updating over tailscale (sometimes!) fails with about gokrazy

[damdo]

Yeah noticed this too.
Thanks for putting up a patch for this @nhanb!

[nhanb]

I consistently get this error when running gok update over tailscale (it works fine using direct LAN IP):

Transferred root file system (47 MiB) at 26.01 MiB/s (total: 2s)s
Transferred boot file system (16 MiB) at 26.20 MiB/s (total: 1s)s
Triggering reboot
[update boot file system] 0.00% of 16 MiB, uploading at 0 B/s                 2023/05/18 15:00:50 reboot: Post "https://gokrazy:***@junk/reboot": read tcp 100.95.42.78:45338->100.90.227.60:443: read: connection reset by peer

To be more specific, it got stuck for some time at [update boot file system] 0.00% of 16 MiB, uploading at 0 B/s, then the rest of the log followed. However, the appliance has actually been updated and rebooted correctly.

I'm using latest gok and tailscale 1.40.1 on both machines. My appliance is an amd64 mini pc, if that matters.

[stapelberg]

Not sure yet what the trigger is, but maybe we need to programmatically ignore this error in the updater for now.

Given that you can reproduce it reliably, could you send a PR to ignore this error in the updater? Thanks!

[nhanb]

I'll try!

[nhanb]

I dug into the gokrazy/* repos and found this snippet which could be a race condition:

gokrazy/update.go

Lines 353 to 362 in a764954

	log.Println("Rebooting")
	w.Write([]byte("Rebooting...\n"))
	rc.Flush()
	go func() {
	time.Sleep(time.Second) // give the http response some time to be sent
	if err := reboot(); err != nil {
	log.Println("could not reboot:", err)
	}
	}()

- the /reboot handler assumes the response always finishes successfully within 1 second before rebooting, which may not always be the case (maybe tailscale just adds enough latency for this bug to surface more easily). If we instead allow the http server to gracefully shut down before rebooting, we might solve this bug without any workaround.

On that note, how do I tell gokrazy/tools to use the gokrazy/gokrazy repo from my local disk instead of github.com/gokrazy/gokrazy? Apparently gokrazy/tools/go.mod does not list gokrazy/gokrazy, so I can't replace it. Consequently, I can't test my local changes to the /reboot handler inside the gokrazy/gokrazy repo.

[stapelberg]

the /reboot handler assumes the response always finishes successfully within 1 second before rebooting, which may not always be the case (maybe tailscale just adds enough latency for this bug to surface more easily).

Hmm. Can you try bumping this to 10s to see if that works around the issue?

On that note, how do I tell gokrazy/tools to use the gokrazy/gokrazy repo from my local disk instead of github.com/gokrazy/gokrazy? Apparently gokrazy/tools/go.mod does not list gokrazy/gokrazy, so I can't replace it. Consequently, I can't test my local changes to the /reboot handler inside the gokrazy/gokrazy repo.

You can just gok add . in your gokrazy/gokrazy working copy, that should set up the replace directive in the corresponding build directory within your gokrazy instance. See also https://gokrazy.org/development/modules/#building-local-code-the-replace-directive

[nhanb]

Bumping the wait time to 10s doesn't seem to do anything. I wanted to try gracefully shutting down the http server too, but honestly couldn't figure out how to do that in the gokrazy codebase. I'll sleep on it for now :(

Meanwhile I've pushed gokrazy/tools#55 to simply ignore the error.