Merge branch '19099-singularity-container-shell'
[arvados.git] / lib / crunchrun / singularity.go
index 741f542454e470ede35cc6f682c64c8a9b1bbf09..1da401f859f94b36655f771e5fea5af750b0cbe7 100644 (file)
@@ -5,11 +5,18 @@
 package crunchrun
 
 import (
+       "bytes"
+       "errors"
        "fmt"
        "io/ioutil"
+       "net"
        "os"
        "os/exec"
+       "os/user"
+       "regexp"
        "sort"
+       "strconv"
+       "strings"
        "syscall"
        "time"
 
@@ -19,6 +26,7 @@ import (
 
 type singularityExecutor struct {
        logf          func(string, ...interface{})
+       fakeroot      bool // use --fakeroot flag, allow --network=bridge when non-root (currently only used by tests)
        spec          containerSpec
        tmpdir        string
        child         *exec.Cmd
@@ -36,6 +44,14 @@ func newSingularityExecutor(logf func(string, ...interface{})) (*singularityExec
        }, nil
 }
 
+func (e *singularityExecutor) Runtime() string {
+       buf, err := exec.Command("singularity", "--version").CombinedOutput()
+       if err != nil {
+               return "singularity (unknown version)"
+       }
+       return strings.TrimSuffix(string(buf), "\n")
+}
+
 func (e *singularityExecutor) getOrCreateProject(ownerUuid string, name string, containerClient *arvados.Client) (*arvados.Group, error) {
        var gp arvados.GroupList
        err := containerClient.RequestAndDecode(&gp,
@@ -101,7 +117,7 @@ func (e *singularityExecutor) checkImageCache(dockerImageID string, container ar
        if len(cl.Items) == 1 {
                imageCollection = cl.Items[0]
        } else {
-               collectionName := collectionName + " " + time.Now().UTC().Format(time.RFC3339)
+               collectionName := "converting " + collectionName
                exp := time.Now().Add(24 * 7 * 2 * time.Hour)
                err = containerClient.RequestAndDecode(&imageCollection,
                        arvados.EndpointCollectionCreate.Method,
@@ -112,6 +128,7 @@ func (e *singularityExecutor) checkImageCache(dockerImageID string, container ar
                                        "name":       collectionName,
                                        "trash_at":   exp.UTC().Format(time.RFC3339),
                                },
+                               "ensure_unique_name": true,
                        })
                if err != nil {
                        return nil, fmt.Errorf("error creating '%v' collection: %s", collectionName, err)
@@ -141,6 +158,12 @@ func (e *singularityExecutor) LoadImage(dockerImageID string, imageTarballPath s
        }
 
        if _, err := os.Stat(imageFilename); os.IsNotExist(err) {
+               // Make sure the docker image is readable, and error
+               // out if not.
+               if _, err := os.Stat(imageTarballPath); err != nil {
+                       return err
+               }
+
                e.logf("building singularity image")
                // "singularity build" does not accept a
                // docker-archive://... filename containing a ":" character,
@@ -151,7 +174,22 @@ func (e *singularityExecutor) LoadImage(dockerImageID string, imageTarballPath s
                        return err
                }
 
+               // Set up a cache and tmp dir for singularity build
+               err = os.Mkdir(e.tmpdir+"/cache", 0700)
+               if err != nil {
+                       return err
+               }
+               defer os.RemoveAll(e.tmpdir + "/cache")
+               err = os.Mkdir(e.tmpdir+"/tmp", 0700)
+               if err != nil {
+                       return err
+               }
+               defer os.RemoveAll(e.tmpdir + "/tmp")
+
                build := exec.Command("singularity", "build", imageFilename, "docker-archive://"+e.tmpdir+"/image.tar")
+               build.Env = os.Environ()
+               build.Env = append(build.Env, "SINGULARITY_CACHEDIR="+e.tmpdir+"/cache")
+               build.Env = append(build.Env, "SINGULARITY_TMPDIR="+e.tmpdir+"/tmp")
                e.logf("%v", build.Args)
                out, err := build.CombinedOutput()
                // INFO:    Starting build...
@@ -217,11 +255,27 @@ func (e *singularityExecutor) Create(spec containerSpec) error {
        return nil
 }
 
-func (e *singularityExecutor) Start() error {
-       args := []string{"singularity", "exec", "--containall", "--no-home", "--cleanenv", "--pwd", e.spec.WorkingDir}
+func (e *singularityExecutor) execCmd(path string) *exec.Cmd {
+       args := []string{path, "exec", "--containall", "--cleanenv", "--pwd=" + e.spec.WorkingDir}
+       if e.fakeroot {
+               args = append(args, "--fakeroot")
+       }
        if !e.spec.EnableNetwork {
                args = append(args, "--net", "--network=none")
+       } else if u, err := user.Current(); err == nil && u.Uid == "0" || e.fakeroot {
+               // Specifying --network=bridge fails unless (a) we are
+               // root, (b) we are using --fakeroot, or (c)
+               // singularity has been configured to allow our
+               // uid/gid to use it like so:
+               //
+               // singularity config global --set 'allow net networks' bridge
+               // singularity config global --set 'allow net groups' mygroup
+               args = append(args, "--net", "--network=bridge")
+       }
+       if e.spec.CUDADeviceCount != 0 {
+               args = append(args, "--nv")
        }
+
        readonlyflag := map[bool]string{
                false: "rw",
                true:  "ro",
@@ -233,7 +287,12 @@ func (e *singularityExecutor) Start() error {
        sort.Strings(binds)
        for _, path := range binds {
                mount := e.spec.BindMounts[path]
-               args = append(args, "--bind", mount.HostPath+":"+path+":"+readonlyflag[mount.ReadOnly])
+               if path == e.spec.Env["HOME"] {
+                       // Singularity treats $HOME as special case
+                       args = append(args, "--home", mount.HostPath+":"+path)
+               } else {
+                       args = append(args, "--bind", mount.HostPath+":"+path+":"+readonlyflag[mount.ReadOnly])
+               }
        }
 
        // This is for singularity 3.5.2. There are some behaviors
@@ -243,21 +302,34 @@ func (e *singularityExecutor) Start() error {
        env := make([]string, 0, len(e.spec.Env))
        for k, v := range e.spec.Env {
                if k == "HOME" {
-                       // $HOME is a special case
-                       args = append(args, "--home="+v)
-               } else {
-                       env = append(env, "SINGULARITYENV_"+k+"="+v)
+                       // Singularity treats $HOME as special case,
+                       // this is handled with --home above
+                       continue
                }
+               env = append(env, "SINGULARITYENV_"+k+"="+v)
        }
 
+       // Singularity always makes all nvidia devices visible to the
+       // container.  If a resource manager such as slurm or LSF told
+       // us to select specific devices we need to propagate that.
+       if cudaVisibleDevices := os.Getenv("CUDA_VISIBLE_DEVICES"); cudaVisibleDevices != "" {
+               // If a resource manager such as slurm or LSF told
+               // us to select specific devices we need to propagate that.
+               env = append(env, "SINGULARITYENV_CUDA_VISIBLE_DEVICES="+cudaVisibleDevices)
+       }
+       // Singularity's default behavior is to evaluate each
+       // SINGULARITYENV_* env var with a shell as a double-quoted
+       // string and pass the result to the contained
+       // process. Singularity 3.10+ has an option to pass env vars
+       // through literally without evaluating, which is what we
+       // want. See https://github.com/sylabs/singularity/pull/704
+       // and https://dev.arvados.org/issues/19081
+       env = append(env, "SINGULARITY_NO_EVAL=1")
+
        args = append(args, e.imageFilename)
        args = append(args, e.spec.Command...)
 
-       path, err := exec.LookPath(args[0])
-       if err != nil {
-               return err
-       }
-       child := &exec.Cmd{
+       return &exec.Cmd{
                Path:   path,
                Args:   args,
                Env:    env,
@@ -265,6 +337,14 @@ func (e *singularityExecutor) Start() error {
                Stdout: e.spec.Stdout,
                Stderr: e.spec.Stderr,
        }
+}
+
+func (e *singularityExecutor) Start() error {
+       path, err := exec.LookPath("singularity")
+       if err != nil {
+               return err
+       }
+       child := e.execCmd(path)
        err = child.Start()
        if err != nil {
                return err
@@ -302,3 +382,118 @@ func (e *singularityExecutor) Close() {
                e.logf("error removing temp dir: %s", err)
        }
 }
+
+func (e *singularityExecutor) InjectCommand(ctx context.Context, detachKeys, username string, usingTTY bool, injectcmd []string) (*exec.Cmd, error) {
+       target, err := e.containedProcess()
+       if err != nil {
+               return nil, err
+       }
+       return exec.CommandContext(ctx, "nsenter", append([]string{fmt.Sprintf("--target=%d", target), "--all"}, injectcmd...)...), nil
+}
+
+var (
+       errContainerHasNoIPAddress = errors.New("container has no IP address distinct from host")
+)
+
+func (e *singularityExecutor) IPAddress() (string, error) {
+       target, err := e.containedProcess()
+       if err != nil {
+               return "", err
+       }
+       targetIPs, err := processIPs(target)
+       if err != nil {
+               return "", err
+       }
+       selfIPs, err := processIPs(os.Getpid())
+       if err != nil {
+               return "", err
+       }
+       for ip := range targetIPs {
+               if !selfIPs[ip] {
+                       return ip, nil
+               }
+       }
+       return "", errContainerHasNoIPAddress
+}
+
+func processIPs(pid int) (map[string]bool, error) {
+       fibtrie, err := os.ReadFile(fmt.Sprintf("/proc/%d/net/fib_trie", pid))
+       if err != nil {
+               return nil, err
+       }
+
+       addrs := map[string]bool{}
+       // When we see a pair of lines like this:
+       //
+       //              |-- 10.1.2.3
+       //                 /32 host LOCAL
+       //
+       // ...we set addrs["10.1.2.3"] = true
+       lines := bytes.Split(fibtrie, []byte{'\n'})
+       for linenumber, line := range lines {
+               if !bytes.HasSuffix(line, []byte("/32 host LOCAL")) {
+                       continue
+               }
+               if linenumber < 1 {
+                       continue
+               }
+               i := bytes.LastIndexByte(lines[linenumber-1], ' ')
+               if i < 0 || i >= len(line)-7 {
+                       continue
+               }
+               addr := string(lines[linenumber-1][i+1:])
+               if net.ParseIP(addr).To4() != nil {
+                       addrs[addr] = true
+               }
+       }
+       return addrs, nil
+}
+
+var (
+       errContainerNotStarted = errors.New("container has not started yet")
+       errCannotFindChild     = errors.New("failed to find any process inside the container")
+       reProcStatusPPid       = regexp.MustCompile(`\nPPid:\t(\d+)\n`)
+)
+
+// Return the PID of a process that is inside the container (not
+// necessarily the topmost/pid=1 process in the container).
+func (e *singularityExecutor) containedProcess() (int, error) {
+       if e.child == nil || e.child.Process == nil {
+               return 0, errContainerNotStarted
+       }
+       lsns, err := exec.Command("lsns").CombinedOutput()
+       if err != nil {
+               return 0, fmt.Errorf("lsns: %w", err)
+       }
+       for _, line := range bytes.Split(lsns, []byte{'\n'}) {
+               fields := bytes.Fields(line)
+               if len(fields) < 4 {
+                       continue
+               }
+               if !bytes.Equal(fields[1], []byte("pid")) {
+                       continue
+               }
+               pid, err := strconv.ParseInt(string(fields[3]), 10, 64)
+               if err != nil {
+                       return 0, fmt.Errorf("error parsing PID field in lsns output: %q", fields[3])
+               }
+               for parent := pid; ; {
+                       procstatus, err := os.ReadFile(fmt.Sprintf("/proc/%d/status", parent))
+                       if err != nil {
+                               break
+                       }
+                       m := reProcStatusPPid.FindSubmatch(procstatus)
+                       if m == nil {
+                               break
+                       }
+                       parent, err = strconv.ParseInt(string(m[1]), 10, 64)
+                       if err != nil {
+                               break
+                       }
+                       if int(parent) == e.child.Process.Pid {
+                               return int(pid), nil
+                       }
+               }
+       }
+       return 0, errCannotFindChild
+}