Category Archives: Technology

Really everything related to technology

6 hours of bliss

December 23, 2016 Daniel Stenberg 4 Comments

I sent out the release announcement for curl 7.52.0 exactly 07:59 in the morning of December 21, 2016. A Wednesday. We typically release curl on Wednesdays out of old habit. It is a good release day.

curl 7.52.0 was just as any other release. Perhaps with a slightly larger set of new features than what’s typical for us. We introduce TLS 1.3 support, we now provide HTTPS-proxy support and the command line tool has this option called –fail-early that I think users will start to appreciate once they start to discover it. We also announced three fixed security vulnerabilities. And some other good things.

I pushed the code to git, signed and uploaded the tarballs, I updated the info on the web site and I sent off that release announcement email and I felt good. Release-time good. That short feeling of relief and starting over on a new slate that I often experience these release days. Release days make me happy.

Any bets?

It is not unusual for someone to find a bug really fast after a release has shipped. As I was feeling good, I had to joke in the #curl IRC channel (42 minutes after that email):

08:41 <bagder> any bets on when the first bug report on the new release shows up? =)

Hours passed and maybe, just maybe there was not going to be any quick bugs filed on this release?

But of course. I wouldn’t write this blog post if it all had been nice and dandy. At 14:03, I got the email. 6 hours and 4 minutes since I wrote the 7.52.0 announcement email.

The email was addressed to the curl project security email list and included a very short patch and explanation how the existing code is wrong and needs “this fix” to work correctly. And it was entirely correct!

Now I didn’t feel that sense of happiness anymore. For some reason it was now completely gone and instead I felt something that involved sensations like rage, embarrassment and general tiredness. How the [beep] could this slip through like this?

I’ve done releases in the past that were broken to various extents but this is a sort of a new record and an unprecedented event. Enough time had passed that I couldn’t just yank the package from the download page either. I had to take it through the correct procedures.

What happened?

As part of a general code cleanup during this last development round, I changed all the internals to use a proper internal API to get random data and if libcurl is built with a TLS library it uses its provided API to get secure and safe random data. As a move to improve our use of random internally. We use this internal API for getting the nonce in authentication mechanisms such as Digest and NTLM and also for generating the boundary string in HTTP multipart formposts and more. (It is not used for any TLS or SSH level protocol stuff though.)

I did the largest part of the random overhaul of this in commit f682156a4f, just a little over a month ago.

Of course I made sure that all test cases kept working and there were no valgrind reports or anything, the code didn’t cause any compiler warnings. It did not generate any reports in the many clang-analyzer or Coverity static code analyzer runs we’ve done since. We run clang-analyzer daily and Coverity perhaps weekly.

But there’s a valgrind report just here!

Kamil Dudka, who sent the 14:03 email, got a valgrind error and that’s what set him off – but how come he got that and I didn’t?

The explanation consists of the following two conditions that together worked to hide the problem for us quite successfully:

I (and I suppose several of the other curl hackers) usually build curl and libcurl “debug enabled”. This allows me to run more tests, do more diagnostics and debug it easier when I run into problems. It also provides a system with “fake random” so that we can actually verify that functions that otherwise use real random values generate the correct output when given a known random value… and yeah, this debug system prevented valgrind from detecting any problem!
In the curl test suite we once had a problem with valgrind generating reports on third party libraries etc which then ended up as false positives. We then introduced a “valgrind report parser” that would detect if the report concerns curl or something else. It turns out this parser doesn’t detect the errors if curl is compiled without the cc’s -g command line option. And of course… curl and libcurl both build without -g by default!

The patch?

The vulnerable function basically uses this simple prototype. It is meant to get an “int” worth of random value stored in the buffer ‘rnd’ points to. That’s 4 bytes.

randit(struct Curl_easy *data, unsigned int *rnd)

But due to circumstances I can’t explain on anything other than my sloppy programming, I managed to write the function store random value in the actual pointer instead of the buffer it points to. So when the function returns, there’s nothing stored in the buffer. No 4 bytes of random. Just the uninitialized value of whatever happened to be there, on the stack.

The patch that fixes this problem looks like this (with some names shortened to simplify but keep the idea):

- res = random(data, (char *)&rnd, sizeof(rnd));
+ res = random(data, (char *)rnd, sizeof(*rnd));

So yeah. I introduced this security flaw in 7.52.0. We had it fixed in 7.52.1, released roughly 48 hours later.

(I really do not need comments on what other languages that wouldn’t have allowed this mistake or otherwise would’ve brought us world peace a long time ago.)

Make it not happen again

The primary way to make this same mistake not happen again easily, is that I’m removing the valgrind report parsing function from the test suite and we will now instead assume that valgrind reports will be legitimate and if not, work on suppressing the false positives in a better way.

References

This flaw is officially known as CVE-2016-9594

The real commit that fixed this problem is here, or as stand-alone patch.

The full security advisory for this flaw is here: https://curl.haxx.se/docs/adv_20161223.html

Facepalm photo by Alex E. Proimos.

xkcd: 221

cURL and libcurl, Mozilla, Security

curl security audit

November 23, 2016 Daniel Stenberg 6 Comments

“the overall impression of the state of security and robustness
of the cURL library was positive.”

I asked for, and we were granted a security audit of curl from the Mozilla Secure Open Source program a while ago. This was done by Mozilla getting a 3rd party company involved to do the job and footing the bill for it. The auditing company is called Cure53.

I applied for the security audit because I feel that we’ve had some security related issues lately and I’ve had the feeling that we might be missing something so it would be really good to get some experts’ eyes on the code. Also, as curl is one of the most used software components in the world a serious problem in curl could have a serious impact on tools, devices and applications everywhere. We don’t want that to happen.

Scans and tests and all

We run static analyzers on the code frequently with a zero warnings tolerance. The daily clang-analyzer scan hasn’t found a problem in a long time and the Coverity once-every-few-weeks occasionally finds something suspicious but we always fix those immediately.

We have thousands of tests and unit tests that we run non-stop on the code on multiple platforms running multiple build combinations. We also use valgrind when running tests to verify memory use and check for potential memory leaks.

Secrecy

The audit itself. The report and the work on fixing the issues were all done on closed mailing lists without revealing to the world what was really going on. All as our fine security process describes.

There are several downsides with fixing things secretly. One of the primary ones is that we get much fewer eyes on the fixes and there aren’t that many people involved when discussing solutions or approaches to the issues at hand. Another is that our test infrastructure is made for and runs only public code so the code can’t really be fully tested until it is merged into the public git repository.

The report

We got the report on September 23, 2016 and it certainly gave us a lot of work.

The audit report has now been made public and is a very interesting work if you’re into security, C code and curl hacking. I find the report very clear, well written and it spells out each problem very accurately and even shows proof of concept code snippets and exploit examples to drive the points home.

Quoted from the report intro:

As for the approach, the test was rooted in the public availability of the source code belonging to the cURL software and the investigation involved five testers of the Cure53 team. The tool was tested over the course of twenty days in August and September of 2016 and main efforts were focused on examining cURL 7.50.1. and later versions of cURL. It has to be noted that rather than employ fuzzing or similar approaches to validate the robustness of the build of the application and library, the latter goal was pursued through a classic source code audit. Sources covering authentication, various protocols, and, partly, SSL/TLS, were analyzed in considerable detail. A rationale behind this type of scoping pointed to these parts of the cURL tool that were most likely to be prone and exposed to real-life attack scenarios. Rounding up the methodology of the classic code audit, Cure53 benefited from certain tools, which included ASAN targeted with detecting memory errors, as well as Helgrind, which was tasked with pinpointing synchronization errors with the threading model.

They identified no less than twenty-three (23) potential problems in the code, out of which nine were deemed security vulnerabilities. But I’d also like to emphasize that they did also actually say this:

At the same time, the overall impression of the state of security and robustness of the cURL library was positive.

Resolving problems

In the curl security team we decided to downgrade one of the 9 vulnerabilities to a “plain bug” since the required attack scenario was very complicated and the risk deemed small, and two of the issues we squashed into treating them as a single one. That left us with 7 security vulnerabilities. Whoa, that’s a lot. The largest amount we’ve ever fixed in a single release before was 4.

I consider handling security issues in the project to be one of my most important tasks; pretty much all other jobs are down-prioritized in comparison. So with a large queue of security work, a lot of bug fixing and work on features basically had to halt.

You can get a fairly detailed description of our work on fixing the issues in the fix and validation log. The report, the log and the advisories we’ve already posted should cover enough details about these problems and associated fixes that I don’t feel a need to write about them much further.

curl and TLS 1.3

November 12, 2016 Daniel Stenberg 1 Comment

Draft 18 of the TLS version 1.3 spec was publi SSL padlock shed at the end of October 2016.

Already now, both Firefox and Chrome have test versions out with TLS 1.3 enabled. Firefox 52 will have it by default, and while Chrome will ship it, I couldn’t figure out exactly when we can expect it to be there by default.

Over the last few days we’ve merged TLS 1.3 support to curl, primarily in this commit by Kamil Dudka. Both the command line tool and libcurl will negotiate TLS 1.3 in the next version (7.52.0 – planned release date at the end of December 2016) if built with a TLS library that supports it and told to do it by the user.

The two current TLS libraries that will speak TLS 1.3 when built with curl right now is NSS and BoringSSL. The plan is to gradually adjust curl over time as the other libraries start to support 1.3 as well. As always we will appreciate your help in making this happen!

Right now, there’s also the minor flux in that servers and clients may end up running implementations of different draft versions of the TLS spec which contributes to a layer of extra fun!

Three TLS current 1.3 test servers to play with: https://enabled.tls13.com/ , https://www.allizom.org/ and https://tls13.crypto.mozilla.org/. I doubt any of these will give you any guarantees of functionality.

TLS 1.3 offers a few new features that allow clients such as curl to do subsequent TLS connections much faster, with only 1 or even 0 RTTs, but curl has no code for any of those features yet.

Technology

Another wget reference was Bourne

October 22, 2016 Daniel Stenberg 4 Comments

wget-is-not-a-crime Back in 2013, it came to light that Wget was used to to copy the files private Manning was convicted for having leaked. Around that time, EFF made and distributed stickers saying wget is not a crime.

Weirdly enough, it was hard to find a high resolution version of that image today but I’m showing you a version of it on the right side here.

In the 2016 movie Jason Bourne, Swedish actress Alicia Vikander is seen working on her laptop at around 1:16:30 into the movie and there’s a single visible sticker on that laptop. Yeps, it is for sure the same EFF sticker. There’s even a very brief glimpse of the top of the red EFF dot below the “crime” word.

Also recall the wget occurance in The Social Network.

c-ares, Security

a single byte write opened a root execution exploit

October 14, 2016 Daniel Stenberg 7 Comments

Thursday, September 22nd 2016. An email popped up in my inbox.

Subject: ares_create_query OOB write

As one of the maintainers of the c-ares project I’m receiving mails for suspected security problems in c-ares and this was such a one. In this case, the email with said subject came from an individual who had reported a ChromeOS exploit to Google.

It turned out that this particular c-ares flaw was one important step in a sequence of necessary procedures that when followed could let the user execute code on ChromeOS from JavaScript – as the root user. I suspect that is pretty much the worst possible exploit of ChromeOS that can be done. I presume the reporter will get a fair amount of bug bounty reward for this. (Update: he got 100,000 USD for it.)

The setup and explanation on how this was accomplished is very complicated and I am deeply impressed by how this was figured out, tracked down and eventually exploited in a repeatable fashion. But bear with me. Here comes a very simplified explanation on how a single byte buffer overwrite with a fixed value could end up aiding running exploit code as root.

The main Google bug for this problem is still not open since they still have pending mitigations to perform, but since the c-ares issue has been fixed I’ve been told that it is fine to talk about this publicly.

c-ares writes a 1 outside its buffer

c-ares has a function called ares_create_query. It was added in 1.10 (released in May 2013) as an updated version of the older function ares_mkquery. This detail is mostly interesting because Google uses an older version than 1.10 of c-ares so in their case the flaw is in the old function. This is the two functions that contain the problem we’re discussing today. It used to be in the ares_mkquery function but was moved over to ares_create_query a few years ago (and the new function got an additional argument). The code was mostly unchanged in the move so the bug was just carried over. This bug was actually already present in the original ares project that I forked and created c-ares from, back in October 2003. It just took this long for someone to figure it out and report it!

I won’t bore you with exactly what these functions do, but we can stick to the simple fact that they take a name string as input, allocate a memory area for the outgoing packet with DNS protocol data and return that newly allocated memory area and its length.

Due to a logic mistake in the function, you could trick the function to allocate a too short buffer by passing in a string with an escaped trailing dot. An input string like “one.two.three\.” would then cause the allocated memory area to be one byte too small and the last byte would be written outside of the allocated memory area. A buffer overflow if you want. The single byte written outside of the memory area is most commonly a 1 due to how the DNS protocol data is laid out in that packet.

This flaw was given the name CVE-2016-5180 and was fixed and announced to the world in the end of September 2016 when c-ares 1.12.0 shipped. The actual commit that fixed it is here.

What to do with a 1?

Ok, so a function can be made to write a single byte to the value of 1 outside of its allocated buffer. How do you turn that into your advantage?

The Redhat security team deemed this problem to be of “Moderate security impact” so they clearly do not think you can do a lot of harm with it. But behold, with the right amount of imagination and luck you certainly can!

Back to ChromeOS we go.

First, we need to know that ChromeOS runs an internal HTTP proxy which is very liberal in what it accepts – this is the software that uses c-ares. This proxy is a key component that the attacker needed to tickle really badly. So by figuring out how you can send the correctly crafted request to the proxy, it would send the right string to c-ares and write a 1 outside its heap buffer.

ChromeOS uses dlmalloc for managing the heap memory. Each time the program allocates memory, it will get a pointer back to the request memory region, and dlmalloc will put a small header of its own just before that memory region for its own purpose. If you ask for N bytes with malloc, dlmalloc will use ( header size + N ) and return the pointer to the N bytes the application asked for. Like this:

With a series of cleverly crafted HTTP requests of various sizes to the proxy, the attacker managed to create a hole of freed memory where he then reliably makes the c-ares allocated memory to end up. He knows exactly how the ChromeOS dlmalloc system works and its best-fit allocator, how big the c-ares malloc will be and thus where the overwritten 1 will end up. When the byte 1 is written after the memory, it is written into the header of the next memory chunk handled by dlmalloc:

The specific byte of that following dlmalloc header that it writes to, is used for flags and the lowest bits of size of that allocated chunk of memory.

Writing 1 to that byte clears 2 flags, sets one flag and clears the lowest bits of the chunk size. The important flag it sets is called prev_inuse and is used by dlmalloc to tell if it can merge adjacent areas on free. (so, if the value 1 simply had been a 2 instead, this flaw could not have been exploited this way!)

When the c-ares buffer that had overflowed is then freed again, dlmalloc gets fooled into consolidating that buffer with the subsequent one in memory (since it had toggled that bit) and thus the larger piece of assumed-to-be-free memory is partly still being in use. Open for manipulations!

Using that memory buffer mess

This freed memory area whose end part is actually still being used opened up the play-field for more “fun”. With doing another creative HTTP request, that memory block would be allocated and used to store new data into.

The attacker managed to insert the right data in that further end of the data block, the one that was still used by another part of the program, mostly since the proxy pretty much allowed anything to get crammed into the request. The attacker managed to put his own code to execute in there and after a few more steps he ran whatever he wanted as root. Well, the user would have to get tricked into running a particular JavaScript but still…

I cannot even imagine how long time it must have taken to make this exploit and how much work and sweat that were spent. The report I read on this was 37 very detailed pages. And it was one of the best things I’ve read in a long while! When this goes public in the future, I hope at least parts of that description will become available for you as well.

A lesson to take away from this?

No matter how limited or harmless a flaw may appear at a first glance, it can serve a malicious purpose and serve as one little step in a long chain of events to attack a system. And there are skilled people out there, ready to figure out all the necessary steps.

Update: A detailed write-up about this flaw (pretty much the report I refer to above) by the researcher who found it was posted on Google’s Project Zero blog on December 14:
Chrome OS exploit: one byte overflow and symlinks.

cURL and libcurl, Network

poll on mac 10.12 is broken

October 11, 2016 Daniel Stenberg 11 Comments

When Mac OS X first launched they did so without an existing poll function. They later added poll() in Mac OS X 10.3, but we quickly discovered that it was broken (it returned a non-zero value when asked to wait for nothing) so in the curl project we added a check in configure for that and subsequently avoided using poll() in all OS X versions to and including Mac OS 10.8 (Darwin 12). The code would instead switch to the alternative solution based on select() for these platforms.

With the release of Mac OS X 10.9 “Mavericks” in October 2013, Apple had fixed their poll() implementation and we’ve built libcurl to use it since with no issues at all. The configure script picks the correct underlying function to use.

Enter macOS 10.12 (yeah, its not called OS X anymore) “Sierra”, released in September 2016. Quickly we discovered that poll() once against did not act like it should and we are back to disabling the use of it in preference to the backup solution using select().

The new error looks similar to the old problem: when there’s nothing to wait for and we ask poll() to wait N milliseconds, the 10.12 version of poll() returns immediately without waiting. Causing busy-loops. The problem has been reported to Apple and its Radar number is 28372390. (There has been no news from them on how they plan to act on this.)

poll() is defined by POSIX and The Single Unix Specification it specifically says:

If none of the defined events have occurred on any selected file descriptor, poll() waits at least timeout milliseconds for an event to occur on any of the selected file descriptors.

We pushed a configure check for this in curl, to be part of the upcoming 7.51.0 release. I’ll also show you a small snippet you can use stand-alone below.

Apple is hardly alone in the broken-poll department. Remember how Windows’ WSApoll is broken?

Here’s a little code snippet that can detect the 10.12 breakage:

#include <poll.h>
#include <stdio.h>
#include <sys/time.h>

int main(void)
{
  struct timeval before, after;
  int rc;
  size_t us;

  gettimeofday(&before, NULL);
  rc = poll(NULL, 0, 500);
  gettimeofday(&after, NULL);

  us = (after.tv_sec - before.tv_sec) * 1000000 +
    (after.tv_usec - before.tv_usec);

  if(us < 400000) {
    puts("poll() is broken");
    return 1;
  }
  else {
    puts("poll() works");
  }
  return 0;
}

Follow-up, January 2017

This poll bug has been confirmed fixed in the macOS 10.12.2 update (released on December 13, 2016), but I’ve found no official mention or statement about this fact.

Firefox, Network, Web

1,000,000 sites run HTTP/2

September 28, 2016 Daniel Stenberg 1 Comment

… out of the top ten million sites that is. So there’s at least that many, quite likely a few more.

This is according to w3techs who runs checks daily. Over the last few months, there have been about 50,000 new sites per month switching it on.

It also shows that the HTTP/2 ratio has increased from a little over 1% deployment a year ago to the 10% today.

HTTP/2 gets more used the more popular site it is. Among the top 1,000 sites on the web, more than 20% of them use HTTP/2. HTTP/2 also just recently (September 9) overcame SPDY among the top-1000 most popular sites.

On September 7, Amazon announced their CloudFront service having enabled HTTP/2, which could explain an adoption boost over the last few days. New CloudFront users get it enabled by default but existing users actually need to go in and click a checkbox to make it happen.

As the web traffic of the world is severely skewed toward the top ones, we can be sure that a significantly larger share than 10% of the world’s HTTPS traffic is using version 2.

Recent usage stats in Firefox shows that HTTP/2 is used in half of all its HTTPS requests!

cURL and libcurl, Network, Open Source

My first 20 years of HTTP

September 6, 2016 Daniel Stenberg 4 Comments

During the autumn 1996 I took my first swim in the ocean known as HTTP. Twenty years ago now.

I had previously worked with writing an IRC bot in C, and IRC is a pretty simple text based protocol over TCP so I could use some experiences from that when I started to look into HTTP. That IRC bot was my first real application distributed to the world that was using TCP/IP. It was portable to most unixes and Amiga and it was open source.

1996 was the year the movie Independence Day premiered and the single hit song that plagued the world more than others that year was called Macarena. AOL, Webcrawler and Netscape were the most popular websites on the Internet. There were less than 300,000 web sites on the Internet (compared to some 900 million today).

I decided I should spice up the bot and make it offer a currency exchange rate service so that people who were chatting could ask the bot what 200 SEK is when converted to USD or what 50 AUD might be in DEM. – Right, there was no Euro currency yet back then!

I simply had to fetch the currency rates at a regular interval and keep them in the same server that ran the bot. I just needed a little tool to download the rates over HTTP. How hard can that be? I googled around (this was before Google existed so that was not the search engine I could use!) and found a tool named ‘httpget’ that made pretty much what I wanted. It truly was tiny – a few hundred lines of code.

I don’t have an exact date saved or recorded for when this happened, only the general time frame. You know, we had no smart phones, no Google calendar and no digital cameras. I sported my first mobile phone back then, the sexy Nokia 1610 – viewed in the picture on the right here.

The HTTP/1.0 RFC had just recently came out – which was the first ever real spec published for HTTP. RFC 1945 was published in May 1996, but I was blissfully unaware of the youth of the standard and I plunged into my little project. This was the first published HTTP spec and it says:

HTTP has been in use by the World-Wide Web global information initiative since 1990. This specification reflects common usage of the protocol referred too as "HTTP/1.0". This specification describes the features that seem to be consistently implemented in most HTTP/1.0 clients and servers.

Many years after that point in time, I have learned that already at this time when I first searched for a HTTP tool to use, wget already existed. I can’t recall that I found that in my searches, and if I had found it maybe history would’ve made a different turn for me. Or maybe I found it and discarded for a reason I can’t remember now.

I wasn’t the original author of httpget; Rafael Sagula was. But I started contributing fixes and changes and soon I was the maintainer of it. Unfortunately I’ve lost my emails and source code history from those earliest years so I cannot easily show my first steps. Even the oldest changelogs show that we very soon got help and contributions from users.

The earliest saved code archive I have from those days, is from after we had added support for Gopher and FTP and renamed the tool ‘urlget’. urlget-3.5.zip was released on January 20 1998 which thus was more than a year later my involvement in httpget started.

The original httpget/urlget/curl code was stored in CVS and it was licensed under the GPL. I did most of the early development on SunOS and Solaris machines as my first experiments with Linux didn’t start until 97/98 something.

The first web page I know we have saved on archive.org is from December 1998 and by then the project had been renamed to curl already. Roughly two years after the start of the journey.

RFC 2068 was the first HTTP/1.1 spec. It was released already in January 1997, so not that long after the 1.0 spec shipped. In our project however we stuck with doing HTTP 1.0 for a few years longer and it wasn’t until February 2001 we first started doing HTTP/1.1 requests. First shipped in curl 7.7. By then the follow-up spec to HTTP/1.1, RFC 2616, had already been published as well.

The IETF working group called HTTPbis was started in 2007 to once again refresh the HTTP/1.1 spec, but it took me a while until someone pointed out this to me and I realized that I too could join in there and do my part. Up until this point, I had not really considered that little me could actually participate in the protocol doings and bring my views and ideas to the table. At this point, I learned about IETF and how it works.

I posted my first emails on that list in the spring 2008. The 75th IETF meeting in the summer of 2009 was held in Stockholm, so for me still working on HTTP only as a spare time project it was very fortunate and good timing. I could meet a lot of my HTTP heroes and HTTPbis participants in real life for the first time.

I have participated in the HTTPbis group ever since then, trying to uphold the views and standpoints of a command line tool and HTTP library – which often is not the same as the web browsers representatives’ way of looking at things. Since I was employed by Mozilla in 2014, I am of course now also in the “web browser camp” to some extent, but I remain a protocol puritan as curl remains my first “child”.

cURL and libcurl, Open Source, Windows

Removing the PowerShell curl alias?

August 19, 2016 Daniel Stenberg 3 Comments

PowerShell is a spiced up command line shell made by Microsoft. According to some people, it is a really useful and good shell alternative.

Already a long time ago, we got bug reports from confused users who couldn’t use curl from their PowerShell prompts and it didn’t take long until we figured out that Microsoft had added aliases for both curl and wget. The alias had the shell instead invoke its own command called “Invoke-WebRequest” whenever curl or wget was entered. Invoke-WebRequest being PowerShell’s own version of a command line tool for fiddling with URLs.

Invoke-WebRequest is of course not anywhere near similar to neither curl nor wget and it doesn’t support any of the command line options or anything. The aliases really don’t help users. No user who would want the actual curl or wget is helped by these aliases, and users who don’t know about the real curl and wget won’t use the aliases. They were and remain pointless. But they’ve remained a thorn in my side ever since. Me knowing that they are there and confusing users every now and then – not me personally, since I’m not really a Windows guy.

Fast forward to modern days: Microsoft released PowerShell as open source on github yesterday. Without much further ado, I filed a Pull-Request, asking the aliases to be removed. It is a minuscule, 4 line patch. It took way longer to git clone the repo than to make the actual patch and submit the pull request!

It took 34 minutes for them to close the pull request:

“Those aliases have existed for multiple releases, so removing them would be a breaking change.”

To be honest, I didn’t expect them to merge it easily. I figure they added those aliases for a reason back in the day and it seems unlikely that I as an outsider would just make them change that decision just like this out of the blue.

But the story didn’t end there. Obviously more Microsoft people gave the PR some attention and more comments were added. Like this:

“You bring up a great point. We added a number of aliases for Unix commands but if someone has installed those commands on WIndows, those aliases screw them up.

We need to fix this.”

So, maybe it will trigger a change anyway? The story is ongoing…

Network, Web

HTTP/2 connection coalescing

August 18, 2016 Daniel Stenberg 6 Comments

Section 9.1.1 in RFC7540 explains how HTTP/2 clients can reuse connections. This is my lengthy way of explaining how this works in reality.

Many connections in HTTP/1

With HTTP/1.1, browsers are typically using 6 connections per origin (host name + port). They do this to overcome the problems in HTTP/1 and how it uses TCP – as each connection will do a fair amount of waiting. Plus each connection is slow at start and therefore limited to how much data you can get and send quickly, you multiply that data amount with each additional connection. This makes the browser get more data faster (than just using one connection).

6 connections

Add sharding

Web sites with many objects also regularly invent new host names to trigger browsers to use even more connections. A practice known as “sharding”. 6 connections for each name. So if you instead make your site use 4 host names you suddenly get 4 x 6 = 24 connections instead. Mostly all those host names resolve to the same IP address in the end anyway, or the same set of IP addresses. In reality, some sites use many more than just 4 host names.

24 connections

The sad reality is that a very large percentage of connections used for HTTP/1.1 are only ever used for a single HTTP request, and a very large share of the connections made for HTTP/1 are so short-lived they actually never leave the slow start period before they’re killed off again. Not really ideal.

One connection in HTTP/2

With the introduction of HTTP/2, the HTTP clients of the world are going toward using a single TCP connection for each origin. The idea being that one connection is better in packet loss scenarios, it makes priorities/dependencies work and reusing that single connections for many more requests will be a net gain. And as you remember, HTTP/2 allows many logical streams in parallel over that single connection so the single connection doesn’t limit what the browsers can ask for.

Unsharding

The sites that created all those additional host names to make the HTTP/1 browsers use many connections now work against the HTTP/2 browsers’ desire to decrease the number of connections to a single one. Sites don’t want to switch back to using a single host name because that would be a significant architectural change and there are still a fair number of HTTP/1-only browsers still in use.

Enter “connection coalescing”, or “unsharding” as we sometimes like to call it. You won’t find either term used in RFC7540, as it merely describes this concept in terms of connection reuse.

Connection coalescing means that the browser tries to determine which of the remote hosts that it can reach over the same TCP connection. The different browsers have slightly different heuristics here and some don’t do it at all, but let me try to explain how they work – as far as I know and at this point in time.

Coalescing by example

Let’s say that this cool imaginary site “example.com” has two name entries in DNS: A.example.com and B.example.com. When resolving those names over DNS, the client gets a list of IP address back for each name. A list that very well may contain a mix of IPv4 and IPv6 addresses. One list for each name.

You must also remember that HTTP/2 is also only ever used over HTTPS by browsers, so for each origin speaking HTTP/2 there’s also a corresponding server certificate with a list of names or a wildcard pattern for which that server is authorized to respond for.

In our example we start out by connecting the browser to A. Let’s say resolving A returns the IPs 192.168.0.1 and 192.168.0.2 from DNS, so the browser goes on and connects to the first of those addresses, the one ending with “1”. The browser gets the server cert back in the TLS handshake and as a result of that, it also gets a list of host names the server can deal with: A.example.com and B.example.com. (it could also be a wildcard like “*.example.com”)

If the browser then wants to connect to B, it’ll resolve that host name too to a list of IPs. Let’s say 192.168.0.2 and 192.168.0.3 here.

Host A: 192.168.0.1 and 192.168.0.2
Host B: 192.168.0.2 and 192.168.0.3

Now hold it. Here it comes.

The Firefox way

Host A has two addresses, host B has two addresses. The lists of addresses are not the same, but there is an overlap – both lists contain 192.168.0.2. And the host A has already stated that it is authoritative for B as well. In this situation, Firefox will not make a second connect to host B. It will reuse the connection to host A and ask for host B’s content over that single shared connection. This is the most aggressive coalescing method in use.

one connection

The Chrome way

Chrome features a slightly less aggressive coalescing. In the example above, when the browser has connected to 192.168.0.1 for the first host name, Chrome will require that the IPs for host B contains that specific IP for it to reuse that connection. If the returned IPs for host B really are 192.168.0.2 and 192.168.0.3, it clearly doesn’t contain 192.168.0.1 and so Chrome will create a new connection to host B.

Chrome will reuse the connection to host A if resolving host B returns a list that contains the specific IP of the connection host A is already using.

The Edge and Safari ways

They don’t do coalescing at all, so each host name will get its own single connection. Better than the 6 connections from HTTP/1 but for very sharded sites that means a lot of connections even in the HTTP/2 case.

curl also doesn’t coalesce anything (yet).

Surprises and a way to mitigate them

Given some comments in the Firefox bugzilla, the aggressive coalescing sometimes causes some surprises. Especially when you have for example one IPv6-only host A and a second host B with both IPv4 and IPv6 addresses. Asking for data on host A can then still use IPv4 when it reuses a connection to B (assuming that host A covers host B in its cert).

In the rare case where a server gets a resource request for an authority (or scheme) it can’t serve, there’s a dedicated error code 421 in HTTP/2 that it can respond with and the browser can then go back and retry that request on another connection.

Starts out with 6 anyway

Before the browser knows that the server speaks HTTP/2, it may fire up 6 connection attempts so that it is prepared to get the remote site at full speed. Once it figures out that it doesn’t need all those connections, it will kill off the unnecessary unused ones and over time trickle down to one. Of course, on subsequent connections to the same origin the client may have the version information cached so that it doesn’t have to start off presuming HTTP/1.

daniel.haxx.se