tstest/integration/vms: test vms as they are ready (#2022)

Instead of testing all the VMs at once when they are all ready, this
patch changes the testing logic so that the vms are tested as soon as
they register with testcontrol. Also limit the amount of VM ram used at
once with the `-ram-limit` flag. That uses a semaphore to guard resource
use.

Also document CentOS' sins.

Updates #1988

Signed-off-by: Christine Dodrill <xe@tailscale.com>

tstest/integration/vms/vms_test.go

@@ -7,6 +7,7 @@
 package vms
 
 import (
+	"context"
 	"crypto/sha256"
 	"encoding/hex"
 	"flag"
@@ -31,6 +32,7 @@ import (
 	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"
@@ -40,7 +42,8 @@ import (
 
 const securePassword = "hunter2"
 
-var runVMTests = flag.Bool("run-vm-tests", false, "if set, run expensive (10G+ ram) VM based integration tests")
+var runVMTests = flag.Bool("run-vm-tests", false, "if set, run expensive VM based integration tests")
+var 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")
 var distroRex *regexValue = func() *regexValue {
 	result := &regexValue{r: regexp.MustCompile(`.*`)}
 	flag.Var(result, "distro-regex", "The regex that matches what distros should be run")
@@ -259,9 +262,8 @@ func mkVM(t *testing.T, n int, d Distro, sshKey, hostURL, tdir string) func() {
 	if err != nil {
 		t.Fatalf("can't find cache dir: %v", err)
 	}
-	cdir = filepath.Join(cdir, "within", "mkvm")
+	cdir = filepath.Join(cdir, "tailscale", "vm-test")
 	os.MkdirAll(filepath.Join(cdir, "qcow2"), 0755)
-	os.MkdirAll(filepath.Join(cdir, "seed"), 0755)
 
 	port := 23100 + n
 
@@ -280,6 +282,7 @@ func mkVM(t *testing.T, n int, d Distro, sshKey, hostURL, tdir string) func() {
 		"-drive", driveArg,
 		"-cdrom", filepath.Join(tdir, d.name, "seed", "seed.iso"),
 		"-vnc", fmt.Sprintf(":%d", n),
+		"-smbios", "type=1,serial=ds=nocloud;h=" + d.name,
 	}
 
 	t.Logf("running: qemu-system-x86_64 %s", strings.Join(args, " "))
@@ -378,7 +381,7 @@ func TestVMIntegrationEndToEnd(t *testing.T) {
 
 	var (
 		ipMu  sync.Mutex
-		ipMap = []ipMapping{}
+		ipMap = map[string]ipMapping{}
 	)
 
 	mux := http.NewServeMux()
@@ -398,7 +401,8 @@ func TestVMIntegrationEndToEnd(t *testing.T) {
 		if err != nil {
 			log.Panicf("bad port: %v", port)
 		}
-		ipMap = append(ipMap, ipMapping{r.UserAgent(), port, host})
+		distro := r.UserAgent()
+		ipMap[distro] = ipMapping{distro, port, host}
 		t.Logf("%s: %v", name, host)
 	})
 
@@ -424,86 +428,86 @@ func TestVMIntegrationEndToEnd(t *testing.T) {
 	loginServer := fmt.Sprintf("http://%s", ln.Addr())
 	t.Logf("loginServer: %s", loginServer)
 
-	var numDistros = 0
+	tstest.FixLogs(t)
+	defer tstest.UnfixLogs(t)
 
-	cancels := make(chan func(), len(distros))
+	ramsem := semaphore.NewWeighted(int64(*vmRamLimit))
 
-	t.Run("mkvm", func(t *testing.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)
-				numDistros++
 			} else {
 				continue
 			}
 
 			t.Run(distro.name, func(t *testing.T) {
-				t.Parallel()
+				ctx, done := context.WithCancel(context.Background())
+				defer done()
 
-				cancel := mkVM(t, n, distro, string(pubkey), loginServer, dir)
+				if distro.name == "opensuse-leap-15-1" {
-				cancels <- cancel
+					t.Skip("OpenSUSE Leap 15.1's cloud-init image just doesn't work for some reason, see https://github.com/tailscale/tailscale/issues/1988")
-			})
 				}
-	})
 
-	close(cancels)
+				t.Parallel()
-	for cancel := range cancels {
-		//lint:ignore SA9001 They do actually get ran
-		defer cancel()
 
-		if len(cancels) == 0 {
+				err := ramsem.Acquire(ctx, int64(distro.mem))
-			t.Log("all VMs started")
+				if err != nil {
-			break
+					t.Fatalf("can't acquire ram semaphore: %v", err)
-		}
 				}
+				defer ramsem.Release(int64(distro.mem))
 
-	t.Run("wait-for-vms", func(t *testing.T) {
+				cancel := mkVM(t, n, distro, string(pubkey), loginServer, dir)
-		t.Log("waiting for VMs to register")
+				defer cancel()
+				var ipm ipMapping
+
+				t.Run("wait-for-start", func(t *testing.T) {
 					waiter := time.NewTicker(time.Second)
 					defer waiter.Stop()
-		n := 0
+					var ok bool
 					for {
 						<-waiter.C
 						ipMu.Lock()
-			if len(ipMap) == numDistros {
+						if ipm, ok = ipMap[distro.name]; ok {
 							ipMu.Unlock()
 							break
-			} else {
-				if n%30 == 0 {
-					t.Logf("ipMap:   %d", len(ipMap))
-					t.Logf("distros: %d", numDistros)
 						}
-			}
-			n++
 						ipMu.Unlock()
 					}
 				})
 
-	ipMu.Lock()
+				testDistro(t, loginServer, signer, ipm)
-	defer ipMu.Unlock()
+			})
-	t.Run("join-net", func(t *testing.T) {
+		}
-		for _, ipm := range ipMap {
+	})
-			ipm := ipm
+}
-			port := ipm.port
-			t.Run(ipm.name, func(t *testing.T) {
-				tstest.FixLogs(t)
-				t.Parallel()
 
+func testDistro(t *testing.T, loginServer string, signer ssh.Signer, ipm ipMapping) {
+	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 retry loop helps to make things a bit faster, centos sometimes is slow at starting its sshd. I don't know why they don't use socket activation.
+	// 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++ {
-					conn, err := net.Dial("tcp", hostport)
+		cli, err := ssh.Dial("tcp", hostport, ccfg)
 		if err == nil {
 			working = true
-						conn.Close()
+			cli.Close()
 			break
 		}
 
-					time.Sleep(5 * time.Second)
+		time.Sleep(10 * time.Second)
 	}
 
 	if !working {
@@ -511,11 +515,7 @@ func TestVMIntegrationEndToEnd(t *testing.T) {
 	}
 
 	t.Logf("about to ssh into 127.0.0.1:%d", port)
-				cli, err := ssh.Dial("tcp", hostport, &ssh.ClientConfig{
+	cli, err := ssh.Dial("tcp", hostport, ccfg)
-					User:            "root",
-					Auth:            []ssh.AuthMethod{ssh.PublicKeys(signer), ssh.Password(securePassword)},
-					HostKeyCallback: ssh.InsecureIgnoreHostKey(),
-				})
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -552,13 +552,6 @@ func TestVMIntegrationEndToEnd(t *testing.T) {
 	if err != nil {
 		t.Fatalf("not successful: %v", err)
 	}
-			})
-		}
-	})
-
-	if numNodes := cs.NumNodes(); numNodes != len(ipMap) {
-		t.Errorf("wanted %d nodes, got: %d", len(ipMap), numNodes)
-	}
 }
 
 func copyBinaries(t *testing.T, conn *ssh.Client) {