I occasionally do talks about curl. In these talks I often include a few slides that say something abut curl’s coverage and presence on different platforms. Mostly to boast of course, but also to help explain to the audience how curl has manged to reach its ten billion installations.
This is current incarnation of those seven slides in November 2022. I am of course also eager to get your feedback on the specific contents, especially if you miss details in them, that I should add so that my future curl presentations include more accurate data.
curl runs in all your devices
curl is used in (almost) every Internet-connected device out there, and I try to visualize that with this packed slide. Cars, servers, game consoles, medical devices, games, apps, operating systems, watches, robots, TVs, speakers, light-bulbs, freezers, printers, motorcycles, music instruments and more.
The intent being to show with images that it runs in quite a few devices.
28 transfer protocols
More strictly speaking these are the 28 URL schemes curl supports right now, including in experimental builds. The image tries to put them in a sort of hierarchical way so that you can see what underlying protocol that is used for them: TCP, SSH, TLS, QUIC etc.
Volunteers have created and maintain libcurl bindings for at least these 60 different environments, making it possible for developers to access and use libcurl powers from virtually every programming language you can think of.
37 third party dependencies
curl’s modular design and ability letting the developer who builds curl to mix and match features and what particular third party dependencies to use, makes it possible to craft exactly the curl builds you want. Device manufacturers get the combination they like for exactly their purposes and needs.
89 operating systems
This list has been worked on and bounced around several times between friends and it always brings out a few questions and people like to argue with me about a few entries I include and about a few entries I do not include. The problem is both that there is not a clear defining line between the definition of an operating systems and a distribution or a different running environment and sometimes it is just branding differences separating X from Y. With a “flexible” attitude to the definition of operating systems, this is the current collection of no less than 89 individual ones on which curl runs or has run on:
24 CPU architectures
Older versions of this slide used to have x86-64 as a separate one, but I think we have concluded that a large amount of the architectures have separate 64 bit versions so if I were to keep x86-64 I should also include a lot of other 64 bit versions so I removed the x86-64 from the slide. Maybe I should rather go the other way and add all the 64 bit version as separate architectures?
Anyway, curl has been made to run on virtually all modern or semi-modern 32 bit or larger CPU architectures we can find:
I admit it. I include this slide in my presentation and in this blog post because it feels like the ultimate show-off. curl was used in the mars 2020 helicopter mission.
The curl project builds on foundations that started in late 1996 with the tool named httpget.
ANSI C became known as C89
In 1996 there were not too many good alternatives for making a small and efficient command line tool for doing Internet transfers. I am not saying that C was the only available language, but for me the choice was easy and frankly I did not even think about any other languages when this journey started. We called the C flavor “ANSI C” back then, as compared to the K&R “old style” C. The ANSI C version would later be renamed to C89 (confusingly enough it is also sometimes known as C90).
In the year 2000 we introduced libcurl, the library that provides Internet transfer super powers to whoever wants it. This made the choice of using C even better. C made it possible for us to provide a stable API/ABI without problems – something not even C++ could offer at the time. It was also a reasonably portable language that made it possible for us to bring curl and libcurl to virtually all modern operating systems.
As I wanted curl and libcurl to be system level options and I aimed for the widest possible adoption, they could not be written in any of the higher level languages like Perl, Python or similar. That would make them too big and require too much “extra baggage”.
I am convinced that the use of (conservative) C for curl is a key factor to its success and its ability to get used “everywhere”.
C99 was published in (surprise!) 1999 but the adoption in compilers took a long time and it remained a blocker for adoption for us. We want curl available “everywhere” so as long some of the major compilers did not support C99 we did not even consider switching C flavor, as it would risk hamper curl adoption.
The slowest of the “big compilers” to adopt C99 was the Microsoft Visual C++ compiler, which did not adopt it properly until 2015 and added more compliance in 2019. A large number of our users/developers are still stuck on older MSVC versions so not even all users of this compiler suite can build C99 programs even today, in late 2022.
C11, C17 and beyond
Meanwhile, the ISO C Working Group continue to crank out updates to the C language. C11 shipped, C17 came and now they are working on the C2x pending version, presumed to end up called C23.
Bump the requirement for curl?
We are aware that other widely popular C projects are moving forward and have raised their requirements to C99 or beyond. Like the Linux kernel, the git project and more.
The discussion about bumping C flavor has been brought up on the libcurl mailing list as well, in particular as we are already planning a version 8 release to happen in the spring of 2023 so in theory it could be a good moment to make some changes like this.
What C99 features would improve a project like curl? The most interesting parts of C99 that could impact curl code that I could think of are:
__func__ predefined identifer
boolean type in <stdbool.h>
designated struct initializers
empty macro arguments
extended integer types in <inttypes.h> and <stdint.h>
flexible array members (zero size arrays)
integer constant type rules
mixed declarations and code
the long long type and library functions
the snprintf() family of functions
trailing comma allowed in enum declaration
So sure, there are lots of cool things we could use. But do we need them?
For several of the features above, we already have decent and functional replacements. Several of the features don’t matter. The rest risk becoming distractions.
Opening up for C99 without conditions in curl code would risk opening the flood gates for people rewriting things, so we would have to go gently and open up for allowing new C99 features slowly. That is also how the git project does it. A challenge with that approach, is that it is hard to verify which features that are allowed vs used as existing tooling normally don’t have that resolution.
The question has also been asked that if we consider bumping the requirement, should we then not bump it to C11 at once instead of staying at C99?
Ultimately, not a single person has yet been able to clearly articulate what benefits such a C flavor requirement bump would provide for the curl project. We mostly see a risk that we all get caught in rather irrelevant discussions and changes that perhaps will not actually bring the project forward very much. Neither in features nor in quality/security.
I think there are still much better things to do and much more worthwhile efforts to spend our energy on that could actually improve the project and bring it forward.
Like improving the test suite, increasing test coverage, making sure more code is exercised by the fuzzers.
A minor requirement change
We have decided that starting with curl 8, we will require that the compiler supports a 64 bit data type. This is not something that existed in the original C89 version but was introduced in C99. However, there is no longer any modern compiler around that does not support this.
This is a way to allow us to stop caring about those odd platforms and write code and checks for when the large types are not very large. It is hard to verify that code nowadays since virtually nobody actually uses such compilers/systems.
Maybe this is the way we can continue to adapt to and use specific post C89 features going forward. By cherry-picking them one by one and adapting to them slowly over time.
It is not a no to C99 forever
I am sure we will bring up this topic for discussion again in the future. We have not closed the door forever or written anything in stone. We have only decided that for the moment we have not been persuaded to switch. Maybe we will in a future.
We do not consider switching or rewriting curl into any other language.
In the curl project, one of the holiest and most sacred rules is:
we do not break the API or ABI
Everything else is a matter of discussion.
More features all the time
We keep adding features and we do improvements at a rather high pace. So much that we actually rarely do a release without introducing something new.
To be able to add features and to keep changing curl and making sure that it keeps up with the world around it and that it provides the features and the abilities that a world of Internet transfers needs, we need to make sure that the internals are written correctly. And by correctly, I mean in a way that allows us to extend and change curl when we want to – that doesn’t break the ABIs nor the tests.
curl is old and choices sometimes need to be reconsidered. Over the years we have refactored and changed the curl internals and design quite drastically several times. Thanks to an extensive test suite and a library API that was designed from the start to hide most internal choices, this has been possible to do without being visible to users. The upside has been that the internals have become easier to maintain and to extend with more features.
This time, we are again on a mission to extend the curl feature set as I blogged about recently, and this time we have Stefan Eissing on board to do it.
So, without changing any API or breaking the ABI and having the large set of test cases remain working in the many CI jobs we have, Stefan introduced a new internal concept for curl: connection filters.
We call them filters but they could also be seen as layers or maybe even domino pieces. Each filter is a piece of network logic and the idea is that we can chain them together at run-time to create protocol cakes (my word). curl can connect to a HTTP proxy, do TLS and speak HTTP/2 over that. That makes three separate filters put together.
Adding for example TLS to the proxy would just be inserting a filter in the right place in the chain, while using the filter-chain is done the same way no matter the filter chain length and independently of which exact filters it consists of.
The previous logic, before the filters, was a more like a vast number of conditional flag checks done in the right order. This new system reduces the amount of conditional checks and it also moves code for handling the different network filters into more localized and compartmentalized functions.
More protocol combos
In addition to the more localized code for specifics features, this new concept more notably makes it easier to build new protocol layer combinations. Adding support for HTTP/2 to the proxy for example, should now ideally be a matter of adding a filter the right way and the transfer pipeline should otherwise “just work”.
Not everything internally is yet converted to filters even if we have merged the first large pull request. Stefan now works on getting more curl code to use this concept before he can get into the actual protocol changes lined up for him.
The filters do not impact transfer performance, I/O works the same as before.
If you long for more technical details and explanations about this, maybe to be able to dig into the curl source code yourself, then an excellent starting-point is the document in the curl source made for this purpose CONNECTION-FILTERS.md.
When setting up a TLS or QUIC connection, a client like curl needs a CA store in order to verify the certificate(s) the server provides in the TLS handshake.
A CA store is a fancy name for a number of certificates. Certificates for the Certificate Authorities (CAs) that a TLS client trusts. On the curl website, we offer a PEM version of the CA store that Mozilla maintains, for download. This set currently contains 142 certificates and while the exact amount vary a little over time, it has been more than a hundred for many years. A fair amount. And there is nothing in the pipe that will bring down the number significantly anytime soon, to my knowledge. These 142 certificates make up a file that is exactly 225,403 bytes. 1587 bytes per certificate on average.
Load and parse
When setting up a TLS connection, the 142 certificates need to be loaded from the external file into memory and parsed so that the server’s certificate can be verified. So that curl knows that the server it has connected to is indeed the correct server and not a man in the middle, an impostor.
This procedure is a rather costly one, in terms of CPU cycles needed.
A classic approach to avoid heavy work is to cache the results from a previous use to be able to reuse them again. Starting in curl 7.87.0 curl introduces a CA store cache.
Now, curl can keep the loaded and parsed CA store in memory associated with the handle and then subsequent requests can avoid re-loading and re-parsing the CA data when new connections are created – if they use the same CA store of course. The performance gain in doing this shortcut can be enormous. After all, most transfers are done using the same single CA store.
To quote the numbers Michael Drake presented in the pull request for this new feature. He measured number of instructions to load and render a particular web page from BBC with the NetSurf browser (which obviously is using libcurl for its HTTPS transfers). With and without this cache.
CA store cache
Total instruction fetch cost
I think a reduction to one fifth of the original cost is significant.
Converted into a little graph they compare like this (smaller is better):
But even in simpler applications and curl command lines this caching should have a measurable impact as soon as multiple TLS connections are done using the same handle. An extremely common usage pattern.
Keeping the data around after use potentially changes the behavior a little, but the huge performance gain made us decide to still do this by default. We compensate this a little by setting the default life-time to 24 hours, so applications that keep handles alive for a very long time will still get the cache flushed and read from file again every day.
The CA store is typically not updated more frequently than once every few months or weeks.
This is a new option for libcurl that allows applications to tweak the life-time and CA cache behavior for when the default as described above is not enough.
This CA cache system is so far only supported when curl is built to use OpenSSL or one of its forks. I hope others will get inspired and bring this support for other TLS backends as well as we go forward.
CA cache support for curl was authored by Michael Drake. Thanks!
curl offered the -d / --data option already in its first release back in 1998. curl 4.0. A trusted old friend.
curl also has some companion versions of this option that work slightly differently, but they all have the common feature that they append data to the the request body. Put simply: with these options users construct the body contents to POST. Very useful and powerful. Still today one of the most commonly used curl options, for apparent reasons.
A few years into curl’s lifetime, in 2001, we introduced the -G / --get option. This option let you use -d to create a data set, but the data is not sent as a POST body anymore but is instead converted to a query string and used in a GET request.
This would make curl send a GET request to this URL: https://example.com/?name=mrsmith&color=blue
The “query” is the part of the URL that sits on the right side of the question mark (but before the fragment that if it exists starts with the first # following the question mark).
In 2008 we added --data-urlencode which made it even easier for users and scripts to use these options correctly as now curl itself can URL-encode the given data instead of relying on the user to do it. Previously, script authors would have to do that encoding before passing the data to curl which was tedious and error prone. This feature also works in combination with -G of course.
How about both?
The -d options family make a POST. The -G converts it to a GET.
If you want convenient curl command line options to both make content to send in the POST body andto create query parameters in the URL you were however out of luck. You would then have to go back to use -d but handcraft and encode the query parameters “manually”.
Until curl 7.87.0. Due to ship on December 21, 2022. (this commit)
--url-query is your new friend
This is curl’s 249th command line option and it lets you append parameters to the query part of the given URL, using the same syntax as --data-urlencode uses.
Using this, a user can now conveniently create a POST request body and at the same time add a set of query parameters for the URL which the request uses.
A basic example that sends the same POST body and URL query:
I told you it uses the data-urlencode syntax, but let me remind you how that works. You use --url-query [data] where [data] can be provided using these different ways:
This will make curl URL-encode the content and pass that on. Just be careful so that the content does not contain any = or @ symbols, as that will then make the syntax match one of the other cases below!
This will make curl URL-encode the content and pass that on. The preceding = symbol is not included in the data.
This will make curl URL-encode the content part and pass that on. Note that the name part is expected to be URL-encoded already.
This will make curl load data from the given file (including any newlines), URL-encode that data and pass it on in the POST.
This will make curl load data from the given file (including any newlines), URL-encode that data and pass it on in the POST. The name part gets an equal sign appended, resulting in name=urlencoded-file-content. Note that the name is expected to be URL-encoded already.
The data is provided as-is unencoded
For each new --url-query, curl will insert an ampersand (&) between the parts it adds to the query.
This new friend we call --url-query makes -G rather pointless, as this is a more powerful option that does everything -G ever did and a lot more. We will of course still keep -G supported and working. Because that is how we work.
A boring fact of life is that new versions of curl trickle out into the world rather slowly to ordinary users. Because of this, we can be certain that scripts and users all over will need to keep using -G for yet another undefined period of time.
Finally: remember that if you want curl to show you what it sends in a POST request, the normal -v / --verbose does not suffice as it will not show you the request body. You then rather need to use --trace or --trace-ascii.
The last day of this edition of the HTTP workshop. Thursday November 3, 2022. A half day only. Many participants at the Workshop are going to continue their UK adventure and attend the IETF 115 in London next week.
We started off the day with a deep dive into connection details. How to make connections for HTTP – in particular on mobile devices. How to decide which IP to use, racing connections, timeouts, when to consider a connection attempt “done” (ie after the TCP SYNACK or after the TLS handshake is complete). QUIC vs TCP vs TLS and early data. IPv4 vs IPv6.
ECH. On testing, how it might work, concerns. Statistics are lies. What is the success expectancy for this and what might be the explanations for failures. What tests should be done and what answers about ECH in the wild would we like to get answered going forward?
Mike Bishop brought up an old favorite subject: a way for a server to provide hints to a client on how to get its content from another host. Is it time to revive this idea? Blind caching the idea and concept was called in an old IETF 95 presentation by Martin Thomson. A host that might be closer the client or faster etc or simply that has the content in question could be used for providing content.
Mark Nottingham addressed the topic of HTTP (interop) testing. Ideas were raised around some kind of generic infrastructure for doing tests for HTTP implementations. I think people were generally positive but I figure time will tell what if anything will actually happen with this.
What’s next for HTTP? Questions were asked to the room, mostly of the hypothetical kind, but I don’t think anyone actually had the nerve to suggest they actually have any clear ideas for that at this moment. At least not in this way. HTTP/4 has been mentioned several times during these days, but then almost always in a joke context.
This is how the workshop ends for this time. Super fun, super informative, packed with information with awesome people. One of these events that give me motivation and a warm fuzzy feeling inside for an extended period of time into the future.
Thanks everyone for your excellence. Thank you organizers for a fine setup!
During the discussions today it was again noticeable that apart from some specific individuals we also lack people in the room from prominent “players” in this area such as Chrome developers or humans with closer knowledge of how larger content deliverer work such as Netflix. Or what about a search engine person?
Years ago Chrome was represented well and the Apple side of the world was weak, but the situation seems to have been totally reversed by now.
Mike Bishop talked about the complexities of current internet. Redirects before, during, after. HTTPS records. Alt-svc. Alt-SvcB. Use the HTTPS record for that alternative name.
This presentation triggered a long discussion on how to do things, how things could be done in a future and how the different TTLs in this scenario should or could interact. How to do multi-CDN, how to interact with DNS and what happens if a CDN wants to disable QUIC?
A very long discussion that mostly took us all back to square one in the end. The alt-svcb proposal as is.
Alan Frindell talked about HTTP priorities. What it is (only within connections, it has to be more than one thing and it something goes faster something else needs to go slower, re-prioritization takes a RTT/2 for a client to take effect), when to prioritize (as late as possible, for h2 just before committing to TCP).
Prioritizing images in Meta’s apps. On screen, close, not close. Guess first, then change priority once it knows better. An experiment going on is doing progressive images. Change priority of image transfers once they have received a certain amount. Results and conclusions are pending.
Priority within a single video. An important clue to successful priority handling seems to be more content awareness in the server side. By knowing what is being delivered, the server can make better decisions without the client necessarily having to say anything.
Who is doing SF libraries and APIs? Do we need an SF schema?
SF compression? Mark’s experiment makes it pretty much on par with HPACK.
Binary SF. Takes ~40% less time parsing in Mark’s experiment. ~10% larger in size (for now).
The was expressed interest in continuing this experimenting going forward. Reducing the CPU time for header parsing is considered valuable.
Sebastian Poreba talked “Networking in your pocket” about the challenges and architectures of mobile phone and smartwatch networking. Connection migration is an important property of QUIC that is attractive in the mobile world. H3 is good.
A challenge is to keep the modem activity off as much as possible and do network activities when the modem is already on.
My very important duties during these days also involved spending time in pubs and drinking beers with this awesome group of people. This not only delayed my blog post publish times a little, but it might also have introduced an ever so slight level of “haziness” into the process and maybe, just maybe, my recollection of the many details from the day is not exactly as detailed as it could otherwise have been.
That’s just the inevitable result of me sacrificing myself for the team. I did it for us all. You’re welcome.
Another full day
Another day fully packed with HTTP details from early to late. From 9am in the morning until 11pm in the evening. I’m having a great time. Tomorrow is the last day. I’ll let you know what happens then.
The HTTP Workshop is an occasional gathering of HTTP experts and other interested parties to discuss the Web’s foundational protocol.
The fifth HTTP Workshop is a three day event that takes place in Oxford, UK. I’m happy to say that I am attending this one as well, as I have all the previous occasions. This is now more than seven years since the first one.
A lot of the people attending this year have attended previous workshops, but in a lot of cases we are now employed by other companies then when we attended our first workshops. Almost thirty HTTP stack implementers, experts and spec authors in a room.
There was a few newcomers and some regulars are absent (and missed). Unfortunately, we also maintain the lack of diversity from previous years. We are of course aware of this, but have still failed to improve things in this area very much.
All the people gather in the same room. A person talks briefly on a specific topic and then we have a free-form discussion about it. When I write this, the slides from today’s presentations have not yet been made available so I cannot link them here. I will add those later.
Lucas Pardue started the morning with a presentation and discussion around logging, tooling and the difficulty of debugging modern HTTP setups. With binary protocols doing streams and QUIC, qlog and qvis are in many ways the only available options but they are not supported by everything and there are gaps in what they offer. Room for improvements.
Anne van Kesteren showed a proposal from Domenic Denicola for a No-Vary-Search response header. An attempt to specify a way for HTTP servers to tell clients (and proxies) that parts of the query string produces the same response/resource and should be cached and treated as the same.
An issue that is more complicated than what if first might seem. The proposal has some fans in the room but I think the general consensus was that it is a bad name for the header…
HTTP versions. HTTP/3 is used in over a third of the requests both as seen by Cloudflare server-side and measured in Firefox telemetry client-side. Extensible and functional protocols. Nobody is talking about or seeing a point in discussing about a HTTP/4.
WebSocket over h2/h3. There does not seem to be any particular usage and nobody mentioned any strong reason or desired to change the status. Web Transport is probably what instead is going to be the future.
Frames. Discussions around the use and non-use of the origin frame. Note widely used. Could help to avoid extra “SNI leakage” and extra connections.
Anne then took a second round in front of the room and questioned us on the topic of cookies. Or perhaps more specifically about details in the spec and how to possible change the spec going forward. At least one person in the room insisted fairly strongly than any such restructures of said documents should be done after the ongoing 6265bis work is done.
A company very generously sponsored a group dinner for us in the evening and I had a great time. I was asked to not reveal the name of said company, but I can tell you that a lot of the conversations at the table, at least in the area where I was parked, kept up the theme from the day and were HTTP oriented. Including oblivious HTTP, IPv4 formatting allowed in URLs and why IP addresses should not be put in the SNI field. Like any good conversion among friends.
When doing HTTP(S) transfers, libcurl might erroneously use the read callback (CURLOPT_READFUNCTION) to ask for data to send, even when the CURLOPT_POSTFIELDS option has been set, if the same handle previously was used to issue a PUT request which used that callback.
This flaw may surprise the application and cause it to misbehave and either send off the wrong data or use memory after free or similar in the subsequent POST request.
The problem exists in the logic for a reused handle when it is changed from a PUT to a POST.
curl can be told to parse a .netrc file for credentials. If that file ends in a line with consecutive non-white space letters and no newline, curl could read past the end of the stack-based buffer, and if the read works, write a zero byte possibly beyond its boundary.
This will in most cases cause a segfault or similar, but circumstances might also cause different outcomes.
If a malicious user can provide a custom .netrc file to an application or otherwise affect its contents, this flaw could be used as denial-of-service.
If curl is told to use an HTTP proxy for a transfer with a non-HTTP(S) URL, it sets up the connection to the remote server by issuing a CONNECT request to the proxy, and then tunnels the rest of protocol through.
An HTTP proxy might refuse this request (HTTP proxies often only allow outgoing connections to specific port numbers, like 443 for HTTPS) and instead return a non-200 response code to the client.
Due to flaws in the error/cleanup handling, this could trigger a double-free in curl if one of the following schemes were used in the URL for the transfer: dict, gopher, gophers, ldap, ldaps, rtmp, rtmps, telnet
curl’s HSTS check could be bypassed to trick it to keep using HTTP.
Using its HSTS support, curl can be instructed to use HTTPS directly instead of using an insecure clear-text HTTP step even when HTTP is provided in the URL. This mechanism could be bypassed if the host name in the given URL uses IDN characters that get replaced to ASCII counterparts as part of the IDN conversion. Like using the character UTF-8 U+3002 (IDEOGRAPHIC FULL STOP) instead of the common ASCII full stop (U+002E) ..
Like this: http://curl?se?
This time around we add one and we remove one.
NPN support removed
curl no longer supports using NPN for negotiating HTTP/2. The standard way for doing this has been ALPN for a long time and the browsers removed their support for NPN several years ago.
There is an experimental WebSocket API included in this release. It comes in the form of three new functions and a new setopt option to control behavior. The possibly best introduction to this new API is in everything curl.
I am very interested in feedback on the API.
Here some of the fixed issues from this cycle that I think are especially worthy to highlight.
aws_sigv4 header computation
The sigv4 code got a significant overhaul and should now do much better than before. This is a fairly complicated setup and there are more improvements coming for future releases.
curl man page details multi-use for each option
Every command line option is documented in its own file, which is then used as input when the huge curl.1 man page is generated. Starting now, each such file needs to specify how the option functions when specified more than once. So from now on, this information is mentioned in the man page for all supported options.
deprecate builds with small curl_off_t
Starting in this release, we deprecate support for building curl for systems without 64 bit data types. Those systems are extremely rare this days and we believe it makes sense to finally simplify a few internals when it hurts virtually no one. This is still only deprecated so users can still build on such systems for a short while longer if they really want to.
the ngtcp2 configure option defaults to ‘no’
You need to explicitly ask for ngtcp2 to be enabled in the build.
reject cookie names or content with TAB characters
for builds with gcc + want warnings, set gnu89 standard
Just to make better sure we maintain compatibility.
use -O2 as default optimize for clang in configure
It was just a mistake that it did not already do this.
warn for –ssl use, considered insecure
To better highlight for users that this option merely suggests for curl that it should use TLS for the protocol, while --ssl-reqd is the one that requires TLS.
ctype functions converted to macros-only
We replaced the entire function family with macros.
100+ documentation spellfixes
After a massive effort and new CI jobs, we now regularly run a spellcheck on most man pages and as a result we fixed lots of typos and we should now be able to better maintain properly spelled documentation going forward.
make nghttp2 less picky about field whitespace in HTTP/2
If built with a new enough nghttp2 library, curl will now ask it to be less picky about trailing white space after header fields. The protocol spec says they should cause failure, but they are simply too prevalent in live servers responses for this to be a sensible behavior by curl.
use the URL-decoded user name for .netrc parsing
This regression made curl not URL decode the user name provided in a URL properly when it later used a .netrc file to find the corresponding password.
When asking curl to use IPv4-only for transfers, curl now only resolves IPv4 names. Out in the wide world there is a significant share of systems causing problems when asking for AAAA addresses so having this option to avoid them is needed.
schannel: when importing PFX, disable key persistence
Some operations when using the Schannel backend caused leftover files on disk afterward. It really makes you wonder who ever thought designing such a thing was a good idea, but now curl no longer triggers this effect.
add and use Curl_timestrcmp
curl now uses this new constant-time function when comparing secrets in the library in an attempt to make it even less likely for an outsider to be able to use timing as a feedback as to how closely a guessed user name or password match the secret ones.
curl: prevent over-queuing in parallel mode
The command line tool would too eagerly create and queue up pending transfers in parallel mode, making a command line with millions of transfers easily use ridiculous amounts of memory.
url parser: extract scheme better when not guessing
A URL has a scheme and we can use that fact to detect it better and more reliable when not using the scheme-guessing mode.
fix parsing URL without slash with CURLU_URLENCODE
When the URL encode option is set when parsing a URL, the parser would not always properly handle URLs with queries that also lacked a slash in the path. Like https://example.com?moo.
url parser: leaner with fewer allocs
The URL parser is now a few percent faster and makes significantly fewer memory allocations and it uses less memory in total.
url parser: reject more bad characters from the host name field
Another step on the journey of making the parser stricter.
wolfSSL: fix session management bug
The session-id cache handling could trigger a crash due to a missing reference counter.
We have several pull-requests in the pipe that will add changes to trigger a minor number bump.
We are planning to remove the following features in a future-:
support for systems without 64 bit data type
support for the NSS TLS library
If you depend on one of those features, yell at us on the mailing list!
I am happy to announce that curl receives funding from The Sovereign Tech Fund. This funding is directed towards three specific projects that we have identified as interesting and worthwhile to push forward as ways to improve curl and the life of curl users.
This “investment” will fund two developers to work on curl over a period of six months: Stefan Eissing and myself. The three projects are explained at some detail below. Of course everyone and anyone is welcome to join in and help out with these projects. Everything will be done in the open, as usual.
At the end of the period, we will produce some kind of report or summary of how things turned out.
The three projects we are getting funded have been especially created and crafted (by me) to be good solid projects that we really want to see done. This funding is different than many others we have gotten over the years in that we got to decide and plan what we wanted done. These are things that are meant to improve curl as a project and to generally make Internet transfers better and more powerful for a vast amount of users.
Project 1 – known bugs cleanup
The curl project currently lists 120+ items as known bugs (up from 77 just two years ago).
The items in that list are reported problems that were recognized as problems at the time of their submission but as nobody worked on the issues at the time they were added to this list. The list includes everything from smaller irks up to big things that will either take a long time to fix or be (almost) impossible to address.
There is a good chance that the list will be extended during the project period just because some new bugs fall into the description mentioned above.
This project is an effort to go through as many as possible to make sure they are correctly categorized/described and work on fixing the issues or whatever is necessary to get them off the list.
The process would entail an initial proper research round to extend the description and increase the understanding of each entry, followed by a rough assessment of the amount of work it would take to fix them. Possibly with a 1 (easy) to 5 (extreme) scale.
The action would then be to address the issues, possibly in an easy to hard order. Addressing the issues could be to fix the code to remove the issue, dismiss it as not actually intended to work or document it as not working or even moving it over to the TODO document if it is more of a good idea for the future.
The goal being to reduce the list to zero entries and thus polish off numerous rough corners and annoyances in the project.
This will be done by Daniel Stenberg over a period of 6 months.
Project 2 – HTTP/3
Make HTTP/3 release-ready.
curl features experimental HTTP/3 and QUIC support already since a few years back, but there are several details still lacking:
proper multiplexing: doing multiple transfers to the same host should be able to reuse an existing connection and multiplex over that, just as curl already does when using HTTP/2
HTTP/3 support for the test suite (and CI jobs) need to be done for us to be able to consider the support release ready. Cooperation can be had with QUIC libraries such as ngtcp2 to consider where/how some of the testing is best performed.
considerations for 0-RTT connection establishments (if anything needs to be done)
support for early data: to send off the HTTP request to the server faster.
connection migration:, a QUIC feature that allows a server to move over a live connection from one server to another without disruption
fallback to h1/h2 if the QUIC connection fails. The failure rate for QUIC connections are still in the 3-7% rate generally, so having a good fallback mechanism or documented for how applications can go back to an older HTTP version instead, is important.
HTTPS RR. This fairly new DNS resource record might contain information about the target server’s support for HTTP/3. If such a record is provided, curl can avoid superfluous round-trips to get the Alt-Svc header and rather connect directly to the HTTP/3 server.
All features and changes need to be documented. Functionality needs to be verified by test cases. Interop with real world servers is of course implied and assumed.
Stefan Eissing will spend 4 months on this project.
Project 3 – HTTP/2 over proxy
curl has provided support for doing network transfers via HTTP proxies since decades, and this is a very commonly used feature and network setup.
curl however only supports using HTTP version 1 over proxies. This makes applications less effective as it sometimes leads to many more TCP connections being used than otherwise would be necessary if HTTP/2 could be used. In particular when applications behind a proxy operate against many different hosts on the other side of the proxy.
It can also be noted that in many enterprise setups, this kind of HTTP proxy is used for all kinds of network operations through the use of the CONNECT method, so this functionality is not limited to plain HTTP(S), it should work for all TCP based protocols libcurl supports and which already work over HTTP/1 proxies.
The project will require adding new options to enable this functionality, to both the library and the command line tool. With associated documentation.
It will require creating or extending the HTTP/2 support in the curl test suite so that this new functionality can be verified and proven to work at a satisfactory level.
Considerations must be taken so that this work does not close the door for future extending this to support HTTP/3 over proxy. Time permitting, work should be taken to pave the road for that or even perhaps gently start the work to support that as well.
Stefan Eissing will spend 2 months on this project.
I hope and presume that the results of these projects will appear as a stream of pull-requests for curl that will be done and managed through-out the project period and not saved up to the end or anything. The review, test and merge process of these pull requests will follow our normal and standard project guidelines and procedures.
The projects are fully packed and (over) ambitious. There is a high risk that we will not be able to complete all the details for these projects within the time frame. But we will try.