPPORTED which causes curl to fail with CURLE_AUTH_ERROR.

 Microsoft does not document supported digest algorithms and that SEC_E error
 code is not a documented error for InitializeSecurityContext (digest).

 https://github.com/curl/curl/issues/6302

6.10 curl never completes Negotiate over HTTP

 Apparently it is not working correctly...?

 See https://github.com/curl/curl/issues/5235

6.11 Negotiate on Windows fails

 When using --negotiate (or NTLM) with curl on Windows, SSL/TLS handshake
 fails despite having a valid kerberos ticket cached. Works without any issue
 in Unix/Linux.

 https://github.com/curl/curl/issues/5881

6.12 cannot use Secure Transport with Crypto Token Kit

 https://github.com/curl/curl/issues/7048

7. FTP

7.1 FTP without or slow 220 response

 If a connection is made to an FTP server but the server then just never sends
 the 220 response or otherwise is dead slow, libcurl will not acknowledge the
 connection timeout during that phase but only the "real" timeout - which may
 surprise users as it is probably considered to be the connect phase to most
 people. Brought up (and is being misunderstood) in:
 https://curl.se/bug/view.cgi?id=856

7.2 FTP with CONNECT and slow server

 When doing FTP over a socks proxy or CONNECT through HTTP proxy and the multi
 interface is used, libcurl will fail if the (passive) TCP connection for the
 data transfer is not more or less instant as the code does not properly wait
 for the connect to be confirmed. See test case 564 for a first shot at a test
 case.

7.3 FTP with NOBODY and FAILONERROR

 It seems sensible to be able to use CURLOPT_NOBODY and CURLOPT_FAILONERROR
 with FTP to detect if a file exists or not, but it is not working:
 https://curl.se/mail/lib-2008-07/0295.html

7.4 FTP with ACCT

 When doing an operation over FTP that requires the ACCT command (but not when
 logging in), the operation will fail since libcurl does not detect this and
 thus fails to issue the correct command:
 https://curl.se/bug/view.cgi?id=635

7.5 ASCII FTP

 FTP ASCII transfers do not follow RFC959. They do not convert the data
 accordingly (not for sending nor for receiving). RFC 959 section 3.1.1.1
 clearly describes how this should be done:

    The sender converts the data from an internal character representation to
    the standard 8-bit NVT-ASCII representation (see the Telnet
    specification). The receiver will convert the data from the standard
    form to his own internal form.

 Since 7.15.4 at least line endings are converted.

7.6 FTP with NULs in URL parts

 FTP URLs passed to curl may contain NUL (0x00) in the RFC 1738 <user>,
 <password>, and <fpath> components, encoded as "%00". The problem is that
 curl_unescape does not detect this, but instead returns a shortened C string.
 From a strict FTP protocol standpoint, NUL is a valid character within RFC
 959 <string>, so the way to handle this correctly in curl would be to use a
 data structure other than a plain C string, one that can handle embedded NUL
 characters. From a practical standpoint, most FTP servers would not
 meaningfully support NUL characters within RFC 959 <string>, anyway (e.g.,
 Unix pathnames may not contain NUL).

7.7 FTP and empty path parts in the URL

 libcurl ignores empty path parts in FTP URLs, whereas RFC1738 states that
 such parts should be sent to the server as 'CWD ' (without an argument). The
 only exception to this rule, is that we knowingly break this if the empty
 part is first in the path, as then we use the double slashes to indicate that
 the user wants to reach the root dir (this exception SHALL remain even when
 this bug is fixed).

7.8 Premature transfer end but healthy control channel

 When 'multi_done' is called before the transfer has been completed the normal
 way, it is considered a "premature" transfer end. In this situation, libcurl
 closes the connection assuming it does not know the state of the connection so
 it cannot be reused for subsequent requests.

 With FTP however, this is not necessarily true but there are a bunch of
 situations (listed in the ftp_done code) where it *could* keep the connection
 alive even in this situation - but the current code does not. Fixing this would
 allow libcurl to reuse FTP connections better.

7.9 Passive transfer tries only one IP address

 When doing FTP operations through a proxy at localhost, the reported spotted
 that curl only tried to connect once to the proxy, while it had multiple
 addresses and a failed connect on one address should make it try the next.

 After switching to passive mode (EPSV), curl should try all IP addresses for
 "localhost". Currently it tries ::1, but it should also try 127.0.0.1.

 See https://github.com/curl/curl/issues/1508

7.10 FTPS needs session reuse

 When the control connection is reused for a subsequent transfer, some FTPS
 servers complain about "missing session reuse" for the data channel for the
 second transfer.

 https://github.com/curl/curl/issues/4654

7.11 FTPS upload data loss with TLS 1.3

 During FTPS upload curl does not attempt to read TLS handshake messages sent
 after the initial handshake. OpenSSL servers running TLS 1.3 may send such a
 message. When curl closes the upload connection if unread data has been
 received (such as a TLS handshake message) then the TCP protocol sends an
 RST to the server, which may cause the server to discard or truncate the
 upload if it has not read all sent data yet, and then return an error to curl
 on the control channel connection.

 Since 7.78.0 this is mostly fixed. curl will do a single read before closing
 TLS connections (which causes the TLS library to read handshake messages),
 however there is still possibility of an RST if more messages need to be read
 or a message arrives after the read but before close (network race condition).

 https://github.com/curl/curl/issues/6149

8. TELNET

8.1 TELNET and time limitations do not work

 When using telnet, the time limitation options do not work.
 https://curl.se/bug/view.cgi?id=846

8.2 Microsoft telnet server

 There seems to be a problem when connecting to the Microsoft telnet server.
 https://curl.se/bug/view.cgi?id=649


9. SFTP and SCP

9.1 SFTP does not do CURLOPT_POSTQUOTE correct

 When libcurl sends CURLOPT_POSTQUOTE commands when connected to a SFTP server
 using the multi interface, the commands are not being sent correctly and
 instead the connection is "cancelled" (the operation is considered done)
 prematurely. There is a half-baked (busy-looping) patch provided in the bug
 report but it cannot be accepted as-is. See
 https://curl.se/bug/view.cgi?id=748

9.2 wolfssh: publickey auth does not work

 When building curl to use the wolfSSH backend for SFTP, the publickey
 authentication does not work. This is simply functionality not written for curl
 yet, the necessary API for make this work is provided by wolfSSH.

 See https://github.com/curl/curl/issues/4820

9.3 Remote recursive folder creation with SFTP

 On this servers, the curl fails to create directories on the remote server
 even when the CURLOPT_FTP_CREATE_MISSING_DIRS option is set.

 See https://github.com/curl/curl/issues/5204


10. SOCKS

10.3 FTPS over SOCKS

 libcurl does not support FTPS over a SOCKS proxy.

10.4 active FTP over a SOCKS

 libcurl does not support active FTP over a SOCKS proxy


11. Internals

11.1 Curl leaks .onion hostnames in DNS

 Curl sends DNS requests for hostnames with a .onion TLD. This leaks
 information about what the user is attempting to access, and violates this
 requirement of RFC7686: https://datatracker.ietf.org/doc/html/rfc7686

 Issue: https://github.com/curl/curl/issues/543

11.2 error buffer not set if connection to multiple addresses fails

 If you ask libcurl to resolve a hostname like example.com to IPv6 addresses
 only. But you only have IPv4 connectivity. libcurl will correctly fail with
 CURLE_COULDNT_CONNECT. But the error buffer set by CURLOPT_ERRORBUFFER
 remains empty. Issue: https://github.com/curl/curl/issues/544

11.3 Disconnects do not do verbose

 Due to how libcurl keeps connections alive in the "connection pool" after use
 to potentially transcend the life-time of the initial easy handle that was
 used to drive the transfer over that connection, it uses a *separate* and
 internal easy handle when it shuts down the connection. That separate
 connection might not have the same settings as the original easy handle, and
 in particular it is often note-worthy that it does not have the same VERBOSE
 and debug callbacks setup so that an application will not get the protocol
 data for the disconnect phase of a transfer the same way it got all the other
 data.

 This is because the original easy handle might have already been freed at that
 point and the application might not at all be prepared that the callback
 would get called again long after the handle was freed.

 See for example https://github.com/curl/curl/issues/6995

11.4 HTTP test server 'connection-monitor' problems

 The 'connection-monitor' feature of the sws HTTP test server does not work
 properly if some tests are run in unexpected order. Like 1509 and then 1525.

 See https://github.com/curl/curl/issues/868

11.5 Connection information when using TCP Fast Open

 CURLINFO_LOCAL_PORT (and possibly a few other) fails when TCP Fast Open is
 enabled.

 See https://github.com/curl/curl/issues/1332 and
 https://github.com/curl/curl/issues/4296

11.7 signal-based resolver timeouts

 libcurl built without an asynchronous resolver library uses alarm() to time
 out DNS lookups. When a timeout occurs, this causes libcurl to jump from the
 signal handler back into the library with a sigsetjmp, which effectively
 causes libcurl to continue running within the signal handler. This is
 non-portable and could cause problems on some platforms. A discussion on the
 problem is available at https://curl.se/mail/lib-2008-09/0197.html

 Also, alarm() provides timeout resolution only to the nearest second. alarm
 ought to be replaced by setitimer on systems that support it.

11.8 DoH leaks memory after followlocation

 https://github.com/curl/curl/issues/4592

11.9 DoH does not inherit all transfer options

 Some options are not inherited because they are not relevant for the DoH SSL
 connections, or inheriting the option may result in unexpected behavior. For
 example the user's debug function callback is not inherited because it would
 be unexpected for internal handles (ie DoH handles) to be passed to that
 callback.

 If an option is not inherited then it is not possible to set it separately for
 DoH without a DoH-specific option. For example: CURLOPT_DOH_SSL_VERIFYHOST,
 CURLOPT_DOH_SSL_VERIFYPEER and CURLOPT_DOH_SSL_VERIFYSTATUS.

 See https://github.com/curl/curl/issues/6605

11.10 Blocking socket operations in non-blocking API

 The list of blocking socket operations is in TODO section "More non-blocking".

11.11 A shared connection cache is not thread-safe

 The share interface offers CURL_LOCK_DATA_CONNECT to have multiple easy
 handle share a connection cache, but due to how connections are used they are
 still not thread-safe when used shared.

 See https://github.com/curl/curl/issues/4915 and lib1541.c

11.12 'no_proxy' string-matches IPv6 numerical addresses

 This has the downside that "::1" for example does not match "::0:1" even
 though they are in fact the same address.

 See https://github.com/curl/curl/issues/5745

11.13 wakeup socket disconnect causes havoc

 waking an iPad breaks the wakeup socket pair, triggering a POLLIN event and
 resulting in SOCKERRNO being set to ENOTCONN.

 This condition, and other possible error conditions on the wakeup socket, are
 not handled, so the condition remains on the FD and curl_multi_poll will
 never block again.

 See https://github.com/curl/curl/issues/6132 and
 https://github.com/curl/curl/pull/6133

11.14 Multi perform hangs waiting for threaded resolver

 If a threaded resolver takes a long time to complete, libcurl can be blocked
 waiting for it for a longer time than expected - and longer than the set
 timeouts.

 See https://github.com/curl/curl/issues/2975 and
 https://github.com/curl/curl/issues/4852

11.15 CURLOPT_OPENSOCKETPAIRFUNCTION is missing

 When libcurl creates sockets with socketpair(), those are not "exposed" in
 CURLOPT_OPENSOCKETFUNCTION and therefore might surprise and be unknown to
 applications that expect and want all sockets known beforehand. One way to
 address this issue is to introduce a CURLOPT_OPENSOCKETPAIRFUNCTION callback.

 https://github.com/curl/curl/issues/5747

11.16 libcurl uses renames instead of locking for atomic operations

 For saving cookies, alt-svc and hsts files. This is bad when for example the
 file is stored in a directory where the application has no write permission
 but it has permission for the file.

 https://github.com/curl/curl/issues/6882
 https://github.com/curl/curl/pull/6884

12. LDAP

12.1 OpenLDAP hangs after returning results

 By configuration defaults, openldap automatically chase referrals on
 secondary socket descriptors. The OpenLDAP backend is asynchronous and thus
 should monitor all socket descriptors involved. Currently, these secondary
 descriptors are not monitored, causing openldap library to never receive
 data from them.

 As a temporary workaround, disable referrals chasing by configuration.

 The fix is not easy: proper automatic referrals chasing requires a
 synchronous bind callback and monitoring an arbitrary number of socket
 descriptors for a single easy handle (currently limited to 5).

 Generic LDAP is synchronous: OK.

 See https://github.com/curl/curl/issues/622 and
     https://curl.se/mail/lib-2016-01/0101.html

12.2 LDAP on Windows does authentication wrong?

 https://github.com/curl/curl/issues/3116

12.3 LDAP on Windows does not work

 A simple curl command line getting "ldap://ldap.forumsys.com" returns an
 error that says "no memory" !

 https://github.com/curl/curl/issues/4261

12.4 LDAPS with NSS is slow

 See https://github.com/curl/curl/issues/5874

13. TCP/IP

13.1 --interface for ipv6 binds to unusable IP address

 Since IPv6 provides a lot of addresses with different scope, binding to an
 IPv6 address needs to take the proper care so that it does not bind to a
 locally scoped address as that is bound to fail.

 https://github.com/curl/curl/issues/686

13.2 Trying local ports fails on Windows

 This makes '--local-port [range]' to not work since curl can't properly
 detect if a port is already in use, so it'll try the first port, use that and
 then subsequently fail anyway if that was actually in use.

 https://github.com/curl/curl/issues/8112

14. DICT

14.1 DICT responses show the underlying protocol

 When getting a DICT response, the protocol parts of DICT are not stripped off
 from the output.

 https://github.com/curl/curl/issues/1809

15. CMake

15.1 use correct SONAME

 The autotools build sets the SONAME properly according to VERSIONINFO in
 lib/Makefile.am and so should cmake to make comparable build.

 See https://github.com/curl/curl/pull/5935

15.2 support build with GnuTLS

15.3 unusable tool_hugehelp.c with MinGW

 see https://github.com/curl/curl/issues/3125

15.4 build docs/curl.1

 The cmake build does not create the docs/curl.1 file and therefore must rely on
 it being there already. This makes the --manual option not work and test
 cases like 1139 cannot function.

15.5 build on Linux links libcurl to libdl

 ... which it should not need to!

 See https://github.com/curl/curl/issues/6165

15.6 uses -lpthread instead of Threads::Threads

 See https://github.com/curl/curl/issues/6166

15.7 generated .pc file contains strange entries

 The Libs.private field of the generated .pc file contains -lgcc -lgcc_s -lc
 -lgcc -lgcc_s

 See https://github.com/curl/curl/issues/6167

15.8 libcurl.pc uses absolute library paths

 The libcurl.pc file generated by cmake contains things like Libs.private:
 /usr/lib64/libssl.so /usr/lib64/libcrypto.so /usr/lib64/libz.so. The
 autotools equivalent would say Libs.private: -lssl -lcrypto -lz

 See https://github.com/curl/curl/issues/6169

15.9 cert paths autodetected when cross-compiling

 The autotools build disables the ca_path/ca_bundle detection when
 cross-compiling. The cmake build keeps doing the detection.

 See https://github.com/curl/curl/issues/6178

15.10 libpsl is not supported

 See https://github.com/curl/curl/issues/6214

15.11 ExternalProject_Add does not set CURL_CA_PATH

 CURL_CA_BUNDLE and CURL_CA_PATH are not set properly when cmake's
 ExternalProject_Add is used to build curl as a dependency.

 See https://github.com/curl/curl/issues/6313

15.12 cannot enable LDAPS on Windows

 See https://github.com/curl/curl/issues/6284

15.13 CMake build with MIT Kerberos does not work

 Minimum CMake version was bumped in curl 7.71.0 (#5358) Since CMake 3.2
 try_compile started respecting the CMAKE_EXE_FLAGS. The code dealing with
 MIT Kerberos detection sets few variables to potentially weird mix of space,
 and ;-separated flags. It had to blow up at some point. All the CMake checks
 that involve compilation are doomed from that point, the configured tree
 cannot be built.

 https://github.com/curl/curl/issues/6904

16. Applications

17. HTTP/2

17.1 Excessive HTTP/2 packets with TCP_NODELAY

 Because of how curl sets TCP_NODELAY by default, HTTP/2 requests are issued
 using more separate TCP packets than it would otherwise need to use. This
 means spending more bytes than it has to. Just disabling TCP_NODELAY for
 HTTP/2 is also not the correct fix because that then makes the outgoing
 packets to get delayed.

 See https://github.com/curl/curl/issues/6363

17.2 HTTP/2 frames while in the connection pool kill reuse

 If the server sends HTTP/2 frames (like for example an HTTP/2 PING frame) to
 curl while the connection is held in curl's connection pool, the socket will
 be found readable when considered for reuse and that makes curl think it is
 dead and then it will be closed and a new connection gets created instead.

 This is *best* fixed by adding monitoring to connections while they are kept
 in the pool so that pings can be responded to appropriately.

17.3 ENHANCE_YOUR_CALM causes infinite retries

 Infinite retries with 2 parallel requests on one connection receiving GOAWAY
 with ENHANCE_YOUR_CALM error code.

 See https://github.com/curl/curl/issues/5119

17.4 Connection failures with parallel HTTP/2

 See https://github.com/curl/curl/issues/5611

17.5 HTTP/2 connections through HTTPS proxy frequently stall

 See https://github.com/curl/curl/issues/6936

18. HTTP/3

18.1 If the HTTP/3 server closes connection during upload curl hangs

 See https://github.com/curl/curl/issues/6606

18.2 Transfer closed with n bytes remaining to read

 HTTP/3 transfers with the Jetty HTTP/3 server seem to not work.

 https://github.com/curl/curl/issues/8523

18.4 timeout when reusing a http3 connection

 HTTP/3 with quiche seems to not work and always timeout a subsequent transfer
 that reuses an already established connection

 https://github.com/curl/curl/issues/8764

18.9 connection migration does not work

 https://github.com/curl/curl/issues/7695