Words

Why?

I wanted to get to my WiFi router from the Internet, but didn’t want it serving the Internet. Putting it on a VPN is a good way to do this, and Tailscale makes VPNs easy enough to run and use that there’s no good reason not to.

When?

This past weekend. I’d just gotten back to Brooklyn and installed OpenWRT 19.07, which included a new enough kernel to support Wireguard.

This was possible before – Tailscale’s been shipping ARM builds for a while, and OpenWRT’s been shipping 19.07 for a while too – but I didn’t try it before now.

How?

I get out a MacBook Pro that I’ve already set up Tailscale on, and read through the Tailscale docs for setting up Tailscale with static binaries to know how the install is supposed to flow.

It probably won’t be possible to follow them exactly, because OpenWRT is different, so I start playing around.

Installing Tailscale on the router

I look up my router’s hardware specs, then find the tarball for my router in Tailscale’s stable track, and download it to my machine:

% curl -O https://pkgs.tailscale.com/stable/tailscale_1.0.5_arm.tgz
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 11.9M  100 11.9M    0     0  3433k      0  0:00:03  0:00:03 --:--:-- 3433k
%

I’m not sure how much free RAM and flash I have on my router, but I know I haven’t installed anything else, and guess that 11.9M is relatively small enough that I don’t need to worry about it.

So I copy the tarball over:

% scp tailscale_1.0.5_arm.tgz root@192.168.1.1:/tmp

then SSH to the router and try to unpack the tarball:

# cd /tmp
# tar xvf tailscale_1.0.5_arm.tgz
tar: invalid tar magic
# 

and it doesn’t work. OpenWRT uses busybox tar, so I fumble around for a bit (xkcd.com/1168 applies) and eventually figure out:

# tar x -zvf tailscale_1.0.5_arm.tgz

Then I try running the binaries to make sure they’re actually the right ones 😛 :

# cd tailscale_1.0.5_arm
# ./tailscale version
1.0.5-g31b5dec0a
# ./tailscaled --help
<...>

Looks it. Then I try to start tailscaled for real to figure out its dependencies.

# ./tailscaled
logtail started
Program starting: v1.0.5-g31b5dec0a, Go 1.14.4-ts56db765: []string{"/usr/sbin/tailscaled"}
LogID: 025f6648f94bcd1d40b0ee005ac472e123f7aeb2813874a71a7ba121da3d1827
logpolicy: using system state directory "/var/lib/tailscale"
2.6M/14.9M Starting userspace wireguard engine with tun device "tailscale0"
2.9M/16.6M Linux kernel version: 4.14.195
2.9M/16.8M is CONFIG_TUN enabled in your kernel? `modprobe tun` failed with:
2.9M/16.8M CreateTUN: can't create TUN device; /dev/net/tun does not exist
2.9M/16.8M wgengine.New: can't create TUN device; /dev/net/tun does not exist
flushing log.
logger closing down
logtail: dialed "log.tailscale.io:443" in 415ms
logtail: upload: log upload of 659 bytes compressed failed: Post "https://log.tailscale.io/c/tailnode.log.tailscale.io/0055d7ffaa6c80ecda7fef05574c36d025a42ead7d7409fce3b7218f84628e4a": x509: certificate signed by unknown authority
logtail: backoff: 12 msec
logtail: dialed "log.tailscale.io:443" in 413ms
logtail: upload: log upload of 659 bytes compressed failed: Post "https://log.tailscale.io/c/tailnode.log.tailscale.io/0055d7ffaa6c80ecda7fef05574c36d025a42ead7d7409fce3b7218f84628e4a": context canceled
#

There are many lines of error messages but really two errors:

• `modprobe tun` failed means that Tailscale asked the kernel for a TUN device and got back “huh?”
• x509: certificate signed by unknown authority means that Tailscale tried to make a TLS connection but couldn’t figure out who was on the other end. When this happens with professional software (like Tailscale) it usually means no certificates are installed at all.

Both of these are pretty standard things to get from a package manager, but everyone calls them different names. I poke around OpenWRT’s packages and determine I probably need kmod-tun and ca-bundle, so I install them and try again.

# opkg update
# opkg install ca-bundle kmod-tun

After this, tailscaled is able to start:

# ./tailscaled
<...>

so I copy it to Flash.

# cp tailscale tailscaled /usr/sbin

Running as a service

Tailscale comes with a simple systemd unit file. OpenWRT can’t use this directly, since it uses procd instead. I read the procd docs and, after about an hour of trial and error, port enough of it to get things working:

It’s my first time working with procd, so I’m pretty happy with how it turned out.

Gists have awkwardly long URLs, so I make a redirect to it that will let me download it with a short URL and upload it to my router easily:

% curl -LO https://willangley.org/tailscale-procd
% less tailscale-procd
% scp tailscale-procd root@192.168.1.1:/tmp

I SSH to the router, install the init script, and run it to start tailscaled.

# cp /tmp/tailscale-procd /etc/init.d/tailscale
# chmod +x /etc/init.d/tailscale
# /etc/init.d/tailscale start

Log in to Tailscale

Once tailscaled is running, I run

# tailscale up

to get a login link, and click it to log in.

After logging in, I go to the Tailscale admin console and look for my router’s hostname; since I’ve never changed it, it’s OpenWRT.

I copy its IP address and make sure I can ping it from the machine I’m working from. (This would fail if the machine wasn’t on Tailscale, or was on a different Tailscale network.)

% ping 100.114.61.77
PING 100.114.61.77 (100.114.61.77): 56 data bytes
64 bytes from 100.114.61.77: icmp_seq=0 ttl=64 time=43.378 ms
64 bytes from 100.114.61.77: icmp_seq=1 ttl=64 time=7.895 ms
64 bytes from 100.114.61.77: icmp_seq=2 ttl=64 time=7.046 ms
64 bytes from 100.114.61.77: icmp_seq=3 ttl=64 time=5.387 ms
^C
--- 100.114.61.77 ping statistics ---
4 packets transmitted, 4 packets received, 0.0% packet loss
round-trip min/avg/max/stddev = 5.387/15.926/43.378/15.875 ms

Once I confirm the connection works, I disable key expiry for the router to avoid it falling off of my network.

Surviving reboot

Now that I know it’s working, I enable the Tailscale service so it launches on every boot.

# /etc/init.d/tailscale enable

And check to make sure it has succeeded:

# ls /etc/rc.d/*tailscale
/etc/rc.d/S80tailscale

This took me several tries to get right, because I forgot to add the START=80 variable to my init script; without it, OpenWRT didn’t know where to link it into the startup order.

But now that it’s there, I’m good to go! At least until the next sysupgrade, when all of this will get wiped away.

Future Work

There’s an open Tailscale issue, tailscale/tailscale#724, to package Tailscale for OpenWRT. If this happens soon, there won’t be much to do here 🙂

If it takes longer, I’ll probably try to get OpenWRT to not wipe this away with a sysupgrade – I think this’ll involve some combination of editing /etc/sysupgrade.conf and using sysupgrade -k on future upgrades.

I thought about building custom OpenWRT images with Tailscale pre-authenticated keys baked in, too; Google Cloud Build makes systems like this easy to spin up nowadays, and I think I could do it in O(days) worth of work from experience. But it’s probably overkill unless I wind up with a fleet of these somehow.

Appendix

I broke my router’s firmware once doing this.

It wasn’t even a step that I needed to do; I was trying to figure out how much space I’d consumed with packages, and ran a command from a snippet in a Google search that was supposed to remove unused packages without clicking it to read the rest of the page.

The snippet looked reasonable, but if I’d clicked in, I would have seen the command I was about to run was going to remove the WiFi drivers.

But as I said, I didn’t, and wound up needing to rummage for a USB-to-Ethernet adapter and Ethernet cable to reinstall my firmware.

It’s easy to search for resume tips, but too many of them come back; it’s unlikely they’re all right.

So I started looking for scientific papers in the area to see if anyone had studied it systematically. Eventually the Google keywords [evidence based job search] turned up a review paper that looked useful and that I had access to:

Liu, Songqi, Jason L. Huang, and Mo Wang. 2014. “Effectiveness of Job Search Interventions: A Meta-Analytic Review.” Psychological Bulletin 140 (4): 1009–41. https://doi.org/10.1037/a0035923.

Working outwards from there, I’ve found that we’re pretty sure about:

Writing

Competency statements (Bright and Hutton, 2000) work. They work well enough that copy-pasting them into resumes meaningfully boosted how managers rated them in a systematic survey.

Length

I’ve not been able to find a decision rule for “how long should I make my resume,” but it’s increasingly seeming like the answer is “as short as possible.”

Recruiters sometimes only turned out to have shortlisted one page resumes (Lipovsky, 2014), and studies that varied design (Wilgenburg, 2017; Uskaurs, 2018) used one page resumes exclusively.

Bright and Hutton, which varied content and left design fixed, used two page resumes and a cover letter.

No one asked recruiters to look at three page resumes, presumably because the response rate would have been too poor to publish.

Design

Design risks rarely paid off when studied, although the relationship between design and outcomes is difficult to describe. Reportedly, it may depend on:

• fashion, with (Arnulf et al., 2010) finding creative resume designs uniformly underperforming traditional ones, and later studies turning up more qualified answers
• intelligibility, with (Lipovsky, 2014) finding recruiters selected organized resumes
• the length of time a recruiter looks at your resume, with shorter review times favoring splashy resumes (Wilgenburg, 2017)
• the position applied for, with creative positions favoring creative resumes, and others favoring restrained, one column layouts (ibid; Uskars, 2018)

The best we can say is there’s more evidence for investing time in writing and proofreading than in visual design. Unless you’re applying for a design position, focus on making an easy to follow resume, with one column or one column and a sidebar, and move on.

References

Arnulf, Jan Ketil, Lisa Tegner, and Øyunn Larssen. 2010. “Impression Making by Résumé Layout: Its Impact on the Probability of Being Shortlisted.” European Journal of Work and Organizational Psychology 19 (2): 221–30. https://doi.org/10.1080/13594320902903613

Azrin, N. H., T. Flores, and S. J. Kaplan. 1975. “Job-Finding Club: A Group-Assisted Program for Obtaining Employment.” Behaviour Research and Therapy 13 (1): 17–27. https://doi.org/10.1016/0005-7967(75)90048-0.

Bright, Jim E. H., and Sonia Hutton. 2000. “The Impact of Competency Statements on Résumés for Short‐listing Decisions.” International Journal of Selection and Assessment 8 (2): 41–53. https://doi.org/10.1111/1468-2389.00132.

Higgins, Chad A., and Timothy A. Judge. 2004. “The Effect of Applicant Influence Tactics on Recruiter Perceptions of Fit and Hiring Recommendations: A Field Study.” The Journal of Applied Psychology 89 (4): 622–32. https://doi.org/10.1037/0021-9010.89.4.622.

Lipovsky, Caroline. 2014. “The CV as a Multimodal Text.” Visual Communication 13 (4): 429–58. https://doi.org/10.1177/1470357213497869.

Liu, Songqi, Jason L. Huang, and Mo Wang. 2014. “Effectiveness of Job Search Interventions: A Meta-Analytic Review.” Psychological Bulletin 140 (4): 1009–41. https://doi.org/10.1037/a0035923.

Uskaurs, Matiss. 2018. “The Impact of a Resume on a Job Interview and First Impression.” Edited by Thomas Sabel. Vaasa University of Applied Sciences. https://www.semanticscholar.org/paper/834c681f76b20b6928f17f3218834c7aa5941e05.

Wilgenburg, Florence van. 2017. “The Impact of Creativity in Resume Shortlisting.” Edited by dr Ruud Koolen. Masters in Communication and Information Sciences, Tilburg University. http://arno.uvt.nl/show.cgi?fid=145069

When I saw the subject lines:

WordPress Email System	[Will Angley] Contact – Name: …
WordPress Email System	[Will Angley] Contact – Name: …

in my inbox I was hopeful for a moment. Then I opened the messages and saw they were offering to submit ads in other sites’ contact forms for me.

This seemed like something that could backfire if I used it to advertise my personal blog; spamming my boss is not a career enhancing move. I passed.

My next reaction, after that, was “omg, this is exciting! I’m having a spam problem now! What cool technology can I use to fight it‽”

Then I remembered I was already using one; my contact form is protected by Akismet. I clicked around through the WordPress Admin until I found Akismet’s stats:

showing that I’ve been getting 75-90 messages a month, almost all blatant spam.

In hindsight I’m not surprised. Email spam is a hard problem, and contact form spam is in some ways harder; Akismet is doing really well here.

I know that big sites often need to do more to fight spam, using techniques like honeypot form fields (which look different to spam-writing robots than they do to humans), CAPTCHAs, and bot defense software that pokes at browsers to see if they look like they’re being used by people.

But willangley.org isn’t that big, and I like knowing I’m not breaking for readers who use content blockers, so I think I’ll leave things alone until I actually want to work on my contact form again 😛