tailscale/tstest/integration/vms/vms_test.go

919 lines
27 KiB
Go
Raw Normal View History

// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// +build linux
package vms
import (
"context"
"crypto/sha256"
"encoding/hex"
"errors"
"flag"
"fmt"
"io"
"log"
"net"
"net/http"
"os"
"os/exec"
"path"
"path/filepath"
"regexp"
"strconv"
"strings"
"sync"
"syscall"
"testing"
"text/template"
"time"
"github.com/aws/aws-sdk-go/aws"
"github.com/aws/aws-sdk-go/aws/session"
"github.com/aws/aws-sdk-go/service/s3"
"github.com/aws/aws-sdk-go/service/s3/s3manager"
expect "github.com/google/goexpect"
"github.com/pkg/sftp"
"golang.org/x/crypto/ssh"
"golang.org/x/sync/semaphore"
"inet.af/netaddr"
"tailscale.com/net/interfaces"
"tailscale.com/tstest"
"tailscale.com/tstest/integration"
"tailscale.com/tstest/integration/testcontrol"
"tailscale.com/types/logger"
)
const (
securePassword = "hunter2"
bucketName = "tailscale-integration-vm-images"
)
var (
runVMTests = flag.Bool("run-vm-tests", false, "if set, run expensive VM based integration tests")
noS3 = flag.Bool("no-s3", false, "if set, always download images from the public internet (risks breaking)")
vmRamLimit = flag.Int("ram-limit", 4096, "the maximum number of megabytes of ram that can be used for VMs, must be greater than or equal to 1024")
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
useVNC = flag.Bool("use-vnc", false, "if set, display guest vms over VNC")
distroRex = func() *regexValue {
result := &regexValue{r: regexp.MustCompile(`.*`)}
flag.Var(result, "distro-regex", "The regex that matches what distros should be run")
return result
}()
)
type Distro struct {
name string // amazon-linux
url string // URL to a qcow2 image
sha256sum string // hex-encoded sha256 sum of contents of URL
mem int // VM memory in megabytes
packageManager string // yum/apt/dnf/zypper
initSystem string // systemd/openrc
}
func (d *Distro) InstallPre() string {
switch d.packageManager {
case "yum":
return ` - [ yum, update, gnupg2 ]
- [ yum, "-y", install, iptables ]`
case "zypper":
return ` - [ zypper, in, "-y", iptables ]`
case "dnf":
return ` - [ dnf, install, "-y", iptables ]`
case "apt":
return ` - [ apt-get, update ]
- [ apt-get, "-y", install, curl, "apt-transport-https", gnupg2 ]`
case "apk":
return ` - [ apk, "-U", add, curl, "ca-certificates", iptables ]
- [ modprobe, tun ]`
}
return ""
}
func TestDownloadImages(t *testing.T) {
if !*runVMTests {
t.Skip("not running integration tests (need --run-vm-tests)")
}
bins := integration.BuildTestBinaries(t)
for _, d := range distros {
distro := d
t.Run(distro.name, func(t *testing.T) {
if !distroRex.Unwrap().MatchString(distro.name) {
t.Skipf("distro name %q doesn't match regex: %s", distro.name, distroRex)
}
if strings.HasPrefix(distro.name, "nixos") {
t.Skip("NixOS is built on the fly, no need to download it")
}
t.Parallel()
fetchDistro(t, distro, bins)
})
}
}
var distros = []Distro{
// NOTE(Xe): If you run into issues getting the autoconfig to work, run
// this test with the flag `--distro-regex=alpine-edge`. Connect with a VNC
// client with a command like this:
//
// $ vncviewer :0
//
// On NixOS you can get away with something like this:
//
// $ env NIXPKGS_ALLOW_UNFREE=1 nix-shell -p tigervnc --run 'vncviewer :0'
//
// Login as root with the password root. Then look in
// /var/log/cloud-init-output.log for what you messed up.
// NOTE(Xe): These images are not official images created by the Alpine Linux
// cloud team because the cloud team hasn't created any official images yet.
// These images were created under the guidance of the cloud team and contain
// few notable differences from what they would end up shipping. The Alpine
// Linux cloud team probably won't have official images up until a year or so
// after this comment is written (2021-06-11), but overall they will be
// compatible with these images. These images were created using the setup in
// this repo: https://github.com/Xe/alpine-image. I hereby promise to not break
// these links.
{"alpine-3-13-5", "https://xena.greedo.xeserv.us/pkg/alpine/img/alpine-3.13.5-cloud-init-within.qcow2", "a2665c16724e75899723e81d81126bd0254a876e5de286b0b21553734baec287", 256, "apk", "openrc"},
{"alpine-edge", "https://xena.greedo.xeserv.us/pkg/alpine/img/alpine-edge-2021-05-18-cloud-init-within.qcow2", "b3bb15311c0bd3beffa1b554f022b75d3b7309b5fdf76fb146fe7c72b83b16d0", 256, "apk", "openrc"},
// NOTE(Xe): All of the following images are official images straight from each
// distribution's official documentation.
{"amazon-linux", "https://cdn.amazonlinux.com/os-images/2.0.20210427.0/kvm/amzn2-kvm-2.0.20210427.0-x86_64.xfs.gpt.qcow2", "6ef9daef32cec69b2d0088626ec96410cd24afc504d57278bbf2f2ba2b7e529b", 512, "yum", "systemd"},
{"arch", "https://mirror.pkgbuild.com/images/v20210515.22945/Arch-Linux-x86_64-cloudimg-20210515.22945.qcow2", "e4077f5ba3c5d545478f64834bc4852f9f7a2e05950fce8ecd0df84193162a27", 512, "pacman", "systemd"},
{"centos-7", "https://cloud.centos.org/centos/7/images/CentOS-7-x86_64-GenericCloud-2003.qcow2c", "b7555ecf90b24111f2efbc03c1e80f7b38f1e1fc7e1b15d8fee277d1a4575e87", 512, "yum", "systemd"},
{"centos-8", "https://cloud.centos.org/centos/8/x86_64/images/CentOS-8-GenericCloud-8.3.2011-20201204.2.x86_64.qcow2", "7ec97062618dc0a7ebf211864abf63629da1f325578868579ee70c495bed3ba0", 768, "dnf", "systemd"},
{"debian-9", "http://cloud.debian.org/images/cloud/OpenStack/9.13.22-20210531/debian-9.13.22-20210531-openstack-amd64.qcow2", "c36e25f2ab0b5be722180db42ed9928476812f02d053620e1c287f983e9f6f1d", 512, "apt", "systemd"},
{"debian-10", "https://cdimage.debian.org/images/cloud/buster/20210329-591/debian-10-generic-amd64-20210329-591.qcow2", "70c61956095870c4082103d1a7a1cb5925293f8405fc6cb348588ec97e8611b0", 768, "apt", "systemd"},
{"fedora-34", "https://download.fedoraproject.org/pub/fedora/linux/releases/34/Cloud/x86_64/images/Fedora-Cloud-Base-34-1.2.x86_64.qcow2", "b9b621b26725ba95442d9a56cbaa054784e0779a9522ec6eafff07c6e6f717ea", 768, "dnf", "systemd"},
{"opensuse-leap-15-1", "https://download.opensuse.org/repositories/Cloud:/Images:/Leap_15.1/images/openSUSE-Leap-15.1-OpenStack.x86_64.qcow2", "40bc72b8ee143364fc401f2c9c9a11ecb7341a29fa84c6f7bf42fc94acf19a02", 512, "zypper", "systemd"},
{"opensuse-leap-15-2", "https://download.opensuse.org/repositories/Cloud:/Images:/Leap_15.2/images/openSUSE-Leap-15.2-OpenStack.x86_64.qcow2", "4df9cee9281d1f57d20f79dc65d76e255592b904760e73c0dd44ac753a54330f", 512, "zypper", "systemd"},
{"opensuse-leap-15-3", "http://mirror.its.dal.ca/opensuse/distribution/leap/15.3/appliances/openSUSE-Leap-15.3-JeOS.x86_64-OpenStack-Cloud.qcow2", "22e0392e4d0becb523d1bc5f709366140b7ee20d6faf26de3d0f9046d1ee15d5", 512, "zypper", "systemd"},
{"opensuse-tumbleweed", "https://download.opensuse.org/tumbleweed/appliances/openSUSE-Tumbleweed-JeOS.x86_64-OpenStack-Cloud.qcow2", "79e610bba3ed116556608f031c06e4b9260e3be2b193ce1727914ba213afac3f", 512, "zypper", "systemd"},
{"ubuntu-16-04", "https://cloud-images.ubuntu.com/xenial/20210429/xenial-server-cloudimg-amd64-disk1.img", "50a21bc067c05e0c73bf5d8727ab61152340d93073b3dc32eff18b626f7d813b", 512, "apt", "systemd"},
{"ubuntu-18-04", "https://cloud-images.ubuntu.com/bionic/20210526/bionic-server-cloudimg-amd64.img", "389ffd5d36bbc7a11bf384fd217cda9388ccae20e5b0cb7d4516733623c96022", 512, "apt", "systemd"},
{"ubuntu-20-04", "https://cloud-images.ubuntu.com/focal/20210603/focal-server-cloudimg-amd64.img", "1c0969323b058ba8b91fec245527069c2f0502fc119b9138b213b6bfebd965cb", 512, "apt", "systemd"},
{"ubuntu-20-10", "https://cloud-images.ubuntu.com/groovy/20210604/groovy-server-cloudimg-amd64.img", "2196df5f153faf96443e5502bfdbcaa0baaefbaec614348fec344a241855b0ef", 512, "apt", "systemd"},
{"ubuntu-21-04", "https://cloud-images.ubuntu.com/hirsute/20210603/hirsute-server-cloudimg-amd64.img", "bf07f36fc99ff521d3426e7d257e28f0c81feebc9780b0c4f4e25ae594ff4d3b", 512, "apt", "systemd"},
// NOTE(Xe): We build fresh NixOS images for every test run, so the URL being
// used here is actually the URL of the NixOS channel being built from and the
// shasum is meaningless. This `channel:name` syntax is documented at [1].
//
// [1]: https://nixos.org/manual/nix/unstable/command-ref/env-common.html
{"nixos-21-05", "channel:nixos-21.05", "lolfakesha", 512, "nix", "systemd"},
// // NOTE(Xe): disabled until https://github.com/NixOS/nixpkgs/issues/128783
// // is fixed.
// {"nixos-unstable", "channel:nixos-unstable", "lolfakesha", 512, "nix", "systemd"},
}
// fetchFromS3 fetches a distribution image from Amazon S3 or reports whether
// it is unable to. It can fail to fetch from S3 if there is either no AWS
// configuration (in ~/.aws/credentials) or if the `-no-s3` flag is passed. In
// that case the test will fall back to downloading distribution images from the
// public internet.
//
// Like fetching from HTTP, the test will fail if an error is encountered during
// the downloading process.
//
// This function writes the distribution image to fout. It is always closed. Do
// not expect fout to remain writable.
func fetchFromS3(t *testing.T, fout *os.File, d Distro) bool {
t.Helper()
if *noS3 {
t.Log("you asked to not use S3, not using S3")
return false
}
sess, err := session.NewSession(&aws.Config{
Region: aws.String("us-east-1"),
})
if err != nil {
t.Logf("can't make AWS session: %v", err)
return false
}
dler := s3manager.NewDownloader(sess, func(d *s3manager.Downloader) {
d.PartSize = 64 * 1024 * 1024 // 64MB per part
})
t.Logf("fetching s3://%s/%s", bucketName, d.sha256sum)
_, err = dler.Download(fout, &s3.GetObjectInput{
Bucket: aws.String(bucketName),
Key: aws.String(d.sha256sum),
})
if err != nil {
fout.Close()
t.Fatalf("can't get s3://%s/%s: %v", bucketName, d.sha256sum, err)
}
err = fout.Close()
if err != nil {
t.Fatalf("can't close fout: %v", err)
}
return true
}
// fetchDistro fetches a distribution from the internet if it doesn't already exist locally. It
// also validates the sha256 sum from a known good hash.
func fetchDistro(t *testing.T, resultDistro Distro, bins *integration.Binaries) string {
t.Helper()
cdir, err := os.UserCacheDir()
if err != nil {
t.Fatalf("can't find cache dir: %v", err)
}
cdir = filepath.Join(cdir, "tailscale", "vm-test")
if strings.HasPrefix(resultDistro.name, "nixos") {
return makeNixOSImage(t, resultDistro, cdir, bins)
}
qcowPath := filepath.Join(cdir, "qcow2", resultDistro.sha256sum)
_, err = os.Stat(qcowPath)
if err == nil {
hash := checkCachedImageHash(t, resultDistro, cdir)
if hash != resultDistro.sha256sum {
t.Logf("hash for %s (%s) doesn't match expected %s, re-downloading", resultDistro.name, qcowPath, resultDistro.sha256sum)
err = errors.New("some fake non-nil error to force a redownload")
if err := os.Remove(qcowPath); err != nil {
t.Fatalf("can't delete wrong cached image: %v", err)
}
}
}
if err != nil {
t.Logf("downloading distro image %s to %s", resultDistro.url, qcowPath)
fout, err := os.Create(qcowPath)
if err != nil {
t.Fatal(err)
}
if !fetchFromS3(t, fout, resultDistro) {
resp, err := http.Get(resultDistro.url)
if err != nil {
t.Fatalf("can't fetch qcow2 for %s (%s): %v", resultDistro.name, resultDistro.url, err)
}
if resp.StatusCode != http.StatusOK {
resp.Body.Close()
t.Fatalf("%s replied %s", resultDistro.url, resp.Status)
}
_, err = io.Copy(fout, resp.Body)
resp.Body.Close()
if err != nil {
t.Fatalf("download of %s failed: %v", resultDistro.url, err)
}
hash := checkCachedImageHash(t, resultDistro, cdir)
if hash != resultDistro.sha256sum {
t.Fatalf("hash mismatch, want: %s, got: %s", resultDistro.sha256sum, hash)
}
}
}
return qcowPath
}
func checkCachedImageHash(t *testing.T, d Distro, cacheDir string) (gotHash string) {
t.Helper()
qcowPath := filepath.Join(cacheDir, "qcow2", d.sha256sum)
fin, err := os.Open(qcowPath)
if err != nil {
t.Fatal(err)
}
hasher := sha256.New()
if _, err := io.Copy(hasher, fin); err != nil {
t.Fatal(err)
}
hash := hex.EncodeToString(hasher.Sum(nil))
if hash != d.sha256sum {
t.Fatalf("hash mismatch, got: %q, want: %q", hash, d.sha256sum)
}
gotHash = hash
return
}
// run runs a command or fails the test.
func run(t *testing.T, dir, prog string, args ...string) {
t.Helper()
t.Logf("running: %s %s", prog, strings.Join(args, " "))
tstest.FixLogs(t)
cmd := exec.Command(prog, args...)
cmd.Stdout = logger.FuncWriter(t.Logf)
cmd.Stderr = logger.FuncWriter(t.Logf)
cmd.Dir = dir
if err := cmd.Run(); err != nil {
t.Fatal(err)
}
}
// mkLayeredQcow makes a layered qcow image that allows us to keep the upstream
// VM images pristine and only do our changes on an overlay.
func mkLayeredQcow(t *testing.T, tdir string, d Distro, qcowBase string) {
t.Helper()
run(t, tdir, "qemu-img", "create",
"-f", "qcow2",
"-o", "backing_file="+qcowBase,
filepath.Join(tdir, d.name+".qcow2"),
)
}
var (
metaDataTempl = template.Must(template.New("meta-data.yaml").Parse(metaDataTemplate))
userDataTempl = template.Must(template.New("user-data.yaml").Parse(userDataTemplate))
)
// mkSeed makes the cloud-init seed ISO that is used to configure a VM with
// tailscale.
func mkSeed(t *testing.T, d Distro, sshKey, hostURL, tdir string, port int) {
t.Helper()
dir := filepath.Join(tdir, d.name, "seed")
os.MkdirAll(dir, 0700)
// make meta-data
{
fout, err := os.Create(filepath.Join(dir, "meta-data"))
if err != nil {
t.Fatal(err)
}
err = metaDataTempl.Execute(fout, struct {
ID string
Hostname string
}{
ID: "31337",
Hostname: d.name,
})
if err != nil {
t.Fatal(err)
}
err = fout.Close()
if err != nil {
t.Fatal(err)
}
}
// make user-data
{
fout, err := os.Create(filepath.Join(dir, "user-data"))
if err != nil {
t.Fatal(err)
}
err = userDataTempl.Execute(fout, struct {
SSHKey string
HostURL string
Hostname string
Port int
InstallPre string
Password string
}{
SSHKey: strings.TrimSpace(sshKey),
HostURL: hostURL,
Hostname: d.name,
Port: port,
InstallPre: d.InstallPre(),
Password: securePassword,
})
if err != nil {
t.Fatal(err)
}
err = fout.Close()
if err != nil {
t.Fatal(err)
}
}
args := []string{
"-output", filepath.Join(dir, "seed.iso"),
"-volid", "cidata", "-joliet", "-rock",
filepath.Join(dir, "meta-data"),
filepath.Join(dir, "user-data"),
}
if hackOpenSUSE151UserData(t, d, dir) {
args = append(args, filepath.Join(dir, "openstack"))
}
run(t, tdir, "genisoimage", args...)
}
// mkVM makes a KVM-accelerated virtual machine and prepares it for introduction
// to the testcontrol server. The function it returns is for killing the virtual
// machine when it is time for it to die.
func mkVM(t *testing.T, n int, d Distro, sshKey, hostURL, tdir string, bins *integration.Binaries) {
t.Helper()
cdir, err := os.UserCacheDir()
if err != nil {
t.Fatalf("can't find cache dir: %v", err)
}
cdir = filepath.Join(cdir, "tailscale", "vm-test")
os.MkdirAll(filepath.Join(cdir, "qcow2"), 0755)
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
port, err := getProbablyFreePortNumber()
if err != nil {
t.Fatal(err)
}
mkLayeredQcow(t, tdir, d, fetchDistro(t, d, bins))
mkSeed(t, d, sshKey, hostURL, tdir, port)
driveArg := fmt.Sprintf("file=%s,if=virtio", filepath.Join(tdir, d.name+".qcow2"))
args := []string{
"-machine", "pc-q35-5.1,accel=kvm,usb=off,vmport=off,dump-guest-core=off",
"-netdev", fmt.Sprintf("user,hostfwd=::%d-:22,id=net0", port),
"-device", "virtio-net-pci,netdev=net0,id=net0,mac=8a:28:5c:30:1f:25",
"-m", fmt.Sprint(d.mem),
"-boot", "c",
"-drive", driveArg,
"-cdrom", filepath.Join(tdir, d.name, "seed", "seed.iso"),
"-smbios", "type=1,serial=ds=nocloud;h=" + d.name,
}
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
if *useVNC {
// test listening on VNC port
ln, err := net.Listen("tcp", net.JoinHostPort("0.0.0.0", strconv.Itoa(5900+n)))
if err != nil {
t.Fatalf("would not be able to listen on the VNC port for the VM: %v", err)
}
ln.Close()
args = append(args, "-vnc", fmt.Sprintf(":%d", n))
} else {
args = append(args, "-display", "none")
}
t.Logf("running: qemu-system-x86_64 %s", strings.Join(args, " "))
cmd := exec.Command("qemu-system-x86_64", args...)
cmd.Stdout = logger.FuncWriter(t.Logf)
cmd.Stderr = logger.FuncWriter(t.Logf)
err = cmd.Start()
if err != nil {
t.Fatal(err)
}
time.Sleep(time.Second)
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
// NOTE(Xe): In Unix if you do a kill with signal number 0, the kernel will do
// all of the access checking for the process (existence, permissions, etc) but
// nothing else. This is a way to ensure that qemu's process is active.
if err := cmd.Process.Signal(syscall.Signal(0)); err != nil {
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
t.Fatalf("qemu is not running: %v", err)
}
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
t.Cleanup(func() {
err := cmd.Process.Kill()
if err != nil {
t.Errorf("can't kill %s (%d): %v", d.name, cmd.Process.Pid, err)
}
cmd.Wait()
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
})
}
// ipMapping maps a hostname, SSH port and SSH IP together
type ipMapping struct {
name string
port int
ip string
}
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
// getProbablyFreePortNumber does what it says on the tin, but as a side effect
// it is a kind of racy function. Do not use this carelessly.
//
// This is racy because it does not "lock" the port number with the OS. The
// "random" port number that is returned here is most likely free to use, however
// it is difficult to be 100% sure. This function should be used with care. It
// will probably do what you want, but it is very easy to hold this wrong.
func getProbablyFreePortNumber() (int, error) {
l, err := net.Listen("tcp", ":0")
if err != nil {
return 0, err
}
defer l.Close()
_, port, err := net.SplitHostPort(l.Addr().String())
if err != nil {
return 0, err
}
portNum, err := strconv.Atoi(port)
if err != nil {
return 0, err
}
return portNum, nil
}
// TestVMIntegrationEndToEnd creates a virtual machine with qemu, installs
// tailscale on it and then ensures that it connects to the network
// successfully.
func TestVMIntegrationEndToEnd(t *testing.T) {
if !*runVMTests {
t.Skip("not running integration tests (need --run-vm-tests)")
}
os.Setenv("CGO_ENABLED", "0")
if _, err := exec.LookPath("qemu-system-x86_64"); err != nil {
t.Logf("hint: nix-shell -p go -p qemu -p cdrkit --run 'go test --v --timeout=60m --run-vm-tests'")
t.Fatalf("missing dependency: %v", err)
}
if _, err := exec.LookPath("genisoimage"); err != nil {
t.Logf("hint: nix-shell -p go -p qemu -p cdrkit --run 'go test --v --timeout=60m --run-vm-tests'")
t.Fatalf("missing dependency: %v", err)
}
dir := t.TempDir()
rex := distroRex.Unwrap()
bindHost := deriveBindhost(t)
ln, err := net.Listen("tcp", net.JoinHostPort(bindHost, "0"))
if err != nil {
t.Fatalf("can't make TCP listener: %v", err)
}
defer ln.Close()
t.Logf("host:port: %s", ln.Addr())
cs := &testcontrol.Server{}
derpMap := integration.RunDERPAndSTUN(t, t.Logf, bindHost)
cs.DERPMap = derpMap
var (
ipMu sync.Mutex
ipMap = map[string]ipMapping{}
)
mux := http.NewServeMux()
mux.Handle("/", cs)
// This handler will let the virtual machines tell the host information about that VM.
// This is used to maintain a list of port->IP address mappings that are known to be
// working. This allows later steps to connect over SSH. This returns no response to
// clients because no response is needed.
mux.HandleFunc("/myip/", func(w http.ResponseWriter, r *http.Request) {
ipMu.Lock()
defer ipMu.Unlock()
name := path.Base(r.URL.Path)
host, _, _ := net.SplitHostPort(r.RemoteAddr)
port, err := strconv.Atoi(name)
if err != nil {
log.Panicf("bad port: %v", port)
}
distro := r.UserAgent()
ipMap[distro] = ipMapping{distro, port, host}
t.Logf("%s: %v", name, host)
})
hs := &http.Server{Handler: mux}
go hs.Serve(ln)
run(t, dir, "ssh-keygen", "-t", "ed25519", "-f", "machinekey", "-N", ``)
pubkey, err := os.ReadFile(filepath.Join(dir, "machinekey.pub"))
if err != nil {
t.Fatalf("can't read ssh key: %v", err)
}
privateKey, err := os.ReadFile(filepath.Join(dir, "machinekey"))
if err != nil {
t.Fatalf("can't read ssh private key: %v", err)
}
signer, err := ssh.ParsePrivateKey(privateKey)
if err != nil {
t.Fatalf("can't parse private key: %v", err)
}
loginServer := fmt.Sprintf("http://%s", ln.Addr())
t.Logf("loginServer: %s", loginServer)
ramsem := semaphore.NewWeighted(int64(*vmRamLimit))
bins := integration.BuildTestBinaries(t)
t.Run("do", func(t *testing.T) {
for n, distro := range distros {
n, distro := n, distro
if rex.MatchString(distro.name) {
t.Logf("%s matches %s", distro.name, rex)
} else {
continue
}
t.Run(distro.name, func(t *testing.T) {
ctx, done := context.WithCancel(context.Background())
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
t.Cleanup(done)
t.Parallel()
tstest/integration/vms: fix flake when testing (#2145) Occasionally the test framework would fail with a timeout due to a virtual machine not phoning home in time. This seems to be happen whenever qemu can't bind the VNC or SSH ports for a virtual machine. This was fixed by taking the following actions: 1. Don't listen on VNC unless the `-use-vnc` flag is passed, this removes the need to listen on VNC at all in most cases. The option to use VNC is still left in for debugging virtual machines, but removing this makes it easier to deal with (VNC uses this odd system of "displays" that are mapped to ports above 5900, and qemu doesn't offer a decent way to use a normal port number, so we just disable VNC by default as a compromise). 2. Use a (hopefully) inactive port for SSH. In an ideal world I'd just have the VM's SSH port be exposed via a Unix socket, however the QEMU documentation doesn't really say if you can do this or not. While I do more research, this stopgap will have to make do. 3. Strictly tie more VM resource lifetimes to the tests themselves. Previously the disk image layers for virtual machines were only cleaned up at the end of the test and existed in the parent test-scoped temporary folder. This can make your tmpfs run out of space, which is not ideal. This should minimize the use of temporary storage as much as I know how to. 4. Strictly tie the qemu process lifetime to the lifetime of the test using testing.T#Cleanup. Previously it used a defer statement to clean up the qemu process, however if the tests timed out this defer was not run. This left around an orphaned qemu process that had to be killed manually. This change ensures that all qemu processes exit when their relevant tests finish. Signed-off-by: Christine Dodrill <xe@tailscale.com>
2021-06-25 18:45:12 +00:00
dir := t.TempDir()
err := ramsem.Acquire(ctx, int64(distro.mem))
if err != nil {
t.Fatalf("can't acquire ram semaphore: %v", err)
}
defer ramsem.Release(int64(distro.mem))
mkVM(t, n, distro, string(pubkey), loginServer, dir, bins)
var ipm ipMapping
t.Run("wait-for-start", func(t *testing.T) {
waiter := time.NewTicker(time.Second)
defer waiter.Stop()
var ok bool
for {
<-waiter.C
ipMu.Lock()
if ipm, ok = ipMap[distro.name]; ok {
ipMu.Unlock()
break
}
ipMu.Unlock()
}
})
testDistro(t, loginServer, distro, signer, ipm, bins)
})
}
})
}
func testDistro(t *testing.T, loginServer string, d Distro, signer ssh.Signer, ipm ipMapping, bins *integration.Binaries) {
t.Helper()
port := ipm.port
hostport := fmt.Sprintf("127.0.0.1:%d", port)
ccfg := &ssh.ClientConfig{
User: "root",
Auth: []ssh.AuthMethod{ssh.PublicKeys(signer), ssh.Password(securePassword)},
HostKeyCallback: ssh.InsecureIgnoreHostKey(),
}
// NOTE(Xe): This deadline loop helps to make things a bit faster, centos
// sometimes is slow at starting its sshd and will sometimes randomly kill
// SSH sessions on transition to multi-user.target. I don't know why they
// don't use socket activation.
const maxRetries = 5
var working bool
for i := 0; i < maxRetries; i++ {
cli, err := ssh.Dial("tcp", hostport, ccfg)
if err == nil {
working = true
cli.Close()
break
}
time.Sleep(10 * time.Second)
}
if !working {
t.Fatalf("can't connect to %s, tried %d times", hostport, maxRetries)
}
t.Logf("about to ssh into 127.0.0.1:%d", port)
cli, err := ssh.Dial("tcp", hostport, ccfg)
if err != nil {
t.Fatal(err)
}
copyBinaries(t, d, cli, bins)
timeout := 30 * time.Second
e, _, err := expect.SpawnSSH(cli, timeout,
expect.Verbose(true),
expect.VerboseWriter(logger.FuncWriter(t.Logf)),
// // NOTE(Xe): if you get a timeout, uncomment this line to have the raw
// output be sent to the test log quicker.
//expect.Tee(nopWriteCloser{logger.FuncWriter(t.Logf)}),
)
if err != nil {
t.Fatalf("%d: can't register a shell session: %v", port, err)
}
defer e.Close()
t.Log("opened session")
_, _, err = e.Expect(regexp.MustCompile(`(\#)`), timeout)
if err != nil {
t.Fatalf("%d: can't get a shell: %v", port, err)
}
t.Logf("got shell for %d", port)
switch d.initSystem {
case "openrc":
// NOTE(Xe): this is a sin, however openrc doesn't really have the concept
// of service readiness. If this sleep is removed then tailscale will not be
// ready once the `tailscale up` command is sent. This is not ideal, but I
// am not really sure there is a good way around this without a delay of
// some kind.
err = e.Send("rc-service tailscaled start && sleep 2\n")
case "systemd":
err = e.Send("systemctl start tailscaled.service\n")
}
if err != nil {
t.Fatalf("can't send command to start tailscaled: %v", err)
}
_, _, err = e.Expect(regexp.MustCompile(`(\#)`), timeout)
if err != nil {
t.Fatalf("%d: can't get a shell: %v", port, err)
}
err = e.Send(fmt.Sprintf("tailscale up --login-server %s\n", loginServer))
if err != nil {
t.Fatalf("%d: can't send tailscale up command: %v", port, err)
}
_, _, err = e.Expect(regexp.MustCompile(`Success.`), timeout)
if err != nil {
t.Fatalf("not successful: %v", err)
}
}
func copyBinaries(t *testing.T, d Distro, conn *ssh.Client, bins *integration.Binaries) {
if strings.HasPrefix(d.name, "nixos") {
return
}
cli, err := sftp.NewClient(conn)
if err != nil {
t.Fatalf("can't connect over sftp to copy binaries: %v", err)
}
mkdir(t, cli, "/usr/bin")
mkdir(t, cli, "/usr/sbin")
mkdir(t, cli, "/etc/default")
mkdir(t, cli, "/var/lib/tailscale")
copyFile(t, cli, bins.Daemon, "/usr/sbin/tailscaled")
copyFile(t, cli, bins.CLI, "/usr/bin/tailscale")
// TODO(Xe): revisit this assumption before it breaks the test.
copyFile(t, cli, "../../../cmd/tailscaled/tailscaled.defaults", "/etc/default/tailscaled")
switch d.initSystem {
case "openrc":
mkdir(t, cli, "/etc/init.d")
copyFile(t, cli, "../../../cmd/tailscaled/tailscaled.openrc", "/etc/init.d/tailscaled")
case "systemd":
mkdir(t, cli, "/etc/systemd/system")
copyFile(t, cli, "../../../cmd/tailscaled/tailscaled.service", "/etc/systemd/system/tailscaled.service")
}
t.Log("tailscale installed!")
}
func mkdir(t *testing.T, cli *sftp.Client, name string) {
t.Helper()
err := cli.MkdirAll(name)
if err != nil {
t.Fatalf("can't make %s: %v", name, err)
}
}
func copyFile(t *testing.T, cli *sftp.Client, localSrc, remoteDest string) {
t.Helper()
fin, err := os.Open(localSrc)
if err != nil {
t.Fatalf("can't open: %v", err)
}
defer fin.Close()
fi, err := fin.Stat()
if err != nil {
t.Fatalf("can't stat: %v", err)
}
fout, err := cli.Create(remoteDest)
if err != nil {
t.Fatalf("can't create output file: %v", err)
}
err = fout.Chmod(fi.Mode())
if err != nil {
fout.Close()
t.Fatalf("can't chmod fout: %v", err)
}
n, err := io.Copy(fout, fin)
if err != nil {
fout.Close()
t.Fatalf("copy failed: %v", err)
}
if fi.Size() != n {
t.Fatalf("incorrect number of bytes copied: wanted: %d, got: %d", fi.Size(), n)
}
err = fout.Close()
if err != nil {
t.Fatalf("can't close fout on remote host: %v", err)
}
}
func deriveBindhost(t *testing.T) string {
t.Helper()
ifName, err := interfaces.DefaultRouteInterface()
if err != nil {
t.Fatal(err)
}
var ret string
err = interfaces.ForeachInterfaceAddress(func(i interfaces.Interface, prefix netaddr.IPPrefix) {
if ret != "" || i.Name != ifName {
return
}
ret = prefix.IP().String()
})
if ret != "" {
return ret
}
if err != nil {
t.Fatal(err)
}
t.Fatal("can't find a bindhost")
return "unreachable"
}
func TestDeriveBindhost(t *testing.T) {
t.Log(deriveBindhost(t))
}
type nopWriteCloser struct {
io.Writer
}
func (nwc nopWriteCloser) Close() error { return nil }
const metaDataTemplate = `instance-id: {{.ID}}
local-hostname: {{.Hostname}}`
const userDataTemplate = `#cloud-config
#vim:syntax=yaml
cloud_config_modules:
- runcmd
cloud_final_modules:
- [users-groups, always]
- [scripts-user, once-per-instance]
users:
- name: root
ssh-authorized-keys:
- {{.SSHKey}}
- name: ts
plain_text_passwd: {{.Password}}
groups: [ wheel ]
sudo: [ "ALL=(ALL) NOPASSWD:ALL" ]
shell: /bin/sh
ssh-authorized-keys:
- {{.SSHKey}}
write_files:
- path: /etc/cloud/cloud.cfg.d/80_disable_network_after_firstboot.cfg
content: |
# Disable network configuration after first boot
network:
config: disabled
runcmd:
{{.InstallPre}}
- [ curl, "{{.HostURL}}/myip/{{.Port}}", "-H", "User-Agent: {{.Hostname}}" ]
`