Not all bugs can easily be reproduced — sometimes, all you have is a core dump from a crashing program, but no idea about the triggering conditions of the bug yet.

When using Go, we can use the delve debugger for core dump debugging, but I had trouble figuring out how to save byte slice contents (for example: the incoming request causing the crash) from memory into a file for further analysis, so this article walks you through how to do it.

Simple Example

Let’s imagine the following scenario: You are working on a performance optimization in Go Protobuf and have accidentally badly broken the proto.Marshal function. The function is now returning an error, so let’s run one of the failing tests with delve:

~/protobuf/proto master % dlv test
(dlv) b ExampleMarshal
(dlv) c
> [Breakpoint 1] google.golang.org/protobuf/proto_test.ExampleMarshal() ./encode_test.go:293 (hits goroutine(1):1 total:1) (PC: 0x9d6c96)
(dlv) next 4
> google.golang.org/protobuf/proto_test.ExampleMarshal() ./encode_test.go:297 (PC: 0xb54495)
   292: // [google.golang.org/protobuf/types/known/durationpb.New].
   293: func ExampleMarshal() {
   294: b, err := proto.Marshal(&durationpb.Duration{
   295: Nanos: 125,
   296: })
=> 297: if err != nil {
   298: panic(err)
   299: }
   300:
   301: fmt.Printf("125ns encoded into %d bytes of Protobuf wire format:\n% x\n", len(b), b)
   302:

Go Protobuf happens to return the already encoded bytes even when returning an error, so we can inspect the b byte slice to see how far the encoding got before the error happened:

(dlv) print b
[]uint8 len: 2, cap: 2, [16,125]

In this case, we can see that the entire (trivial) message was encoded, so our error must happen at a later stage — this allows us to rule out a large chunk of code in our search for the bug.

But what would we do if a longer part of the message was displayed and we wanted to load it into a different tool for further analysis, e.g. the excellent protoscope?

The low-tech approach is to print the contents and copy&paste from the delve output into an editor or similar. This stops working as soon as your data contains non-printable characters.

We have multiple options to export the byte slice to a file:

1. We could add os.WriteFile("/tmp/b.raw", b, 0644) to the source code and re-run the test. This is definitely the simplest option, as it works with or without a debugger.

2. As long as delve is connected to a running program, we can use delve’s call command to just execute the same code without having to add it to our source:

   (dlv) call os.WriteFile("/tmp/b.raw", b, 0644)
   (dlv)
   

Notably, both options only work when you can debug interactively. For the first option, you need to be able to change the source. The second option requires that delve is attached to a running process that you can afford to pause and interactively control.

These are trivial requirements when running a unit tests on your local machine, but get much harder when debugging an RPC service that crashes with specific requests, as you need to only run your changed debugging code for the troublesome requests, skipping the unproblematic requests that should still be handled normally.

Core dump debugging with Go

So let’s switch example: we are no longer working on Go Protobuf. Instead, we now need to debug an RPC service where certain requests crash the service. We’ll use core dump debugging!

In case you’re wondering: The name “core dump” comes from magnetic-core memory. These days we should probably say “memory dump” instead. The picture above shows an exhibit from the MIT Museum (Core Memory Unit, Bank C (from Project Whirlwind, 1953-1959)), a core memory unit with 4 KB of capacity.

To make Go write a core dump when panicing, run your program with the environment variable GOTRACEBACK=crash set (all possible values are documented in the runtime package).

You also need to ensure your system is set up to collect core dumps, as they are typically discarded by default:

• On Linux, the easiest way is to install systemd-coredump(8) , after which core dumps will automatically be collected. You can use coredumpctl(1) to list and work with them.
• On macOS, you can enable core dump collection, but delve cannot open macOS core dumps. Luckily, macOS is rarely used for production servers.
• I don’t know about Windows and other systems.

You can find more details and options in the CoreDumpDebugging page of the Go wiki. For this article, we will stick to the coredumpctl route:

We’ll use the gRPC Go Quick start example, a greeter client/server program, and add a panic() call to the server SayHello handler:

% cd greeter_server
% go build -gcflags=all="-N -l"  # disable optimizations
% GOTRACEBACK=crash ./greeter_server
2024/10/19 21:48:01 server listening at [::]:50051
2024/10/19 21:48:03 Received: world
panic: oh no!

goroutine 5 gp=0xc000007c00 m=5 mp=0xc000100008 [running]:
panic({0x83ca60?, 0x9a3710?})
	/home/michael/sdk/go1.23.0/src/runtime/panic.go:804 +0x168 fp=0xc000169850 sp=0xc0001697a0 pc=0x46fe88
main.(*server).SayHello(0xcbb840?, {0x877200?, 0xc000094900?}, 0x4a6f25?)
	/home/michael/go/src/github.com/grpc/grpc-go/examples/helloworld/greeter_server/main.go:45 +0xbf fp=0xc0001698c0 sp=0xc000169850 pc=0x8037ff
[…]
signal: aborted (core dumped)

The last line is what we want to see: it should say “core dumped”.

We can now use coredumpctl(1) to launch delve for this program + core dump:

% coredumpctl debug --debugger=dlv --debugger-arguments=core
           PID: 1729467 (greeter_server)
           UID: 1000 (michael)
           GID: 1000 (michael)
        Signal: 6 (ABRT)
     Timestamp: Sat 2024-10-19 21:50:12 CEST (1min 49s ago)
  Command Line: ./greeter_server
    Executable: /home/michael/go/src/github.com/grpc/grpc-go/examples/helloworld/greeter_server/greeter_server
 Control Group: /user.slice/user-1000.slice/session-1.scope
          Unit: session-1.scope
         Slice: user-1000.slice
       Session: 1
     Owner UID: 1000 (michael)
       Storage: /var/lib/systemd/coredump/core.greeter_server.1000.zst (present)
  Size on Disk: 204.7K
       Message: Process 1729467 (greeter_server) of user 1000 dumped core.
                
                Module /home/michael/go/src/github.com/grpc/grpc-go/examples/helloworld/greeter_server/greeter_server without build-id.
                Stack trace of thread 1729470:
                #0  0x0000000000479461 n/a (greeter_server + 0x79461)
[…]
                ELF object binary architecture: AMD x86-64

Type 'help' for list of commands.
(dlv) bt
 0  0x0000000000479461 in runtime.raise
    at /home/michael/sdk/go1.23.0/src/runtime/sys_linux_amd64.s:154
 1  0x0000000000451a85 in runtime.dieFromSignal
    at /home/michael/sdk/go1.23.0/src/runtime/signal_unix.go:942
 2  0x00000000004520e6 in runtime.sigfwdgo
    at /home/michael/sdk/go1.23.0/src/runtime/signal_unix.go:1154
 3  0x0000000000450a85 in runtime.sigtrampgo
    at /home/michael/sdk/go1.23.0/src/runtime/signal_unix.go:432
 4  0x0000000000479461 in runtime.raise
    at /home/michael/sdk/go1.23.0/src/runtime/sys_linux_amd64.s:153
 5  0x0000000000451a85 in runtime.dieFromSignal
    at /home/michael/sdk/go1.23.0/src/runtime/signal_unix.go:942
 6  0x0000000000439551 in runtime.crash
    at /home/michael/sdk/go1.23.0/src/runtime/signal_unix.go:1031
 7  0x0000000000439551 in runtime.fatalpanic
    at /home/michael/sdk/go1.23.0/src/runtime/panic.go:1290
 8  0x000000000046fe88 in runtime.gopanic
    at /home/michael/sdk/go1.23.0/src/runtime/panic.go:804
 9  0x00000000008037ff in main.(*server).SayHello
    at ./main.go:45
10  0x00000000008033a6 in google.golang.org/grpc/examples/helloworld/helloworld._Greeter_SayHello_Handler
    at /home/michael/go/src/github.com/grpc/grpc-go/examples/helloworld/helloworld/helloworld_grpc.pb.go:115
11  0x00000000007edeeb in google.golang.org/grpc.(*Server).processUnaryRPC
    at /home/michael/go/src/github.com/grpc/grpc-go/server.go:1394
12  0x00000000007f2eab in google.golang.org/grpc.(*Server).handleStream
    at /home/michael/go/src/github.com/grpc/grpc-go/server.go:1805
13  0x00000000007ebbff in google.golang.org/grpc.(*Server).serveStreams.func2.1
    at /home/michael/go/src/github.com/grpc/grpc-go/server.go:1029
14  0x0000000000477c21 in runtime.goexit
    at /home/michael/sdk/go1.23.0/src/runtime/asm_amd64.s:1700
(dlv) 

Alright! Now let’s switch to frame 9 (our server’s SayHello handler) and inspect the Name field of the incoming RPC request:

(dlv) frame 9
> runtime.raise() /home/michael/sdk/go1.23.0/src/runtime/sys_linux_amd64.s:154 (PC: 0x482681)
Warning: debugging optimized function
Frame 9: ./main.go:45 (PC: aaabf8)
    40:	}
    41:	
    42:	// SayHello implements helloworld.GreeterServer
    43:	func (s *server) SayHello(_ context.Context, in *pb.HelloRequest) (*pb.HelloReply, error) {
    44:		log.Printf("Received: %v", in.GetName())
=>  45:		panic("oh no!")
    46:		return &pb.HelloReply{Message: "Hello " + in.GetName()}, nil
    47:	}
    48:	
    49:	func main() {
    50:		flag.Parse()
(dlv) p in
("*google.golang.org/grpc/examples/helloworld/helloworld.HelloRequest")(0xc000120100)
*google.golang.org/grpc/examples/helloworld/helloworld.HelloRequest {
[…]
	unknownFields: []uint8 len: 0, cap: 0, nil,
	Name: "world",}

In this case, it’s easy to see that the Name field was set to world in the incoming request, but let’s assume the request contained lots of binary data that was not as easy to read or copy.

How do we write the byte slice contents to a file? In this scenario, we cannot modify the source code and delve’s call command does not work on core dumps (only when delve is attached to a running process):

(dlv) call os.WriteFile("/tmp/name.raw", in.Name, 0644)
> runtime.raise() /home/michael/sdk/go1.23.0/src/runtime/sys_linux_amd64.s:154 (PC: 0x482681)
Warning: debugging optimized function
Command failed: can not continue execution of core process

Luckily, we can extend delve with a custom Starlark function to write byte slice contents to a file.

Exporting byte slices with writebytestofile

You need a version of dlv that contains commit https://github.com/go-delve/delve/commit/52405ba86bd9e14a2e643db391cbdebdcbdb3368. Until the commit is part of a released version, you can install the latest dlv directly from git:

% go install github.com/go-delve/delve/cmd/dlv@master

Save the following Starlark code to a file, for example ~/dlv_writebytestofile.star:

# Syntax: writebytestofile <byte slice var> <output file path>
def command_writebytestofile(args):
	var_name, filename = args.split(" ")
	s = eval(None, var_name).Variable
	mem = examine_memory(s.Base, s.Len).Mem
	write_file(filename, mem)

Then, in delve, load the Starlark code and run the function to export the byte slice contents of in.Name to /tmp/name.raw:

% coredumpctl debug --debugger=dlv --debugger-arguments=core
(dlv) frame 9
(dlv) source ~/dlv_writebytestofile.star
(dlv) writebytestofile in.Name /tmp/name.raw

Let’s verify that we got the right contents:

% hexdump -C /tmp/name.raw
00000000  77 6f 72 6c 64                                    |world|
00000005

Core dump debugging with net/http servers

When you want to apply the core dump debugging technique on a net/http server (instead of a gRPC server, as above), you will notice that panics in your HTTP handlers do not actually result in a core dump! This code in go/src/net/http/server.go recovers panics and logs a stack trace:

defer func() {
    if err := recover(); err != nil && err != ErrAbortHandler {
        const size = 64 << 10
        buf := make([]byte, size)
        buf = buf[:runtime.Stack(buf, false)]
        c.server.logf("http: panic serving %v: %v\n%s", c.remoteAddr, err, buf)
    }
}()

Or, in other words: the GOTRACEBACK=crash environment variable configures what happens for unhandled signals, but this signal is handled with the recover() call, so no core is dumped.

This default behavior of net/http servers is now considered regrettable but cannot be changed for compatibility. (We probably can add a struct field to optionally not recover panics, though. I’ll update this paragraph once there is a proposal.)

So, what options do we have in the meantime?

We could recover panics in our own code (before net/http’s panic handler is called), but then how do we produce a core dump from our own handler?

A closer look reveals that the Go runtime’s crash function is defined in signal_unix.go and sends signal SIGABRT with the dieFromSignal function to the current thread:

//go:nosplit
func crash() {
        dieFromSignal(_SIGABRT)
}

The default action for SIGABRT is to “terminate the process and dump core”, see signal(7) .

We can follow the same strategy and send SIGABRT to our process:

func main() {
	http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
		defer func() {
			if err := recover(); err != nil {
				proc, err := os.FindProcess(syscall.Getpid())
				if err != nil {
					panic(fmt.Sprintf("could not find own process (pid %d): %v", syscall.Getpid(), err))
				}
				proc.Signal(syscall.SIGABRT)
				// Ensure the stack triggering the core dump sticks around
				proc.Wait()
			}
		}()
		// …buggy handler code goes here; for illustration we panic
		panic("this should result in a core dump")
	})
	log.Fatal(http.ListenAndServe(":8080", nil))
}

There is one caveat: If you have any non-Go threads running in your program, e.g. by using cgo, they might pick up the signal, so ensure they do not install a SIGABRT handler (see also: cgo-related documentation in os/signal).

If this is a concern, you can make the above code more platform-specific and use the tgkill(2) syscall to direct the signal to the current thread, as the Go runtime does.

Conclusion

Core dump debugging can be a very useful technique to quickly make progress on otherwise hard-to-debug problems. In small environments (single to few Linux servers), core dumps are easy enough to turn on and work with, but in larger environments you might need to invest into central core dump collection.

I hope the technique shown above comes in handy when you need to work with core dumps.

I have a couple of people who are best reachable on the Signal messaging app, but not that many. This exposes me to an awkward edge case of Signal’s design decisions: Whenever I get a message (on my phone), I want to reply to it (on my laptop) only to discover that Signal has un-linked my laptop because of inactivity and won’t sync the message history from my phone to my laptop, making it impossible to quote-reply to messages.

After complaining about this on Social Media for the n-th time, I figured I’d write a quick program to run Signal once a day, so that it won’t un-link my devices because of too long a period of inactivity. (Of course, the easiest solution would be to just run Signal in the background all the time. But I don’t use Signal often enough to justify letting it drain my battery and mobile data.)

In this article, I want to share the program in case it’s useful to anyone else, and also explain how to install it on a Mac, as this kind of “do a task once a day” automation is a useful pattern.

High-level sketch

1. Run Signal for, say, 5 minutes.
2. Ensure at-most-once semantics regardless of the task scheduler. For example, if I wanted to start this program from an @reboot hook and restart my computer a few times, I don’t want the program to do anything after the first run of the day. (Similarly, an on-online hook of NetworkManager or similar software might fire once per network interface, or something like that.)
3. Depending on the specifics of the activation mechanism, the computer might be online or not. The program should wait for a little while, say, 10 minutes, until internet connectivity was established.
4. I would like to log the program’s output (and Signal’s output) for debugging.

Checking connectivity

The easiest option is to just… not do a connectivity check at all, and hope for the best. This would probably work well enough in practice, but I would like the debug logs to have a high signal-to-noise ratio: If I have to debug why Signal was unlinked despite my automation attempts, I don’t want to comb through tons of spurious log messages that were a result from being offline. So, I want to check that I’m online before even starting Signal.

The most thorough option would be to somehow ask Signal programmatically whether it can connect to its servers and then wait until it can. I don’t think Signal has such an interface, so we’ll chose a middle-ground solution and work with a stand-in.

Using HTTP for connectivity checks is an easy way in today’s world. We just need a target website that doesn’t go offline unless I want it to. So let’s just use this website! Go’s net/http package that is included in Go’s standard library makes this super easy:

func checkConnectivity() error {
	_, err := http.Get("https://michael.stapelberg.ch/")
	return err
}

Now we just need to loop around this single connectivity check:

func waitForConnectivity(timeout time.Duration) error {
	const freq = 1 * time.Second
	for start := time.Now(); time.Since(start) < timeout; time.Sleep(freq) {
		if err := checkConnectivity(); err != nil {
			log.Printf("connectivity check failed: %v", err)
			continue
		}
		return nil // connectivity check succeeded
	}
	return fmt.Errorf("no connectivity established within %v", timeout)
}

We could improve this code to be more generally applicable by adding Exponential Backoff, but for this particular connectivity check, we should be fine even without Exponential Backoff.

Ensuring at-most-once semantics

An easy way to implement at-most-once semantics is to delegate to the file system: we can specify the O_EXCL flag when creating our program’s log file to make the first creation attempt proceed, but any further creation attempt fail because the file already exists. We’ll then redirect the standard library’s log package output to the log file:

logFn := filepath.Join(home, "signal-keepalive", "_logs", time.Now().Format("2006-01-02")+".txt")
f, err := os.OpenFile(logFn, os.O_RDWR|os.O_CREATE|os.O_EXCL, 0666)
if err != nil {
	if os.IsExist(err) {
		return nil // nothing to do, already ran today
	}
	return err
}
// Intentionally not closing this file so that even the log.Fatal()
// call in the calling function will end up in the log file.

log.SetOutput(f) // redirect standard library logging into this file
log.Printf("signal-keepalive, waiting for internet connectivity")

Not closing the file might seem weird at first, but remember that this is a short-lived program and the operating system closes all file handles of a process when it exits.

Full program code

For your convenience, here is the full program code. It contains a bunch of file system paths that you might want or need to adjust.

package main

import (
	"fmt"
	"log"
	"net/http"
	"os"
	"os/exec"
	"path/filepath"
	"time"
)

func checkConnectivity() error {
	_, err := http.Get("https://michael.stapelberg.ch/")
	return err
}

func waitForConnectivity(timeout time.Duration) error {
	for start := time.Now(); time.Since(start) < timeout; time.Sleep(1 * time.Second) {
		if err := checkConnectivity(); err != nil {
			log.Printf("connectivity check failed: %v", err)
			continue
		}
		return nil // connectivity check succeeded
	}
	return fmt.Errorf("no connectivity established within %v", timeout)
}

func keepalive() error {
	// Limit to one attempt per day by exclusively creating a logfile.
	home := os.Getenv("HOME")
	if home == "" {
		home = "/Users/michael"
	}
	logFn := filepath.Join(home, "signal-keepalive", "_logs", time.Now().Format("2006-01-02")+".txt")
	f, err := os.OpenFile(logFn, os.O_RDWR|os.O_CREATE|os.O_EXCL, 0666)
	if err != nil {
		if os.IsExist(err) {
			return nil // nothing to do, already ran today
		}
		return err
	}
	// Intentionally not closing this file so that even the log.Fatal()
	// call in the calling function will end up in the log file.

	log.SetOutput(f) // redirect standard library logging into this file
	log.Printf("signal-keepalive, waiting for internet connectivity")

	// Wait for network connectivity
	if err := waitForConnectivity(10 * time.Minute); err != nil {
		return err
	}

	// Start signal
	log.Printf("connectivity verified, starting signal")
	signal := exec.Command("/Applications/Signal.app/Contents/MacOS/Signal", "--start-in-tray")
	signal.Stdout = f
	signal.Stderr = f
	if err := signal.Start(); err != nil {
		return err
	}

	// Wait for some time to give Signal a chance to synchronize messages.
	const signalWaitTime = 5 * time.Minute
	log.Printf("giving signal %v to sync messages", signalWaitTime)
	time.Sleep(signalWaitTime)

	// Stop signal
	log.Printf("killing signal")
	if err := signal.Process.Kill(); err != nil {
		return err
	}
	log.Printf("waiting for signal")
	log.Printf("signal returned: %v", signal.Wait())
	log.Printf("all done")

	return f.Sync()
}

func main() {
	if err := keepalive(); err != nil {
		log.Fatal(err)
	}
}

(Use go build keepalive.go to compile if you’re unfamiliar with Go.)

macOS installation: launchd

The corresponding piece of infrastructure to systemd on Linux is called launchd on macOS. Aside from managing daemon processes, launchd also supports time-triggered program execution, specifically via the StartCalendarInterval configuration option.

I followed Alvin Alexander’s blog post about launchd StartCalendarInterval examples and decided to configure my program to run at 08:03 each day:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple Computer//DTD PLIST 1.0//EN"
"http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>StartCalendarInterval</key>
    <dict>
      <key>Hour</key>
      <integer>8</integer>
      <key>Minute</key>
      <integer>3</integer>
    </dict>
    <key>Label</key>
    <string>net.zekjur.signalkeepalive</string>
    <key>Program</key>
    <string>/Users/michael/signal-keepalive/signalkeepalive</string>
</dict>
</plist>

What happens when my computer isn’t running at 08:03, for example because the lid is closed? Apple documents the behavior in the launchd.plist(5) man page:

Unlike cron which skips job invocations when the computer is asleep, launchd will start the job the next time the computer wakes up.

To install and test this configuration:

1. Copy the plist file to ~/Library/LaunchAgents
2. Run launchctl load ~/Library/LaunchAgents/net.zekjur.signalkeepalive.plist
3. Run launchctl start net.zekjur.signalkeepalive

In practice

It’s interesting to see this behavior in practice. Take note of the time stamps in the following log. The computer was not running at 08:03. At 08:18, it woke up to update background information (Apple calls this “Power Nap”), and then it suspended again (while Signal was running) until it woke up at 08:47 again:

2024/09/07 08:18:10 signal-keepalive, waiting for internet connectivity
2024/09/07 08:18:11 connectivity verified, starting signal
2024/09/07 08:18:11 giving signal 5m0s to sync messages
Set Windows Application User Model ID (AUMID) { AUMID: 'org.whispersystems.signal-desktop' }
NODE_ENV production
NODE_CONFIG_DIR /Applications/Signal.app/Contents/Resources/app.asar/config
NODE_CONFIG {}
ALLOW_CONFIG_MUTATIONS undefined
HOSTNAME m1a.fritz.box
NODE_APP_INSTANCE undefined
SUPPRESS_NO_CONFIG_WARNING undefined
SIGNAL_ENABLE_HTTP undefined
userData: /Users/michael/Library/Application Support/Signal
config/get: Successfully read user config file
config/get: Successfully read ephemeral config file
2024/09/07 08:47:31 killing signal
2024/09/07 08:47:31 waiting for signal
2024/09/07 08:47:31 signal returned: signal: killed
2024/09/07 08:47:31 all done

Linux installation: systemd

With systemd, we need two units. First, a signal-keepalive.service unit to declare which program should be run:

cat > ~/.config/systemd/user/signal-keepalive.service <<'EOT'
[Unit]
Description=signal keepalive
After=network.target

[Service]
Type=oneshot
ExecStart=/home/michael/signal-keepalive/signalkeepalive

[Install]
WantedBy=default.target
EOT

And secondly, a signal-keepalive.timer unit which automatically starts the signal-keepalive.service every day:

cat > ~/.config/systemd/user/signal-keepalive.timer <<'EOT'
[Unit]
Description=signal keepalive

[Timer]
Persistent=true
OnCalendar=daily

[Install]
WantedBy=timers.target
EOT

The Persistent=true line is important so that the program will be run even when the computer is asleep when the timer would have fired.

Let’s enable the timer:

systemctl --user enable --now signal-keepalive.timer

For an initial test run, we can start the .service directly:

systemctl --user restart signal-keepalive.service

Conclusion

It’s silly that I need to go through so much trouble just because I don’t use Signal enough.

I also don’t understand why Signal can’t just sync message history from my phone to my computer when linking. WhatsApp and Telegram have no trouble doing it.

Either way, I thought this was a fun little refresher on automating periodic jobs.

When I saw the first reviews of the ASRock DeskMini X600 barebone, I was immediately interested in building a home-lab hypervisor (VM host) with it. Apparently, the DeskMini X600 uses less than 10W of power but supports latest-generation AMD CPUs like the Ryzen 7 8700G!

Sounds like the perfect base for a power-efficient, always-on VM host that still provides enough compute power (and fast disks!) to be competitive with commercial VM offerings. In this article, I’ll show how I built and set up my DIY self-hosting VM host.

Component List

The term “barebone” means that the machine comes without CPU, RAM and disk. You only get a case with a mainboard and power supply, the rest is up to you. I chose the following parts:

Total cost: 969 CHF

The CPU fan is not strictly required (the DeskMini X600 already comes with a fan), but I wanted the best cooling performance at lowest noise levels, so Noctua it is.

I read that the machine should support ECC RAM, too. Update: The Ryzen 8700G does not support ECC-RAM after all. Only the Ryzen 7 PRO 8700G supports ECC-RAM.

It took me about an hour to assemble the parts. Note that the M.2 SSD screws might seem a little hard to screw in, but don’t be deterred by that. When first powering on the system, be patient as the memory training will take a minute or so, during which the screen will stay black.

UEFI Setup

The UEFI on the DeskMini X600 comes with reasonable defaults.

The CPU fan setting alreadys defaults to “Silent Mode”, for example.

I changed the following option, which is typical for server usage:

• Advanced → ACPI Configuration → Restore on AC/Power Loss: Power On

And I disabled the onboard devices I know I won’t need, just in case it saves power:

• Advanced → Onboard Devices Configuration → Onboard HD Audio: Disabled
• SATA3 Controller: Disabled

Operating System Setup

I want to run this machine as a VM hypervisor. The easiest way that I know to set up such a hypervisor is to install Proxmox, an open source virtualization appliance based on Debian.

I booted the machine with the Proxmox installer copied to a USB memory stick, then selected ZFS in a RAID-1 configuration. The setup worked smoothly and was done in a few minutes.

Then, I set up Tailscale as recommended and used tailscale serve so that I can access the Proxmox web interface on its Tailscale hostname via HTTPS, instead of having to deal with certificates and custom ports:

pve# curl -fsSL https://tailscale.com/install.sh | sh
pve# tailscale up
[…]
  follow instructions and disable key expiration
[…]
pve# tailscale serve --bg https+insecure://localhost:8006

(Of course I’ll also install Tailscale on each VM running on the host.)

Now I can log into the Proxmox web interface from anywhere without certificate warnings:

In this screenshot, I have already created 2 VMs (“batch” and “web”) using the “Create VM” button at the top right. Proxmox allows controlling the installer via its “Console” tab and once set up, the VM shows up in the same network that the hypervisor is connected to with a MAC address from the “Proxmox Server Solutions GmbH” range. That’s pretty much all there is to it.

I don’t have enough nodes for advanced features like clustering, but I might investigate whether I want to set up backups on the Proxmox layer or keep doing them on the OS layer.

Power Usage

The power usage values I measure are indeed excellent: The DeskMini X600 with Ryzen 7 8700G consumes less than 10W (idle)! When the machine has something to do, it spikes up to 50W:

Noise

ASRock explicitly lists the Noctua NH-L9a-AM5 as compatible with the DeskMini X600, which was one of the factors that made me select this barebone. Installing the fan was easy.

Fan noise is very low, as expected with Noctua. I can’t hear the device even when it is standing in front of me on my desk. Of course, under heavy load, the fan will be audible. This is an issue with all small form-factor PCs, as they just don’t have enough case space to swallow more noise.

Aside from the fan noise, if you hold your ear directly next to the X600, you can hear the usual electrical component noise (not coil whine per se, but that sort of thing).

I recommend positioning this device under a desk, or on a shelf, or similar.

Performance comparison

You can find synthetic benchmark results for the Ryzen 8700G elsewhere, so as usual, I will write about the specific angle I care about: How fast can this machine handle Go workloads?

Compiling Go 1.22.4

On the Ryzen 8700G, we can compile Go 1.22.4 in a little under 40 seconds:

% time ./make.bash
[…]
./make.bash  208,55s user 36,96s system 631% cpu 38,896 total

For comparison, my 2022 high-end Linux PC with Core i9-12900K is only a few seconds faster:

% time ./make.bash
[…]
./make.bash  207,33s user 29,55s system 685% cpu 34,550 total

Go HTTP and JSON benchmarks

I also ran the HTTP and JSON benchmarks from Go’s x/benchmarks repository.

Compared to the Virtual Server I’m currently renting, the Ryzen 8700G is more than twice as fast:

% benchstat rentedvirtual ryzen8700g 
name    old time/op                  new time/op                  delta
HTTP-2  28.5µs ± 2%                  10.2µs ± 1%  -64.17%  (p=0.008 n=5+5)
JSON-2  24.1ms ±29%                   9.4ms ± 1%  -61.06%  (p=0.008 n=5+5)

Of course, the Intel i9 12900K is still a bit faster — how much depends on the specific workload:

% benchstat ryzen8700g i9_12900k 
name    old time/op                  new time/op                  delta
HTTP-2  10.2µs ± 1%                   7.6µs ± 1%  -25.13%  (p=0.008 n=5+5)
JSON-2  9.40ms ± 1%                  9.23ms ± 1%   -1.82%  (p=0.008 n=5+5)

Conclusion

What a delightful little Mini-PC! It’s modern enough to house the current generation of CPUs, compact enough to fit in well anywhere, yet just large enough to fit a Noctua CPU cooler for super-quiet operation. The low power draw makes it acceptable to run this machine 24/7.

Paired with 64 GB of RAM and large, fast NVMe disks, this machine packs a punch and will easily power your home automation, home lab, hobby project, small office server, etc.

If a Raspberry Pi isn’t enough for your needs, check out the DeskMini X600, or perhaps its larger variant, the DeskMeet X600 which is largely identical, but comes with a PCIe slot.

If this one doesn’t fit your needs, keep looking: there are many more mini PCs on the market. Check out ServeTheHome’s “Project TinyMiniMicro” for a lot more reviews.

Update: Apparently ASRock is releasing their X600 mainboard as a standalone product, too, if you like the electronics but not the form factor.

Sometimes, you need to be able to constrain a type-parameter with a method, but that method should be defined on the pointer type. For example, say you want to parse some bytes using JSON and pass the result to a handler. You might try to write this as

func Handle[M json.Unmarshaler](b []byte, handler func(M) error) error {
	var m M
	if err := m.UnmarshalJSON(b); err != nil {
		return err
	}
	return handler(m)
}

However, this code does not work. Say you have a type Message, which implements json.Unmarshaler with a pointer receiver (it needs to use a pointer receiver, as it needs to be able to modify data):

1. If you try to call Handle[Message], you get a compiler error (playground). That is because Message does not implement json.Unmarshal, only *Message does.
2. If you try to call Handle[*Message], the code panics (playground), because var m M creates a *Message and initializes that to nil. You then call UnmarshalJSON with a nil receiver.

Neither of these options work. You really want to rewrite Handle, so that it says that the pointer to its type parameter implements json.Unmarshaler. And this is how to do that (playground):

type Unmarshaler[M any] interface {
	*M
	json.Unmarshaler
}

func Handle[M any, PM Unmarshaler[M]](b []byte, handler func(M) error) error {
	var m M
	// note: you need PM(&m), as the compiler can not infer (yet) that you can
	// call the method of PM on a pointer to M.
	if err := PM(&m).UnmarshalJSON(b); err != nil {
		return err
	}
	return handler(m)
}

I maintain two builds of the Linux kernel, a linux/arm64 build for gokrazy, my Go appliance platform, which started out on the Raspberry Pi, and then a linux/amd64 one for router7, which runs on PCs.

The update process for both of these builds is entirely automated, meaning new Linux kernel releases are automatically tested and merged, but recently the continuous integration testing failed to automatically merge Linux 6․7 — this article is about tracking down the root cause of that failure.

Background info on the bootloader

gokrazy started out targeting only the Raspberry Pi, where you configure the bootloader with a plain text file on a FAT partition, so we did not need to include our own UEFI/MBR bootloader.

When I ported gokrazy to work on PCs in BIOS mode, I decided against complicated solutions like GRUB — I really wasn’t looking to maintain a GRUB package. Just keeping GRUB installations working on my machines is enough work. The fact that GRUB consists of many different files (modules) that can go out of sync really does not appeal to me.

Instead, I went with Sebastian Plotz’s Minimal Linux Bootloader because it fits entirely into the Master Boot Record (MBR) and does not require any files. In bootloader lingo, this is a stage1-only bootloader. You don’t even need a C compiler to compile its (Assembly) code. It seemed simple enough to integrate: just write the bootloader code into the first sector of the gokrazy disk image; done. The bootloader had its last release in 2012, so no need for updates or maintenance.

You can’t really implement booting a kernel and parsing text configuration files in 446 bytes of 16-bit 8086 assembly instructions, so to tell the bootloader where on disk to load the kernel code and kernel command line from, gokrazy writes the disk offset (LBA) of vmlinuz and cmdline.txt to the last bytes of the bootloader code. Because gokrazy generates the FAT partition, we know there is never any fragmentation, so the bootloader does not need to understand the FAT file system.

Symptom

The symptom was that the rtr7/kernel pull request #434 for updating to Linux 6.7 failed.

My continuous integration tests run in two environments: a physical embedded PC from PC Engines (apu2c4) in my living room, and a virtual QEMU PC. Only the QEMU test failed.

On the physical PC Engines apu2c4, the pull request actually passed the boot test. It would be wrong to draw conclusions like “the issue only affects QEMU” from this, though, as later attempts to power on the apu2c4 showed the device boot-looping. I made a mental note that something is different about how the problem affects the two environments, but both are affected, and decided to address the failure in QEMU first, then think about the PC Engines failure some more.

In QEMU, the output I see is:

SeaBIOS (version Arch Linux 1.16.3-1-1)

iPXE (http://ipxe.org) 00:03.0 C900 PCI2.10 PnP PMM+06FD3360+06F33360 C900

Booting from Hard Disk...

Notably, the kernel doesn’t even seem to start — no “Decompressing linux” message is printed, the boot just hangs. I tried enabling debug output in SeaBIOS and eventually succeeded, but only with an older QEMU version:

Booting from Hard Disk...
Booting from 0000:7c00
In resume (status=0)
In 32bit resume
Attempting a hard reboot

This doesn’t tell me anything unfortunately.

Okay, so something about introducing Linux 6.7 into my setup breaks MBR boot.

I figured using Git Bisection should identify the problematic change within a few iterations, so I cloned the currently working Linux 6.6 source code, applied the router7 config and compiled it.

To my surprise, even my self-built Linux 6.6 kernel would not boot! 😲

Why does the router7 build work when built inside the Docker container, but not when built on my Linux installation? I decided to rebase the Docker container from Debian 10 (buster, from 2019) to Debian 12 (bookworm, from 2023) and that resulted in a non-booting kernel, too!

We have two triggers: building Linux 6.7 or building older Linux, but in newer environments.

Meta: Following Along

% mkdir -p go/src/github.com/rtr7/
% cd go/src/github.com/rtr7/
% git clone --depth=1 https://github.com/rtr7/kernel
% cd kernel
% sed -i 's,CONFIG_KERNEL_ZSTD,#CONFIG_KERNEL_ZSTD,g' cmd/rtr7-build-kernel/config.addendum.txt
% go run ./cmd/rtr7-rebuild-kernel
# takes a few minutes to compile Linux
% ls -l vmlinuz
-rw-r--r-- 1 michael michael 15885312 2024-01-28 16:18 vmlinuz

% gok -i mbr new
% gok -i mbr add .
% gok -i mbr edit
# Adjust to contain:
    "KernelPackage": "github.com/rtr7/kernel",
    "FirmwarePackage": "github.com/rtr7/kernel",
    "EEPROMPackage": "",

% GOARCH=amd64 gok -i mbr overwrite \
  --full /tmp/gokr-boot.img \
  --target_storage_bytes=1258299392
% qemu-system-i386 \
  -nographic \
  -drive file=/tmp/gokr-boot.img,format=raw

Up/Downgrade Versions

Unlike application programs, the Linux kernel doesn’t depend on shared libraries at runtime, so the dependency footprint is a little smaller than usual. The most significant dependencies are the components of the build environment, like the C compiler or the linker.

So let’s look at the software versions of the known-working (Debian 10) environment and the smallest change we can make to that (upgrading to Debian 11):

• Debian 10 (buster) contains gcc-8 (8.3.0-6) and binutils 2.31.1-16.
• Debian 11 (bullseye) contains gcc-10 (10.2.1-6) and binutils 2.35.2-2.

To figure out if the problem is triggered by GCC, binutils, or something else entirely, I checked:

Debian 10 (buster) with its gcc-8, but with binutils 2.35 from bullseye still works. (Checked by updating /etc/apt/sources.list, then upgrading only the binutils package.)

Debian 10 (buster), but with gcc-10 and binutils 2.35 results in a non-booting kernel.

So it seems like upgrading from GCC 8 to GCC 10 triggers the issue.

Instead of working with a Docker container and Debian’s packages, you could also use Nix. The instructions aren’t easy, but I used nix-shell to quickly try out GCC 8 (works), GCC 9 (works) and GCC 10 (kernel doesn’t boot) on my machine.

New Hypothesis

To recap, we have two triggers: building Linux 6.7 or building older Linux, but with GCC 10.

Two theories seemed most plausible to me at this point: Either a change in GCC 10 (possibly enabled by another change in Linux 6.7) is the problem, or the size of the kernel is the problem.

To verify the file size hypothesis, I padded a known-working vmlinuz file to the size of a known-broken vmlinuz:

% ls -l vmlinuz
% dd if=/dev/zero bs=108352 count=1 >> vmlinuz

But, even though it had the same file size as the known-broken kernel, the padded kernel booted!

So I ruled out kernel size as a problem and started researching significant changes in GCC 10.

I read that GCC 10 changed behavior with regards to stack protection.

Indeed, building the kernel with Debian 11 (bullseye), but with CONFIG_STACKPROTECTOR=n makes it boot. So, I suspected that our bootloader does not set up the stack correctly, or similar.

I sent an email to Sebastian Plotz, the author of the Minimal Linux Bootloader, to ask if he knew about any issues with his bootloader, or if stack protection seems like a likely issue with his bootloader to him.

To my surprise (it has been over 10 years since he published the bootloader!) he actually replied: He hadn’t received any problem reports regarding his bootloader, but didn’t really understand how stack protection would be related.

Debugging with QEMU

At this point, we have isolated at least one trigger for the problem, and exhausted the easy techniques of upgrading/downgrading surrounding software versions and asking upstream.

It’s time for a Tooling Level Up! Without a debugger you can only poke into the dark, which takes time and doesn’t result in thorough explanations. Particularly in this case, I think it is very likely that any source modifications could have introduced subtle issues. So let’s reach for a debugger!

Luckily, QEMU comes with built-in support for the GDB debugger. Just add the -s -S flags to your QEMU command to make QEMU stop execution (-s) and set up a GDB stub (-S) listening on localhost:1234.

If you wanted to debug the Linux kernel, you could connect GDB to QEMU right away, but for debugging a boot loader we need an extra step, because the boot loader runs in Real Mode, but QEMU’s GDB integration rightfully defaults to the more modern Protected Mode.

When GDB is not configured correctly, it decodes addresses and registers with the wrong size, which throws off the entire disassembly — compare GDB’s output with our assembly source:

(gdb) b *0x7c00
(gdb) c
(gdb) x/20i $pc                         ; [expected (bootloader.asm)]
=> 0x7c00: cli                          ; => 0x7c00: cli
   0x7c01: xor    %eax,%eax             ;    0x7c01: xor %ax,%ax
   0x7c03: mov    %eax,%ds              ;    0x7c03: mov %ax,%ds
   0x7c05: mov    %eax,%ss              ;    0x7c05: mov %ax,%ss
   0x7c07: mov    $0xb87c00,%esp        ;    0x7c07: mov $0x7c00,%sp
   0x7c0c: adc    %cl,-0x47990440(%esi) ;    0x7c0a: mov $0x1000,%ax
   0x7c12: add    %eax,(%eax)           ;    0x7c0d: mov %ax,%es
   0x7c14: add    %al,(%eax)            ;    0x7c0f: sti
   0x7c16: xor    %ebx,%ebx

So we need to ensure we use qemu-system-i386 (qemu-system-x86_64 prints Remote 'g' packet reply is too long) and configure the GDB target architecture to 16-bit 8086:

(gdb) set architecture i8086
(gdb) target remote localhost:1234

Unfortunately, the above doesn’t actually work in QEMU 2.9 and newer: https://gitlab.com/qemu-project/qemu/-/issues/141.

On the web, people are working around this bug by using a modified target.xml file. I tried this, but must have made a mistake — I thought modifying target.xml didn’t help, but when I wrote this article, I found that it does actually seem to work. Maybe I didn’t use qemu-system-i386 but the x86_64 variant or something like that.

Using an older QEMU

It is typically an exercise in frustration to get older software to compile in newer environments.

It’s much easier to use an older environment to run old software.

By querying packages.debian.org, we can see the QEMU versions included in current and previous Debian versions.

Unfortunately, the oldest listed version (QEMU 3.1 in Debian 10 (buster)) isn’t old enough. By querying snapshot.debian.org, we can see that Debian 9 (stretch) contained QEMU 2.8.

So let’s run Debian 9 — the easiest way I know is to use Docker:

% docker run --net=host -v /tmp:/tmp -ti debian:stretch

Unfortunately, the debian:stretch Docker container does not work out of the box anymore, because its /etc/apt/sources.list points to the deb.debian.org CDN, which only serves current versions and no longer serves stretch.

So we need to update the sources.list file to point to archive.debian.org. To correctly install QEMU you need both entries, the debian line and the debian-security line, because the Docker container has packages from debian-security installed and gets confused when these are missing from the package list:

root@650a2157f663:/# cat > /etc/apt/sources.list <<'EOT'
deb http://archive.debian.org/debian/ stretch contrib main non-free
deb http://archive.debian.org/debian-security/ stretch/updates main
EOT
root@650a2157f663:/# apt update

Now we can just install QEMU as usual and start it to debug our boot process:

root@650a2157f663:/# apt install qemu-system-x86
root@650a2157f663:/# qemu-system-i386 \
  -nographic \
  -drive file=/tmp/gokr-boot.img,format=raw \
  -s -S

Now let’s start GDB and set a breakpoint on address 0x7c00, which is the address to which the BIOS loades the MBR code and starts execution:

% gdb
(gdb) set architecture i8086
The target architecture is set to "i8086".
(gdb) target remote localhost:1234
Remote debugging using localhost:1234
0x0000fff0 in ?? ()
(gdb) break *0x7c00
Breakpoint 1 at 0x7c00
(gdb) continue
Continuing.

Breakpoint 1, 0x00007c00 in ?? ()
(gdb)

Debug symbols

Okay, so we have GDB attached to QEMU and can step through assembly instructions. Let’s start debugging!?

Not so fast. There is another Tooling Level Up we need first: debug symbols. Yes, even for a Minimal Linux Bootloader, which doesn’t use any libraries or local variables. Having proper names for functions, as well as line numbers, will be hugely helpful in just a second.

Before debug symbols, I would directly build the bootloader using nasm bootloader.asm, but to end up with a symbol file for GDB, we need to instruct nasm to generate an ELF file with debug symbols, then use ld to link it and finally use objcopy to copy the code out of the ELF file again.

After commit d29c615 in gokrazy/internal/mbr, I have bootloader.elf.

Back in GDB, we can load the symbols using the symbol-file command:

(gdb) set architecture i8086
The target architecture is set to "i8086".
(gdb) target remote localhost:1234
Remote debugging using localhost:1234
0x0000fff0 in ?? ()
(gdb) symbol-file bootloader.elf
Reading symbols from bootloader.elf...
(gdb) break *0x7c00
Breakpoint 1 at 0x7c00: file bootloader.asm, line 48.
(gdb) continue
Continuing.

Breakpoint 1, ?? () at bootloader.asm:48
48		cli
(gdb)

Automation with .gdbinit

At this point, we need 4 commands each time we start GDB. We can automate these by writing them to a .gdbinit file:

% cat > .gdbinit <<'EOT'
set architecture i8086
target remote localhost:1234
symbol-file bootloader.elf
break *0x7c00
EOT

% gdb
The target architecture is set to "i8086".
0x0000fff0 in ?? ()
Breakpoint 1 at 0x7c00: file bootloader.asm, line 48.
(gdb) 

Understanding program flow

The easiest way to understand program flow seems to be to step through the program.

But Minimal Linux Bootloader (MLB) contains loops that run through thousands of iterations. You can’t use gdb’s stepi command with that.

Because MLB only contains a few functions, I eventually realized that placing a breakpoint on each function would be the quickest way to understand the high-level program flow:

(gdb) b read_kernel_setup
Breakpoint 2 at 0x7c38: file bootloader.asm, line 75.
(gdb) b check_version
Breakpoint 3 at 0x7c56: file bootloader.asm, line 88.
(gdb) b read_protected_mode_kernel
Breakpoint 4 at 0x7c8f: file bootloader.asm, line 105.
(gdb) b read_protected_mode_kernel_2
Breakpoint 5 at 0x7cd6: file bootloader.asm, line 126.
(gdb) b run_kernel
Breakpoint 6 at 0x7cff: file bootloader.asm, line 142.
(gdb) b error
Breakpoint 7 at 0x7d51: file bootloader.asm, line 190.
(gdb) b reboot
Breakpoint 8 at 0x7d62: file bootloader.asm, line 204.

With the working kernel, we get the following transcript:

(gdb)
Continuing.

Breakpoint 2, read_kernel_setup () at bootloader.asm:75
75		xor	eax, eax
(gdb)
Continuing.

Breakpoint 3, check_version () at bootloader.asm:88
88		cmp	word [es:0x206], 0x204		; we need protocol version >= 2.04
(gdb)
Continuing.

Breakpoint 4, read_protected_mode_kernel () at bootloader.asm:105
105		mov	edx, [es:0x1f4]			; edx stores the number of bytes to load
(gdb)
Continuing.

Breakpoint 5, read_protected_mode_kernel_2 () at bootloader.asm:126
126		mov	eax, edx
(gdb)
Continuing.

Breakpoint 6, run_kernel () at bootloader.asm:142
142		cli
(gdb)

With the non-booting kernel, we get:

(gdb) c
Continuing.

Breakpoint 1, ?? () at bootloader.asm:48
48		cli
(gdb)
Continuing.

Breakpoint 2, read_kernel_setup () at bootloader.asm:75
75		xor	eax, eax
(gdb)
Continuing.

Breakpoint 3, check_version () at bootloader.asm:88
88		cmp	word [es:0x206], 0x204		; we need protocol version >= 2.04
(gdb)
Continuing.

Breakpoint 4, read_protected_mode_kernel () at bootloader.asm:105
105		mov	edx, [es:0x1f4]			; edx stores the number of bytes to load
(gdb)
Continuing.

Breakpoint 1, ?? () at bootloader.asm:48
48		cli
(gdb)

Okay! Now we see that the bootloader starts loading the kernel from disk into RAM, but doesn’t actually get far enough to call run_kernel, meaning the problem isn’t with stack protection, with loading a working command line or with anything inside the Linux kernel.

This lets us rule out a large part of the problem space. We now know that we can focus entirely on the bootloader and why it cannot load the Linux kernel into memory.

Let’s take a closer look…

Wait, this isn’t GDB!

In the example above, using breakpoints was sufficient to narrow down the problem.

You might think we used GDB, and it looked like this:

But that’s not GDB! It’s an easy mistake to make. After all, GDB starts up with just a text prompt, and as you can see from the example above, we can just enter text and achieve a good result.

To see the real GDB, you need to start it up fully, meaning including its user interface.

You can either use GDB’s text user interface (TUI), or a graphical user interface for gdb, such as the one available in Emacs.

The GDB text-mode user interface (TUI)

You’re already familiar with the architecture, target and breakpoint commands from above. To also set up the text-mode user interface, we run a few layout commands:

(gdb) set architecture i8086
(gdb) target remote localhost:1234
(gdb) symbol-file bootloader.elf
(gdb) layout split
(gdb) layout src
(gdb) layout regs
(gdb) break *0x7c00
(gdb) continue

The layout split command loads the text-mode user interface and splits the screen into a register window, disassembly window and command window.

With layout src we disregard the disassembly window in favor of a source listing window. Both are in assembly language in our case, but the source listing contains comments as well.

The layout src command also got rid of the register window, which we’ll get back using layout regs. I’m not sure if there’s an easier way.

The result looks like this:

The source window will highlight the next line of code that will be executed. On the left, the B+ marker indicates an enabled breakpoint, which will become helpful with multiple breakpoints. Whenever a register value changes, the register and its new value will be highlighted.

The up and down arrow keys scroll the source window.

Use C-x o to switch between the windows.

If you’re familiar with Emacs, you’ll recognize the keyboard shortcut. But as an Emacs user, you might prefer the GDB Emacs user interface:

The GDB Emacs user interface (M-x gdb)

This is M-x gdb with gdb-many-windows enabled:

Debugging the failing loop

Let’s take a look at the loop that we know the bootloader is entering, but not leaving (neither read_protected_mode_kernel_2 nor run_kernel are ever called):

read_protected_mode_kernel:
    mov  edx, [es:0x1f4]              ; edx stores the number of bytes to load
    shl  edx, 4

.loop:
    cmp  edx, 0
    je   run_kernel

    cmp  edx, 0xfe00                  ; less than 127*512 bytes remaining?
    jb   read_protected_mode_kernel_2

    mov  eax, 0x7f                    ; load 127 sectors (maximum)
    xor  bx, bx                       ; no offset
    mov  cx, 0x2000                   ; load temporary to 0x20000
    mov  esi, current_lba
    call read_from_hdd

    mov  cx, 0x7f00                   ; move 65024 bytes (127*512 byte)
    call do_move

    sub  edx, 0xfe00                  ; update the number of bytes to load
    add  word [gdt.dest], 0xfe00
    adc  byte [gdt.dest+2], 0
    jmp  short read_protected_mode_kernel.loop

The comments explain that the code loads chunks of FE00h == 65024 (127*512) bytes at a time.

Loading means calling read_from_hdd, then do_move. Let’s take a look at do_move:

do_move:
    push edx
    push es
    xor  ax, ax
    mov  es, ax
    mov  ah, 0x87
    mov  si, gdt
    int  0x15     ; line 182
    jc   error
    pop  es
    pop  edx
    ret

int 0x15 is a call to the BIOS Service Interrupt, which will dispatch the call based on AH == 87H to the Move Memory Block (techhelpmanual.com) function.

This function moves the specified amount of memory (65024 bytes in our case) from source/destination addresses specified in a Global Descriptor Table (GDT) record.

We can use GDB to show the addresses of each of do_move’s memory move calls by telling it to stop at line 182 (the int 0x15 instruction) and print the GDT record’s destination descriptor:

(gdb) break 182
Breakpoint 2 at 0x7d49: file bootloader.asm, line 176.

(gdb) command 2
Type commands for breakpoint(s) 2, one per line.
End with a line saying just "end".
>x/8bx gdt+24
>end

(gdb) continue
Continuing.

Breakpoint 1, ?? () at bootloader.asm:48
42		cli

(gdb)
Continuing.

Breakpoint 2, do_move () at bootloader.asm:182
182		int	0x15
0x7d85:	0xff	0xff	0x00	0x00	0x10	0x93	0x00	0x00

(gdb)
Continuing.

Breakpoint 2, do_move () at bootloader.asm:182
182		int	0x15
0x7d85:	0xff	0xff	0x00	0xfe	0x10	0x93	0x00	0x00

(gdb)

The destination address is stored in byte 2..4. Remember to read these little endian entries “back to front”.

• Address #1 is 0x100000.

• Address #2 is 0x10fe00.

If we press Return long enough, we eventually end up here:

Breakpoint 2, do_move () at bootloader.asm:182
182		int	0x15
0x7d85:	0xff	0xff	0x00	0x1e	0xff	0x93	0x00	0x00
(gdb)
Continuing.

Breakpoint 2, do_move () at bootloader.asm:182
182		int	0x15
0x7d85:	0xff	0xff	0x00	0x1c	0x00	0x93	0x00	0x00

(gdb)
Continuing.

Breakpoint 1, ?? () at bootloader.asm:48
42		cli
(gdb)

Program received signal SIGTRAP, Trace/breakpoint trap.
0x000079b0 in ?? ()
(gdb)

Now that execution left the bootloader, let’s take a look at the last do_move call parameters: We notice that the destination address overflowed its 24 byte data type:

• Address #y is 0xff1e00
• Address #z is 0x001c00

Root cause

At this point I reached out to Sebastian again to ask him if there was an (undocumented) fundamental architectural limit to his Minimal Linux Bootloader — with 24 bit addresses, you can address at most 16 MB of memory.

He replied explaining that he didn’t know of this limit either! He then linked to Move Memory Block (techhelpmanual.com) as proof for the 24 bit limit.

Speculation

So, is it impossible to load larger kernels into memory from Real Mode? I’m not sure.

The current bootloader code prepares a GDT in which addresses are 24 bits long at most. But note that the techhelpmanual.com documentation that Sebastian referenced is apparently for the Intel 286 (a 16 bit CPU), and some of the GDT bytes are declared reserved.

Today’s CPUs are Intel 386-compatible (a 32 bit CPU), which seems to use one of the formerly reserved bytes to represent bit 24..31 of the address, meaning we might be able to pass 32 bit addresses to BIOS functions in a GDT after all!

I wasn’t able to find clear authoritative documentation on the Move Memory Block API on 386+, or whether BIOS functions in general are just expected to work with 32 bit addresses.

But Microsoft’s 1989 HIMEM.SYS source contains a struct that documents this 32-bit descriptor usage. A more modern reference is this Operating Systems Class from FAU 2023 (page 71/72).

Hence I’m thinking that most BIOS implementations should actually support 32 bit addresses for their Move Memory Block implementation — provided you fill the descriptor accordingly.

If that doesn’t work out, there’s also “Unreal Mode”, which allows using up to 4 GB in Real Mode, but is a change that is a lot more complicated. See also Julio Merino’s “Beyond the 1 MB barrier in DOS” post to get an idea of the amount of code needed.

Update: a fix!

Lobsters reader abbeyj pointed out that the following code change should fix the truncation and result in a GDT with all address bits in the right place:

--- i/mbr/bootloader.asm
+++ w/mbr/bootloader.asm
@@ -119,6 +119,7 @@ read_protected_mode_kernel:
 	sub	edx, 0xfe00			; update the number of bytes to load
 	add	word [gdt.dest], 0xfe00
 	adc	byte [gdt.dest+2], 0
+	adc	byte [gdt.dest+5], 0
 	jmp	short read_protected_mode_kernel.loop

 read_protected_mode_kernel_2:

…and indeed, in my first test this seems to fix the problem! It’ll take me a little while to clean this up and submit it. You can follow gokrazy issue #248 if you’re interested.

Bonus: reading BIOS source

There are actually a couple of BIOS implementations that we can look into to get a better understanding of how Move Memory Block works.

We can look at DOSBox, an open source DOS emulator. Its Move Memory Block implementation does seem to support 32 bit addresses:

PhysPt dest	= (mem_readd(data+0x1A) & 0x00FFFFFF) +
              (mem_readb(data+0x1E)<<24);

Another implementation is SeaBIOS. Contrary to DOSBox, SeaBIOS is not just used in emulation: The PC Engines apu uses coreboot with SeaBIOS. QEMU also uses SeaBIOS.

The SeaBIOS handle_1587 source code is a little harder to follow, because it requires knowledge of Real Mode assembly. The way I read it, SeaBIOS doesn’t truncate or otherwise modify the descriptors and just passes them to the CPU. On 386 or newer, 32 bit addresses should work.

Mitigation

While it’s great to understand the limitation we’re running into, I wanted to unblock the pull request as quickly as possible, so I needed a quick mitigation instead of investigating if my speculation can be developed into a proper fix.

When I started router7, we didn’t support loadable kernel modules, so everything had to be compiled into the kernel. We now do support loadable kernel modules, so I could have moved functionality into modules.

Instead, I found an even easier quick fix: switching from gzip to zstd compression. This saved about 1.8 MB and will buy us some time to implement a proper fix while unblocking automated new Linux kernel version merges.

Conclusion

I wanted to share this debugging story because it shows a couple of interesting lessons:

1. Being able to run older versions of various parts of your software stack is a very valuable debugging tool. It helped us isolate a trigger for the bug (using an older GCC) and it helped us set up a debugging environment (using an older QEMU).

2. Setting up a debugger can be annoying (symbol files, learning the UI) but it’s so worth it.

3. Be on the lookout for wrong turns during debugging. Write down every conclusion and challenge it.

4. The BIOS can seem mysterious and “too low level” but there are many blog posts, lectures and tutorials. You can also just read open-source BIOS code to understand it much better.

Enjoy poking at your BIOS!

Appendix: Resources

I found the following resources helpful:

• linux-insides: From bootloader to kernel
• The MS-DOS Encyclopedia
• Ben Eater: A simple BIOS for my breadboard computer (22 minutes)

When a service fails to start up enough times in a row, systemd gives up on it.

On servers, this isn’t what I want — in general it’s helpful for automated recovery if daemons are restarted indefinitely. As long as you don’t have circular dependencies between services, all your services will eventually come up after transient failures, without having to specify dependencies.

This is particularly useful because specifying dependencies on the systemd level introduces footguns: when interactively stopping individual services, systemd also stops the dependents. And then you need to remember to restart the dependent services later, which is easy to forget.

Enabling indefinite restarts for a service

To make systemd restart a service indefinitely, I first like to create a drop-in config file like so:

cat > /etc/systemd/system/restart-drop-in.conf <<'EOT'
[Unit]
StartLimitIntervalSec=0

[Service]
Restart=always
RestartSec=1s
EOT

Then, I can enable the restart behavior for individual services like prometheus-node-exporter, without having to modify their .service files (which needs manual effort when updating):

cd /etc/systemd/system
mkdir prometheus-node-exporter.service.d
cd prometheus-node-exporter.service.d
ln -s ../restart-drop-in.conf
systemctl daemon-reload

Changing the defaults for all services

If most of your services set Restart=always or Restart=on-failure, you can change the system-wide defaults for RestartSec and StartLimitIntervalSec like so:

mkdir /etc/systemd/system.conf.d
cat > /etc/systemd/system.conf.d/restartdefaults.conf <<'EOT'
[Manager]
DefaultRestartSec=1s
DefaultStartLimitIntervalSec=0
EOT
systemctl daemon-reload

What do the default settings do?

So why do we need to change these settings to begin with?

The default systemd settings (as of systemd 255) are:

DefaultRestartSec=100ms
DefaultStartLimitIntervalSec=10s
DefaultStartLimitBurst=5

This means that services which specify Restart=always are restarted 100ms after they crash, and if the service crashes more than 5 times in 10 seconds, systemd does not attempt to restart the service anymore.

It’s easy to see that for a service which takes, say, 100ms to crash, for example because it can’t bind on its listening IP address, this means:

Why does systemd give up by default?

I’m not sure. If I had to speculate, I would guess the developers wanted to prevent laptops running out of battery too quickly because one CPU core is permanently busy just restarting some service that’s crashing in a tight loop.

That same goal could be achieved with a more relaxed DefaultRestartSec= value, though: With DefaultRestartSec=5s, for example, we would sufficiently space out these crashes over time.

There is some recent discussion upstream regarding changing the default. Let’s see where the discussion goes.

[2024-01-13: I added a section with an option I forgot to put into my talk and thus elided from the initial post as well.]

I gave a talk at GopherConAU 2023 about a particular problem we encountered when designing generics for Go and what we might do about it.

This blog post is meant as a supplement to that talk. It mostly reproduces its content, while giving some supplementary information and more detailed explanations where necessary.

So if you prefer to ingest your information from text, then this blog post should serve you well. If you prefer a talk, you can watch the recording and use it to get some additional details in the relevant sections.

The talk (and hence this post) is also a follow-up to a previous blog post of mine. But I believe the particular explanation I give here should be a bit more approachable and is also more general. If you have read that post and are just interested in the differences, feel free to skip to the Type Parameter Problem.

With all that out of the way, let us get into it.

The Problem

If you are using Go generics, you are probably aware that it’s possible to constrain type parameters. This makes sure that a type argument has all the operations that your generic function expects available to it.

One particular way to constrain a type parameter is using union elements, which allow you to say that a type has to be from some list of types. The most common use of this is to allow you to use Go’s operators on a generic parameter:

// Allows any type argument that has underlying type int, uint or string.
type Ordered interface {
    ~int | ~uint | ~string
}

func Max[T Ordered](a, b T) T {
    // As all int, uint and string types support the > operator, our generic
    // function can use it:
    if a > b {
        return a
    }
    return b
}

Another case this would be very useful for would be to allow us to call a method as a fallback:

type Stringish interface {
    fmt.Stringer | ~string
}

func Stringify[T Stringish](v T) string {
    if s, ok := any(v).(fmt.Stringer); ok {
        return s.String()
    }
    return reflect.ValueOf(v).String()
}

However, if we try this, the compiler will complain:

cannot use fmt.Stringer in union (fmt.Stringer contains methods)

And if we check the spec, we find a specific exception for this:

Implementation restriction: A union (with more than one term) cannot contain the predeclared identifier comparable or interfaces that specify methods, or embed comparable or interfaces that specify methods.

To explain why this restriction is in place, we will dive into a bit of theory.

Some Theory

You have probably heard about the P versus NP problem. It concerns two particular classes of computational problems:

• P is the class of problems that can be solved efficiently¹. An example of this is multiplication of integers: If I give you two integers, you can write an algorithm that quickly multiplies them.
• NP is the class of problems that can be verified efficiently: If you have a candidate for a solution, you can write an efficient algorithm that verifies it. An example is factorization: If you give me an integer \(N\) and a prime \(p\), you can efficiently check whether or not it is a factor of \(N\). You just divide \(N\) by \(p\) and check whether there is any remainder.

Every problem in P is also in NP: If you can efficiently solve a problem, you can also easily verify a solution, by just doing it yourself and comparing the answers.

However, the opposite is not necessarily true. For example, if I give you an integer \(N\) and tell you to give me a non-trivial factor of it, the best you could probably do is try out all possible candidates until you find one. This is exponential in the size of the input (an integer with \(k\) digits has on the order of \(10^k\) candidate factors).

We generally assume that there are in fact problems which are in NP but not in P - but we have not actually proven so. Doing that is the P versus NP problem.

While we have not proven that there are such “hard” problems, we did prove that there are some problems which are “at least as hard as any other problem in NP”. This means that if you can solve them efficiently, you can solve any problem in NP efficiently. These are called “NP-hard” or “NP-complete”².

One such problem is the Boolean Satisfiability Problem. It asks you to take in a boolean formula - a composition of some boolean variables, connected with “and”, “or” and “not” operators - and determine an assignment to the variables that makes the formula true.

So, for example, I could ask you to find me a satisfying assignment for this function:

func F(x, y, z bool) bool {
    return (!x || z) && (y || z) && (x || !z)
}

For example, F(true, true, false) is false, so it is not a satisfying assignment. But F(false, true, false) is true, so that is a satisfying assignment.

It is easy to verify whether any given assignment satisfies your formula - you just substitute all the variables and evaluate it. But to find one, you probably have to try out all possible inputs. And for \(n\) variables, you have \(2^n\) different options, so this takes exponential time.

In practice, this means that if you can show that solving a particular problem would allow you to solve SAT, your problem is itself NP-hard: It would be at least as hard as solving SAT, which is at least as hard as solving any other NP problem. And as we assume that NP≠P, this means your problem can probably not be solved efficiently.

The last thing we need to mention is co-NP, the class of complements of problems in NP. The complement of a (decision) problem is simply the same problem, with the answer is inverted: You have to answer “yes” instead of “no” and vice versa. And where with NP, a “yes” answer should have an efficiently verifiable proof, with co-NP, a “no” answer should have an efficiently verifiable proof.

Notably, the actual difficulty of solving the problem does not change. To decide between “yes” and “no” is just as hard, you just turn around the answer. So, in a way, this is a technicality.

A co-NP complete problem is simply a problem that is the complement of an NP complete problem and as you would expect, it is just as hard and it is at least as hard as any other problem in co-NP.

Now, with the theory out of the way, let’s look at Go again.

The Type Parameter Problem

When building a Go program, the compiler has to solve a couple of computational problems as well. For example, it has to be able to answer “does a given type argument satisfy a given constraint”. This happens if you instantiate a generic function with a concrete type:

func F[T C]() {} // where C is some constraint
func G() {
    F[int]() // Allowed if and only if int satisfies C.
}

This problem is in P: The compiler can just evaluate the constraint as if it was a logical formula, with | being an “or” operator, multiple lines being an “and” operator and checking if the type argument has the right methods or underlying types on the way.

Another problem it has to be able to solve is whether a given constraint C1 implies another constraint C2: Does every type satisfying C1 also satisfy C2? This comes up if you instantiate a generic function with a type parameter:

func F[T C1]() {
    G[T]() // Allowed if and only if C1 implies C2
}
func G[T C2]() {}

My claim now is that this problem (which I will call the “Type Parameter Problem” for the purposes of this post) is co-NP complete³.

To prove this claim, we reduce SAT to the (complement of the) Type Parameter Problem. We show that if we had a Go compiler which solves this problem, we can use it so solve the SAT problem as well. And we do that, by translating an arbitrary boolean formula into a Go program and then check whether it compiles.

On a technical note, we are going to assume that the fomula is in Conjunctive Normal Form (CNF): A list of terms connected with “and” operators, where each term is a list of (possibly negated) variables connected with “or” terms. The example I used above is in CNF and we use it as an example to demonstrate the translation:

func F(x, y, z bool) bool {
    return (!x || z) && (y || z) && (x || !z)
}

This assumption may seem like a cheat, but importantly, SAT is still NP-complete with it.

The first step in our reduction is to model our boolean variables. Every variable can be either true or false and it can appear negated or not negated. We encode that by defining two interfaces per variable³:

type X interface { X() }     // X is assigned "true"
type NotX interface{ NotX()} // X is assigned "false"

This allows us to translate our formula directly, using union elements for “or” and interface-embedding for “and”:

// Represents (!x || z) && (y || z) && (x || !z)
type Formula interface {
    NotX | Z
    Y | Z
    X | NotZ
}

There are, however, two issues with this:

1. A type could have neither of X() and NotX().
2. A type could have both of X() and NotX().

This breaks our representation, because a boolean variable always has to be exactly true or false - it can’t be neither and it can’t be both.

To address the first point, we define another interface:

type AtLeastOne interface {
    X | NotX
    Y | NotY
    Z | NotZ
}

Any type satisfying AtLeastOne has to assign at least one of true and false to each variable.

Similarly, we define an interface to address the second problem:

type Both_X interface { X; NotX }
type Both_Y interface { Y; NotY }
type Both_Z interface { Z; NotZ }
type Both interface {
    Both_X | Both_Y | Both_Z
}

Any type satisfying Both now assigns both true and false to at least one variable.

To represent a valid, satisfying assignment, a type thus has to

1. satisfy Formula
2. satisfy AtLeastOne
3. not satisfy Both

Now, we ask our compiler to type-check this Go program⁴:

func G[T Both]() {}
func F[T interface{ Formula; AtLeastOne }]() {
    G[T]() // Allowed if and only if (Formula && AtLeastOne) => Both
}

This program should compile, if and only if any type satisfying Formula and AtLeastOne also satisfies Both. Because we are looking at the complement of SAT, we invert this, to get our final answer:

    !( (Formula && AtLeastOne) =>  Both )
<=> !(!(Formula && AtLeastOne) ||  Both ) // "A => B" is equivalent to "!A || B"
<=> !(!(Formula && AtLeastOne  && !Both)) // De Morgan's law
<=>     Formula && AtLeastOne  && !Both   // Double negation

This finishes our reduction: The compiler should reject the program, if and only if the formula has a satisfying assignment. The Type Parameter Problem is at least as hard as the complement of SAT.

Going forward

So the restriction on methods in union elements is in place, because we are concerned about type checking Go would become a very hard problem if we allowed them. But that is, of course, a deeply dissatisfying situation.

Our Stringish example would clearly be a very useful constraint - so useful, in fact, that it was used an example in the original design doc. More generally, this restriction prevents us from having a good way to express operator constraints for generic functions and types. We currently end up writing multiple versions of the same functions, one that uses operators and one that takes functions to do the operations. This leads to boilerplate and extra API surface⁵.

The slices package contains a bunch of examples like that (look for the Func suffix to the name):

// Uses the == operator. Useful for predeclared types (int, string,…) and
// structs/arrays of those.
func Contains[S ~[]E, E comparable](s S, v E) bool
// Uses f. Needed for slices, maps, comparing by pointer-value or other notions
// of equality.
func ContainsFunc[S ~[]E, E any](s S, f func(E) bool) bool

So we should consider compromises, allowing us to get some of the power of removing this restriction at least.

Option 1: Ignore the problem

This might be a surprising option to consider after spending all these words on demonstrating that this problem is hard to solve, but we can at least consider it: We simply say that a Go compiler has to include some form of (possibly limited) SAT solver and is allowed to just give up after some time, if it can not find a proof that a program is safe.

C++ concepts do this. A C++ compiler has to determine if one constraint implies another one, when it has to decide which of multiple overloaded generic functions to invoke. And it does so using a simple SAT solver. In particular, if it wants to prove \(P ⇒ Q\), it first converts \(P\) into Disjunctive Normal Form (DNF) and then convert \(Q\) into Conjunctive Normal Form (CNF).

With \(P\) in DNF and \(Q\) in CNF, \(P ⇒ Q\) is easy to prove (and disprove). But this normalization into DNF or CNF itself requires exponential time in general. And you can indeed create C++ programs that crash C++ compilers.

Personally, I find all versions of this option very dissatisfying:

• Leaving the heuristic up to the implementation feels like too much wiggle-room for what makes a valid Go program.
• Describing an explicit heuristic in the spec takes up a lot of the complexity budget of the spec.
• Allowing the compiler to try and give up after some time feels antithetical to the pride Go takes in fast compilation.

Option 2: Limit the expressiveness of interfaces

For the interfaces as they exist today, we actually can solve the SAT problem: Any interface can ultimately be represented in the form (with some elements perhaps being empty):

interface {
    A | … | C | ~X | … | ~Z // for some concrete types
    comparable
    M1(…) (…)
    // …
    Mn(…) (…)
}

And it is straight-forward to use this representation to do the kind of inference we need.

This tells us that there are some restrictions we can put on the kinds of interfaces we can write down, while still not running into the kinds of problems discussed in this post. That’s because every such kind of interfaces gives us a restricted sub problem of SAT, which only looks at formulas conforming to some extra restrictions.

One example of such a sub problem we actually used above, where we assumed that our formula is in Conjunctive Normal Form. Another important such sub problem is the one where the formulas are in Disjunctive Normal Form instead: Where we have a list of terms linked with “or” operators and each term is a list of (possibly negated) variables linked with “and” operators. For DNF, the SAT problem is efficiently solvable.

We could take advantage of that by allowing union elements to contain methods - but only if

1. There is exactly one union in the top-level interface.
2. The interfaces embedded in that union are “easy” interfaces, i.e. ones we allow today.

So, for example

type Stringish interface {
    // Allowed: fmt.Stringer and ~string are both allowed today
    fmt.Stringer | ~string
}
type A interface {
    // Not Allowed: Stringish is not allowed today, so we have more than one level
    Stringish | ~int
}
type B interface {
    // Allowed: Same as A, but we "flattened" it, so each element is an
    // "easy" interface.
    fmt.Stringer | ~string | ~int
}
type C interface {
    // Not Allowed: Can only have a single union (or must be an "easy" interface)
    fmt.Stringer | ~string
    comparable
}

This restriction makes our interfaces be in DNF, in a sense. It’s just that every “variable” of our DNF is itself an “easy” interface. If we need to solve SAT for one of these, we first solve it on the SAT formula to determine which “easy” interfaces need to be satisfied and then use our current algorithms to check which of those can be satisfied.

Of course, this restriction is somewhat hard to explain. But it would allow us to write at least some of the useful programs we want to use this feature for. And we might find another set of restrictions that are easier to explain but still allow that.

We should probably try to collect some useful programs that we would want to write with this feature and then see, for some restricted interface languages if they allow us to write them.

Option 3: Make the type-checker conservative

For our reduction, we assumed that the compiler should allow the program if and only if it can prove that every type satisfying C1 also satisfies C2.

We could allow it to reject some programs that would be valid, though. Wec could describe an algorithm for determining if C1 implies C2 that can have false negatives: Rejecting a theoretically safe program, just because it cannot prove that it is safe with that algorithm, requiring you to re-write your program into something it can handle more easily.

Ultimately, this is kind of what a type system does: It gives you a somewhat limited language to write a proof to the compiler that your program is “safe”, in the sense that it satisfies certain invariants. And if you accidentally pass a variable of the wrong type - even if your program would still be perfectly valid - you might have to add a conversion or call some function that verifies its invariants, before being allowed to do so.

For this route, we still have to decide which false negatives we are willing to accept though: What is the algorithm the compiler should use?

For some cases, this is trivial. For example, this should obviously compile:

func StringifyAll[T Stringish](vals ...T) []string {
    out := make([]string, len(vals))
    for i, v := range vals {
        // Stringify as above. Should be allowed, as T uses the same constraint
        // as Stringify.
        out[i] = Stringify(v)
    }
    return out
}

But other cases are not as straight forward and require some elaboration:

func Marshal[T Stringish | ~bool | constraints.Integer](v T) string { /* … */ }

// Stringish appears in the union of the target constraint.
func F[T Stringish](v T) string { return Marshal[T](v) }

// string has underlying type string and fmt.Stringer is the Stringish union.
func G[T string|fmt.Stringer](v T) string { return Marshal[T](v) }

// The method name is just a different representation of fmt.Stringer
func H[T interface{ String() string }](v T) string { return Marshal[T](v) }

These examples are still simple, but they are useful, so should probably be allowed. But they already show that there is somewhat complex inference needed: Some terms on the left might satisfy some terms on the right, but we can not simply compare them as a subset relation, we actually have to take into account the different cases.

And remember that converting to DNF or CNF takes exponential time, so the simple answer of “convert the left side into DNF and the right side into CNF, then check each term individually” does not solve our problem.

In practice, this option has a large intersection with the previous one: The algorithm would probably reject programs that use interfaces with too complex a structure on either side, to guarantee that it terminates quickly. But it would allow us, in principle, to use different restrictions for the left and the right hand side: Allow you to write any interface and only check the structure if you actually use them in a way that would make inference impossible.

We have to decide whether we would find that acceptable though, or whether it seems to confusing in practice. Describing the algorithm also would take quite a lot of space and complexity budget in the spec.

Option 4: Delay constraint checking until instantiation

One option I forgot to bring up in my talk is essentially the opposite of the previous one: We could have the compiler skip checking the constraints of generic function calls in generic functions altogether. So, for example, this code would be valid:

func G[T fmt.Stringer](v T) string {
    return v.String()
}

func F[T any](v T) string {
    // T constrained on any does not satisfy fmt.Stringer.
    // But we allow the call anyways, for now.
    return G(v)
}

To retain type-safety, we would instead check the constraints only when F is instantiated with a concrete type:

func main() {
    F(time.Second) // Valid: time.Duration implements fmt.Stringer
    F(42)          // Invalid: int does not implement fmt.Stringer
}

The upside is that this seems very easy to implement. It means we completely ignore any questions that require us to do inference on “sets of all types”. We only ever need to answer whether a specific type satisfies a specific constraint. Which we know we can do efficiently.

The downside is that this effectively introduces new constraints on the type-parameter of F implicitly. The signature says that F can be instantiated with any type, but it actually requires a fmt.Stringer.

One consequence of that is that it becomes harder to figure out what type arguments are allowed for a generic function or type. An instantiation might fail and the only way to understand why is to look into the code of the function you are calling. Potentially multiple layers of dependency deep.

Another consequence is that it means your program might break because of a seemingly innocuous change in a dependency. A library author might add a generic call to one of their functions. Because it only changes the implementation and not the API, they assume that this is a backwards compatible change. Their tests pass, because none of the types they use in their tests triggers this change in behavior. So they release a new minor version of their library. Then you upgrade (perhaps by way of upgrading another library that also depends on it) and your code no longer compiles, because you use a different type - conforming to the actual constraints from the signature, but not the implicit ones several layers of dependency down.

Because of this breakage in encapsulation, Go generics have so far eschewed this idea of delayed constraint checking. But it is possible that we could find a compromise here: Check the most common and easy to handle cases statically, while delaying some of the more complex and uncommon ones until instantiation. Where to draw that line would then be open for discussion.

Personally, just like with Option 1, I dislike this idea. But we should keep it in mind.

Future-proofing

Lastly, when we talk about this we should keep in mind possible future extensions to the generics design.

For example, there is a proposal by Rog Peppe to add a type-switch on type parameters. The proposal is to add a new type switch syntax for type parameters, where every case has a new constraint and in that branch, you could use the type parameter as if it was further constrained by that. So, for example, it would allow us to rewrite Stringify without reflect:

func Stringify[T Stringish](v T) string {
    switch type T {
    case fmt.Stringer:
        // T is constrained by Stringish *and* fmt.Stringer. So just fmt.Stringer
        // Calling String on a fmt.Stringer is allowed.
        return v.String()
    case ~string:
        // T is consrtained by Stringish *and* ~string. So just ~string
        // Converting a ~string to string is allowed.
        return string(v)
    }
}

The crux here is, that this proposal allows us to create new, implicit interfaces out of old ones.

If we restrict the structure of our interfaces, these implicit interfaces might violate this structure. And if we make the type checker more conservative, a valid piece of code might no longer be valid if copied into a type parameter switch, if the implicit constraints would lead to a generic all the compiler can’t prove to be safe.

Of course it is impossible to know what extension we really want to add in the future. But we should at least consider some likely candidates during the discussion.

Summary

I hope I convinced you that

1. Simply allowing methods in unions would make type-checking Go code co-NP hard.
2. But we might be able to find some compromise that still allows us to do some of the things we want to use this for.
3. The devil is in the details and we still have to think hard and carefully about those.

For over 10 years now, I run two self-built NAS (Network Storage) devices which serve media (currently via Jellyfin) and run daily backups of all my PCs and servers.

In this article, I describe my goals, which hardware I picked for my new build (and why) and how I set it up.

Design Goals

I use my network storage devices primarily for archival (daily backups), and secondarily as a media server.

There are days when I don’t consume any media (TV series and movies) from my NAS, because I have my music collection mirrored to another server that’s running 24/7 anyway. In total, my NAS runs for a few hours in some evenings, and for about an hour (daily backups) in the mornings.

This usage pattern is distinctly different than, for example, running a NAS as a file server for collaborative video editing that needs to be available 24/7.

The goals of my NAS setup are:

1. Save power: each NAS build only runs when needed.
   • They must support Wake-on-LAN or similar (ESP32 remote power button).
   • Scheduling of backups is done separately, on a Raspberry Pi with gokrazy.
   • Convenient power off (tied to our all-lights-out button) and power on (with webwake).
2. Use Off-the-shelf hardware and software.
   • When hardware breaks, I can get replacements from the local PC store the same day.
   • Even when only the data disk(s) survive, I should be able to access my data when booting a standard live Linux system.
   • Minimal application software risk: I want to minimize risk for manual screw-ups or software bugs, meaning I use the venerable rsync for my backup needs (not Borg, restic, or similar).
   • Minimal system software risk: I use reliable file systems with the minimal feature set — no LVM or btrfs snapshots, no ZFS replication, etc. To achieve redundancy, I don’t use a cluster file system with replication, instead I synchronize my two NAS builds using rsync, without the --delete flag.
3. Minimal failure domains: when one NAS fails, the other one keeps working.
   • Having N+1 redundancy here takes the stress out of repairing your NAS.
   • I run each NAS in a separate room, so that accidents like fires or spilled drinks only affect one machine.

File System: ZFS

In this specific build, I am trying out ZFS. Because I have two NAS builds running, it is easy to change one variable of the system (which file system to use) in one build, without affecting the other build.

My main motivation for using ZFS instead of ext4 is that ZFS does data checksumming, whereas ext4 only checksums metadata and the journal, but not data at rest. With large enough datasets, the chance of bit flips increases significantly, and I would prefer to know about them so that I can restore the affected files from another copy.

Hardware

Each of the two storage builds has (almost) the same components. This makes it easy to diagnose one with the help of the other. When needed, I can swap out components of the second build to temporarily repair the first one, or vice versa.

Base Components

The total price of 476 CHF makes this not a cheap build.

But, I think each component is well worth its price. Here’s my thinking regarding the components:

• Why not a cheaper system disk? I wanted to use an M.2 NVMe disk so that I could mount it on the bottom of the mainboard instead of having to mount another SATA disk in the already-crowded case. Instead of chosing the cheapest M.2 disk I could find, I went with WD Red as a brand I recognize. While it’s not a lot of effort to re-install the system disk, it’s still annoying and something I want to avoid if possible. If spending 20 bucks saves me one disk swap + re-install, that’s well worth it for me!
• Why not skip the system disk entirely and install on the data disks? That makes the system harder to (re-)install, and easier to make manual errors when recovering the system. I like to physically disconnect the data disks while re-installing a NAS, for example. (I’m a fan of simple precautions that prevent drastic mistakes!)
• Why not a cheaper CPU cooler? In one of my earlier NAS builds, I used a (cheaper) passive CPU fan, which was directly in the air stream of the Noctua 120mm case fan. This setup was spec’ed for the CPU I used, and yet said CPU died as the only CPU to die on me in many many years. I want a reliable CPU fan, but also an absolutely silent build, so I went with the Noctua CPU cooler.
• Why not skip the case fan, or go with the Silverstone-supplied one? You might argue that the airflow of the CPU cooler is sufficient for this entire build. Maybe that’s true, but I don’t want to risk it. Also, there are 3 disks (two data disks and one system disk) that can benefit from additional airflow.
• Regarding the CPU, I chose the cheapest AMD CPU for Socket AM4, with a 35W TDP and built-in graphics. The built-in graphics means I can connect an HDMI monitor for setup and troubleshooting, without having to use the mainboard’s valuable one and only PCIe slot.
  
  Unfortunately, AMD CPUs with 35W TDP are not readily available right now. My tip is to look around for a bit, and maybe buy a used one. Chose either the predecessor Athlon 200GE, or the newer generation Ryzen APU series, whichever you can get your hands on.
• Regarding the mainboard, I went with the AsRock Mini ITX series, which have served me well over the years. I started with an AsRock AM1H-ITX in 2016, then bought two AsRock AB350 Gaming ITX/ac in 2019, and recently an AsRock B450 Gaming ITX/ac.

As a disclaimer: the two builds I use are very similar to the component list above, with the following differences:

1. On storage2, I use an old AMD Ryzen 5 5600X CPU instead of the listed Athlon 3000G. The extra performance isn’t needed, and the lack of integrated graphics is annoying. But, I had the CPU lying around and didn’t want it to go to waste.
2. On storage3, I use an old AMD Athlon 200GE CPU on an AsRock AB350 mainboard.

I didn’t describe the exact builds I use because a component list is more useful if the components on it are actually available :-).

16 TB SSD Data Disks

It used to be that Solid State Drives (SSDs) were just way too expensive compared to spinning hard disks when talking about terabyte sizes, so I used to put the largest single disk drive I could find into each NAS build: I started with 8 TB disks, then upgraded to 16 TB disks later.

Luckily, the price of flash storage has come down quite a bit: the Samsung SSD 870 QVO (8 TB) costs “only” 42 CHF per TB. For a total of 658 CHF, I can get 16 TB of flash storage in 2 drives:

Of course, spinning hard disks are at 16 CHF per TB, so going all-flash is over 3x as expensive.

I decided to pay the premium to get a number of benefits:

• My NAS devices are quieter because there are no more spinning disks in them. This gives me more flexibility in where to physically locate each storage machine.
• My daily backups run quicker, meaning each NAS needs to be powered on for less time. The effect was actually quite pronounced, because figuring out which files need backing up requires a lot of random disk access. My backups used to take about 1 hour, and now finish in less than 20 minutes.
• The quick access times of SSDs solve the last remaining wrinkle in my backup scheme: deleting backups and measuring used disk space is finally fast!

Power Usage

The choice of CPU, Mainboard and Network Card all influence the total power usage of the system. Here are a couple of measurements to give you a rough idea of the power usage:

These values were measured using a myStrom WiFi Switch.

Operating System

Previously: CoreOS

Before this build, I ran my NAS using Docker containers on CoreOS (later renamed to Container Linux), which was a light-weight Linux distribution focused on containers. There are two parts about CoreOS that I liked most.

The most important part was that CoreOS updated automatically, using an A/B updating scheme, just like I do in gokrazy. I want to run as many of my devices as possible with A/B updates.

The other bit I like is that the configuration is very clearly separated from the OS. I managed the configuration (a cloud-init YAML file) on my main PC, so when swapping out the NAS system disk with a blank disk, I could just plug my config file into the CoreOS installer, and be done.

When CoreOS was bought by Red Hat and merged into Project Atomic, there wasn’t a good migration path and cloud-init wasn’t supported anymore. As a short-term solution, I switched from CoreOS to Flatcar Linux, a spiritual successor.

Now: Ubuntu Server

For this build, I wanted to try out ZFS. I always got the impression that ZFS was a pain to run because its kernel modules are not included in the upstream Linux kernel source.

Then, in 2016, Ubuntu decided to include ZFS by default. There are a couple of other Linux distributions on which ZFS seems easy enough to run, like Gentoo, Arch Linux or NixOS.

I wanted to spend my “innovation tokens” on ZFS, and keep the rest boring and similar to what I already know and work with, so I chose Ubuntu Server over NixOS. It’s similar enough to Debian that I don’t need to re-learn.

Luckily, the migration path from Flatcar’s cloud-init config to Ubuntu Server is really easy: just copy over parts of the cloud-config until you’re through the entire thing. It’s like a checklist!

Maybe later? gokrazy

In the future, it might be interesting to build a NAS setup using gokrazy. In particular since we now can run Docker containers on gokrazy, which makes running Samba or Jellyfin quite easy!

Using gokrazy instead of Ubuntu Server would get rid of a lot of moving parts. The current blocker is that ZFS is not available on gokrazy. Unfortunately that’s not easy to change, in particular also from a licensing perspective.

Setup

UEFI

I changed the following UEFI settings:

• Advanced → ACPI Configuration → PCIE Devices Power On: Enabled

  • This setting is needed (but not sufficient) for Wake On LAN (WOL). You also need to enable WOL in your operating system.

• Advanced → Onboard Devices Configuration → Restore on AC/Power Loss: Power On

  • This setting ensures the machine turns back on after a power loss. Without it, WOL might not work after a power loss.

Operating System

Network preparation

I like to configure static IP addresses for devices that are a permanent part of my network.

I have come to prefer configuring static addresses as static DHCP leases in my router, because then the address remains the same no matter which operating system I boot — whether it’s the installed one, or a live USB stick for debugging.

Ubuntu Server

1. Download Ubuntu Server from https://ubuntu.com/download/server

   • I initially let the setup program install Docker, but that’s a mistake. The setup program will get you Docker from snap (not apt), which can’t work with the whole file system.

2. Disable swap:

   • swapoff -a
   • $EDITOR /etc/fstab # delete the swap line

3. Automatically load the corresponding sensors kernel module for the mainboard so that the Prometheus node exporter picks up temperature values and fan speed values:

   • echo nct6775 | sudo tee /etc/modules

4. Enable unattended upgrades:

   • dpkg-reconfigure -plow unattended-upgrades

   • Edit /etc/apt/apt.conf.d/50unattended-upgrades — I like to make the following changes:

     Unattended-Upgrade::MinimalSteps "true";
     Unattended-Upgrade::Mail "michael@example.net";
     Unattended-Upgrade::MailReport "only-on-error";
     Unattended-Upgrade::Automatic-Reboot "true";
     Unattended-Upgrade::Automatic-Reboot-Time "08:00";
     Unattended-Upgrade::SyslogEnable "true";
     

Network

Tailscale Mesh VPN

I have come to like Tailscale. It’s a mesh VPN (data flows directly between the machines) that allows me access to and from my PCs, servers and storage machines from anywhere.

Specifically, I followed the install Tailscale on Ubuntu 22.04 guide.

Prometheus Node Exporter

For monitoring, I have an existing Prometheus setup. To add a new machine to my setup, I need to configure it as a new target on my Prometheus server. In addition, I need to set up Prometheus on the new machine.

First, I installed the Prometheus node exporter using apt install prometheus-node-exporter.

Then, I modified /etc/default/prometheus-node-exporter to only listen on the Tailscale IP address:

ARGS="--web.listen-address=100.85.3.16:9100"

Lastly, I added a systemd override to ensure the node exporter keeps trying to start until tailscale is up: the command systemctl edit prometheus-node-exporter opens an editor, and I configured the override like so:

# /etc/systemd/system/prometheus-node-exporter.service.d/override.conf
[Unit]
# Allow infinite restarts, even within a short time.
StartLimitIntervalSec=0

[Service]
RestartSec=1

Static IPv6 address

Similar to the static IPv4 address, I like to give my NAS a static IPv6 address as well. This way, I don’t need to reconfigure remote systems when I (sometimes temporarily) switch my NAS to a different network card with a different MAC address. Of course, this point becomes moot if I ever switch all my backups to Tailscale.

Ubuntu Server comes with Netplan by default, but I don’t know Netplan and don’t want to use it.

To switch to systemd-networkd, I ran:

apt remove --purge netplan.io
systemctl enable --now systemd-networkd

Then, I created a systemd-networkd config file with a static IPv6 token, resulting in a predictable IPv6 address:

$EDITOR /etc/systemd/network/enp.network

My config file looks like this:

[Match]
Name=enp*

[Network]
DHCP=yes
IPv6Token=0:0:0:0:10::253
IPv6AcceptRouterAdvertisements=yes

IPv6 firewall setup

An easy way to configure Linux’s netfilter firewall is to apt install iptables-persistent. That package takes care of saving firewall rules on shutdown and restoring them on the next system boot.

My rule setup is very simple: allow ICMP (IPv6 needs it), then set up ACCEPT rules for the traffic I expect, and DROP the rest.

Here’s my resulting /etc/iptables/rules.v6 from such a setup:

# Generated by ip6tables-save v1.4.14 on Fri Aug 26 19:57:51 2016
*filter
:INPUT DROP [0:0]
:FORWARD ACCEPT [0:0]
:OUTPUT ACCEPT [0:0]
-A INPUT -p ipv6-icmp -m comment --comment "IPv6 needs ICMPv6 to work" -j ACCEPT
-A INPUT -m state --state RELATED,ESTABLISHED -m comment --comment "Allow packets for outgoing connections" -j ACCEPT
-A INPUT -s fe80::/10 -d fe80::/10 -m comment --comment "Allow link-local traffic" -j ACCEPT
-A INPUT -s 2001:db8::/64 -m comment --comment "local traffic" -j ACCEPT
-A INPUT -p tcp -m tcp --dport 22 -m comment --comment "SSH" -j ACCEPT
COMMIT
# Completed on Fri Aug 26 19:57:51 2016

Encrypted ZFS

Before you can use ZFS, you need to install the ZFS tools using apt install zfsutils-linux.

Then, we create a zpool that spans both SSDs:

zpool create \
  -o ashift=12 \
  srv \
  /dev/disk/by-id/ata-Samsung_SSD_870_QVO_8TB_S5SSNF0TC06121Z \
  /dev/disk/by-id/ata-Samsung_SSD_870_QVO_8TB_S5SSNF0TC06787P

The -o ashift=12 ensures proper alignment on disks with a sector size of either 512B or 4KB.

On that zpool, we now create our datasets:

(echo -n on-device-secret && \
 wget -qO - https://autounlock.zekjur.net:8443/nascrypto) | zfs create \
  -o encryption=on \
  -o compression=off \
  -o atime=off \
  -o keyformat=passphrase \
  -o keylocation=file:///dev/stdin \
  srv/data

The key I’m piping into zfs create is constructed from two halves: the on-device secret and the remote secret, which is a setup I’m using to implement an automated crypto unlock that is remotely revokable. See the next section for the corresponding unlock.service.

I repeated this same command (adjusting the dataset name) for each dataset: I currently have one for data and one for backup, just so that the used disk space of each major use case is separately visible:

df -h /srv /srv/backup /srv/data   
Filesystem      Size  Used Avail Use% Mounted on
srv             4,2T  128K  4,2T   1% /srv
srv/backup      8,1T  3,9T  4,2T  49% /srv/backup
srv/data         11T  6,4T  4,2T  61% /srv/data

ZFS maintenance

To detect errors on your disks, ZFS has a feature called “scrubbing”. I don’t think I need to scrub more often than monthly, but maybe your scrubbing requirements are different.

I enabled monthly scrubbing on my zpool srv:

systemctl enable --now zfs-scrub-monthly@srv.timer

On this machine, a scrub takes a little over 4 hours and keeps the disks busy:

  scan: scrub in progress since Wed Oct 11 16:32:05 2023
	808G scanned at 909M/s, 735G issued at 827M/s, 10.2T total
	0B repaired, 7.01% done, 03:21:02 to go

We can confirm by looking at the Prometheus Node Exporter metrics:

The other maintenance-related setting I changed is to enable automated TRIM:

zpool set autotrim=on srv

Auto Crypto Unlock

To automatically unlock the encrypted datasets at boot, I’m using a custom unlock.service systemd service file.

My unlock.service constructs the crypto key from two halves: the on-device secret and the remote secret that’s downloaded over HTTPS.

This way, my NAS can boot up automatically, but in an emergency I can remotely stop this mechanism.

[Unit]
Description=unlock hard drive
Wants=network.target
After=systemd-networkd-wait-online.service
Before=samba.service

[Service]
Type=oneshot
RemainAfterExit=yes
# Wait until the host is actually reachable.
ExecStart=/bin/sh -c "c=0; while [ $c -lt 5 ]; do /bin/ping6 -n -c 1 autounlock.zekjur.net && break; c=$((c+1)); sleep 1; done"
ExecStart=/bin/sh -c "(echo -n secret && wget --retry-connrefused -qO - https://autounlock.zekjur.net:8443/nascrypto) | zfs load-key srv/data"
ExecStart=/bin/sh -c "(echo -n secret && wget --retry-connrefused -qO - https://autounlock.zekjur.net:8443/nascrypto) | zfs load-key srv/backup"
ExecStart=/bin/sh -c "zfs mount srv/data"
ExecStart=/bin/sh -c "zfs mount srv/backup"

[Install]
WantedBy=multi-user.target

Backup

For the last 10 years, I have been doing my backups using rsync.

Each machine pushes an incremental backup of its entire root file system (and any mounted file systems that should be backed up, too) to the backup destination (storage2/3).

All the machines I’m backing up run Linux and the ext4 file system. I verified that my backup destination file systems support all the features of the backup source file system that I care about, i.e. extended attributes and POSIX ACLs.

The scheduling of backups is done by “dornröschen”, a Go program that wakes up the backup sources and destination machines and starts the backup by triggering a command via SSH.

SSH configuration

The backup scheduler establishes an SSH connection to the backup source.

On the backup source, I authorized the scheduler like so, meaning it will run /root/backup.pl when connecting:

command="/root/backup.pl",no-port-forwarding,no-X11-forwarding ssh-ed25519 AAAAC3Nzainvalidkey backup-scheduler

backup.pl runs rsync, which establishes another SSH connection, this time from the backup source to the backup destination.

On the backup destination (storage2/3), I authorize the backup source’s SSH public key to run rrsync(1) , a script that only permits running rsync in the specified directory:

command="/usr/bin/rrsync /srv/backup/server.zekjur.net",no-port-forwarding,no-X11-forwarding ssh-ed25519 AAAAC3Nzainvalidkey server.zekjur.net

Signaling Readiness after Wake-Up

I found it easiest to signal readiness by starting an empty HTTP server gated on After=unlock.service in systemd:

[Unit]
Description=nginx for /srv health check
Wants=network.target
After=unlock.service
Requires=unlock.service
StartLimitInterval=0

[Service]
Restart=always
# https://itectec.com/unixlinux/restarting-systemd-service-on-dependency-failure/
ExecStartPre=/bin/sh -c 'systemctl is-active docker.service'
# Stay on the same major version in the hope that nginx never decides to break
# the config file syntax (or features) without doing a major version bump.
ExecStartPre=/usr/bin/docker pull nginx:1
ExecStartPre=-/usr/bin/docker kill nginx-healthz
ExecStartPre=-/usr/bin/docker rm -f nginx-healthz
ExecStart=/usr/bin/docker run \
  --name nginx-healthz \
  --publish 10.0.0.253:8200:80 \
  --log-driver=journald \
nginx:1

[Install]
WantedBy=multi-user.target

My wake program then polls that port and returns once the server is up, i.e. the file system has been unlocked and mounted.

Auto Shutdown

Instead of explicitly triggering a shutdown from the scheduler program, I run “dramaqueen”, which shuts down the machine after 10 minutes, but will be inhibited while a backup is running. Optionally, shutting down can be inhibited while there are active samba sessions.

[Unit]
Description=dramaqueen
After=docker.service
Requires=docker.service

[Service]
Restart=always
StartLimitInterval=0

# Always pull the latest version (bleeding edge).
ExecStartPre=-/usr/bin/docker pull stapelberg/dramaqueen
ExecStartPre=-/usr/bin/docker rm -f dramaqueen
ExecStartPre=/usr/bin/docker create --name dramaqueen stapelberg/dramaqueen
ExecStartPre=/usr/bin/docker cp dramaqueen:/usr/bin/dramaqueen /tmp/
ExecStartPre=/usr/bin/docker rm -f dramaqueen
ExecStart=/tmp/dramaqueen -net_command=

[Install]
WantedBy=multi-user.target

Enabling Wake-on-LAN

Luckily, the network driver of the onboard network card supports WOL by default. If that’s not the case for your network card, see the Arch wiki Wake-on-LAN article.

Conclusion

I have been running a PC-based few-large-disk Network Storage setup for years at this point, and I am very happy with all the properties of the system. I expect to run a very similar setup for years to come.

The low-tech approach to backups of using rsync has worked well — without changes — for years, and I don’t see rsync going away anytime soon.

The upgrade to all-flash is really nice in terms of random access time (for incremental backups) and to eliminate one of the largest sources of noise from my builds.

ZFS seems to work fine so far and is well-integrated into Ubuntu Server.

Related Options

There are solutions for almost everyone’s NAS needs. This build obviously hits my personal sweet spot, but your needs and preferences might be different!

Here are a couple of related solutions:

• If you would like a more integrated solution, you could take a look at the Odroid H3 (Celeron).
• If you’re okay with less compute power, but want more power efficiency, you could use an ARM64-based Single Board Computer.
• If you want to buy a commercial solution, buy a device from qnap and fill it with SSD disks.
  • There are even commercial M.2 flash storage devices like the ASUSTOR Flashstor becoming available! If not for the “off the shelf hardware” goal of my build, this would probably be the most interesting commercial alternative to me.
• If you want more compute power, consider a Thin Client (perhaps used) instead of a Single Board Computer.
  • ServeTheHome has a nice series called Project TinyMiniMicro (introduction, blog posts)
  • If you’re a heise+ subscriber, they have a (German) article about building a NAS from a thin client.
• Very similar to thin clients is the Intel NUC (“Next Unit of Computing”): (German) article comparing different NUC 12 devices

For the last 10 years, I have been interested in hi-DPI monitors, and recently I read about an interesting new monitor: Dell’s 32-inch 6K monitor (U3224KBA), a productivity monitor that offers plenty of modern connectivity options like DisplayPort 2, HDMI 2 and Thunderbolt 4.

My current monitor is a Dell 32-inch 8K monitor (UP3218K), which has a brilliant picture, but a few annoying connectivity limitations and quirks — it needs two (!) DisplayPort cables on a GPU with MST support, meaning that in practice, it only works with nVidia graphics cards.

I was curious to try out the new 6K monitor to see if it would improve the following points:

• Does the 6K monitor work well with most (all?) of my PCs and laptops?
• Is 6K resolution enough, or would I miss the 8K resolution?
• Is a matte screen the better option compared to the 8K monitor’s glossy finish?
• Do the built-in peripherals work with Linux out of the box?

I read a review on heise+ (also included in their c’t magazine), but the review can’t answer these subjective questions of mine.

So I ordered one and tried it out!

Compatibility

The native resolution of this monitor is 6144x3456 pixels.

To drive that resolution at 60 Hz, about 34 Gbps of data rate is needed.

DisplayPort 1.4a only offers a data rate of 25 Gbps, so your hardware and driver need to support Display Stream Compression (DSC) to reach the full resolution at 60 Hz. I tried using DisplayPort 2.0, which supports 77 Gbps of data rate, but the only GPU I have that supports DisplayPort 2 is the Intel A380, which I could not get to work well with this monitor (see the next section).

HDMI 2.1 offers 42 Gbps of data rate, but in my setup, the link would still always use DSC.

Here are the combinations I have successfully tried:

The MacBook Air is the only device in my test that reaches full resolution without using DSC.

Compatibility issues

Let’s talk about the combinations that did not work well.

Too old nVidia driver (< 535.54.03): not at native resolution

You need a quite recent version of the nVidia driver, as they just recently shipped support for DSC at high resolutions. I successfully used DSC with 535.54.03.

With the “older” 530.41.03, I could only select 6016x3384 at 60 Hz, which is not the native resolution of 6144x3456 at 60 Hz.

Intel GPU: no picture or only 4K?!

I was so excited when Intel announced that they are entering the graphics card business. With all the experience and driver support for their integrated graphics, I hoped for good Linux support.

Unfortunately, the Intel A380 I bought months ago continues to disappoint.

I could not get the 6K monitor to work at any resolution higher than 4K, not even under Windows. Worse, when connecting the monitor using DisplayPort, I wouldn’t get a picture at all (in Linux)!

No picture after resume from suspend-to-RAM

I suspend my PC to RAM at least once per day, sometimes even more often.

With my current 8K monitor, I have nailed the suspend/wakeup procedure. With the help of a smart plug, I’m automatically turning the monitor off (on suspend) and on (on wakeup). After a couple of seconds of delay, I configure the correct resolution using xrandr.

I had hoped that the 6K monitor would make any sort of intricate automation superfluous.

Unfortunately, when I resumed my PC, I noticed that the monitor would not show a picture at all! I had to log in from my laptop via SSH to change the resolution with xrandr to 4K, then power the monitor off and on again, then change resolution back to the native 6K.

Scaling

Once you have a physical connection established, how do you configure your computer? With 6K at 32 inches, you’ll need to enable some kind of scaling in order to comfortably read text.

This section shows what options Linux and macOS offer.

i3 (X11)

Just like many other programs on Linux, you configure i3’s scaling by setting the Xft.dpi X resource. The default is 96 dpi, so to get 200% scaling, set Xft.dpi: 192.

Personally, I found 240% scaling more comfortable, i.e. Xft.dpi: 230.

This corresponds to a logical resolution of 2560x1440 pixels.

GNOME (Wayland)

I figured I’d also give Wayland a shot, so I ran GNOME in Fedora 38 on my ThinkPad X1 Extreme.

Here’s what the settings app shows in its “Displays” tab:

I tried enabling fractional scaling, but then GNOME froze until I disconnected the Dell monitor.

macOS

When connecting the monitor to my MacBook Air M1 (2020), it defaults to a logical resolution of 3072x1728, i.e. 200% scaling.

For comparison, with Apple’s (5K) Studio Display, the default setting is 2560x1440 (200% scaling), or 2880x1620 (“More Space”, 177% scaling).

Observations

Matte screen

I remember the uproar when Lenovo introduced ThinkPads with glossy screens. At the time, I thought I prefer matte screens, but over the years, I heard that glossy screens are getting better and better, and consumers typically prefer them for their better picture quality.

The 8K monitor I’m using has a glossy screen on which reflections are quite visible. The MacBook Air’s screen shows fewer reflections in comparison.

Dell’s 6K monitor offers me a nice opportunity to see which option I prefer.

Surprisingly, I found that I don’t like the matte screen better!

It’s hard to describe, but somehow the picture seems more “dull”, or less bright (independent of the actual brightness of the monitor), or more toned down. The colors don’t pop as much.

Philosophical question: peripherals powered on by default?

One thing that I did not anticipate beforehand is the difference in how peripherals are treated when they are built into the monitor vs. when they are plugged into a USB hub.

I like to have my peripherals off-by-default, with “on” being the exceptional state. In fact, I leave my microphone disconnected and only plug its USB cable in when I need it. I also recently realized that I want sound to only be played on headphones, so I disconnected my normal speakers in favor of my Bluetooth dongle.

The 6K monitor, on the other hand, has all of its peripherals on-by-default, and bright red LEDs light up when the speaker or microphone is muted.

This is the opposite of how I want my peripherals to behave, but of course I understand why Dell developed the monitor with on-by-default peripherals.

Conclusion

Let’s go back to the questions I started the article with and answer them one by one:

1. Does the 6K monitor work well with most (all?) of my PCs and laptops?

   → Answer: The 6K monitor works a lot better than the 8K monitor, but that’s a low bar to clear. I would still call the 6K monitor finicky. Even when you run a latest-gen GPU with latest drivers, the monitor does not reliably show a picture after a suspend/resume cycle.

2. Is 6K resolution enough, or would I miss the 8K resolution?

   → Answer: I had really hoped that 6K would turn out to be enough, but the difference to 8K is visible with the bare eye. Just like 200% scaling is a nice step up from working at 96 dpi, 300% scaling (what I use on 8K) is another noticeable step up.

3. Is a matte screen the better option compared to the 8K monitor’s glossy finish?

   → Answer: While I don’t like the reflections in Dell’s 8K monitor, the picture quality is undeniably better compared to a matte screen. The 6K monitor just doesn’t look as good, and it’s not just about the difference in text sharpness.

4. Do the built-in peripherals work with Linux out of the box?

   → Answer: Yes, as far as I can tell. The webcam works fine with the generic uvcvideo USB webcam driver, the microphone and speakers work out of the box. I have not tested the presence sensor.

So, would I recommend the monitor? Depends on what you’re using as your current monitor and as the device you want to connect!

If you’re coming from a 4K display, the 6K resolution will be a nice step up. Connecting a MacBook Air M1 or newer is a great experience. If you want to connect PCs, be sure to use a new-enough nVidia GPU with latest drivers. Even under these ideal conditions, you might run into quirks like the no picture after resume problem. If you don’t mind early adopter pains like that, and are looking for a monitor that includes peripherals, go for it!

For me, switching from my 8K monitor would be a downgrade without enough benefits.

The ideal monitor for me would be a mixture between Dell’s 8K and 6K models:

• 8K resolution
  • …but with more modern connectivity options (one cable! works out of the box!).
• without built-in peripherals like webcam, microphone and speaker
  • …but with the USB KVM switch concept (monitor input coupled to USB upstream).
• glossy finish for best picture quality
  • …but with fewer reflections.

Maybe they’ll develop an updated version of the 8K monitor at some point?

With the team surrounding our previous paper on reduced-basis methods for quantum spin systems, Matteo Rizzi, Benjamin Stamm and Stefan Wessel and myself, we recently worked on a follow-up, extending our approach to tensor-network methods. Most of the work was done by Paul Brehmer, a master student in Stefan's group, whom I had the pleasure to co-supervise. Paul did an excellent job in cleaning up and extending the original code we had, which we have now released in open-source form as the ReducedBasis.jl Julia package.

The extension towards tensor-network methods and the integration with libraries such as ITensor.jl following the standard density-matrix renormalisation group (DMRG) approach, finally allows us to treat larger quantum spin systems, closer or at the level of the state of the art. In this work we demonstrate this by a number of different one-dimensional quantum spin-1 models, where our approach allowed us even to identify a few new phases, which have not been studied so far.

The full abstract of our paper reads

Within the reduced basis methods approach, an effective low-dimensional subspace of a quantum many-body Hilbert space is constructed in order to investigate, e.g., the ground-state phase diagram. The basis of this subspace is built from solutions of snapshots, i.e., ground states corresponding to particular and well-chosen parameter values. Here, we show how a greedy strategy to assemble the reduced basis and thus to select the parameter points can be implemented based on matrix-product-states (MPS) calculations. Once the reduced basis has been obtained, observables required for the computation of phase diagrams can be computed with a computational complexity independent of the underlying Hilbert space for any parameter value. We illustrate the efficiency and accuracy of this approach for different one-dimensional quantum spin-1 models, including anisotropic as well as biquadratic exchange interactions, leading to rich quantum phase diagrams.

In the past month I have renovated my appartment. Because of this I had to redo my entire desk setup. If you know me that means spending a lot time managing cables 😅. But I am really happy with the result. See for yourself …

I always wanted to be flexible in how I use the devices on my desk. I want to switch between using my laptop and desktop without having to replug everything. But I also want to be able to use certain devices from both at the same time. I have been using USB Hubs and the like. But I always was left wanting. To be fair my current solution is still not as perfect as in my dreams, but it is damn close.

So lets begin with the easy things. The monitors have multiple inputs, so I just connect those to my desktop and the docking station and voila. Well switching still requires me to use the monitor menus, but that I don’t really need to do that because I set them to “automatic mode” meaning the just show which ever device starts sending data first. And I don’t really need to use all monitors with the laptop when my desktop is running anyway so switching does not happen much.

For the keyboard and mouse I am using the “Logitech MX” keyboard and “Logitech MX Master” mouse. The can be paired with multiple Logitech wireless receivers. The devices can the be switch with the press of a button. Sadly switching one does not switch the other automatically which is still a little annyoing but I have seen some scripts that could be used to automate that as well. Maybe I will give that a shot. I still have a “USB Switch” which is connected to the desktop and laptop, it has a switch to toggle which device is “connected”. I mostly use it for my yubikey now. It also was fine for switching my previous mouse and keyboard.

There is still some room for improvements here, but that is not what has been bugging me. The parts I really wanted to be better are the Speakers, Microphone and the Webcam. In an ideal world they should be accessible on either device or both at the same time. Hence the USB Switch is not a good solution since that only enables operation with a single device at a time. Also using the USB switch is annyoing for other reasons. It means the audio dac is reset when switching devices resulting in an unpleasent noise coming out of my speakers. And also having all devices connected to a switch takes away the ability to attach usb sticks or other devices that i really only need temporarily and daisy chaning usb hubs often results in inconsistent behavior.

What would be a better solution?

Enter everybodies favorite single board computer the Raspberry Pi 🥧. Luckly I still have one lying around since getting one online is next to impossible if you don’t want to pay a scalper an unreasonable amount of money. Hopefully this will change. But anyway how can it help me acomplish my goal.

Thanks to a little something called networking computers can talk to each other. So it should be possible to attach the audio dac and webcam to the pi and then stream the video and audio data to both the laptop and desktop. What do we need to accomplish that.

1. Configure the Network
2. Setup Pipewire to run as system service
3. Enable audio streaming with the pipewire pulse server implementation
4. Enable laptop and desktop to discover audio devices
5. Setup USBIP for sharing the webcam

Network Setup

I do not want to share the devices with my entire home network, I just want to share with devices attached to the desk. Since the raspberry only has one network jack and using wireless for streaming data is not a great idea because of increased latency, the first thing I did was setup VLAN that is only availible to the devices on my desk.

First of the network switch needs to support VLANs, there are a lot of switches capable of doing this. They are a little more expensive then unmanged switches, but a basic models are availible starting at around 30€. I opted for a more expensive model from microtik (CSS610-8G-2S+) that is also able to support fibre glas connections instead of just RJ45. Then I configured the switch to setup the home network on each port in untagged mode. Then I created a VLAN with ID 2668 which will only be availible on the ports attached to the raspberry pi, desktop and laptop in tagged mode. The choise of the ID is arbirary, just make sure to not have clashes with other VLAN if you already have a more elaborate network setup at home.

Next the devices need to be configured to know about the VLAN and the IP address range needs to be configured. I like to use systemd-networkd for this. The configuration is done with three files in the /etc/systemd/network directory.

[root@pi ~]# tree /etc/systemd/network
/etc/systemd/network
|-- 0-audio.netdev
|-- 1-audio.network
`-- eth.network

The file 0-audio.netdev defines the VLAN:

[NetDev]
Name=audio
Kind=vlan

[VLAN]
Id=2668

The file eth.network configures the normal home network on the pi, here we need to add a line specifing that the VLAN is availible on this port:

[Match]
Name=eth*

[Network]
DHCP=yes
IPv6PrivacyExtensinos=true
VLAN=audio

Lastly the VLAN network needs to be configured. Since the pi is running continously it is useful to configure its ip statically and setup a DHCP server. All of this is configured with just a few lines in the 1-audio.network file.

[Match]
Name=audio

[Network]
Address=172.16.128.1/24
DHCPServer=true

[DHCPServer]
PoolOffset=100
PoolSize=100
EmitRouter=false

The same steps are used to configure the VLAN on the desktop and laptop, the only things that change are the interface names for the home network and that the audio vlans network can use the configured DHCP server to obtain a lease. Resulting in the follwing 1-audio.network file on the clients.

[Match]
Name=audio

[Network]
DHCP=yes

Of course to use systemd-networkd the service needs to be enabled: systemctl enable --now systemd-networkd.

Pipewire System Service

Pipewire intends to be a modern linux media deamon. It is still in active development. For now it already can be used as a replacement for pulseaudio or jack. Normally pipewire starts when you login to your user session. But since there is no desktop running on the pi pipewire needs to be configured to run as a system service.

First of the software packages need to be installed. I am more of a minimalist when it comes to the systems I configure, means I am running archlinux on the raspberry pi. The packages names might vary if you are running raspbian. For me doning

pacman -S pipewire pipewire-alsa pipewire-jack pipewire-pulse pipewire-zeroconf wireplumber pipewire-docs pipewire-audio realtime

installed all desired packages. There is not a lot of documentation on how to setup pipewire as a system service. I found this issue thread which lists all the steps required. Maybe the process will get simpler in the future, but for now a lot of steps are required.

First a pipewire user and group needs to be created with a statically assigned uid and gid. This is important to correctly set the environment variables in the service files created later. The pipewire user needs to be added to the audio and realtime group.

addgroup --gid 901 pipewire 
adduser --system  --uid 091 --gid 901 pipewire
for g in audio realtime; do sudo adduser pipewire ${g}; done

Next we need to add a configuration file /etc/security/limits.d/99-realtime-privileges.conf to allow the realtime group to change the process priorities to the levels recommended by pipewire.

@realtime - rtprio 98
@realtime - memlock unlimited
@realtime - nice -11

With the limits in place, the next step is to setup systemd units for pipewire, pipewire-pulse and wireplumber. In total 5 files need to be created:

• /etc/systemd/system/pipewire.socket
• /etc/systemd/system/pipewire.service
• /etc/systemd/system/pipewire-pulse.socket
• /etc/systemd/system/pipewire-pulse.service
• /etc/systemd/system/wireplumber.service

The content of these files is as follows.

#/etc/systemd/system/pipewire.socket
[Unit]
Description=PipeWire Multimedia System Socket

[Socket]
Priority=6
ListenStream=%t/pipewire/pipewire-0
SocketUser=pipewire
SocketGroup=pipewire
SocketMode=0660

[Install]
WantedBy=sockets.target

#/etc/systemd/system/pipewire.service
[Unit]
Description=PipeWire Multimedia Service
Before=gdm.service

# We require pipewire.socket to be active before starting the daemon, because
# while it is possible to use the service without the socket, it is not clear
# why it would be desirable.
#
# Installing pipewire and doing `systemctl start pipewire` will not get the
# socket started, which might be confusing and problematic if the server is to
# be restarted later on, as the client autospawn feature might kick in. Also, a
# start of the socket unit will fail, adding to the confusion.
#
# After=pipewire.socket is not needed, as it is already implicit in the
# socket-service relationship, see systemd.socket(5).
Requires=pipewire.socket

[Service]
User=pipewire
Type=simple
ExecStart=/usr/bin/pipewire
Restart=on-failure
RuntimeDirectory=pipewire
RuntimeDirectoryPreserve=yes
Environment=PIPEWIRE_RUNTIME_DIR=%t/pipewire
# Add if you need debugging
# Environment=PIPEWIRE_DEBUG=4

# These hardcoded runtime and dbus paths must stay this way for a system service
# as the User= is not resolved here 8(
## NOTE we do not change PIPEWIRE_RUNTIME_DIR as this is the system socket dir...
#Environment=PIPEWIRE_RUNTIME_DIR=/run/user/91/pipewire
Environment=XDG_RUNTIME_DIR=/run/user/91
Environment=DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/91/bus

#/etc/systemd/system/pipewire-pulse.socket
[Unit]
Description=PipeWire PulseAudio
Conflicts=pulseaudio.socket

[Socket]
Priority=6
ListenStream=%t/pulse/native
SocketUser=pipewire
SocketGroup=pipewire
SocketMode=0660

[Install]
WantedBy=sockets.target

#/etc/systemd/system/pipewire-pulse.service
[Unit]
Description=PipeWire PulseAudio

# We require pipewire-pulse.socket to be active before starting the daemon, because
# while it is possible to use the service without the socket, it is not clear
# why it would be desirable.
#
# A user installing pipewire and doing `systemctl --user start pipewire-pulse`
# will not get the socket started, which might be confusing and problematic if
# the server is to be restarted later on, as the client autospawn feature
# might kick in. Also, a start of the socket unit will fail, adding to the
# confusion.
#
# After=pipewire-pulse.socket is not needed, as it is already implicit in the
# socket-service relationship, see systemd.socket(5).
Requires=pipewire-pulse.socket
Wants=pipewire.service pipewire-session-manager.service
After=pipewire.service pipewire-session-manager.service
Conflicts=pulseaudio.service
# To ensure that multiple user instances are not created. May not be requiered
Before=gdm.service

[Service]
User=pipewire
Type=simple
ExecStart=/usr/bin/pipewire-pulse
Restart=on-failure
Slice=session.slice

# These hardcoded runtime and dbus paths must stay this way for a system service
# as the User= is not resolved here 8(
Environment=PULSE_RUNTIME_PATH=/home/pipewire
Environment=DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/91/bus

[Install]
Also=pipewire-pulse.socket
WantedBy=multi-user.target

#/etc/systemd/system/wireplumber.service   
[Unit]
Description=Multimedia Service Session Manager
After=pipewire.service
BindsTo=pipewire.service
Conflicts=pipewire-media-session.service

[Service]
User=pipewire

Type=simple
ExecStart=/usr/bin/wireplumber
Restart=on-failure
Slice=session.slice

# These hardcoded runtime and dbus paths must stay this way for a system service
# as the User= is not resolved here 8(
Environment=XDG_RUNTIME_DIR=/run/user/91
Environment=DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/91/bus

[Install]
WantedBy=pipewire.service
Alias=pipewire-session-manager.service

For the services to work correctly we need a running user session with dbus. This can be acomplished by telling loginctl to start a pipewire user session at system boot:

loginctl enable-linger pipewire 

Since running pipewire on the pi as a user is undesired the user services need to be masked.

systemctl --user --global mask pipewire.socket pipewire.service pipewire-pulse.socket pipewire-pulse.service wireplumber.service

After this the pipewire system services we just created can be enabled:

systemctl enable --now pipewire.socket pipewire.service pipewire-pulse.socket pipewire-pulse.service wireplumber.service

Configure Pipewire for Network Streaming

At this point piperwire is running on the raspberry after boot up. The next step is to setup network streaming. Thankfully that is easly done in two steps:

1. Setup Pipewire on the Raspberry Pi to be reachable via the VLAN and enable publishing of its devices via zeroconf
2. Setup clients (laptop, desktop) to listen for zeroconf announcements

For compatibility with the existing playback methods and to be a “drop-in” replacement pipewire has implementated a full pulseaudio server on top of itself. This way existing tools for managing audio playback and recording can still be used like pavucontrol. Pulseaudio supported being used over a network. This is not low latency so doing this over wifi is not really recommended, but over a wired connection the latencies are so low that it is not noticable. Pipewire supports this as well. So all we need to do to create a configuration file to configure network access:

# /etc/pipewire/pipewire-pulse.conf.d/network.conf 
pulse.properties = {
    # the addresses this server listens on
    pulse.min.frag = 32/48000           #0.5ms
    pulse.default.frag = 256/48000       #5ms 
    pulse.min.quantum = 32/48000        #0.5ms
    server.address = [
        "unix:native"
        #"unix:/tmp/something"              # absolute paths may be used
        #"tcp:4713"                         # IPv4 and IPv6 on all addresses
        #"tcp:[::]:9999"                    # IPv6 on all addresses
        #"tcp:127.0.0.1:8888"               # IPv4 on a single address
        #
        { address = "tcp:172.16.128.1:4713"             # address
          max-clients = 64                 # maximum number of clients
          listen-backlog = 32              # backlog in the server listen queue
          client.access = "allowed"     # permissions for clients
        }
    ]
}

Per default piperwire-pulse only enables the “unix:native” socket for access via dbus. To enable the network streaming the last 4 lines starting with address are of interest. In order to restict access to the VLAN the Ip address of the raspberry pi in the audio network needs to be specified. Also the client.access value needs to be set to “allowed” in order to enable all devices on that network to use it.

I also had to decrease the default values for pulse.min.frag, pulse.default.frag and pulse.min.quantum quite a bit in order for the latency of the mircophone to be usable while in a video call. Otherwise video and audio would be very out of sink. The pipewire documentation warns that this will increase CPU usage. I have not noticed a big impact on the raspberry pi 4 I am using to do this.

Next enabling the publishing of the pipewire server via zeroconf needs to be enabled. This could be done in the same configuration file. But for better overview over the configuration a created an extra configuration file:

# /etc/pipewire/pipewire-pulse.conf.d/publish.conf 
context.exec = [
  { path = "pactl"        args = "load-module module-zeroconf-publish" }
]

Thats really short. All we are doing is to tell the pulseaudio server to enable the zeroconf publish module. And on the clients we need to enable zeroconf discovery like this:

# /etc/pipewire/pipewire-pulse.conf.d/zeroconf-discover.conf 
context.exec = [
  { path = "pactl"        args = "load-module module-zeroconf-discover" }
]

For this to work the zeroconf deamon needs to be running. On linux the zeroconf implementation is provided by avahi. Most systems probably have it running already. On archlinux enable the avaih-daemon via systemd. The daemon also needs to be running on the raspberry pi for the publishing to work.

If everything worked correctly you should see the audio devices attached to the pi pop up in pavucontrol (after restarting the pipewire-pulse service for the configuration to apply):

Selecting the playback device or mircophone phone should now just work like with a locally attached the device. The really nice thing about this is that you can even use the devices from multiple clients at the same time!!!

Webcam

In theory pipewire is was written for camera device sharing between multiple applications. For example the webcam software cheese is already using pipewire. But I have found absolutly zero infromation if it whould be possible to do that via a network. Im not even really sure that this is on the roadmap. If it is I will definitly revisit this topic. The only other option I could think of was to somehow use some form of continous webcam broadcast that I could then somehow attach as a camera, but I also do not want the webcam to be active all the time.

So the solution I have come up with for now is to use USBIP. Which is a client server application to speak the USB protocol via the network. This comes with the drawback that the webcam can only be used by one device at a time, but at least I do not have to physically replug the device. Just issue a command to attach and detach it.

This can be done in a few simple steps:

1. Install usbip on server (pi) and client (laptop, desktop)
2. Enable the Service on both devices.
3. On pi bind webcam to usbip daemon
4. Attach/detach webcam via usbip daemon on the client

So the first to steps are the same for the client and server: Install the usbip package. Depending on your distribution it might be named differently. The enable the service using systemd: systemctl enable --now usbipd.

The next step is to bind the webcam to the usbipd daemon on the raspberry pi. For this the busid of the device needs to be found. This can be done by using the usbip utility:

$ usbip list -l
 - busid 1-1.1 (08bb:2902)
   Texas Instruments : PCM2902 Audio Codec (08bb:2902)

 - busid 1-1.2 (046d:08b6)
   Logitech, Inc. : unknown product (046d:08b6)

The webcam is the logitech device. Binding it to the daemon is as simple as running:

$ usbip bind -b 1-1.2
usbip: info: bind device on busid 1-1.2: complete

Now the device can be attached to the client. First we can also check that the device is availible to be attached:

$ usbip list -r 172.16.128.1
Exportable USB devices
======================
 - 172.16.128.1
      1-1.2: Logitech, Inc. : unknown product (046d:08b6)
           : /sys/devices/platform/scb/fd500000.pcie/pci0000:00/0000:00:00.0/0000:01:00.0/usb1/1-1/1-1.2
           : Miscellaneous Device / ? / Interface Association (ef/02/01)

The -r option is used to specify the remote server running usbip in this case the raspberry pi. Attaching/detaching is done with the commands:

$ sudo usbip attach -r 172.16.128.1 -b 1-1.2
$ sudo usbip detach -p 0

With the webcam attached it can be used like any other webcam. For example you could open cheese and take a picture:

After usage the webcam should be detached again, to make it possible for other clients to connect to it. If you forget to detach before powering of the device currently using the camera. You will login to the pi to unbind and rebind the device again, since usbip does not seem to have a timeout mechanism. A few other things to note about this setup are:

1. It is still not possible to use the device from multiple clients at the same time 😥
2. To make sure that the camera can only be used via the local VLAN a firewall configuration on the pi is required, since usbip is not confuriable to only listen on a certain network interface.
3. If you are getting an error when attaching the camera, you might also need to make sure the vhci-hcd kernel module is loaded!

I hope you enjoyed this post. If you have any further thoughts or questions. Feel free to reach out to me.

gokrazy is an appliance platform for Go programs: with just a few commands, you can deploy your Go program(s) on a Raspberry Pi or a (typically small) PC.

I’m excited to let you know that gokrazy now comes with a re-designed gok command line tool and gokrazy instance configuration mechanism!

Context: gokrazy in a few words

The traditional way to run Go software on a Raspberry Pi would be to install Raspbian or some other Linux distribution onto the SD card, copy over your program(s) and then maintain that installation (do regular updates).

I thought it would be nicer to run my Raspberry Pis such that only Go software is run by the Linux kernel on it, without any traditional Linux distribution programs like package managers or even the usual GNU Core Utilities.

gokrazy builds Go programs into a read-only SquashFS root file system image. When that image is started on a Raspberry Pi, a minimal init system supervises the Go programs, and a DHCP and NTP client configure the IP address and synchronize the time, respectively. After the first installation, all subsequent updates can be done over the network, with an A/B partitioning scheme.

I use gokrazy to, for example:

• Scan incoming paper mail into PDF files on Google Drive.

• Connect to the internet using router7, my small home internet router written in Go, running on a fast router PC build that handles a 25 Gbit/s Fiber To The Home connection.

• Automate the lights in my home, and control and monitor the heating.

• Offer Tailscale access to a Raspberry Pi Zero 2 W in my home network to then send Wake On Lan (WOL) packets before SSH’ing into my normally-suspended computers. See also my post DIY out-of-band management: remote console server (2022).

Before and after

Previously, the concept of gokrazy instance configuration was only a convention. Each gokrazy build was created using the gokr-packer CLI tool, and configured by the packer’s command-line flags, parameters, config files in ~/.config and per-package config files in the current directory (e.g. flags/github.com/gokrazy/breakglass/flags.txt).

Now, all gokrazy commands and tools understand the --instance flag (or -i for short), which determines the directory from which the Instance Config is read. For a gokrazy instance named “hello”, the default directory is ~/gokrazy/hello, which contains the config.json file.

Example: creating an instance for a Go working copy

Let’s say I have the evcc repository cloned as ~/src/evcc. evcc is an electric vehicle charge controller with PV integration, written in Go.

Now I want to run evcc on my Raspberry Pi using gokrazy. First, I create a new instance:

% gok -i evcc new
gokrazy instance configuration created in /home/michael/gokrazy/evcc/config.json
(Use 'gok -i evcc edit' to edit the configuration interactively.)

Use 'gok -i evcc add' to add packages to this instance

To deploy this gokrazy instance, see 'gok help overwrite'

Now let’s add our working copy of evcc to the instance:

% gok -i evcc add .
2023/01/15 18:55:39 Adding the following package to gokrazy instance "evcc":
  Go package  : github.com/evcc-io/evcc
  in Go module: github.com/evcc-io/evcc
  in local dir: /tmp/evcc
2023/01/15 18:55:39 Creating gokrazy builddir for package github.com/evcc-io/evcc
2023/01/15 18:55:39 Creating go.mod with replace directive
go: creating new go.mod: module gokrazy/build/github.com/evcc-io/evcc
2023/01/15 18:55:39 Adding package to gokrazy config
2023/01/15 18:55:39 All done! Next, use 'gok overwrite' (first deployment), 'gok update' (following deployments) or 'gok run' (run on running instance temporarily)

We might want to monitor this Raspberry Pi’s stats later, so let’s add the Prometheus node exporter to our gokrazy instance, too:

% gok -i evcc add github.com/prometheus/node_exporter
2023/01/15 19:04:05 Adding github.com/prometheus/node_exporter as a (non-local) package to gokrazy instance evcc
2023/01/15 19:04:05 Creating gokrazy builddir for package github.com/prometheus/node_exporter
2023/01/15 19:04:05 Creating go.mod before calling go get
go: creating new go.mod: module gokrazy/build/github.com/prometheus/node_exporter
2023/01/15 19:04:05 running [go get github.com/prometheus/node_exporter@latest]
go: downloading github.com/prometheus/node_exporter v1.5.0
[…]
2023/01/15 19:04:07 Adding package to gokrazy config

It’s time to insert an SD card (/dev/sdx), which we will overwrite with a gokrazy build:

% gok -i evcc overwrite --full /dev/sdx

See gokrazy quickstart for more detailed instructions.

Automation

The new gok subcommands (add, update, etc.) are much easier to manage than long gokr-packer command lines.

The new Automation page shows how to automate common tasks, be it daily updates via cron, or automated building in Continuous Integration environments like GitHub Actions.

Migration

Are you already a gokrazy user? If so, see the Instance Config Migration Guide for how to switch from the old gokr-packer tool to the new gok command.

Feedback / Questions?

If you have any questions, please feel free to reach out at gokrazy GitHub Discussions 👋

Similar to most people in their second PostDoc a considerable chunk of time in the past year has been devoted to job hunting, i.e. writing applications, preparing and attending interviews for junior research group positions. As the year is closing I am finally able to make a positive announcement in this regard: The Swiss ETH board has appointed me as Tenure Track Assistant Professor of Mathematics and of Materials Science and Engineering at EPF Lausanne, a position I am more than happy to take up. From March 2023 I will thus join this school and as part of this interdisciplinary appointment establish a research group located in both the mathematics and materials science institutes.

I am very grateful to the search committee as well as the ETH board and the university for this opportunity to start my own group and to be able to continue my research agenda combining ideas from mathematics and computer science to make materials simulations more robust and efficient. I look forward to becoming a part of the EPFL research environment and being able to contribute to the training of next generation researchers.

Along the lines of this appointment I now also have a few vacancies at PhD and PostDoc level to fill. Further information will be posted here as well as standard channels of the community early next year.

Note: If you don’t want to read the exposition and explanations and just want to know the steps I did, scroll to the summary at the bottom.

For a couple of years I have (with varying degrees of commitment) participated in Advent of Code, a yearly programming competition. It consists of fun little daily challenges. It is great to exercise your coding muscles and can provide opportunity to learn new languages and technologies.

So far I have created a separate repository for each year, with a directory per day. But I decided that I’d prefer to have a single repository, containing my solutions for all years. The main reason is that I tend to write little helpers that I would like to re-use between years.

When merging the repositories it was important to me to preserve the history of the individual years as well, though. I googled around for how to do this and the solutions I found didn’t quite work for me. So I thought I should document my own solution, in case anyone finds it useful.

You can see the result here. As you can see, there are four cleanly disjoint branches with separate histories. They then merge into one single commit.

One neat effect of this is that the merged repository functions as a normal remote for all the four old repositories. It involves no rewrites of history and all the previous commits are preserved exactly as-is. So you can just git pull from this new repository and git will fast-forward the branch.

Step 1: Prepare individual repositories

First I went through all repositories and prepared them. I wanted to have the years in individual directories. In theory, it is possible to use git-filter-repo and similar tooling to automate this step. For larger projects this might be worth it.

I found it simpler to manually make the changes in the individual repositories and commit them. In particular, I did not only need to move the files to the sub directory, I also had to fix up Go module and import paths. Figuring out how to automate that seemed like a chore. But doing it manually is a quick and easy sed command.

You can see an example of that in this commit. While that link points at the final, merged repository, I created the commit in the old repository. You can see that a lot of files simply moved. But some also had additional changes.

You can also see that I left the go.mod in the top-level directory. That was intentional - I want the final repository to share a single module, so that’s where the go.mod belongs.

After this I was left with four repositories, each of which had all the solutions in their own subdirectory, with a go.mod/go.sum file with the shared module path. I tested that all solutions still compile and appeared to work and moved on.

Step 2: Prepare merged repository

The next step is to create a new repository which can reference commits and objects in all the other repos. After all, it needs to contain the individual histories. This is simple by setting the individual repositories as remotes:

$ mkdir ~/src/github.com/Merovius/AdventOfCode
$ cd ~/src/github.com/Merovius/AdventOfCode
$ git init
$ git remote add 2018 ~/src/github.com/Merovius/aoc18
$ git remote add 2020 ~/src/github.com/Merovius/aoc_2020
$ git remote add 2021 ~/src/github.com/Merovius/aoc_2021
$ git remote add 2022 ~/src/github.com/Merovius/aoc_2022
$ git fetch --multiple 2018 2020 2021 2022
$ git branch -a
remotes/2018/master
remotes/2020/main
remotes/2021/main
remotes/2022/main

One thing worth pointing out is that at this point, the merged AdventOfCode repository does not have any branches itself. The only existing branches are remotes/ references. This is relevant because we don’t want our resulting histories to share any common ancestor. And because git behaves slightly differently in an empty repository. A lot of commands operate on HEAD (the “current branch”), so they have special handling if there is no HEAD.

Step 3: Create merge commit

A git commit can have an arbitrary number of “parents”:

• If a commit has zero parents, it is the start of the history. This is what happens if you run git commit in a fresh repository.
• If a commit has exactly one parent, it is a regular commit. This is what happens when you run git commit normally.
• If a parent has more than one parent, it is a merge commit. This is what happens when you use git merge or merge a pull request in the web UI of a git hoster (like GitHub or Gitlab).

Normally merge commits have two parents - one that is the “main” branch and one that is being “merged into”. However, git does not really distinguish between “main” and “merged” branch. And it also allows a branch to have more than two parents.

We want to create a new commit with four parents: The HEADs of our four individual repositories. I expected this to be simple, but:

$ git merge --allow-unrelated-histories remotes/2018/master remotes/2020/main remotes/2021/main remotes/2022/main
fatal: Can merge only exactly one commit into empty head

This command was supposed to create a merge commit with four parents. We have to pass --allow-unrelated-histories, as git otherwise tries to find a common ancestor between the parents and complains if it can’t find any.

But the command is failing. It seems git is unhappy using git merge with multiple parents if we do not have any branch yet.

I suspect the intended path at this point would be to check out one of the branches and then merge the others into that. But that creates merge conflicts and it also felt… asymmetric to me. I did not want to give any of the base repositories preference. So instead I opted for a more brute-force approach: Dropping down to the plumbing layer.

First, I created the merged directory structure:

$ cp -r ~/src/github.com/Merovius/aoc18/* .
$ cp -r ~/src/github.com/Merovius/aoc_2020/* .
$ cp -r ~/src/github.com/Merovius/aoc_2021/* .
$ cp -r ~/src/github.com/Merovius/aoc_2022/* .
$ vim go.mod # fix up the merged list of dependencies
$ go mod tidy
$ git add .

Note: The above does not copy hidden files (like .gitignore). If you do copy hidden files, take care not to copy any .git directories.

At this point the working directory contains the complete directory layout for the merged commit and it is all in the staging area (or “index”). This is where we normally run git commit. Instead we do the equivalent steps manually, allowing us to override the exact contents:

$ TREE=$(git write-tree)
$ COMMIT=$(git commit-tree $TREE \
    -p remotes/2018/master \
    -p remotes/2020/main \
    -p remotes/2021/main \
    -p remotes/2022/main \
    -m "merge history of all years")
$ git branch main $COMMIT

The write-tree command takes the content of the index and writes it to a “Tree Object” and then returns a reference to the Tree it has written.

A Tree is an immutable representation of a directory in git. It (essentially) contains a list of file name and ID pairs, where each ID points either to a “Blob” (an immutable file) or another Tree.

A Commit in git is just a Tree (describing the state of the files in the repository at that commit), a list of parents, a commit message and some meta data (like who created the commit and when).

The commit-tree command is a low-level command to create such a Commit object. We give it the ID of the Tree the Commit should contain and a list of parents (using -p) as well as a message (using -m). It then writes out that Commit to storage and returns its ID.

At this point we have a well-formed Commit, but it is just loosely stored in the repository. We still need a Branch to point at it, so it doesn’t get lost and we have a memorable handle.

You probably used the git branch command before. In the form above, it creates a new branch main (remember: So far our repository had no branches) pointing at the Commit we created.

And that’s it. We can now treat the repository as a normal git repo. All that is left is to publish it:

$ git remote add origin git@github.com:Merovius/AdventOfCode
$ git push --set-upstream origin main

Executive Summary

To summarize the steps I did:

1. Create commits in each of the old repositories to move files around and fixing anticipated merge conflicts as needed.
2. Create a pristine new repository without any branches:

   $ git init merged
   $ cd merged
   

3. Add the old repositories as remotes for the merged repo:

   $ git remote add <repo1> /path/to/repo1
   $ git fetch repo1
   $ git remote add <repo2> /path/to/repo2
   $ git fetch repo2
   $ # …
   

4. Copy files from old repositories into merged repo:

   $ cp -r /path/to/repo1/* .
   $ cp -r /path/to/repo2/* .
   $ # …
   

5. Create commit using plumbing commands:

   $ git add .
   $ TREE=$(git write-tree)
   $ COMMIT=$(git commit-tree $TREE \
       -m "merge repositories" \
       -p remotes/repo1/main \
       -p remotes/repo2/main)
   $ git branch main $COMMIT
   

This post serves as a summary for a live code I did at our local hacker space. For the full experience please refer to the recording. Though I probably should warn that the live coding was done in German (and next time I should make sure to increase the font size everywhere for the recording 🙈).

From zero to a working rust project for the raspberry pi. These are the required steps:

• Setup Rust Project with cargo
• Install Rust Arm + Raspberry Pi Toolchain
• Configure Rust Project for cross compilation
• Import crate for GPIO Access
• Profit 💰

Setting up a Rust Project

The first step is to setup a rust project. This is easily accomplished by using the rust tooling. Using cargo it is possible it initialize a hello world rust project:

> mkdir pi_project
> cd pi_project
> cargo init

This results in the following project structure:

pi_project
├── Cargo.toml
├── .gitignore
└── src
    └── main.rs

Building and running the code is now as simple as running:

> cargo build
> ./target/debug/pi_project
Hello, world!

Looking at the executable we see that the code was build for the x86 Architecture.

> file ./target/debug/pi_project
target/debug/pi_project: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, BuildID[sha1]=0461b95d992ecda8488ad610bb1818344c1eeb8d, for GNU/Linux 4.4.0, with debug_info, not stripped

To be able to run this code on the raspberry pi the target architecture needs to change to ARM.

Rust Arm Toolchain Setup

Installing a different target architecture is easy. All that is required is to use rustup. Warning the following list does not mean that your specific pi revision will work, you need to make extra sure to select the correct architecture based on the model of pi you are using! There are differences per revision of the pi.

# for raspberry pi 3/4
> rustup target add aarch64-unknown-linux-gnu
# for raspberry pi 1/zero
> rustup target add arm-unknown-linux-gnueabihf 

This allows telling the cargo to generate ARM machine code. This would be all we need if the goal was to write bare metal code. But just running cargo build --target arm-unknown-linux-gnueabihf results in an error. This because we still need a linker and the matching system libraries to be able to interface correctly with the Linux kernel running on the pi.

This problem is solved by installing a raspberry pi toolchain. The toolchain can be downloaded from here. They are compatible with the official “Raspian OS” for the pi. If you are running a different OS on your PI, you may need to look further to find the matching toolchain for your OS.

In this case the pi is running the newest Raspian, which is based on Debian 11:

> wget https://sourceforge.net/projects/raspberry-pi-cross-compilers/files/Raspberry%20Pi%20GCC%20Cross-Compiler%20Toolchains/Bullseye/GCC%2010.3.0/Raspberry%20Pi%201%2C%20Zero/cross-gcc-10.3.0-pi_0-1.tar.gz/download -O toolchain.tar.gz
> tar -xvf toolchain.tar.gz 

Configure cross compilation

Now the rust build system needs to be configured to use the toolchain. This is done by placing a config file in the project root:

pi_project
├── .cargo
│   └── config
├── Cargo.lock
├── Cargo.toml
├── .gitignore
└── src
    └── main.rs

The configuration instructs the cargo build system to use the cross compiler gcc as linker and sets the directory where arm system libraries are located.

# content of .cargo/config
[build]
target = "arm-unknown-linux-gnueabihf" #set default target

#for raspberry pi 1/zero
[target.arm-unknown-linux-gnueabihf]
linker = "/home/judge/.toolchains/cross-pi-gcc-10.3.0-0/bin/arm-linux-gnueabihf-gcc"
rustflags = [
    "-C", "link-arg=--sysroot=/home/judge/.toolchains/cross-pi-gcc-10.3.0-0/arm-linux-gnueabihf/libc"
]

#for raspberry pi 3/4
[target.aarch64-unknown-linux-gnu]
linker = "/home/judge/.toolchains/cross-pi-gcc-10.3.0-64/bin/aarch64-linux-gnu-gcc"
rustflags = [
    "-C", "link-arg=--sysroot=/home/judge/.toolchains/cross-pi-gcc-10.3.0-0/aarch64-linux-gnu/libc"
]

This sets the default target of the project to arm-unknown-linux-gnueabihf, now running cargo build results in the following ARM binary being created.

file target/arm-unknown-linux-gnueabihf/debug/pi_project
target/arm-unknown-linux-gnueabihf/debug/pi_project: ELF 32-bit LSB pie executable, ARM, EABI5 version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux-armhf.so.3, for GNU/Linux 3.2.0, with debug_info, not stripped

It can now be copied to the raspberry pi and be executed.

GPIO Access

Until this point the source of the application was not touched. This changes now because just executing

// contents of src/main.rs
fn main() {
    println!("Hello World!");
}

is boring! If we have a raspberry pi it would be much more fun to use it to control some hardware 💪. Thankfully there already is a library that we can use to do just that. rppal enables access to the GPIO pins of the pi. Including the library in the project requires declaring it as a dependency in the Cargo.toml.

[dependencies]
rppal = "0.14.0"

Now we can use the library to make an led blink.

use std::thread;
use std::time::Duration;

use rppal::gpio::Gpio;

// Gpio uses BCM pin numbering. BCM GPIO 23 is tied to physical pin 16.
const GPIO_LED: u8 = 23;

fn main() {
    let gpio = Gpio::new().expect("Unable to access GPIO!");
    let mut pin = gpio.get(GPIO_LED).unwrap().into_output();

    loop {
        pin.toggle();
        thread::sleep(Duration::from_millis(500));
    }
}

And that’s basically it. Now we can use rust to program the raspberry pi to do any task we want. We can even get fancy and use an async runtime to execute many tasks in parallel.

I hope this summary is useful to you and feel free to contact me if you have questions or find this post useful.

Happy coding 🧑‍💻 …

The goal of quantum-chemical calculations is the simulation of materials and molecules. In density-functional theory (DFT) the first step along this line is obtaining the electron density minimising an energy functional. However, since energies and the density are usually not very tractable quantities in an experimental setup, comparison to experiment and scientific intuition also requires the computation of properties. Important properties include the forces (i.e. the energetic change due to a displacement of the structure) polarisabilities (change in dipole moment due to an external electric field) or phonon spectra (which can be measured using infrared spectroscopy). Therefore an efficient and reliable property computation is crucial to make quantum-chemical simulations interpretable and to close the loop back to experimentalists.

In DFT property calculations are done using density-functional perturbation theory (DFPT), which essentially computes the linear response of the electronic structure to the aforementioned changes in external conditions (external field, nuclear displacements etc.). Solving the equations underlying DFPT can become numerically challenging as (especially for metallic systems) the equations are ill-conditioned.

In a collaboration with my former PostDoc advisor Benjamin Stamm and my old group at the CERMICS at École des Ponts, including Eric Cancès, Antoine Levitt, Gaspard Kemlin, we just published an article, where we provide a more mathematical take on DFPT. In our work we provide an extensive review of various practical setups employed in main-stream codes such as ABINIT and QuantumEspresso from a numerical analysis point of view, highlighting the differences and similarities of these approaches. Moreover we develop a novel approach approach to solve the so-called Sternheimer equations (a key component of DFPT), which allows to make better use of the byproducts available in standard SCF schemes (the algorithm used to obtain the DFT ground state). With our approach we show savings up to 40% in the number of matrix-vector products required to solve the response equations. Since these are the most expensive step in DFPT this implies a similar saving in computational cost overall. Naturally our algorithm has been implemented as the default response solver in our DFTK code, starting from version 0.5.9.

Most of this work was done during a two-month visit of Gaspard Kemlin with Benjamin and myself here in Aachen. I think I speak for the both of us when I say that it has been a great pleasure to have Gaspard around, both on a professional as well as a personal level.

The full abstract of the paper reads

Response calculations in density functional theory aim at computing the change in ground-state density induced by an external perturbation. At finite temperature these are usually performed by computing variations of orbitals, which involve the iterative solution of potentially badly-conditioned linear systems, the Sternheimer equations. Since many sets of variations of orbitals yield the same variation of density matrix this involves a choice of gauge. Taking a numerical analysis point of view we present the various gauge choices proposed in the literature in a common framework and study their stability. Beyond existing methods we propose a new approach, based on a Schur complement using extra orbitals from the self-consistent-field calculations, to improve the stability and efficiency of the iterative solution of Sternheimer equations. We show the success of this strategy on nontrivial examples of practical interest, such as Heusler transition metal alloy compounds, where savings of around 40% in the number of required cost-determining Hamiltonian applications have been achieved.

I was pleasantly surprised by how easy it was to make it possible to push a PC’s power button remotely via MQTT by wiring up an ESP32 microcontroller, a MOSFET, a resistor, and a few jumper wires.

While a commercial solution like IPMI offers many more features like remote serial, or remote image mounting, this DIY solution feels really magical, and has great price performance if all you need is power management.

Motivation

To save power, I want to shut down my network storage PC when it isn’t currently needed.

For this plan to work out, my daily backup automation needs to be able to turn on the network storage PC, and power it back off when done.

Usually, I implement that via Wake On LAN (WOL). But, for this particular machine, I don’t have an ethernet network link, I only have a fiber link. Unfortunately, it seems like none of the 3 different 10 Gbit/s network cards I tested has functioning Wake On LAN, and when I asked on Twitter, none of my followers had ever seen functioning WOL on any 10 Gbit/s card. I suppose it’s not a priority for the typical target audience of these network cards, which go into always-on servers.

I didn’t want to run an extra 10 Gbit/s switch just for WOL over an ethernet connection, because switches like the MikroTik CRS305-1G-4S+IN consume at least 10W. As the network storage PC only consumes about 20W overall, I wanted a more power-efficient option.

Hardware and Wiring

The core of this DIY remote power button is a WiFi-enabled micro controller such as the ESP32. To power the micro controller, I use the 5V standby power on the mainboard’s USB 2.0 pin headers, which is also available when the PC is turned off and only the power supply (PSU) is turned on. A micro controller with an on-board 5V voltage regulator is convenient for this.

Aside from the micro controller, we also need a transistor or logic-level MOSFET to simulate a push of the power button, and a resistor to control the transistor. An opto coupler is not needed, since the ESP32 is powered from the mainboard, not from a separate power supply.

The mainboard’s front panel header contains a POWERBTN# signal (3.3V), and a GND signal. When connecting a typical PC case power button to the header, you don’t need to pay attention to the polarity. This is because the power button just physically connects the two signals.

In our case, the polarity matters, because we need the 3.3V on the transistor’s drain pin, otherwise we won’t be able to control the transistor via its base pin. The POWERBTN# 3.3V signal is typically labeled + on the mainboard (or in the manual), whereas GND is labeled -. If you are unsure, double-check the voltage using a multimeter.

Bill of Materials

• WiFi-enabled microcontroller with 5V power input, e.g. the Espressif ESP32 Pico Kit
• transistor or logic-level MOSFET for working with 3.3V, e.g. 2N7000 (→digikey)
• 1K resistor for controlling the transistor, e.g. CF14JT1K00
• a bread board and/or case for mounting, e.g. Adafruit Perma-Proto.

Schematic

Software: ESPHome

I wanted a quick solution (with ideally no custom firmware development) and was already familiar with ESPHome, which turns out to very easily implement the functionality I wanted :)

In addition to a standard ESPHome configuration, I have added the following lines to make the GPIO pin available through MQTT, and make it a momentary switch instead of a toggle switch, so that it briefly presses the power button and doesn’t hold the power button:

switch:
  - platform: gpio
    pin: 25
    id: powerbtn
    name: "powerbtn"
    restore_mode: ALWAYS_OFF
    on_turn_on:
    - delay: 500ms
    - switch.turn_off: powerbtn

I have elided the full configuration for brevity, but you can click here to see it:

esphome:
  name: poweresp

esp32:
  board: pico32
  framework:
    type: arduino

# Enable logging
logger:

mqtt:
  broker: 10.0.0.54

ota:
  password: ""

wifi:
  ssid: "essid"
  password: "secret"

  # Enable fallback hotspot (captive portal) in case wifi connection fails
  ap:
    ssid: "Poweresp Fallback Hotspot"
    password: "secret2"

captive_portal:

switch:
  - platform: gpio
    pin: 25
    id: powerbtn
    name: "powerbtn"
    restore_mode: ALWAYS_OFF
    on_turn_on:
    - delay: 500ms
    - switch.turn_off: powerbtn

For the first flash, I used:

docker run --rm \
  -v "${PWD}":/config \
  --device=/dev/ttyUSB0 \
  -it \
  esphome/esphome \
    run poweresp.yaml

To update over the network after making changes (serial connection no longer needed), I used:

docker run --rm \
  -v "${PWD}":/config \
  -it \
  esphome/esphome \
    run poweresp.yaml

In case you want to learn more about the relevant ESPHome concepts, here are a few pointers:

• https://esphome.io/components/wifi.html might need to set use_address
• https://esphome.io/components/switch/index.html
  • and https://esphome.io/components/switch/gpio.html
• https://esphome.io/components/mqtt.html

Integration into automation

To push the power button remotely from Go, I’m using the following code:

func pushMainboardPower(mqttBroker, clientID string) error {
	opts := mqtt.NewClientOptions().AddBroker(mqttBroker)
	if hostname, err := os.Hostname(); err == nil {
		clientID += "@" + hostname
	}
	opts.SetClientID(clientID)
	opts.SetConnectRetry(true)
	mqttClient := mqtt.NewClient(opts)
	if token := mqttClient.Connect(); token.Wait() && token.Error() != nil {
		return fmt.Errorf("connecting to MQTT: %v", token.Error())
	}

	const topic = "poweresp/switch/powerbtn/command"
	const qos = 0 // at most once (no re-transmissions)
	const retained = false
	token := mqttClient.Publish(topic, qos, retained, string("on"))
	if token.Wait() && token.Error() != nil {
		return fmt.Errorf("publishing to MQTT: %v", token.Error())
	}

	return nil
}

Conclusion

I hope this small project write-up is useful to others in a similar situation!

If you need more features than that, check out the next step on the feature and complexity ladder: PiKVM or TinyPilot. See also this comparison by Jeff Geerling.

For the guest WiFi at an event that eventually fell through, we wanted to tunnel all the traffic through my internet connection via my home router.

Because the event is located in another country, many hours of travel away, there are a couple of scenarios where remote control of my home router can be a life-saver. For example, should my home router crash, remotely turning power off and on again gets the event back online.

But, power-cycling a machine is a pretty big hammer. For some cases, like locking yourself out with a configuration mistake, a more precise tool like a remote serial console might be nicer.

In this article, I’ll present two cheap and pragmatic DIY out-of-band management solutions that I have experimented with in the last couple of weeks and wanted to share:

• Variant 1: Remote Power Management only
• Variant 2: a full Remote Console Server

You can easily start with the first variant and upgrade it into the second variant later.

Variant 1: Remote Power Management

Architecture Diagram

Here is the architecture of the system at a glance. The right-hand side is the existing router I want to control, the left-hand side shows the out of band management system:

Let’s go through the hardware components from top to bottom.

Hardware: 4G WiFi Router (Out Of Band Network)

The easiest way to have another network connection for projects like this one is the digitec iot subscription. They offer various different options, and their cheapest one, a 0.4 Mbps flatrate for 4 CHF per month, is sufficient for our use-case.

A convenient way of making the digitec iot subscription available to other devices is to use a mobile WiFi router such as the TP-Link M7350 4G/LTE Mobile Wi-Fi router (68 CHF). You can power it via USB, and it has a built-in battery that will last for a few hours.

By default, the device turns itself off after a while when it thinks it is unused, which is undesired for us — if the smart plug drops out of the WiFi, we don’t want the whole system to go offline. You can turn off this behavior in the web interface under Advanced → Power Saving → Power Saving Mode.

Hardware: WiFi Smart Plug

With the out of band network connection established, all you need to remotely toggle power is a smart plug such as the Sonoff S26 WiFi Smart Plug.

The simplest setup is to connect the Smart Plug to the 4G router via WiFi, and control it using Sonoff’s mobile app via Sonoff’s cloud.

Non-cloud solution

Alternatively, if you want to avoid the Sonoff cloud, the device comes with a “DIY mode”, but the DIY mode wouldn’t work reliably for me when I tried it. Instead, I flashed the Open Source Tasmota firmware and connected it to a self-hosted MQTT server via the internet.

Of course, now your self-hosted MQTT server is a single point of failure, but perhaps you prefer that over the Sonoff cloud being a single point of failure.

Variant 2: Remote Console Server

Turning power off and on remotely is a great start, but what if you need actual remote access to a system? In my case, I’m using a serial port to see log messages and run a shell on my router. This is also called a “serial console”, and any device that allows accessing a serial console without sitting physically in front of the serial port is called a “remote console server”.

Commercially available remote console servers typically offer lots of ports (up to 48) and cost lots of money (many thousand dollars or equivalent), because their target application is to be installed in a rack full of machines in a lab or data center. A few years ago, I built freetserv, an open source, open hardware solution for this problem.

For the use-case at hand, we only need a single serial console, so we’ll do it with a Raspberry Pi.

Architecture Diagram

The architecture for this variant looks similar to the other variant, but adds the consrv Raspberry Pi Zero 2 W and a USB-to-serial adapter:

Hardware: Raspberry Pi Zero 2 W

We’ll use a Raspberry Pi Zero 2 W as our console server. While the device is a little slower than a Raspberry Pi 3 B, it is still plenty fast enough for providing a serial console, and it only consumes 0.8W of power (see gokrazy → Supported platforms for a comparison):

If the Pi Zero 2 W is not available, you can try using any other Raspberry Pi supported by gokrazy, or even an older Pi Zero with the community-supported Pi OS 32-bit kernel (I didn’t test that).

Our Pi will have at least two tasks:

1. With a USB-to-serial adapter, the Pi will provide a serial console.
2. The Pi will run Tailscale mesh networking, which will transparently use either the wired network or fail over to the Out Of Band network. Tailscale also frees us from setting up port forwardings, dynamic DNS or anything like that.
3. Optionally, the Pi can run a local MQTT server if you want to avoid the Sonoff cloud.

Hardware: USB-to-serial adapter

You can use any USB-to-serial adapter supported by Linux. Personally, I like the Adafruit FT232H adapter, which I like to re-program with FTDI’s FT_Prog so that it has a unique serial number.

In my router, I plugged in an Longshine LCS-6321M serial PCIe card to add a serial port. Before you ask: no, using USB serial consoles for the kernel console does not cut it.

Hardware: USB ethernet adapter

Because we not only want this Raspberry Pi to be available via the Out Of Band network (via WiFi), but also on the regular home network, we need a USB ethernet adapter.

Originally I was going to use the Waveshare ETH-USB-HUB-BOX: Ethernet / USB HUB BOX for Raspberry Pi Zero Series, but it turned out to be unreliable.

Instead, I’m now connecting a USB hub (as the Pi Zero 2 W has only one USB port), a Linksys USB3GIG network adapter I had lying around, and my USB-to-serial adapter.

gokrazy setup

Just like in the gokrazy quickstart, we’re going to create a directory for this gokrazy instance:

INSTANCE=gokrazy/consrv
mkdir -p ~/${INSTANCE?}
cd ~/${INSTANCE?}
go mod init consrv

You could now directly run gokr-packer, but personally, I like putting the gokr-packer command into a Makefile right away:

# The consrv hostname resolves to the device’s Tailscale IP address,
# once Tailscale is set up.
PACKER := gokr-packer -hostname=consrv

PKGS := \
	github.com/gokrazy/breakglass \
	github.com/gokrazy/timestamps \
	github.com/gokrazy/serial-busybox \
	github.com/gokrazy/stat/cmd/gokr-webstat \
	github.com/gokrazy/stat/cmd/gokr-stat \
	github.com/gokrazy/mkfs \
	github.com/gokrazy/wifi \
	tailscale.com/cmd/tailscaled \
	tailscale.com/cmd/tailscale \
	github.com/mdlayher/consrv/cmd/consrv

all:

.PHONY: update overwrite

update:
	${PACKER} -update=yes ${PKGS}

overwrite:
	${PACKER} -overwrite=/dev/sdx ${PKGS}

For the initial install, plug the SD card into your computer, put its device name into the overwrite target, and run make overwrite.

For subsequent changes, you can use make update.

Tailscale

Tailscale is a peer-to-peer mesh VPN, meaning we can use it to connect to our consrv Raspberry Pi from anywhere in the world, without having to set up port forwardings, dynamic DNS, or similar.

As an added bonus, Tailscale also transparently fails over between connections, so while the fast ethernet/fiber connection works, Tailscale uses that, otherwise it uses the Out Of Band network.

Follow the gokrazy guide on Tailscale to include the device in your Tailscale mesh VPN.

WiFi internet connection and dual homing

Setup WiFi:

mkdir -p extrafiles/github.com/gokrazy/wifi/etc
cat '{"ssid": "oob", "psk": "secret"}' \
  > extrafiles/github.com/gokrazy/wifi/etc/wifi.json

consrv should use the Out Of Band mobile uplink to reach the internet. At the same time, it should still be usable from my home network, too, to make gokrazy updates go quickly.

We accomplish this using route priorities.

I arranged for the WiFi interface to have higher route priority (5) than the ethernet interface (typically 1, but 11 in our setup thanks to the -extra_route_priority=10 flag):

mkdir -p flags/github.com/gokrazy/gokrazy/cmd/dhcp
echo '-extra_route_priority=10' \
  > flags/github.com/gokrazy/gokrazy/cmd/dhcp/flags.txt
make update

Now, tailscale netcheck shows an IPv4 address belonging to Sunrise, the mobile network provider behind the digitec iot subscription.

The consrv Console Server

consrv is an SSH serial console server written in Go that Matt Layher and I developed. If you’re curious, you can watch the two of us creating it in this twitch stream recording:

The installation of consrv consists of two steps.

Step 1 is done: we already included consrv in the Makefile earlier in gokrazy setup.

So, we only need to configure the desired serial ports in consrv.toml (in gokrazy extrafiles):

mkdir -p extrafiles/github.com/mdlayher/consrv/cmd/consrv/etc/consrv
cat > extrafiles/github.com/mdlayher/consrv/cmd/consrv/etc/consrv/consrv.toml <<'EOT'
[server]
address = ":2222"

[[devices]]
serial = "01716A92"
name = "router7"
baud = 115200
logtostdout = true

[[identities]]
name = "michael"
public_key = "ssh-ed25519 AAAAC3… michael@midna"
EOT

Run make update to deploy the configuration to your device.

If everything is set up correctly, we can now start a serial console session via SSH:

midna% ssh -p 2222 router7@consrv.lan
Warning: Permanently added '[consrv.lan]:2222' (ED25519) to the list of known hosts.
consrv> opened serial connection "router7": path: "/dev/ttyUSB0", serial: "01716A92", baud: 115200
2022/06/19 20:50:47 dns.go:175: probe results: [{upstream: [2001:4860:4860::8888]:53, rtt: 999.665µs} {upstream: [2001:4860:4860::8844]:53, rtt: 2.041079ms} {upstream: 8.8.8.8:53, rtt: 2.073279ms} {upstream: 8.8.4.4:53, rtt: 16.200959ms}]
[…]

I’m using the logtostdout option to make consrv continuously read the serial port and send it to stdout, which gokrazy in turn sends via remote syslog to the gokrazy syslog daemon, running on another machine. You could also run it on the same machine if you want to log to file.

Controlling Tasmota from breakglass

You can use breakglass to interactively log into your gokrazy installation.

If you flashed your Smart Plug with Tasmota, you can easily turn power on from a breakglass shell by directly calling Tasmota’s HTTP API with curl:

% breakglass consrv
consrv# curl -v -X POST --data 'cmnd=power on' http://tasmota_68462f-1583/cm

The original Sonoff firmware offers a DIY mode which should also offer an HTTP API, but the DIY mode did not work in my tests. Hence, I’m only describing how to do it with Tasmota.

Optional: Local MQTT Server

Personally, I like having the Smart Plug available both on the local network (via Tasmota’s HTTP API) and via the internet with an external MQTT server. That way, even if either option fails, I still have a way to toggle power remotely.

But, maybe you want to obtain usage stats by listening to MQTT or similar, and you don’t want to use an extra server for this. In that situation, you can easily run a local MQTT server on your Pi.

In the gokrazy Makefile, add github.com/fhmq/hmq to the list of packages to install, and configure Tasmota to connect to consrv on port 1883.

To check that everything is working, use mosquitto_sub from another machine:

midna% mosquitto_sub --verbose -h consrv.monkey-turtle.ts.net -t '#'

Conclusion

digitec’s IOT mobile internet subscription makes remote power management delightfully easy with a smart plug and 4G WiFi router, and affordable enough. The subscription is flexible enough that you can decide to only book it while you’re traveling.

We can elevate the whole setup in functionality (but also complexity) by combining Tailscale, consrv and gokrazy, running on a Raspberry Pi Zero 2 W, and connecting a USB-to-serial adapter.

If you need more features than that, check out the next step on the feature and complexity ladder: PiKVM or TinyPilot. See also this comparison by Jeff Geerling.

Appendix A: Unstable Apple USB ethernet adapter

The first USB ethernet adapter I tried was the Apple USB Ethernet Adapter.

Unfortunately, after a few days of uptime, I experienced the following kernel driver crash (with the asix Linux driver), and the link remained down until I rebooted.

I then switched to a Linksys USB3GIG network adapter (supported by the r8152 Linux driver) and did not see any problems with that so far.

dwc2 3f980000.usb: dwc2_hc_chhltd_intr_dma: Channel 5 - ChHltd set, but reason is unknown
dwc2 3f980000.usb: hcint 0x00000002, intsts 0x04600009
dwc2 3f980000.usb: dwc2_update_urb_state_abn(): trimming xfer length
asix 1-1.4:1.0 eth0: Failed to read reg index 0x0000: -71
------------[ cut here ]------------
WARNING: CPU: 1 PID: 7588 at drivers/net/phy/phy.c:942 phy_error+0x10/0x58
Modules linked in: brcmfmac brcmutil
CPU: 1 PID: 7588 Comm: kworker/u8:2 Not tainted 5.18.3 #1
Hardware name: Raspberry Pi Zero 2 W Rev 1.0 (DT)
Workqueue: events_power_efficient phy_state_machine
pstate: 80000005 (Nzcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
pc : phy_error+0x10/0x58
lr : phy_state_machine+0x258/0x2b0
sp : ffff800009fe3d40
x29: ffff800009fe3d40 x28: 0000000000000000 x27: ffff6c7ac300c078
x26: ffff6c7ac300c000 x25: ffff6c7ac4390000 x24: 00000000ffffffb9
x23: 0000000000000004 x22: ffff6c7ac4019cd8 x21: ffff6c7ac4019800
x20: ffffce5c97f6f000 x19: ffff6c7ac4019800 x18: 0000000000000010
x17: 0000000400000000 x16: 0000000000000000 x15: 0000000000001007
x14: ffff800009fe3810 x13: 00000000ffffffea x12: 00000000fffff007
x11: fffffffffffe0290 x10: fffffffffffe0240 x9 : ffffce5c988e1018
x8 : c0000000fffff007 x7 : 00000000000000a8 x6 : ffffce5c98889280
x5 : 0000000000000268 x4 : ffff6c7acf392b80 x3 : ffff6c7ac4019cd8
x2 : 0000000000000000 x1 : 0000000000000000 x0 : ffff6c7ac4019800
Call trace:
 phy_error+0x10/0x58
 phy_state_machine+0x258/0x2b0
 process_one_work+0x1e4/0x348
 worker_thread+0x48/0x418
 kthread+0xf4/0x110
 ret_from_fork+0x10/0x20
---[ end trace 0000000000000000 ]---
asix 1-1.4:1.0 eth0: Link is Down

(Cross-post from our report published in the Psi-k blog)

From 20th until 24th June 2022 I co-organised a workshop on the theme of Error control in first-principles modelling at the CECAM Headquarters in Lausanne (workshop website). For one week the workshop unified like-minded researchers from a range of communities, including quantum chemistry, materials sciences, scientific computing and mathematics to jointly discuss the determination of errors in atomistic modelling. The main goal was to obtain a cross-community overview of ongoing work and to establish new links between the disciplines.

Amongst others we discussed topics such as: the determination of errors in observables, which are the result of long molecular dynamics simulations, the reliability and efficiency of numerical procedures and how to go beyond benchmarking or convergence studies via a rigorous mathematical understanding of errors. We further explored interactions with the field of uncertainty quantification to link numerical and modelling errors in electronic structure calculations or to understand error propagation in interatomic potentials via statistical inference.

Organisers

• Gabor Csanyi (University of Cambridge)
• Genevieve Dusson (CNRS & Université Bourgogne Franche-Comté)
• Michael Herbst (RWTH Aachen University)
• Youssef Marzouk (Massachusetts Institute of Technology)

Participants

A primary objective of the conference was to facilitate networking and exchange across communities. Thanks to the funds provided by CECAM and Psi-k we managed to get a crowd of 30 researchers, including about 15 junior researchers, to come to Lausanne in person. Moreover we made an effort to enable a virtual participation to the smoothest extent possible. For example we provided a conference-specific Slack space, which grew into a platform for discussion involving both in-person as well as virtual participants during the conference. In this way in total about 70 researchers from 18 countries could participate in the workshop. The full list of participants is available on the workshop website.

Workshop programme

The workshop programme was split between the afternoon sessions, in which we had introductory and topic-specific lectures, as well as the morning sessions, which were focussed on informal discussion and community brainstorming.

Afternoon lectures

Monday June 20th 2022

• Uncertainty quantification for atomic-scale machine learning. (Michele Ceriotti, EPFL)
  [slides] [recording]
• Testing the hell out of DFT codes with virtual oxides. (Stefaan Cottenier, Ghent University)
  [slides] [recording]
• Prediction uncertainty validation for computational chemists. (Pascal Pernot, Université Paris-Saclay)
  [slides] [recording]
• Uncertainty driven active learning of interatomic potentials for molecular dynamics (Boris Kozinsky, Harvard University)
  [recording]
• Interatomic Potentials from First Principles (Christoph Ortner, University of British Columbia)
  [slides] [recording]

Tuesday June 21st 2022

• Numerical integration in the Brillouin zone (Antoine Levitt, Inria Paris)
  [slides] [recording]
• Sensitivity analysis for assessing and controlling errors in theoretical spectroscopy and computational biochemistry (Christoph Jacob,
  TU Braunschweig)
  [slides]
• Uncertainty quantification and propagation in multiscale materials modelling (James Kermode, University of Warwick)
  [slides] [recording]
• Uncertainty Quantification and Active Learning in Atomistic Computations
  (Habib Najm, Sandia National Labs)
• Nuances in Bayesian estimation and active learning for data-driven interatomic potentials for propagation of uncertainty through molecular dynamics
  (Dallas Foster, MIT)
  [slides] [recording]

Wednesday June 22nd 2022

• The BEEF class of xc functionals (Thomas Bligaard, DTU)
  [recording]
• A Bayesian Approach to Uncertainty Quantification for Density Functional Theory (Kate Fisher, MIT)
  [slides] [recording]
• Dielectric response with short-ranged electrostatics (Stephen Cox, Cambridge)
  [slides]
• Fully guaranteed and computable error bounds for clusters of eigenvalues (Genevieve Dusson, CNRS)
  [slides] [recording]
• Practical error bounds for properties in plane-wave electronic structure calculations (Gaspard Kemlin, Ecole des Ponts)
  [slides] [recording]
• The transferability limits of static benchmarks (Thomas Weymuth, ETH)
  [slides] [recording]

Thursday June 23rd 2022

• An information-theoretic approach to uncertainty quantification in atomistic modelling of crystalline materials (Maciej Buze, Birmingham)
  [slides] [recording]
• Hyperactive Learning (Cas van der Oord, Cambridge)
  [slides] [recording]
• Benchmarking under uncertainty (Jonny Proppe, TU Braunschweig)
• Model Error Estimation and Uncertainty Quantification of Machine Learning Interatomic Potentials (Khachik Sargsyan, Sandia National Labs)
  [slides] [recording]
• Committee neural network potentials control generalization errors and enable active learning (Christoph Schran, Cambridge)
  [slides] [recording]

Morning discussion sessions

The discussion sessions were centred around broad multi-disciplinary topics to stimulate cross-fertilisation. Key topics were active learning techniques for obtaining interatomic potentials on the fly as well as opportunities to connect numerical and statistical approaches for error estimation.

A central topic of the session on Thursday morning was the development of a common cross-community language and guidelines for error estimation. This included the question how to establish a minimal standard for error control and make the broader community aware of such techniques to ensure published results can be validated and are more reproducible. Initial ideas from this discussion are summarised in a public github repository. With this repository we invite everyone to contribute concrete examples of the error control strategies taken in their research context. In the future we hope to community guidelines for error control in first-principle modelling based on these initial ideas.

Feedback from participants

Overall we received mostly positive feedback about the event. Virtual participants enjoyed the opportunity to interact with in-person participants via the zoom sessions and Slack. For several in-person participants this meeting was the first physical meeting since the pandemic and the ample opportunities for informal interchange we allocated in the programme (discussion sessions, poster sessions, social dinner, boat trip excursion) have been much appreciated.

A challenge was to keep the meeting accessible for both researchers from foreign fields as well as junior participants entering this interdisciplinary field. With respect to the discussion sessions we got several suggestions for improvement in this regard. For example it has been suggested to (i) set and communicate the discussion subject well in advance to allow people to get prepared, (ii) motivate postdocs to coordinate the discussion, which would be responsible to curate material and formulate stimulating research questions and (iii) get these postdocs to start the session with an introductory presentation on open problems.

Conclusions and outlook

During the event it became apparent that the meaning associated to the term “error control” deviates between communities, in particular between mathematicians and application scientists. Not only did this result in a considerable language barrier and some communication problems during the workshop, but it also made communities to appear to move at different paces. On a first look this sometimes made it difficult to see the applicability of research results from another community. But the heterogeneity of participants also offered opportunities to learn from each other's viewpoint: for example during the discussion sessions we actively worked towards obtaining a joint language and cross-community standards for error control. Our initial ideas on this point are available in a public github repository, where we invite everyone to participate via opening issues and pull requests to continue the discussion.

A number of applications under Linux provide a “Browse Files” button that is intended to pull up a file manager in a specific directory. While this is convenient for most users, some might want a little more flexibility, so let’s hook up a terminal emulator to that button instead of a file manager.

First, we need a command that starts a terminal emulator in a specific directory, in my case this will be

foot -D <path to directory>

which will start foot in the specified <path to directory>.

As this button is implemented leveraging the XDG MIME Applications specification, we now need to define a new desktop entry, let’s call it TermFM.desktop, which we place under either ~/.local/share/applications or /usr/local/share/applications, depending on preference. The file using a foot terminal should read

[Desktop Entry]
Type=Application
Name=TermFM
Exec=foot -D %U
MimeType=inode/directory;

where %U will be the placeholder for the path that is handed over by the calling application. The MimeType line is optional, but given that the above terminal command only works for directories anyways, it doesn’t hurt to constrain this desktop file to this file type only.

Afterwards, we need to configure this as the default applications for the file type inode/directory, which we do by adding

inode/directory=TermFM.desktop

to the [Default Applications] section in ~/.config/mimeapps.list. Should this file not yet exist, you can create it to contain

[Default Applications]
inode/directory=TermFM.desktop

Once that is done, you should from now on get your terminal at the according location when you click “Browse Files” in an application supporting this.

This post is the third article in a series of blog posts about rsync, see the Series Overview.

With rsync up and running, it’s time to take a peek under the hood of rsync to better understand how it works.

How does rsync work?

When talking about the rsync protocol, we need to distinguish between:

• protocol-level roles: “sender” and “receiver”
• TCP roles: “client” and “server”

All roles can be mixed and matched: both rsync clients (or servers!) can either send or receive.

Now that you know the terminology, let’s take a high-level look at the rsync protocol. We’ll look at protocol version 27, which is older but simpler, and which is the most widely supported protocol version, implemented by openrsync and other third-party implementations:

The rsync protocol can be divided into two phases:

1. In the first phase, the sender walks the local file tree to generate and send the file list to the receiver. The file list must be transferred in full, because both sides sort it by filename (later rsync protocol versions eliminate this synchronous sorting step).

2. In the second phase, concurrently:

   • The receiver compares and requests each file in the file list. The receiver requests the full file when it didn’t exist on disk yet, or it will send checksums for the rsync hash search algorithm when the file already existed.
   • The receiver receives file data from the sender. The sender answers the requests with just enough data to reconstruct the current file contents based on what’s already on the receiver.

The architecture makes it easy to implement the second phase in 3 separate processes, each of which sending to the network as fast as possible using heavy pipelining. This results in utilizing the available hardware resources (I/O, CPU, network) on sender and receiver to the fullest.

Observing rsync’s transfer phases

When starting an rsync transfer, looking at the resource usage of both machines allows us to confirm our understanding of the rsync architecture, and to pin-point any bottlenecks:

1. phase: The rsync sender needs 17 seconds to walk the file system and send the file list. The rsync receiver reads from the network and writes into RAM during that time.
   • This phase is random I/O (querying file system metadata) for the sender.
2. phase: Afterwards, the rsync sender reads from disk and sends to the network. The rsync receiver receives from the network and writes to disk.
   • The receiver does roughly the same amount of random I/O as the sender did in phase 1, as it needs to create directories and request missing files.
   • The sender does sequential disk reads and possibly checksum calculation, if the file(s) existed on the receiver side.

(Again, the above was captured using rsync protocol version 27, later rsync protocol versions don’t synchronize after completing phase 1, but instead interleave the phases more.)

rsync hash search

Up until now, we have described the rsync protocol at a high level. Let’s zoom into the hash search step, which is what many people might associate with the term “rsync algorithm”.

When a file exists on both sides, rsync sender and receiver, the receiver first divides the file into blocks. The block size is a rounded square root of the file’s length. The receiver then sends the checksums of all blocks to the sender. In response, the sender finds matching blocks in the file and sends only the data needed to reconstruct the file on the receiver side.

Specifically, the sender goes through each byte of the file and tries to match existing receiver content. To make this less computationally expensive, rsync combines two checksums.

rsync first calculates what it calls the “sum1”, or “fast signature”. This is a small checksum (two uint16) that can be calculated with minimal effort for a rolling window over the file data. tridge rsync comes with SIMD implementations to further speed this up where possible.

Only if the sum1 matches will “sum2” (or “strong signature”) be calculated, a 16-byte MD4 hash. Newer protocol versions allow negotiating the hash algorithm and support the much faster xxhash algorithms.

If sum2 matches, the block is considered equal on both sides.

Hence, the best case for rsync is when a file has either not changed at all, or shares as many full blocks of content as possible with the old contents.

Changing data sets

Now that we know how rsync works on the file level, let’s take a step back to the data set level.

The easiest situation is when you transfer a data set that is not currently changing. But what happens when the data set changes while your rsync transfer is running? Here are two examples.

debiman, the manpage generator powering manpages.debian.org is running on a Debian VM on which an rsync job periodically transfers the static manpage archive to different static web servers across the world. The rsync job and debiman are not sequenced in any way. Instead, debiman is careful to only ever atomically swap out files in its output directory, or add new files before it swaps out an updated index.

The second example, the PostgreSQL database management system, is the opposite situation: instead of having full control over how files are laid out, here I don’t have control over how files are written (this generalizes to any situation where the model of only ever replacing files is not feasible). The data files which my Postgres installation keeps on disk are not great to synchronize using rsync: they are large and frequently change. Instead, I now exempt them from my rsync transfer and use pg_dump(1) to create a snapshot of my databases instead.

To confirm rsync’s behavior regarding changing data sets in detail, I modified rsync to ask for confirmation between generating the file list and transferring the files. Here’s what I found:

• If files are added after rsync has transferred the file list, the new files will just not be part of the transfer.
• If a file vanishes between generating the file list and transfering the file, rsync exits with status code 24, which its manpage documents as “Partial transfer due to vanished source files”. My rsyncprom monitoring wrapper offers a flag to treat exit code 24 like exit code 0, because depending on the data set, vanishing files are expected.
• If a file’s contents change (no matter whether the file grows, shrinks, or is modified in-place) between generating the file list and the actual file transfer, that’s not a problem — rsync will transfer the file contents as it reads them once the transfer starts. Note that this might be an inconsistent view of the data, depending on the application.
  • Ideally, don’t ever modify files within a data set that is rsynced. Instead, atomically move complete files into the data set.

Another way of phrasing the above is that data consistency is not something that rsync can in any way guarantee. It’s up to you to either live with the inconsistency (often a good-enough strategy!), or to add an extra step that ensures the data set you feed to rsync is consistent.

Next up

The fourth article in this series is rsync, article 4: My own rsync implementation (To be published.)

Appendix A: rsync confirmation hack

For verifying rsync’s behavior with regards to changing data sets, I checked out the following version:

% git clone https://github.com/WayneD/rsync/ rsync-changing-data-sets
% cd rsync-changing-data-sets
% git checkout v3.2.4
% ./configure
% make

Then, I modified flist.c to add a confirmation step between sending the file list and doing the actual file transfers:

diff --git i/flist.c w/flist.c
index 1ba306bc..98981f34 100644
--- i/flist.c
+++ w/flist.c
@@ -20,6 +20,8 @@
  * with this program; if not, visit the http://fsf.org website.
  */
 
+#include <stdio.h>
+
 #include "rsync.h"
 #include "ifuncs.h"
 #include "rounding.h"
@@ -2516,6 +2518,17 @@ struct file_list *send_file_list(int f, int argc, char *argv[])
 	if (DEBUG_GTE(FLIST, 2))
 		rprintf(FINFO, "send_file_list done\n");
 
+	char *line = NULL;
+	size_t llen = 0;
+	ssize_t nread;
+	printf("file list sent. enter 'yes' to continue: ");
+	while ((nread = getline(&line, &llen, stdin)) != -1) {
+	  if (nread == strlen("yes\n") && strcasecmp(line, "yes\n") == 0) {
+	    break;
+	  }
+	  printf("enter 'yes' to continue: ");
+	}
+
 	if (inc_recurse) {
 		send_dir_depth = 1;
 		add_dirs_to_tree(-1, flist, stats.num_dirs);

My rsync invocation is:

./rsync -av --debug=all4 --protocol=27 ~/i3/src /tmp/DEST/

It’s necessary to use an older protocol version to make rsync generate a full file list before starting the transfer. Later protocol versions interleave these parts of the protocol.

This post is the second article in a series of blog posts about rsync, see the Series Overview.

Now that we know what to use rsync for, how can we best integrate rsync into monitoring and alerting, and on which operating systems does it work?

Monitoring and alerting for rsync jobs using Prometheus

Once you have one or two important rsync jobs, it might make sense to alert when your job has not completed as expected.

I’m using Prometheus for all my monitoring and alerting.

Because Prometheus pulls metrics from its (typically always-running) targets, we need an extra component: the Prometheus Pushgateway. The Pushgateway stores metrics pushed by short-lived jobs like rsync transfers and makes them available to subsequent Prometheus pulls.

To integrate rsync with the Prometheus Pushgateway, I wrote rsyncprom, a small tool that wraps rsync, or parses rsync output supplied by you. Once rsync completes, rsyncprom pushes the rsync exit code and parsed statistics about the transfer to your Pushgateway.

Prometheus server-side setup

First, I set up the Prometheus Pushgateway (via Docker and systemd) on my server.

Then, in my prometheus.conf file, I instruct Prometheus to pull data from my Pushgateway:

# prometheus.conf

rule_files:
- backups.rules.yml

scrape_configs:
# […]
- job_name: pushgateway
  honor_labels: true
  static_configs:
  - targets: ['pushgateway:9091']

Finally, in backups.rules.yml, I configure an alert on the time series rsync_exit_code:

# backups.rules.yml

groups:
- name: backups.rules
  rules:
  - alert: RsyncFailing
    expr: rsync_exit_code{job="rsync"} > 0
    for: 1m
    labels:
      job: rsync
    annotations:
      description: rsync {{ $labels.instance }} is failing
      summary: rsync {{ $labels.instance }} is failing

This alert will fire any time an rsync job monitored via rsyncprom exits with a non-zero exit code.

rsync client-side setup

On each machine that runs rsync jobs I want to monitor, I first install rsyncprom:

go install github.com/stapelberg/rsyncprom/cmd/rsync-prom@latest

Then, I just wrap rsync transfers where it’s most convenient, for example in my crontab(5) :

# crontab -e
9 9 * * * /home/michael/go/bin/rsync-prom --job="cron" --instance="gphotos-sync@midna" -- /home/michael/gphotos-sync/sync.sh

The same wrapper technique works in shell scripts or systemd service files.

You can also provide rsync output from Go code (this example runs rsync via SSH).

Monitoring architecture

Here’s how the whole setup looks like architecturally:

The rsync scheduler runs on a Raspberry Pi running gokrazy. The scheduler invokes the rsync job to back up websrv.zekjur.net via SSH and sends the output to Prometheus, which is running on a (different) server at an ISP.

Monitoring dashboard

The grafana dashboard looks like this in action:

• The top left table shows the most recent rsync exit code, green means 0 (success).
• The top right graph shows rsync runtime (wall-clock time) over time. Long runtime can have any number of bottlenecks as the reason: network connections, storage devices, slow CPUs.
• The bottom left graph shows rsync dataset size over time. This allows you to quickly pinpoint transfers that are filling your disk up.
• The bottom right graph shows transferred bytes per rsync over time. The higher the value, the higher the amount of change in your data set between synchronization runs.

rsync operating system availability

Now that we have learnt about a couple of typical use-cases, where can you use rsync to implement these use-cases? The answer is: in most environments, as rsync is widely available on different Linux and BSD versions.

Macs come with rsync available by default (but it’s an old, patched version), and OpenBSD comes with a BSD-licensed implementation called openrsync by default.

On Windows, you can use the Windows Subsystem for Linux.

Next Up

The third article in this series is rsync, article 3: How does rsync work?. With rsync up and running, it’s time to take a peek under the hood of rsync to better understand how it works.

This post is the first article in a series of blog posts about rsync, see the Series Overview.

To motivate why it makes sense to look at rsync, I present three scenarios for which I have come to appreciate rsync: DokuWiki transfers, Software deployment and Backups.

Scenario: DokuWiki transfers using rsync

Recently, I set up a couple of tools for a website that is built on DokuWiki, such as a dead link checker and a statistics program. To avoid overloading the live website (and possibly causing spurious requests that interfere with statistics), I decided it would be best to run a separate copy of the DokuWiki installation locally. This requires synchronizing:

1. The PHP source code files of DokuWiki itself (including plugins and configuration)
2. One text file per wiki page, and all uploaded media files

A DokuWiki installation is exactly the kind of file tree that scp(1) cannot efficiently transfer (too many small files), but rsync(1) can! The rsync transfer only takes a few seconds, no matter if it’s a full download (can be simpler for batch jobs) or an incremental synchronization (more efficient for regular synchronizations like backups).

Scenario: Software deployment using rsync

For smaller projects where I don’t publish new versions through Docker, I instead use a shell script to transfer and run my software on the server.

rsync is a great fit here, as it transfers many small files (static assets and templates) efficiently, only transfers the binaries that actually changed, and doesn’t mind if the binary file it’s uploading is currently running (contrary to scp(1) , for example).

To illustrate how such a script could look like, here’s my push script for Debian Code Search:

#!/bin/zsh
set -ex

# Asynchronously transfer assets while compiling:
(
    ssh root@dcs 'for i in $(seq 0 5); do mkdir -p /srv/dcs/shard${i}/{src,idx}; done'
    ssh root@dcs "adduser --disabled-password --gecos 'Debian Code Search' dcs || true"
    rsync -r systemd/ root@dcs:/etc/systemd/system/ &
    rsync -r cmd/dcs-web/templates/ root@dcs:/srv/dcs/templates/ &
    rsync -r static/ root@dcs:/srv/dcs/static/ &
    wait
) &

# Compile a new Debian Code Search version:
tmp=$(mktemp -d)
mkdir $tmp/bin
GOBIN=$tmp/bin \
GOAMD64=v3 \
  go install \
  -ldflags '-X github.com/Debian/dcs/cmd/dcs-web/common.Version=$version' \
  github.com/Debian/dcs/cmd/...

# Transfer the Debian Code Search binaries:
rsync \
  $tmp/bin/dcs-{web,source-backend,package-importer,compute-ranking,feeder} \
  $tmp/bin/dcs \
  root@dcs:/srv/dcs/bin/

# Wait for the asynchronous asset transfer to complete:
wait

# Restart Debian Code Search on the server:
UNITS=(dcs-package-importer.service dcs-source-backend.service dcs-compute-ranking.timer dcs-web.service)
ssh root@dcs systemctl daemon-reload \&\& \
  systemctl enable ${UNITS} \; \
  systemctl reset-failed ${UNITS} \; \
  systemctl restart ${UNITS} \; \
  systemctl reload nginx

rm -rf "${tmp?}"

Scenario: Backups using rsync

The first backup system I used was bacula, which Wikipedia describes as an enterprise-level backup system. That certainly matches my impression, both in positive and negative ways: while bacula is very powerful, some seemingly common operations turn out quite complicated in bacula. Restoring a single file or directory tree from a backup was always more effort than I thought reasonable. For some reason, I often had to restore backup catalogs before I was able to access the backup contents (I don’t remember the exact details).

When moving apartment last time, I used the opportunity to change my backup strategy. Instead of using complicated custom software with its own volume file format (like bacula), I wanted backed-up files to be usable on the file system level with standard tools like rm, ls, cp, etc.

Working with files in a regular file system makes day-to-day usage easier, and also ensures that when my network storage hardware dies, I can just plug the hard disk into any PC, boot a Linux live system, and recover my data.

To back up machines onto my network storage PC’s file system, I ended up with a hand-written rsync wrapper script that copies the full file system of each machine into dated directory trees:

storage2# ls -l backup/midna/2022-05-27
bin   boot  etc  home  lib  lib64  media  opt
proc  root  run  sbin  sys  tmp    usr    var

storage2# ls -l backup/midna/2022-05-27/home/michael/configfiles/zshrc
-rw-r--r--. 7 1000 1000 14554 May  9 19:37 backup/midna/2022-05-27/home/michael/configfiles/zshrc

To revert my ~/.zshrc to an older version, I can scp(1) the file:

midna% scp storage2:/srv/backup/midna/2022-05-27/home/michael/configfiles/zshrc ~/configfiles/zshrc

To compare a whole older source tree, I can mount it using sshfs(1) :

midna% mkdir /tmp/2022-05-27-i3
midna% sshfs storage2:/srv/backup/midna/2022-05-27/$HOME/i3 /tmp/2022-05-27-i3
midna% diff -ur /tmp/2022-05-27-i3 ~/i3/

Incremental backups

Of course, the idea is not to transfer the full machine contents every day, as that would quickly fill up my network storage’s 16 TB disk! Instead, we can use rsync’s --link-dest option to elegantly deduplicate files using file system hard links:

backup/midna/2022-05-26
backup/midna/2022-05-27 # rsync --link-dest=2022-05-26

To check the de-duplication level, we can use du(1) , first on a single directory:

storage2# du -hs 2022-05-27 
113G	2022-05-27

…and then on two subsequent directories:

storage2# du -hs 2022-05-25 2022-05-27
112G	2022-05-25
7.3G	2022-05-27

As you can see, the 2022-05-27 backup took 7.3 GB of disk space, and 104.7 GB were re-used from the previous backup(s).

To print all files which have changed since the last backup, we can use:

storage2# find 2022-05-27 -type f -links 1 -print

Limitation: file system compatibility

A significant limitation of backups at the file level is that the destination file system (network storage) needs to support all the file system features used on the machines you are backing up.

For example, if you use POSIX ACLs or Extended attributes (possibly for Capabilities or SELinux), you need to ensure that your backup file system has these features enabled, and that you are using rsync(1) ’s --xattrs (or -X for short) option.

This can turn from a pitfall into a dealbreaker as soon as multiple operating systems are involved. For example, the rsync version on macOS has Apple-specific code to work with Apple resource forks and other extended attributes. It’s not clear to me whether macOS rsync can send files to Linux rsync, restore them, and end up with the same system state.

Luckily, I am only interested in backing up Linux systems, or merely home directories of non-Linux systems, where no extended attributes are used.

Downside: slow bulk operations (disk usage, deletion)

The biggest downside of this architecture is that working with the directory trees in bulk can be very slow, especially when using a hard disk instead of an SSD. For example, deleting old backups can easily take many hours to multiple days (!). Sure, you can just let the rm command run in the background, but it’s annoying nevertheless.

Even merely calculating the disk space usage of each directory tree is a painfully slow operation. I tried using stateful disk usage tools like duc, but it didn’t work reliably on my backups.

In practice, I found that for tracking down large files, using ncdu(1) on any recent backup typically quickly shows the large file. In one case, I found var/lib/postgresql to consume many gigabytes. I excluded it in favor of using pg_dump(1) , which resulted in much smaller backups!

Unfortunately, even when using an SSD, determining which files take up most space of a full backup takes a few minutes:

storage2# time du -hs backup/midna/2022-06-09
742G	backup/midna/2022-06-09

real	8m0.202s
user	0m11.651s
sys	2m0.731s

Backup transport (SSH) and scheduling

To transfer data via rsync from the backup host to my network storage, I’m using SSH.

Each machine’s SSH access is restricted in my network storage’s SSH authorized_keys(5) config file to not allow arbitrary commands, but to perform just a specific operation. The only allowed operation in my case is running rrsync (“restricted rsync”) in a container whose file system only contains the backup host’s sub directory, e.g. .websrv.zekjur.net:

command="/bin/docker run --log-driver none -i -e SSH_ORIGINAL_COMMAND -v /srv/backup/websrv.zekjur.net:/srv/backup/websrv.zekjur.net stapelberg/docker-rsync /srv/backup/websrv.zekjur.net",no-port-forwarding,no-X11-forwarding ssh-ed25519 AAAAC3…

(The corresponding Dockerfile can be found in my Gigabit NAS article.)

To trigger such an SSH-protected rsync transfer remotely, I’m using a small custom scheduling program called dornröschen. The program arranges for all involved machines to be powered on (using Wake-on-LAN) and then starts rsync via another operation-restricted SSH connection.

You could easily replace this with a cron job if you don’t care about WOL.

The architecture looks like this:

The operation-restricted SSH connection on each backup host is configured in SSH’s authorized_keys(5) config file:

command="/root/backup-remote.pl",no-port-forwarding,no-X11-forwarding ssh-ed25519 AAAAC3…

Next up

The second article in this series is rsync, article 2: Surroundings. Now that we know what to use rsync for, how can we best integrate rsync into monitoring and alerting, and on which operating systems does it work?

For many years, I was only a casual user of rsync and used it mostly for one-off file transfers.

Over time, I found rsync useful in more and more cases, and would recommend every computer user put this great tool into their toolbox 🛠 🧰 !

I’m publishing a series of blog posts about rsync:

• rsync, article 1: Scenarios. To motivate why it makes sense to look at rsync, I present three scenarios for which I have come to appreciate rsync: DokuWiki transfers, Software deployment and Backups.
• rsync, article 2: Surroundings. Now that we know what to use rsync for, how can we best integrate rsync into monitoring and alerting, and on which operating systems does it work?
• rsync, article 3: How does rsync work?. With rsync up and running, it’s time to take a peek under the hood of rsync to better understand how it works.
• rsync, article 4: My own rsync implementation (To be published.)

Let’s say you want to implement a sorting function in Go. Or perhaps a data structure like a binary search tree, providing ordered access to its elements. Because you want your code to be re-usable and type safe, you want to use type parameters. So you need a way to order user-provided types.

There are multiple methods of doing that, with different trade-offs. Let’s talk about four in particular here:

1. constraints.Ordered
2. A method constraint
3. Taking a comparison function
4. Comparator types

constraints.Ordered

Go 1.18 has a mechanism to constrain a type parameter to all types which have the < operator defined on them. The types which have this operator are exactly all types whose underlying type is string or one of the predeclared integer and float types. So we can write a type set expressing that:

type Integer interface {
  ~int | ~int8 | ~int16 | ~int32 | ~int64 | ~uint | ~uint8 | ~uint16 | ~uint32 | ~uint64 | ~uintptr
}

type Float interface {
  ~float32 | ~float64
}

type Ordered interface {
  Integer | Float | ~string
}

Because that’s a fairly common thing to want to do, there is already a package which contains these kinds of type sets.

With this, you can write the signature of your sorting function or the definition of your search tree as:

func Sort[T constraints.Ordered](s []T) {
  // …
}

type SearchTree[T constraints.Ordered] struct {
  // …
}

The main advantage of this is that it works directly with predeclared types and simple types like time.Duration. It also is very clear.

The main disadvantage is that it does not allow composite types like structs. And what if a user wants a different sorting order than the one implied by <? For example if they want to reverse the order or want specialized string collation. A multimedia library might want to sort “The Expanse” under E. And some letters sort differently depending on the language setting.

constraints.Ordered is simple, but it also is inflexible.

Method constraints

We can use method constraints to allow more flexibility. This allows a user to implement whatever sorting order they want as a method on their type.

We can write that constraint like this:

type Lesser[T any] interface {
  // Less returns if the receiver is less than v.
  Less(v T) bool
}

The type parameter is necessary because we have to refer to the receiver type itself in the Less method. This is hopefully clearer when we look at how this is used:

func Sort[T Lesser[T]](s []T) {
  // …
}

func SearchTree[T Lesser[T]](s []T) {
  // …
}

This allows the user of our library to customize the sorting order by defining a new type with a Less method:

type ReverseInt int

func (i ReverseInt) Less(j ReverseInt) bool {
  return j < i // order is reversed
}

The disadvantage of this is that it requires some boiler plate on part of your user. Using a custom sorting order always requires defining a type with a method.

They can’t use your code with predeclared types like int or string but always have to wrap it into a new type.

Likewise if a type already has a natural comparison method but it is not called Less. For example time.Time is naturally sorted by time.Time.Before. For cases like that there needs to be a wrapper to rename the method.

Whenever one of these wrappings happens your user might have to convert back and forth when passing data to or from your code.

It also is a little bit more confusing than constraints.Ordered, as your user has to understand the purpose of the extra type parameter on Lesser.

Passing a comparison function

A simple way to get flexibility is to have the user pass us a function used for comparison directly:

func Sort[T any](s []T, less func(T, T) bool) {
  // …
}

type SearchTree[T any] struct {
  Less func(T, T) bool
  // …
}

func NewSearchTree(less func(T, T) bool) *SearchTree[T] {
  // …
  return &SearchTree[T]{
    Less: less,
    // …
  }
}