// Copyright (C) The Arvados Authors. All rights reserved. // // SPDX-License-Identifier: AGPL-3.0 package crunchrun import ( "bytes" "context" "errors" "fmt" "io/ioutil" "net" "os" "os/exec" "os/user" "regexp" "sort" "strconv" "strings" "syscall" "time" "git.arvados.org/arvados.git/sdk/go/arvados" ) 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 imageFilename string // "sif" image } func newSingularityExecutor(logf func(string, ...interface{})) (*singularityExecutor, error) { tmpdir, err := ioutil.TempDir("", "crunch-run-singularity-") if err != nil { return nil, err } return &singularityExecutor{ logf: logf, tmpdir: tmpdir, }, 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, arvados.EndpointGroupList.Method, arvados.EndpointGroupList.Path, nil, arvados.ListOptions{Filters: []arvados.Filter{ arvados.Filter{"owner_uuid", "=", ownerUuid}, arvados.Filter{"name", "=", name}, arvados.Filter{"group_class", "=", "project"}, }, Limit: 1}) if err != nil { return nil, err } if len(gp.Items) == 1 { return &gp.Items[0], nil } var rgroup arvados.Group err = containerClient.RequestAndDecode(&rgroup, arvados.EndpointGroupCreate.Method, arvados.EndpointGroupCreate.Path, nil, map[string]interface{}{ "group": map[string]string{ "owner_uuid": ownerUuid, "name": name, "group_class": "project", }, }) if err != nil { return nil, err } return &rgroup, nil } func (e *singularityExecutor) checkImageCache(dockerImageID string, container arvados.Container, arvMountPoint string, containerClient *arvados.Client) (collection *arvados.Collection, err error) { // Cache the image to keep cacheGroup, err := e.getOrCreateProject(container.RuntimeUserUUID, ".cache", containerClient) if err != nil { return nil, fmt.Errorf("error getting '.cache' project: %v", err) } imageGroup, err := e.getOrCreateProject(cacheGroup.UUID, "auto-generated singularity images", containerClient) if err != nil { return nil, fmt.Errorf("error getting 'auto-generated singularity images' project: %s", err) } collectionName := fmt.Sprintf("singularity image for %v", dockerImageID) var cl arvados.CollectionList err = containerClient.RequestAndDecode(&cl, arvados.EndpointCollectionList.Method, arvados.EndpointCollectionList.Path, nil, arvados.ListOptions{Filters: []arvados.Filter{ arvados.Filter{"owner_uuid", "=", imageGroup.UUID}, arvados.Filter{"name", "=", collectionName}, }, Limit: 1}) if err != nil { return nil, fmt.Errorf("error querying for collection '%v': %v", collectionName, err) } var imageCollection arvados.Collection if len(cl.Items) == 1 { imageCollection = cl.Items[0] } else { collectionName := "converting " + collectionName exp := time.Now().Add(24 * 7 * 2 * time.Hour) err = containerClient.RequestAndDecode(&imageCollection, arvados.EndpointCollectionCreate.Method, arvados.EndpointCollectionCreate.Path, nil, map[string]interface{}{ "collection": map[string]string{ "owner_uuid": imageGroup.UUID, "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) } } return &imageCollection, nil } // LoadImage will satisfy ContainerExecuter interface transforming // containerImage into a sif file for later use. func (e *singularityExecutor) LoadImage(dockerImageID string, imageTarballPath string, container arvados.Container, arvMountPoint string, containerClient *arvados.Client) error { var imageFilename string var sifCollection *arvados.Collection var err error if containerClient != nil { sifCollection, err = e.checkImageCache(dockerImageID, container, arvMountPoint, containerClient) if err != nil { return err } imageFilename = fmt.Sprintf("%s/by_uuid/%s/image.sif", arvMountPoint, sifCollection.UUID) } else { imageFilename = e.tmpdir + "/image.sif" } 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, // as in "/path/to/sha256:abcd...1234.tar". Workaround: make a // symlink that doesn't have ":" chars. err := os.Symlink(imageTarballPath, e.tmpdir+"/image.tar") if err != nil { 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... // Getting image source signatures // Copying blob ab15617702de done // Copying config 651e02b8a2 done // Writing manifest to image destination // Storing signatures // 2021/04/22 14:42:14 info unpack layer: sha256:21cbfd3a344c52b197b9fa36091e66d9cbe52232703ff78d44734f85abb7ccd3 // INFO: Creating SIF file... // INFO: Build complete: arvados-jobs.latest.sif e.logf("%s", out) if err != nil { return err } } if containerClient == nil { e.imageFilename = imageFilename return nil } // update TTL to now + two weeks exp := time.Now().Add(24 * 7 * 2 * time.Hour) uuidPath, err := containerClient.PathForUUID("update", sifCollection.UUID) if err != nil { e.logf("error PathForUUID: %v", err) return nil } var imageCollection arvados.Collection err = containerClient.RequestAndDecode(&imageCollection, arvados.EndpointCollectionUpdate.Method, uuidPath, nil, map[string]interface{}{ "collection": map[string]string{ "name": fmt.Sprintf("singularity image for %v", dockerImageID), "trash_at": exp.UTC().Format(time.RFC3339), }, }) if err == nil { // If we just wrote the image to the cache, the // response also returns the updated PDH e.imageFilename = fmt.Sprintf("%s/by_id/%s/image.sif", arvMountPoint, imageCollection.PortableDataHash) return nil } e.logf("error updating/renaming collection for cached sif image: %v", err) // Failed to update but maybe it lost a race and there is // another cached collection in the same place, so check the cache // again sifCollection, err = e.checkImageCache(dockerImageID, container, arvMountPoint, containerClient) if err != nil { return err } e.imageFilename = fmt.Sprintf("%s/by_id/%s/image.sif", arvMountPoint, sifCollection.PortableDataHash) return nil } func (e *singularityExecutor) Create(spec containerSpec) error { e.spec = spec return nil } 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", } var binds []string for path, _ := range e.spec.BindMounts { binds = append(binds, path) } sort.Strings(binds) for _, path := range binds { mount := e.spec.BindMounts[path] 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 // that will change in singularity 3.6, please see: // https://sylabs.io/guides/3.7/user-guide/environment_and_metadata.html // https://sylabs.io/guides/3.5/user-guide/environment_and_metadata.html env := make([]string, 0, len(e.spec.Env)) for k, v := range e.spec.Env { if k == "HOME" { // 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...) return &exec.Cmd{ Path: path, Args: args, Env: env, Stdin: e.spec.Stdin, 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 } e.child = child return nil } func (e *singularityExecutor) Pid() int { // see https://dev.arvados.org/issues/17244#note-21 return 0 } func (e *singularityExecutor) Stop() error { if err := e.child.Process.Signal(syscall.Signal(0)); err != nil { // process already exited return nil } return e.child.Process.Signal(syscall.SIGKILL) } func (e *singularityExecutor) Wait(context.Context) (int, error) { err := e.child.Wait() if err, ok := err.(*exec.ExitError); ok { return err.ProcessState.ExitCode(), nil } if err != nil { return 0, err } return e.child.ProcessState.ExitCode(), nil } func (e *singularityExecutor) Close() { err := os.RemoveAll(e.tmpdir) if err != nil { 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 }