# TODO intro Things to do in project curl. Please tell us what you think, contribute and send us patches that improve things. Be aware that these are things that we could do, or have once been considered things we could do. If you want to work on any of these areas, please consider bringing it up for discussions first on the mailing list so that we all agree it is still a good idea for the project. All bugs documented in the [known_bugs document](https://curl.se/docs/knownbugs.html) are subject for fixing. # libcurl ## Consult `%APPDATA%` also for `.netrc` `%APPDATA%\.netrc` is not considered when running on Windows. Should not it? See [curl issue 4016](https://github.com/curl/curl/issues/4016) ## `struct lifreq` Use `struct lifreq` and `SIOCGLIFADDR` instead of `struct ifreq` and `SIOCGIFADDR` on newer Solaris versions as they claim the latter is obsolete. To support IPv6 interface addresses for network interfaces properly. ## alt-svc sharing The share interface could benefit from allowing the alt-svc cache to be possible to share between easy handles. See [curl issue 4476](https://github.com/curl/curl/issues/4476) 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 [curl issue 4915](https://github.com/curl/curl/issues/4915) and lib1541.c The share interface offers CURL_LOCK_DATA_HSTS to have multiple easy handle share an HSTS cache, but this is not thread-safe. ## thread-safe sharing Using the share interface users can share some data between easy handles but several of the sharing options are documented as not safe and supported to share between multiple concurrent threads. Fixing this would enable more users to share data in more powerful ways. ## updated DNS server while running If `/etc/resolv.conf` gets updated while a program using libcurl is running, it is may cause name resolves to fail unless `res_init()` is called. We should consider calling `res_init()` + retry once unconditionally on all name resolve failures to mitigate against this. Firefox works like that. Note that Windows does not have `res_init()` or an alternative. [curl issue 2251](https://github.com/curl/curl/issues/2251) ## c-ares and CURLOPT_OPENSOCKETFUNCTION curl creates most sockets via the CURLOPT_OPENSOCKETFUNCTION callback and close them with the CURLOPT_CLOSESOCKETFUNCTION callback. However, c-ares does not use those functions and instead opens and closes the sockets itself. This means that when curl passes the c-ares socket to the CURLMOPT_SOCKETFUNCTION it is not owned by the application like other sockets. See [curl issue 2734](https://github.com/curl/curl/issues/2734) ## Monitor connections in the connection pool libcurl's connection cache or pool holds a number of open connections for the purpose of possible subsequent connection reuse. It may contain a few up to a significant amount of connections. Currently, libcurl leaves all connections as they are and first when a connection is iterated over for matching or reuse purpose it is verified that it is still alive. Those connections may get closed by the server side for idleness or they may get an HTTP/2 ping from the peer to verify that they are still alive. By adding monitoring of the connections while in the pool, libcurl can detect dead connections (and close them) better and earlier, and it can handle HTTP/2 pings to keep such ones alive even when not actively doing transfers on them. ## Try to URL encode given URL Given a URL that for example contains spaces, libcurl could have an option that would try somewhat harder than it does now and convert spaces to %20 and perhaps URL encoded byte values over 128 etc (basically do what the redirect following code already does). [curl issue 514](https://github.com/curl/curl/issues/514) ## Add support for IRIs IRIs (RFC 3987) allow localized, non-ASCII, names in the URL. To properly support this, curl/libcurl would need to translate/encode the given input from the input string encoding into percent encoded output "over the wire". To make that work smoothly for curl users even on Windows, curl would probably need to be able to convert from several input encodings. ## try next proxy if one does not work Allow an application to specify a list of proxies to try, and failing to connect to the first go on and try the next instead until the list is exhausted. Browsers support this feature at least when they specify proxies using `PAC`. [curl issue 896](https://github.com/curl/curl/issues/896) ## provide timing info for each redirect curl and libcurl provide timing information via a set of different time-stamps (CURLINFO_*_TIME). When curl is following redirects, those returned time value are the accumulated sums. An improvement could be to offer separate timings for each redirect. [curl issue 6743](https://github.com/curl/curl/issues/6743) ## CURLINFO_PAUSE_STATE Return information about the transfer's current pause state, in both directions. See [curl issue 2588](https://github.com/curl/curl/issues/2588) ## Expose tried IP addresses that failed When libcurl fails to connect to a host, it could offer the application the addresses that were used in the attempt. Source + destination IP, source + destination port and protocol (UDP or TCP) for each failure. Possibly as a callback. Perhaps also provide reason. [curl issue 2126](https://github.com/curl/curl/issues/2126) ## erase secrets from heap/stack after use Introducing a concept and system to erase secrets from memory after use, it could help mitigate and lessen the impact of (future) security problems etc. However: most secrets are passed to libcurl as clear text from the application and then clearing them within the library adds nothing... [curl issue 7268](https://github.com/curl/curl/issues/7268) ## make DoH inherit more transfer properties 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 (i.e 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 [curl issue 6605](https://github.com/curl/curl/issues/6605) # libcurl - multi interface ## More non-blocking Make sure we do not ever loop because of non-blocking sockets returning `EWOULDBLOCK` or similar. Blocking cases include: - Name resolves on non-Windows unless c-ares or the threaded resolver is used. - The threaded resolver may block on cleanup: [curl issue 4852](https://github.com/curl/curl/issues/4852) - `file://` transfers - TELNET transfers - GSSAPI authentication for FTP transfers - The "DONE" operation (post transfer protocol-specific actions) for the protocols SFTP, SMTP, FTP. Fixing `multi_done()` for this is a worthy task. - `curl_multi_remove_handle()` for any of the above. - Calling `curl_ws_send()` from a callback ## Better support for same name resolves If a name resolve has been initiated for a given name and a second easy handle wants to resolve that same name as well, make it wait for the first resolve to end up in the cache instead of doing a second separate resolve. This is especially needed when adding many simultaneous handles using the same hostname when the DNS resolver can get flooded. ## Non-blocking `curl_multi_remove_handle()` The multi interface has a few API calls that assume a blocking behavior, like `add_handle()` and `remove_handle()` which limits what we can do internally. The multi API need to be moved even more into a single function that "drives" everything in a non-blocking manner and signals when something is done. A remove or add would then only ask for the action to get started and then `multi_perform()` etc still be called until the add/remove is completed. ## Split connect and authentication process The multi interface treats the authentication process as part of the connect phase. As such any failures during authentication does not trigger the relevant QUIT or LOGOFF for protocols such as IMAP, POP3 and SMTP. ## Edge-triggered sockets should work The multi_socket API should work with edge-triggered socket events. One of the internal actions that need to be improved for this to work perfectly is the `maxloops` handling in `transfer.c:readwrite_data()`. ## multi upkeep In libcurl 7.62.0 we introduced `curl_easy_upkeep`. It unfortunately only works on easy handles. We should introduces a version of that for the multi handle, and also consider doing `upkeep` automatically on connections in the connection pool when the multi handle is in used. See [curl issue 3199](https://github.com/curl/curl/issues/3199) ## Virtual external sockets libcurl performs operations on the given file descriptor that presumes it is a socket and an application cannot replace them at the moment. Allowing an application to fully replace those would allow a larger degree of freedom and flexibility. See [curl issue 5835](https://github.com/curl/curl/issues/5835) ## dynamically decide to use socketpair For users who do not use `curl_multi_wait()` or do not care for `curl_multi_wakeup()`, we could introduce a way to make libcurl NOT create a socketpair in the multi handle. See [curl issue 4829](https://github.com/curl/curl/issues/4829) # Documentation ## Improve documentation about fork safety See [curl issue 6968](https://github.com/curl/curl/issues/6968) # FTP ## A fixed directory listing format Since listing the contents of a remove directory with FTP is returning the list in a format and style the server likes without any established or even defacto standard existing, it would be a feature to users if curl could parse the directory listing and output a general curl format that is fixed and the same, independent of the server's choice. This would allow users to better and more reliably extract information about remote content via FTP directory listings. ## GSSAPI via Windows SSPI In addition to currently supporting the SASL GSSAPI mechanism (Kerberos V5) via third-party GSS-API libraries, such as MIT Kerberos, also add support for GSSAPI authentication via Windows SSPI. ## STAT for LIST without data connection Some FTP servers allow STAT for listing directories instead of using LIST, and the response is then sent over the control connection instead of as the otherwise used data connection. This is not detailed in any FTP specification. ## Passive transfer could try other IP addresses 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 could try all IP addresses for `localhost`. Currently it tries `::1`, but it should also try `127.0.0.1`. See [curl issue 1508](https://github.com/curl/curl/issues/1508) # HTTP ## Provide the error body from a CONNECT response When curl receives a body response from a CONNECT request to a proxy, it always just reads and ignores it. It would make some users happy if curl instead optionally would be able to make that responsible available. Via a new callback? Through some other means? See [curl issue 9513](https://github.com/curl/curl/issues/9513) ## Obey `Retry-After` in redirects The `Retry-After` response header is said to dictate "the minimum time that the user agent is asked to wait before issuing the redirected request" and libcurl does not obey this. See [curl issue 11447](https://github.com/curl/curl/issues/11447) ## Rearrange request header order Server implementers often make an effort to detect browser and to reject clients it can detect to not match. One of the last details we cannot yet control in libcurl's HTTP requests, which also can be exploited to detect that libcurl is in fact used even when it tries to impersonate a browser, is the order of the request headers. I propose that we introduce a new option in which you give headers a value, and then when the HTTP request is built it sorts the headers based on that number. We could then have internally created headers use a default value so only headers that need to be moved have to be specified. ## Allow SAN names in HTTP/2 server push curl only allows HTTP/2 push promise if the provided :authority header value exactly matches the hostname given in the URL. It could be extended to allow any name that would match the Subject Alternative Names in the server's TLS certificate. See [curl pull request 3581](https://github.com/curl/curl/pull/3581) ## `auth=` in URLs Add the ability to specify the preferred authentication mechanism to use by using `;auth=` in the login part of the URL. For example: `http://test:pass;auth=NTLM@example.com` would be equivalent to specifying `--user test:pass;auth=NTLM` or `--user test:pass --ntlm` from the command line. Additionally this should be implemented for proxy base URLs as well. ## alt-svc should fallback if alt-svc does not work The `alt-svc:` header provides a set of alternative services for curl to use instead of the original. If the first attempted one fails, it should try the next etc and if all alternatives fail go back to the original. See [curl issue 4908](https://github.com/curl/curl/issues/4908) ## Require HTTP version X or higher curl and libcurl provide options for trying higher HTTP versions (for example HTTP/2) but then still allows the server to pick version 1.1. We could consider adding a way to require a minimum version. See [curl issue 7980](https://github.com/curl/curl/issues/7980) # TELNET ## ditch stdin Reading input (to send to the remote server) on stdin is a crappy solution for library purposes. We need to invent a good way for the application to be able to provide the data to send. ## ditch telnet-specific select Move the telnet support's network `select()` loop go away and merge the code into the main transfer loop. Until this is done, the multi interface does not work for telnet. ## feature negotiation debug data Add telnet feature negotiation data to the debug callback as header data. ## exit immediately upon connection if stdin is /dev/null If it did, curl could be used to probe if there is an server there listening on a specific port. That is, the following command would exit immediately after the connection is established with exit code 0: curl -s --connect-timeout 2 telnet://example.com:80 NOTIFY=SUCCESS,FAILURE");`. [curl issue 8232](https://github.com/curl/curl/issues/8232) ## Enhanced capability support Add the ability, for an application that uses libcurl, to obtain the list of capabilities returned from the EHLO command. ## Add `CURLOPT_MAIL_CLIENT` option Rather than use the URL to specify the mail client string to present in the `HELO` and `EHLO` commands, libcurl should support a new `CURLOPT` specifically for specifying this data as the URL is non-standard and to be honest a bit of a hack. Please see the following thread for more information: https://curl.se/mail/lib-2012-05/0178.html # POP3 ## Enhanced capability support Add the ability, for an application that uses libcurl, to obtain the list of capabilities returned from the CAPA command. # IMAP ## Enhanced capability support Add the ability, for an application that uses libcurl, to obtain the list of capabilities returned from the CAPABILITY command. # LDAP ## SASL based authentication mechanisms Currently the LDAP module only supports `ldap_simple_bind_s()` in order to bind to an LDAP server. However, this function sends username and password details using the simple authentication mechanism (as clear text). However, it should be possible to use `ldap_bind_s()` instead specifying the security context information ourselves. ## `CURLOPT_SSL_CTX_FUNCTION` for LDAPS `CURLOPT_SSL_CTX_FUNCTION` works perfectly for HTTPS and email protocols, but it has no effect for LDAPS connections. [curl issue 4108](https://github.com/curl/curl/issues/4108) ## Paged searches on LDAP server [curl issue 4452](https://github.com/curl/curl/issues/4452) ## Certificate-Based Authentication LDAPS not possible with macOS and Windows with Certificate-Based Authentication [curl issue 9641](https://github.com/curl/curl/issues/9641) # SMB ## Support modern versions curl only supports version 1, which barely anyone is using anymore. ## File listing support Add support for listing the contents of an SMB share. The output should probably be the same as/similar to FTP. ## Honor file timestamps The timestamp of the transferred file should reflect that of the original file. ## Use NTLMv2 Currently the SMB authentication uses NTLMv1. ## Create remote directories Support for creating remote directories when uploading a file to a directory that does not exist on the server, just like `--ftp-create-dirs`. # FILE ## Directory listing on non-POSIX Listing the contents of a directory accessed with FILE only works on platforms with `opendir()`. Support could be added for more systems, like Windows. # TLS ## `TLS-PSK` with OpenSSL Transport Layer Security pre-shared key cipher suites (`TLS-PSK`) is a set of cryptographic protocols that provide secure communication based on pre-shared keys (`PSK`). These pre-shared keys are symmetric keys shared in advance among the communicating parties. [curl issue 5081](https://github.com/curl/curl/issues/5081) ## TLS channel binding TLS 1.2 and 1.3 provide the ability to extract some secret data from the TLS connection and use it in the client request (usually in some sort of authentication) to ensure that the data sent is bound to the specific TLS connection and cannot be successfully intercepted by a proxy. This functionality can be used in a standard authentication mechanism such as GSS-API or SCRAM, or in custom approaches like custom HTTP Authentication headers. For TLS 1.2, the binding type is usually `tls-unique`, and for TLS 1.3 it is `tls-exporter`. - https://datatracker.ietf.org/doc/html/rfc5929 - https://datatracker.ietf.org/doc/html/rfc9266 - [curl issue 9226](https://github.com/curl/curl/issues/9226) ## Defeat TLS fingerprinting By changing the order of TLS extensions provided in the TLS handshake, it is sometimes possible to circumvent TLS fingerprinting by servers. The TLS extension order is of course not the only way to fingerprint a client. ## Consider OCSP stapling by default Treat a negative response a reason for aborting the connection. Since OCSP stapling is presumed to get used much less in the future when Let's Encrypt drops the OCSP support, the benefit of this might however be limited. [curl issue 15483](https://github.com/curl/curl/issues/15483) ## Provide callback for cert verification OpenSSL supports a callback for customized verification of the peer certificate, but this does not seem to be exposed in the libcurl APIs. Could it be? There is so much that could be done if it were. ## Less memory massaging with Schannel The Schannel backend does a lot of custom memory management we would rather avoid: the repeated allocation + free in sends and the custom memory + realloc system for encrypted and decrypted data. That should be avoided and reduced for 1) efficiency and 2) safety. ## Support DANE [DNS-Based Authentication of Named Entities (DANE)](https://datatracker.ietf.org/doc/html/rfc6698) is a way to provide SSL keys and certs over DNS using DNSSEC as an alternative to the CA model. A patch was posted on March 7 2013 (https://curl.se/mail/lib-2013-03/0075.html) but it was a too simple approach. See Daniel's comments: https://curl.se/mail/lib-2013-03/0103.html Björn Stenberg once wrote a separate initial take on DANE that was never completed. ## TLS record padding TLS (1.3) offers optional record padding and OpenSSL provides an API for it. I could make sense for libcurl to offer this ability to applications to make traffic patterns harder to figure out by network traffic observers. See [curl issue 5398](https://github.com/curl/curl/issues/5398) ## Support Authority Information Access certificate extension (AIA) AIA can provide various things like certificate revocation lists but more importantly information about intermediate CA certificates that can allow validation path to be fulfilled when the HTTPS server does not itself provide them. Since AIA is about downloading certs on demand to complete a TLS handshake, it is probably a bit tricky to get done right and a serious privacy leak. See [curl issue 2793](https://github.com/curl/curl/issues/2793) ## Some TLS options are not offered for HTTPS proxies Some TLS related options to the command line tool and libcurl are only provided for the server and not for HTTPS proxies. `--proxy-tls-max`, `--proxy-tlsv1.3`, `--proxy-curves` and a few more. For more Documentation on this see: https://curl.se/libcurl/c/tls-options.html [curl issue 12286](https://github.com/curl/curl/issues/12286) ## Make sure we forbid TLS 1.3 post-handshake authentication RFC 8740 explains how using HTTP/2 must forbid the use of TLS 1.3 post-handshake authentication. We should make sure to live up to that. See [curl issue 5396](https://github.com/curl/curl/issues/5396) ## Support the `clienthello` extension Certain stupid networks and middle boxes have a problem with SSL handshake packets that are within a certain size range because how that sets some bits that previously (in older TLS version) were not set. The `clienthello` extension adds padding to avoid that size range. - https://datatracker.ietf.org/doc/html/rfc7685 - [curl issue 2299](https://github.com/curl/curl/issues/2299) ## Share the CA cache For TLS backends that supports CA caching, it makes sense to allow the share object to be used to store the CA cache as well via the share API. Would allow multiple easy handles to reuse the CA cache and save themselves from a lot of extra processing overhead. ## Add missing features to TLS backends The feature matrix at https://curl.se/libcurl/c/tls-options.html shows which features are supported by which TLS backends, and thus also where there are feature gaps. # Proxy ## Retry SOCKS handshake on address type not supported When curl resolves a hostname, it might get a mix of IPv6 and IPv4 returned. curl might then use an IPv6 address with a SOCKS5 proxy, which - if it does not support IPv6 - returns "Address type not supported" and curl exits with that error. Perhaps it is preferred if curl would in this situation instead first retry the SOCKS handshake again for this case and then use one of the IPv4 addresses for the target host. See [curl issue 17222](https://github.com/curl/curl/issues/17222) # Schannel ## Extend support for client certificate authentication The existing support for the `-E`/`--cert` and `--key` options could be extended by supplying a custom certificate and key in PEM format, see: [Getting a Certificate for Schannel](https://learn.microsoft.com/windows/win32/secauthn/getting-a-certificate-for-schannel) ## Extend support for the `--ciphers` option The existing support for the `--ciphers` option could be extended by mapping the OpenSSL/GnuTLS cipher suites to the Schannel APIs, see [Specifying Schannel Ciphers and Cipher Strengths](https://learn.microsoft.com/windows/win32/secauthn/specifying-schannel-ciphers-and-cipher-strengths). ## Add option to allow abrupt server closure libcurl with Schannel errors without a known termination point from the server (such as length of transfer, or SSL "close notify" alert) to prevent against a truncation attack. Really old servers may neglect to send any termination point. An option could be added to ignore such abrupt closures. [curl issue 4427](https://github.com/curl/curl/issues/4427) # SASL ## Other authentication mechanisms Add support for other authentication mechanisms such as `OLP`, `GSS-SPNEGO` and others. ## Add `QOP` support to GSSAPI authentication Currently the GSSAPI authentication only supports the default `QOP` of auth (Authentication), whilst Kerberos V5 supports both `auth-int` (Authentication with integrity protection) and `auth-conf` (Authentication with integrity and privacy protection). # SSH protocols ## Multiplexing SSH is a perfectly fine multiplexed protocols which would allow libcurl to do multiple parallel transfers from the same host using the same connection, much in the same spirit as HTTP/2 does. libcurl however does not take advantage of that ability but does instead always create a new connection for new transfers even if an existing connection already exists to the host. To fix this, libcurl would have to detect an existing connection and "attach" the new transfer to the existing one. ## Handle growing SFTP files The SFTP code in libcurl checks the file size *before* a transfer starts and then proceeds to transfer exactly that amount of data. If the remote file grows while the transfer is in progress libcurl does not notice and does not adapt. The OpenSSH SFTP command line tool does and libcurl could also just attempt to download more to see if there is more to get... [curl issue 4344](https://github.com/curl/curl/issues/4344) ## Read keys from `~/.ssh/id_ecdsa`, `id_ed25519` The libssh2 backend in curl is limited to only reading keys from `id_rsa` and `id_dsa`, which makes it fail connecting to servers that use more modern key types. [curl issue 8586](https://github.com/curl/curl/issues/8586) ## Support `CURLOPT_PREQUOTE` The two other `QUOTE` options are supported for SFTP, but this was left out for unknown reasons. ## SSH over HTTPS proxy for libssh The SSH based protocols SFTP and SCP did not work over HTTPS proxy at all until [curl pull request 6021](https://github.com/curl/curl/pull/6021) brought the functionality with the libssh2 backend. Presumably, this support can/could be added for the libssh backend as well. ## SFTP with `SCP://` OpenSSH 9 switched their `scp` tool to speak SFTP under the hood. Going forward it might be worth having curl or libcurl attempt SFTP if SCP fails to follow suite. # Command line tool ## multi-threading When asked to do transfers in parallel, the curl tool could be extended to use a number of independent worker threads. This would allow faster transfers in situations where curl becomes CPU bound. Ideally, curl would (with permission) fire up new threads on demand when it deems that it might be helpful. Perhaps, if it has more transfers to add and the existing transfers make the CPU busy enough and there are more cores available. ## sync `curl --sync http://example.com/feed[1-100].rss` or `curl --sync http://example.net/{index,calendar,history}.html` Downloads a range or set of URLs using the remote name, but only if the remote file is newer than the local file. A `Last-Modified` HTTP date header should also be used to set the mod date on the downloaded file. ## glob posts Globbing support for `-d` and `-F`, as in `curl -d "name=foo[0-9]" URL`. This is easily scripted though. ## `--proxycommand` Allow the user to make curl run a command and use its stdio to make requests and not do any network connection by itself. Example: curl --proxycommand 'ssh pi@raspberrypi.local -W 10.1.1.75 80' \ http://some/otherwise/unavailable/service.php See [curl issue 4941](https://github.com/curl/curl/issues/4941) ## UTF-8 filenames in Content-Disposition RFC 6266 documents how UTF-8 names can be passed to a client in the `Content-Disposition` header, and curl does not support this. [curl issue 1888](https://github.com/curl/curl/issues/1888) ## Option to make `-Z` merge lined based outputs on stdout When a user requests multiple lined based files using `-Z` and sends them to stdout, curl does not *merge* and send complete lines fine but may send partial lines from several sources. [curl issue 5175](https://github.com/curl/curl/issues/5175) ## specify which response codes that make `-f`/`--fail` return error Allows a user to better specify exactly which error code(s) that are fine and which are errors for their specific uses cases ## Choose the name of file in braces for complex URLs When using braces to download a list of URLs and you use complicated names in the list of alternatives, it could be handy to allow curl to use other names when saving. Consider a way to offer that. Possibly like `{partURL1:name1,partURL2:name2,partURL3:name3}` where the name following the colon is the output name. See [curl issue 221](https://github.com/curl/curl/issues/221) ## improve how curl works in a Windows console window If you pull the scroll bar when transferring with curl in a Windows console window, the transfer is interrupted and can get disconnected. This can probably be improved. See [curl issue 322](https://github.com/curl/curl/issues/322) ## Windows: set attribute 'archive' for completed downloads The archive bit (`FILE_ATTRIBUTE_ARCHIVE, 0x20`) separates files that shall be backed up from those that are either not ready or have not changed. Downloads in progress are neither ready to be backed up, nor should they be opened by a different process. Only after a download has been completed it is sensible to include it in any integer snapshot or backup of the system. See [curl issue 3354](https://github.com/curl/curl/issues/3354) ## keep running, read instructions from pipe/socket Provide an option that makes curl not exit after the last URL (or even work without a given URL), and then make it read instructions passed on a pipe or over a socket to make further instructions so that a second subsequent curl invoke can talk to the still running instance and ask for transfers to get done, and thus maintain its connection pool, DNS cache and more. ## Acknowledge `Ratelimit` headers Consider a command line option that can make curl do multiple serial requests while acknowledging server specified [rate limits](https://datatracker.ietf.org/doc/draft-ietf-httpapi-ratelimit-headers/). See [curl issue 5406](https://github.com/curl/curl/issues/5406) ## `--dry-run` A command line option that makes curl show exactly what it would do and send if it would run for real. See [curl issue 5426](https://github.com/curl/curl/issues/5426) ## `--retry` should resume When `--retry` is used and curl actually retries transfer, it should use the already transferred data and do a resumed transfer for the rest (when possible) so that it does not have to transfer the same data again that was already transferred before the retry. See [curl issue 1084](https://github.com/curl/curl/issues/1084) ## consider filename from the redirected URL with `-O` ? When a user gives a URL and uses `-O`, and curl follows a redirect to a new URL, the filename is not extracted and used from the newly redirected-to URL even if the new URL may have a much more sensible filename. This is clearly documented and helps for security since there is no surprise to users which filename that might get overwritten, but maybe a new option could allow for this or maybe `-J` should imply such a treatment as well as `-J` already allows for the server to decide what filename to use so it already provides the "may overwrite any file" risk. This is extra tricky if the original URL has no filename part at all since then the current code path does error out with an error message, and we cannot *know* already at that point if curl is redirected to a URL that has a filename... See [curl issue 1241](https://github.com/curl/curl/issues/1241) ## retry on network is unreachable The `--retry` option retries transfers on *transient failures*. We later added `--retry-connrefused` to also retry for *connection refused* errors. Suggestions have been brought to also allow retry on *network is unreachable* errors and while totally reasonable, maybe we should consider a way to make this more configurable than to add a new option for every new error people want to retry for? [curl issue 1603](https://github.com/curl/curl/issues/1603) ## hostname sections in config files config files would be more powerful if they could set different configurations depending on used URLs, hostname or possibly origin. Then a default `.curlrc` could a specific user-agent only when doing requests against a certain site. ## retry on the redirected-to URL When curl is told to `--retry` a failed transfer and follows redirects, it might get an HTTP 429 response from the redirected-to URL and not the original one, which then could make curl decide to rather retry the transfer on that URL only instead of the original operation to the original URL. Perhaps extra emphasized if the original transfer is a large POST that redirects to a separate GET, and that GET is what gets the 529 See [curl issue 5462](https://github.com/curl/curl/issues/5462) ## Set the modification date on an uploaded file For SFTP and possibly FTP, curl could offer an option to set the modification time for the uploaded file. See [curl issue 5768](https://github.com/curl/curl/issues/5768) ## Use multiple parallel transfers for a single download To enhance transfer speed, downloading a single URL can be split up into multiple separate range downloads that get combined into a single final result. An ideal implementation would not use a specified number of parallel transfers, but curl could: - First start getting the full file as transfer A - If after N seconds have passed and the transfer is expected to continue for M seconds or more, add a new transfer (B) that asks for the second half of A's content (and stop A at the middle). - If splitting up the work improves the transfer rate, it could then be done again. Then again, etc up to a limit. This way, if transfer B fails (because Range: is not supported) it lets transfer A remain the single one. N and M could be set to some sensible defaults. See [curl issue 5774](https://github.com/curl/curl/issues/5774) ## Prevent terminal injection when writing to terminal curl could offer an option to make escape sequence either non-functional or avoid cursor moves or similar to reduce the risk of a user getting tricked by clever tricks. See [curl issue 6150](https://github.com/curl/curl/issues/6150) ## `-J` and `-O` with %-encoded filenames `-J`/`--remote-header-name` does not decode %-encoded filenames. RFC 6266 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. See 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 do not decode `-O` without the user asking for it with some other means, since `-O` has always been documented to use the name exactly as specified in the URL. ## `-J` with `-C -` 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 filename (and thus its pre-transfer size) has been figured out. This can be improved. https://curl.se/bug/view.cgi?id=1169 ## `--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 truncates and restarts at the original position where it was at before the previous failed attempt. See https://curl.se/mail/lib-2008-01/0080.html # Build ## Enable `PIE` and `RELRO` by default Especially when having programs that execute curl via the command line, `PIE` renders the exploitation of memory corruption vulnerabilities a lot more difficult. This can be attributed to the additional information leaks being required to conduct a successful attack. `RELRO`, on the other hand, masks different binary sections like the `GOT` as read-only and thus kills a handful of techniques that come in handy when attackers are able to arbitrarily overwrite memory. A few tests showed that enabling these features had close to no impact, neither on the performance nor on the general functionality of curl. ## Do not use GNU libtool on OpenBSD When compiling curl on OpenBSD with `--enable-debug` it gives linking errors when you use GNU libtool. This can be fixed by using the libtool provided by OpenBSD itself. However for this the user always needs to invoke make with `LIBTOOL=/usr/bin/libtool`. It would be nice if the script could have some magic to detect if this system is an OpenBSD host and then use the OpenBSD libtool instead. See [curl issue 5862](https://github.com/curl/curl/issues/5862) ## Package curl for Windows in a signed installer See [curl issue 5424](https://github.com/curl/curl/issues/5424) ## make configure use `--cache-file` more and better The configure script can be improved to cache more values so that repeated invokes run much faster. See [curl issue 7753](https://github.com/curl/curl/issues/7753) # Test suite ## SSL tunnel Make our own version of stunnel for simple port forwarding to enable HTTPS and FTP-SSL tests without the stunnel dependency, and it could allow us to provide test tools built with either OpenSSL or GnuTLS ## more protocols supported Extend the test suite to include more protocols. The telnet could just do FTP or http operations (for which we have test servers). ## more platforms supported Make the test suite work on more platforms. OpenBSD and macOS. Remove fork()s and it should become even more portable. ## write an SMB test server to replace impacket This would allow us to run SMB tests on more platforms and do better and more covering tests. See [curl issue 15697](https://github.com/curl/curl/issues/15697) ## Use the RFC 6265 test suite A test suite made for HTTP cookies (RFC 6265) by Adam Barth [is available](https://github.com/abarth/http-state/tree/master/tests). It would be good if someone would write a script/setup that would run curl with that test suite and detect deviance. Ideally, that would even be incorporated into our regular test suite. ## Run web-platform-tests URL tests Run web-platform-tests URL tests and compare results with browsers on `wpt.fyi`. It would help us find issues to fix and help us document where our parser differs from the WHATWG URL spec parsers. See [curl issue 4477](https://github.com/curl/curl/issues/4477) # MQTT ## Support rate-limiting The rate-limiting logic is done in the PERFORMING state in multi.c but MQTT is not (yet) implemented to use that. ## Support MQTTS ## Handle network blocks Running test suite with `CURL_DBG_SOCK_WBLOCK=90 ./runtests.pl -a mqtt` makes several MQTT test cases fail where they should not. ## large payloads libcurl unnecessarily allocates heap memory to hold the entire payload to get sent, when the data is already perfectly accessible where it is when `CURLOPT_POSTFIELDS` is used. This is highly inefficient for larger payloads. Additionally, libcurl does not support using the read callback for sending MQTT which is yet another way to avoid having to hold large payload in memory. # TFTP ## TFTP does not convert LF to CRLF for `mode=netascii` RFC 3617 defines that an TFTP transfer can be done using `netascii` mode. curl does not support extracting that mode from the URL nor does it treat such transfers specifically. It should probably do LF to CRLF translations for them. See [curl issue 12655](https://github.com/curl/curl/issues/12655) # Gopher ## Handle network blocks Running test suite with `CURL_DBG_SOCK_WBLOCK=90 ./runtests.pl -a 1200 to 1300` makes several Gopher test cases fail where they should not. # Signals ## SIGPIPE Since we control the IO functions for most protocols and disable SIGPIPE on sends, libcurl could skip the special SIGPIPE ignore handling for those transfers.