Mar 5, 2020 - Tarbombs considered harmful

So, one day you hear about this great new open source project, and visit the company’s web site and download the latest version of their software tofu-wonder.tar.gz, and extract it in your home directory:

$ tar xvf tofu-wonder.tar.gz

You just got tarbombed. In older versions of tar, tarballs could even contain absolute paths and potentially overwrite existing files on your file system. These days, most versions of tar prevent this unless explicitly allowed, so the worst that happens is a particular tar archive litters it’s files in whatever unfortunate directory you were in when you extracted it. Have fun cleaning that up.

Ok - so how to avoid it? I now include this line in my .zshrc:

export TAR_OPTIONS="--one-top-level"

This option extracts all files into a directory named by the basename. In the example above, it’d now look like this:

$ tar xvf tofu-wonder.tar.gz

Perfect! But, it’s better not to make users do this. The first way to prevent this is to include the top-level directory when you’re creating a tarball:

tar czvf tofu-wonder.tar.gz tofu-wonder/

Another option is to use transform and replace . with something else:

tar czvf tofu-wonder.tar.gz --transform "s?^\.?tofu-wonder-0.1.1?"  .

Sep 3, 2019 - golang: finding a race condition in Docker


In OpenShift’s client utilities, we use some vendored Docker code to extract data from a container image. Several images could be extracted concurrently, and we were running into an issue where only on RHEL 8 clients, occassionally a user would see a panic:

panic: runtime error: slice bounds out of range

goroutine 163 [running]:
        /opt/rh/go-toolset-1.12/root/usr/lib/go-toolset-1.12-golang/src/bufio/bufio.go:89 +0x211

We didn’t know why we only saw it on RHEL 8 clients, and why it only happened sometimes. I wanted a better traceback than the original bug report gave us, and maybe a coredump so I could poke around in gdb. To be honest, I didn’t really know what I’d be looking for in gdb. I’d only ever used it with C, and even in C, I’m generally a printf debuggerer.

However, I hadn’t been able to reproduce the problem myself, so I wanted to get as much information as I could.

We added export GOTRACEBACK=crash to our development scripts, and waited until someone saw it again. It wasn’t too much longer when we got a report of it again, and I was able to see a much longer stack trace that showed me all of the running goroutines, as well as getting a coredump.

It looked like code in go itself was reading past the end of it’s own buffer: what? Was there a bug in go? I started researching this some more, and I was still a bit lost, until I stumbled upon an entry in the longer stack trace that pointed me to Docker’s code using a pool of buffers.

Docker maintains a pool of *bufio.Reader to reduce memory usage. If these were being recycled, and some previous holder of it tried to write to it after giving it back, and someone else got a hold of it very quickly – this all sounded somewhat familiar, and reminded me of my Operating System’s class. Was this a race condition?

Identifying what kind of problem I was dealing with made things a lot easier. In retrospect, maybe I should’ve realized it was a race condition sooner, but now that I knew what it was, I wanted to know how people might uncover a race condition in golang.

Go includes tools for detecting these cases, by simply building or running your go code with the -race argument. After doing that, and running locally, my program exited successfully with no warnings about any kind of race condition. Theoretically, this tooling was supposed to identify the potential race even if it wasn’t causing a panic.

I even tried it on a RHEL 8 virtual machine, just like the reporters of the bugs were using. Nothing.

As a last resort, I asked a coworker if I could experiment in an environment that he seemed to encounter the problem once a day or so. I wrote a script that would run the command over and over again, hoping that it crashed. I used the binary that had been built with the -race flag.

Sure enough, on his system, go enthusiastically reported “WARNING: DATA RACE”, with a traceback telling me exactly where.

Write at 0x00c00115b320 by goroutine 94:
      /usr/local/go/src/bufio/bufio.go:75 +0xe0
      /usr/local/go/src/bufio/bufio.go:71 +0xd1*BufioReaderPool).Put()
      /go/src/ +0x5b*BufioReaderPool).NewReadCloserWrapper.func1()
      /go/src/ +0x140*ReadCloserWrapper).Close()
      /go/src/ +0x5e
      /go/src/ +0x80*ReadCloserWrapper).Close()
      /go/src/ +0x5e
      /go/src/ +0x975*Options).Run.func1.1.2()
      /go/src/ +0xa0f*Options).Run.func1.1()
      /go/src/ +0x31f8*worker).Try.func1()
      /go/src/ +0x6d*workQueue).run.func1()
      /go/src/ +0x35d

Previous read at 0x00c00115b320 by goroutine 8:
      /usr/local/go/src/bufio/bufio.go:525 +0xc7
      /usr/local/go/src/bufio/bufio.go:506 +0x5e1
      /usr/local/go/src/io/io.go:384 +0x13c
      /usr/local/go/src/io/io.go:364 +0x10a
      /usr/local/go/src/os/exec/exec.go:243 +0xfa
      /usr/local/go/src/os/exec/exec.go:409 +0x3d

Goroutine 94 (running) created at:*workQueue).run()
      /go/src/ +0xd8

Goroutine 8 (running) created at:
      /usr/local/go/src/os/exec/exec.go:408 +0x16c2
      /go/src/ +0x243
      /go/src/ +0x52e
      /go/src/ +0x806
      /go/src/ +0xa1*Options).Run.func1.1.2()
      /go/src/ +0xa0f*Options).Run.func1.1()
      /go/src/ +0x31f8*worker).Try.func1()
      /go/src/ +0x6d*workQueue).run.func1()
      /go/src/ +0x35d

Ok: why did his system do it and not mine? After examining the traceback, I noticed that this was happening in the code that Docker uses to decompress a stream of compressed data. And in that code for gzipped files, it can use the native Golang gzip library, or shell out to unpigz which is a super fast, parallel utility. unpgiz was not present on any of my test systems; however it was there on his. Installing the package on my server instantly reproduced the problem.

What was different? The code running unpigz was using one of those shared buffers I mentioned earlier. There was a case where the context for a command was cancelled, and the buffer was returned to the pool. However, with contexts and CommandContext in Go, merely cancelling the context does not guarantee the command is fully done. You also need to wait for cmd.Wait() to finish before returning any buffers to the pool. Writing a fix that ensured that happened resolved our problem.

May 14, 2019 - Replacing Nest


The big announcement at Google I/O ‘19 is that Google is discontinuing the Works with Nest API program. Nest is the only no-local-control ecosystem I have in my house, and it was so very clearly a mistake in hindsight. In order for an end user to integrate Nest with a third party product like Home Assistant, one needs a developer account. Right off the bat, Nest assumed that the only people who would want to do such a thing is other cloud-based services like Hue, or Amazon who would interact with Nest on behalf of a user, rather than a user themselves.

I currently have a developer account, and I use it with Home Assistant to do a number of automations tied to my Nest devices. There’s a lot of customized control over when the HVAC goes into away (eco) or home mode based on the arrival and departure - or expected arrival - of my home’s residents. More interestingly, a number of automations around Nest Protect enhance the safety of my home by unlocking the doors if people are home, turning on lights, and disabling the HVAC system in the event of an incident.

None of these would be possible if Nest shuts down the Works with Nest program, other than whatever specific partner integrations they allow. This is all in the name of “privacy,” which in Google’s mind is that your private information is only known by you and Google. Why, then, have they not assauged the hobbyist communitys concerns by announcing a local API option?

The protocols that Nest devices use are open source-ish, called Weave and Thread. Certainly they could open them up to allow me to control my own devices. Anyway, even if Google eventually figures it out and does the right not-evil thing, I figured I should look at alternatives.

For now, I’ve already replaced my thermostat with a Venstar ColorTouch T7850. It features a local API, and a home assistant component.

It’s not as pretty as a Nest’s round, sleek appearance but it’s works well and has full local control. The Nest Protects’ most likely replacement is a Z-wave device from First Alert, but for now I’ve kept them. I do like their voice announcements of the problem - in Spanish and English - and the pathlight feature when one walks down the hallway at night.