curl/docs/ECH.md

Building curl with HTTPS-RR and ECH support

We've added support for ECH to in this curl build. That can use HTTPS RRs published in the DNS, if curl is using DoH, or else can accept the relevant ECHConfigList values from the command line. That works with OpenSSL, WolfSSL or boringssl as the TLS provider, depending on how you build curl.

This feature is EXPERIMENTAL. DO NOT USE IN PRODUCTION.

This should however provide enough of a proof-of-concept to prompt an informed discussion about a good path forward for ECH support in curl, when using OpenSSL, or other TLS libraries, as those add ECH support.

OpenSSL Build

To build our ECH-enabled OpenSSL fork:

    cd $HOME/code
    git clone https://github.com/defo-project/openssl
    cd openssl
    ./config --libdir=lib --prefix=$HOME/code/openssl-local-inst
    ...stuff...
    make -j8
    ...stuff (maybe go for coffee)...
    make install_sw
    ...a little bit of stuff...

To build curl ECH-enabled, making use of the above:

    cd $HOME/code
    git clone https://github.com/curl/curl
    cd curl
    autoreconf -fi
    LDFLAGS="-Wl,-rpath,$HOME/code/openssl-local-inst/lib/" ./configure --with-ssl=$HOME/code/openssl-local-inst --enable-ech --enable-httpsrr
    ...lots of output...
    WARNING: ech ECH HTTPSRR enabled but marked EXPERIMENTAL...
    make
    ...lots more output...

If you do not get that WARNING at the end of the configure command, then ECH is not enabled, so go back some steps and re-do whatever needs re-doing:-) If you want to debug curl then you should add --enable-debug to the configure command.

In a recent (2024-05-20) build on one machine, configure failed to find the ECH-enabled SSL library, apparently due to the existence of $HOME/code/openssl-local-inst/lib/pkgconfig as a directory containing various settings. Deleting that directory worked around the problem but may not be the best solution.

Using ECH and DoH

Curl supports using DoH for A/AAAA lookups so it was relatively easy to add retrieval of HTTPS RRs in that situation. To use ECH and DoH together:

    cd $HOME/code/curl
    LD_LIBRARY_PATH=$HOME/code/openssl ./src/curl --ech true --doh-url https://one.one.one.one/dns-query https://defo.ie/ech-check.php
    ...
    SSL_ECH_STATUS: success <img src="greentick-small.png" alt="good" /> <br/>
    ...

The output snippet above is within the HTML for the webpage, when things work.

The above works for these test sites:

    https://defo.ie/ech-check.php
    https://draft-13.esni.defo.ie:8413/stats
    https://draft-13.esni.defo.ie:8414/stats
    https://crypto.cloudflare.com/cdn-cgi/trace
    https://tls-ech.dev

The list above has 4 different server technologies, implemented by 3 different parties, and includes a case (the port 8414 server) where HelloRetryRequest (HRR) is forced.

We currently support the following new curl command line arguments/options:

• --ech <config> - the config value can be one of:
  • false says to not attempt ECH
  • true says to attempt ECH, if possible
  • grease if attempting ECH is not possible, then send a GREASE ECH extension
  • hard hard-fail the connection if ECH cannot be attempted
  • ecl:<b64value> a base64 encoded ECHConfigList, rather than one accessed from the DNS
  • pn:<name> over-ride the public_name from an ECHConfigList

Note that in the above "attempt ECH" means the client emitting a TLS ClientHello with a "real" ECH extension, but that does not mean that the relevant server can succeed in decrypting, as things can fail for other reasons.

Supplying an ECHConfigList on the command line

To supply the ECHConfigList on the command line, you might need a bit of cut-and-paste, e.g.:

    dig +short https defo.ie
    1 . ipv4hint=213.108.108.101 ech=AED+DQA8PAAgACD8WhlS7VwEt5bf3lekhHvXrQBGDrZh03n/LsNtAodbUAAEAAEAAQANY292ZXIuZGVmby5pZQAA ipv6hint=2a00:c6c0:0:116:5::10

Then paste the base64 encoded ECHConfigList onto the curl command line:

    LD_LIBRARY_PATH=$HOME/code/openssl ./src/curl --ech ecl:AED+DQA8PAAgACD8WhlS7VwEt5bf3lekhHvXrQBGDrZh03n/LsNtAodbUAAEAAEAAQANY292ZXIuZGVmby5pZQAA https://defo.ie/ech-check.php
    ...
    SSL_ECH_STATUS: success <img src="greentick-small.png" alt="good" /> <br/>
    ...

The output snippet above is within the HTML for the webpage.

If you paste in the wrong ECHConfigList (it changes hourly for defo.ie) you should get an error like this:

    LD_LIBRARY_PATH=$HOME/code/openssl ./src/curl -vvv --ech ecl:AED+DQA8yAAgACDRMQo+qYNsNRNj+vfuQfFIkrrUFmM4vogucxKj/4nzYgAEAAEAAQANY292ZXIuZGVmby5pZQAA https://defo.ie/ech-check.php
    ...
    * OpenSSL/3.3.0: error:0A00054B:SSL routines::ech required
    ...

There is a reason to want this command line option - for use before publishing an ECHConfigList in the DNS as per the Internet-draft A well-known URI for publishing ECHConfigList values.

If you do use a wrong ECHConfigList value, then the server might return a good value, via the retry_configs mechanism. You can see that value in the verbose output, e.g.:

    LD_LIBRARY_PATH=$HOME/code/openssl ./src/curl -vvv --ech ecl:AED+DQA8yAAgACDRMQo+qYNsNRNj+vfuQfFIkrrUFmM4vogucxKj/4nzYgAEAAEAAQANY292ZXIuZGVmby5pZQAA https://defo.ie/ech-check.php
    ...
* ECH: retry_configs AQD+DQA8DAAgACBvYqJy+Hgk33wh/ZLBzKSPgwxeop7gvojQzfASq7zeZQAEAAEAAQANY292ZXIuZGVmby5pZQAA/g0APEMAIAAgXkT5r4cYs8z19q5rdittyIX8gfQ3ENW4wj1fVoiJZBoABAABAAEADWNvdmVyLmRlZm8uaWUAAP4NADw2ACAAINXSE9EdXzEQIJZA7vpwCIQsWqsFohZARXChgPsnfI1kAAQAAQABAA1jb3Zlci5kZWZvLmllAAD+DQA8cQAgACASeiD5F+UoSnVoHvA2l1EifUVMFtbVZ76xwDqmMPraHQAEAAEAAQANY292ZXIuZGVmby5pZQAA
* ECH: retry_configs for defo.ie from cover.defo.ie, 319
    ...

At that point, you could copy the base64 encoded value above and try again. For now, this only works for the OpenSSL and boringssl builds.

Default settings

Curl has various ways to configure default settings, e.g. in $HOME/.curlrc, so one can set the DoH URL and enable ECH that way:

    cat ~/.curlrc
    doh-url=https://one.one.one.one/dns-query
    silent
    ech=true

Note that when you use the system's curl command (rather than our ECH-enabled build), it is liable to warn that ech is an unknown option. If that is an issue (e.g. if some script re-directs stdout and stderr somewhere) then adding the silent line above seems to be a good enough fix. (Though of course, yet another script could depend on non-silent behavior, so you may have to figure out what you prefer yourself.) That seems to have changed with the latest build, previously silent=TRUE was what I used in ~/.curlrc but now that seems to cause a problem, so that the following line(s) are ignored.

If you want to always use our OpenSSL build you can set LD_LIBRARY_PATH in the environment:

    export LD_LIBRARY_PATH=$HOME/code/openssl

When you do the above, there can be a mismatch between OpenSSL versions for applications that check that. A git push for example fails so you should unset LD_LIBRARY_PATH before doing that or use a different shell.

    git push
    OpenSSL version mismatch. Built against 30000080, you have 30200000
    ...

With all that setup as above the command line gets simpler:

    ./src/curl https://defo.ie/ech-check.php
    ...
    SSL_ECH_STATUS: success <img src="greentick-small.png" alt="good" /> <br/>
    ...

The --ech true option is opportunistic, so tries to do ECH but does not fail if the client for example cannot find any ECHConfig values. The --ech hard option hard-fails if there is no ECHConfig found in DNS, so for now, that is not a good option to set as a default. Once ECH has really been attempted by the client, if decryption on the server side fails, then curl fails.

Code changes for ECH support when using DoH

Code changes are #ifdef protected via USE_ECH or USE_HTTPSRR:

• USE_HTTPSRR is used for HTTPS RR retrieval code that could be generically used should non-ECH uses for HTTPS RRs be identified, e.g. use of ALPN values or IP address hints.

• USE_ECH protects ECH specific code.

There are various obvious code blocks for handling the new command line arguments which aren't described here, but should be fairly clear.

As shown in the configure usage above, there are configure.ac changes that allow separately dis/enabling USE_HTTPSRR and USE_ECH. If USE_ECH is enabled, then USE_HTTPSRR is forced. In both cases USE_DOH is required. (There may be some configuration conflicts available for the determined:-)

The main functional change, as you would expect, is in lib/vtls/openssl.c where an ECHConfig, if available from command line or DNS cache, is fed into the OpenSSL library via the new APIs implemented in our OpenSSL fork for that purpose. This code also implements the opportunistic (--ech true) or hard-fail (--ech hard) logic.

Other than that, the main additions are in lib/doh.c where we re-use dohprobe() to retrieve an HTTPS RR value for the target domain. If such a value is found, that is stored using a new store_https() function in a new field in the dohentry structure.

The qname for the DoH query is modified if the port number is not 443, as defined in the SVCB specification.

When the DoH process has worked, Curl_doh_is_resolved() now also returns the relevant HTTPS RR value data in the Curl_dns_entry structure. That is later accessed when the TLS session is being established, if ECH is enabled (from lib/vtls/openssl.c as described above).

Limitations

Things that need fixing, but that can probably be ignored for the moment:

• We could easily add code to make use of an alpn= value found in an HTTPS RR, passing that on to OpenSSL for use as the "inner" ALPN value, but have yet to do that.

Current limitations (more interesting than the above):

• Only the first HTTPS RR value retrieved is actually processed as described above, that could be extended in future, though picking the "right" HTTPS RR could be non-trivial if multiple RRs are published - matching IP address hints versus A/AAAA values might be a good basis for that. Last I checked though, browsers supporting ECH did not handle multiple HTTPS RRs well, though that needs re-checking as it has been a while.

• It is unclear how one should handle any IP address hints found in an HTTPS RR. It may be that a bit of consideration of how "multi-CDN" deployments might emerge would provide good answers there, but for now, it is not clear how best curl might handle those values when present in the DNS.

• The SVCB/HTTPS RR specification supports a new "CNAME at apex" indirection ("aliasMode") - the current code takes no account of that at all. One could envisage implementing the equivalent of following CNAMEs in such cases, but it is not clear if that'd be a good plan. (As of now, chrome browsers do not seem to have any support for that "aliasMode" and we've not checked Firefox for that recently.)

• We have not investigated what related changes or additions might be needed for applications using libcurl, as opposed to use of curl as a command line tool.

• We have not yet implemented tests as part of the usual curl test harness as doing so would seem to require re-implementing an ECH-enabled server as part of the curl test harness. For now, we have a ./tests/ech_test.sh script that attempts ECH with various test servers and with many combinations of the allowed command line options. While that is a useful test and has find issues, it is not comprehensive and we're not (as yet) sure what would be the right level of coverage. When running that script you should not have a $HOME/.curlrc file that affects ECH or some of the negative tests could produce spurious failures.

Building with cmake

To build with cmake, assuming our ECH-enabled OpenSSL is as before:

    cd $HOME/code
    git clone https://github.com/curl/curl
    cd curl
    mkdir build
    cd build
    cmake -DOPENSSL_ROOT_DIR=$HOME/code/openssl -DUSE_ECH=1 -DUSE_HTTPSRR=1 ..
    ...
    make
    ...
    [100%] Built target curl

The binary produced by the cmake build does not need any ECH-specific LD_LIBRARY_PATH setting.

boringssl build

BoringSSL is also supported by curl and also supports ECH, so to build with that, instead of our ECH-enabled OpenSSL:

    cd $HOME/code
    git clone https://boringssl.googlesource.com/boringssl
    cd boringssl
    cmake -DCMAKE_INSTALL_PREFIX:PATH=$HOME/code/boringssl/inst -DBUILD_SHARED_LIBS=1
    make
    ...
    make install

Then:

    cd $HOME/code
    git clone https://github.com/curl/curl
    cd curl
    autoreconf -fi
    LDFLAGS="-Wl,-rpath,$HOME/code/boringssl/inst/lib" ./configure --with-ssl=$HOME/code/boringssl/inst --enable-ech --enable-httpsrr
    ...lots of output...
    WARNING: ech ECH HTTPSRR enabled but marked EXPERIMENTAL. Use with caution!
    make

The boringssl APIs are fairly similar to those in our ECH-enabled OpenSSL fork, so code changes are also in lib/vtls/openssl.c, protected via #ifdef OPENSSL_IS_BORINGSSL and are mostly obvious API variations.

The boringssl APIs however do not support the --ech pn: command line variant as of now.

WolfSSL build

WolfSSL also supports ECH and can be used by curl, so here's how:

    cd $HOME/code
    git clone https://github.com/wolfSSL/wolfssl
    cd wolfssl
    ./autogen.sh
    ./configure --prefix=$HOME/code/wolfssl/inst --enable-ech --enable-debug --enable-opensslextra
    make
    make install

The install prefix (inst) in the above causes WolfSSL to be installed there and we seem to need that for the curl configure command to work out. The --enable-opensslextra turns out (after much faffing about;-) to be important or else we get build problems with curl below.

    cd $HOME/code
    git clone https://github.com/curl/curl
    cd curl
    autoreconf -fi
    ./configure --with-wolfssl=$HOME/code/wolfssl/inst --enable-ech --enable-httpsrr
    make

There are some known issues with the ECH implementation in WolfSSL:

• The main issue is that the client currently handles HelloRetryRequest incorrectly. HRR issue.) The HRR issue means that the client does not work for this ECH test web site and any other similarly configured sites.
• There is also an issue related to so-called middlebox compatibility mode. middlebox compatibility issue

Code changes to support WolfSSL

There are what seem like oddball differences:

• The DoH URL in$HOME/.curlrc can use "1.1.1.1" for OpenSSL but has to be "one.one.one.one" for WolfSSL. The latter works for both, so OK, we'll change to that.
• There seems to be some difference in CA databases too - the WolfSSL version does not like defo.ie, whereas the system and OpenSSL ones do. We can ignore that for our purposes via --insecure/-k but would need to fix for a real setup. (Browsers do like those certificates though.)

Then there are some functional code changes:

• tweak to configure.ac to check if WolfSSL has ECH or not
• added code to lib/vtls/wolfssl.c mirroring what's done in the OpenSSL equivalent above.
• WolfSSL does not support --ech false or the --ech pn: command line argument.

The lack of support for --ech false is because wolfSSL has decided to always at least GREASE if built to support ECH. In other words, GREASE is a compile time choice for wolfSSL, but a runtime choice for OpenSSL or boringssl. (Both are reasonable.)

Additional notes

Supporting ECH without DoH

All of the above only applies if DoH is being used. There should be a use-case for ECH when DoH is not used by curl - if a system stub resolver supports DoT or DoH, then, considering only ECH and the network threat model, it would make sense for curl to support ECH without curl itself using DoH. The author for example uses a combination of stubby+unbound as the system resolver listening on localhost:53, so would fit this use-case. That said, it is unclear if this is a niche that is worth trying to address. (The author is just as happy to let curl use DoH to talk to the same public recursive that stubby might use:-)

Assuming for the moment this is a use-case we'd like to support, then if DoH is not being used by curl, it is not clear at this time how to provide support for ECH. One option would seem to be to extend the c-ares library to support HTTPS RRs, but in that case it is not now clear whether such changes would be attractive to the c-ares maintainers, nor whether the "tag=value" extensibility inherent in the HTTPS/SVCB specification is a good match for the c-ares approach of defining structures specific to decoded answers for each supported RRtype. We're also not sure how many downstream curl deployments actually make use of the c-ares library, which would affect the utility of such changes. Another option might be to consider using some other generic DNS library that does support HTTPS RRs, but it is unclear if such a library could or would be used by all or almost all curl builds and downstream releases of curl.

Our current conclusion is that doing the above is likely best left until we have some experience with the "using DoH" approach, so we're going to punt on this for now.

Debugging

Just a note to self as remembering this is a nuisance:

LD_LIBRARY_PATH=$HOME/code/openssl:./lib/.libs gdb ./src/.libs/curl

Localhost testing

It can be useful to be able to run against a localhost OpenSSL s_server for testing. We have published instructions for such localhost tests in another repository. Once you have that set up, you can start a server and then run curl against that:

    cd $HOME/code/ech-dev-utils
    ./scripts/echsvr.sh -d
    ...

The echsvr.sh script supports many ECH-related options. Use echsvr.sh -h for details.

In another window:

    cd $HOME/code/curl/
    ./src/curl -vvv --insecure  --connect-to foo.example.com:8443:localhost:8443  --ech ecl:AD7+DQA6uwAgACBix2B78sX+EQhEbxMspDOc8Z3xVS5aQpYP0Cxpc2AWPAAEAAEAAQALZXhhbXBsZS5jb20AAA==

Automated use of retry_configs not supported so far...

As of now we have not added support for using retry_config handling in the application - for a command line tool, one can just use dig (or kdig) to get the HTTPS RR and pass the ECHConfigList from that on the command line, if needed, or one can access the value from command line output in verbose more and then re-use that in another invocation.

Both our OpenSSL fork and boringssl have APIs for both controlling GREASE and accessing and logging retry_configs, it seems WolfSSL has neither.