cmd/enter, cmd/run: Add --arch flag for cross-arch containers support#1789
Open
DaliborKr wants to merge 15 commits into
Open
cmd/enter, cmd/run: Add --arch flag for cross-arch containers support#1789DaliborKr wants to merge 15 commits into
DaliborKr wants to merge 15 commits into
Conversation
The existing RunContextWithExitCode() wraps all errors from exec.Command in generic "failed to invoke" messages, which prevents callers from distinguishing between actual error types. Add RunContextWithExitCode2() and RunWithExitCode2() that return errors with their original types intact, including for ExitError. This allows callers to use errors.Is() and errors.As() to handle specific failure modes. For example, detecting a missing skopeo binary (exec.ErrNotFound) or an ENOEXEC error from inside non native containers, when an emulator is not set correctly. These new functions are meant to replace its original versions in the future. containers#1780 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for non-native architecture Toolbx containers by leveraging QEMU user-mode emulation and binfmt_misc. It adds an --arch flag to the create, enter, and run commands, and implements the necessary logic to pull non-native images via skopeo and initialize the emulation environment within the container. Feedback highlights a logical flaw in the container initialization sequence where architecture registration might be skipped or fail due to existing registrations, and suggests improving the binfmt_misc mounting logic to handle cases where the filesystem is already mounted.
Comment on lines
+286
to
+295
| err := architecture.IsArchSupportedOnInitialization(initContainerFlags.archID, interpreterPath) | ||
| if err != nil { | ||
| errNotSupported := fmt.Errorf("Cannot run container for architecture %s:\n%s", archName, err) | ||
| return errNotSupported | ||
| } | ||
|
|
||
| logrus.Debugf("Registering QEMU emulator for architecture %s in binfmt_mist", archName) | ||
| if err := architecture.RegisterBinfmtMisc(initContainerFlags.archID, interpreterPath); err != nil { | ||
| return err | ||
| } |
There was a problem hiding this comment.
The logic for architecture emulator registration is problematic. IsArchSupportedOnInitialization returns an error if the architecture is not already registered, which causes the function to exit at line 289, preventing RegisterBinfmtMisc from ever running. If the architecture is already registered, IsArchSupportedOnInitialization succeeds, but then RegisterBinfmtMisc is called and will likely fail with an 'already exists' error when trying to write to the binfmt_misc register. The code should check if registration is needed and only attempt it if the architecture is not yet supported.
Comment on lines
+130
to
+132
| if err := shell.Run("mount", nil, &stdout, nil, args...); err != nil { | ||
| return fmt.Errorf("failed to mount binfmt_misc: %w", err) | ||
| } |
There was a problem hiding this comment.
The mount command will fail if binfmt_misc is already mounted at /proc/sys/fs/binfmt_misc. In many container environments, this filesystem might already be available. It is recommended to check if the path is already a mount point or to handle the 'already mounted' exit code from the mount command to avoid unnecessary failures during container initialization.
|
Build succeeded. ✔️ unit-test SUCCESS in 3m 34s |
In /src/cmd/create.go, the same pattern of spinner creation and nil-safe stopping is repeated. Extract this into startSpinner() and stopSpinner() helper functions so that future callers can use spinners without duplicating the code. Replace the existing inline spinner code in createContainer() and pullImage() with calls to these new helpers. containers#1781 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
…atching Add IsSupportedDistroImage(), which iterates over all supported distros and checks if the image basename matches any of them. This will be used by the architecture resolution code to decide whether to parse architecture suffixes from image tags, as this should be done only for natively supported images [1]. [1] Toolbx supported distributions: https://containertoolbx.org/distros/ containers#1781 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Introduce the architecture package that represents the core of the Toolbx cross-architecture support, which is based on user-mode emulation using QEMU and binfmt_misc. The Architecture struct collects all per-architecture data (ELF magic/mask, OCI and binfmt naming, aliases, binfmt registration parameters) into a single map. Architectures present in the supportedArchitectures map represent the set of supported architectures within Toolbx. Define architecture ID constants NotSpecified, Aarch64, Ppc64le, and X86_64, along with their supportedArchitectures entries. Add core query functions: - ParseArgArchValue() for resolving user-supplied architecture strings - GetArchNameBinfmt() and GetArchNameOCI() for name lookups (one architecture can have multiple valid names [1]) - HasContainerNativeArch() for comparing against the host - ImageReferenceGetArchFromTag() for extracting architecture from image tag suffixes like "42-aarch64" for architecture detection Expose the HostArchID package variable, which is set in the init() function, so the variable can be accessed in the early init() state from every inheritor that utilizes the architecture package (HostArchID serves as a default value for initContainer --arch flag), and the Config struct for preserving the architecture ID and the QEMU emulator path, through the call chain. [1] https://itsfoss.com/arm-aarch64-x86_64/ containers#1782 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Cross-architecture containers need QEMU binfmt_misc handlers registered within the container so that non-native architecture binaries can be executed via the host's kernel. Add the Registration struct that models a binfmt_misc registration entry, including name, magic type, offset, ELF magic/mask bytes, interpreter path, and flags. Add functions: - MountBinfmtMisc() to mount the sanboxed binfmt_misc filesystem inside a container, which enables setting the C flag in binfmt_misc registration without affecting the host system. The C flag presents a threat of privilege escalation when registered on the host, that why we want to have it isolated [1] - getDefaultRegistration() to fill a Registration struct containing all necessary binfmt_misc information taken from the architecture.supportedArchitectures data - RegisterBinfmtMisc() to write the registration string to /proc/sys/fs/binfmt_misc/register, which makes the non-native binary perception active - bytesToEscapedString() helper that formats byte slices into the \xHH-escaped string format required by the binfmt_misc register interface [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=21ca59b365c0 containers#1782 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Before creating or initializing a cross-architecture container, the system must be checked for the required QEMU emulator and binfmt_misc registration. This prevents users from creating or running non-native containers when their host system doesn't meet the requirements, and provides users with an informative error message referring to the problem. Add IsArchSupportedOnCreation(), which searches for a statically linked QEMU binary on the host using exec.LookPath() and verifies that a matching binfmt_misc registration exists. It returns the path to the QEMU binary for use during container creation, which is meant to be passed to the init-container and registered through sandboxed binfmt_misc within the container. Add IsArchSupportedOnInitialization() which performs similar checks from inside the container, looking at the interpreter path passed from the host and falling back to standard host-mounted locations under /run/host/usr/bin/. Add isStaticallyLinkedELF() helper that uses debug/elf to verify a binary is statically linked. Only a statically linked QEMU interpreter can be used, because a dynamically linked one would cause the kernel to attempt to resolve its host-native shared libraries (such as libc.so) within the container, resulting in an immediate crash. Add validateBinfmtRegistration(), which checks for the presence of qemu-<arch> entries in binfmt_misc (or qemu-<arch>-static, since it can differ based on the system). containers#1783 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Add Architecture and NameFull fields to the Skopeo Image struct so that callers can inspect the architecture of a remote image. Move the image size computation from the /cmd layer into GetSize() and GetSizeHuman() methods on Image, since the skopeo package owns the layer data. Add VerifyArchitectureMatch() method to Image that validates the image's architecture field against an expected architecture ID. The purpose of this function is to check whether the image architecture matches the demanded architecture before it is pulled. Specifically, this verification applies to the images that support only a single architecture (they are not part of a multi-platform manifest list), because the skopeo inspect proceeds successfully even when the value of a flag --override-arch does not match the actual image architecture (for a multi-architecture image the skopeo inspect with not-matching --override-arch would fail). Like this, the user can be prevented from incompatible images. containers#1784 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Change Inspect() to accept archID and authfile parameters. When the requested architecture differs from the host's, --override-arch is passed to skopeo, which then inspects the correct manifest in a multi-arch image (if it exists for the given architecture, otherwise the inspection fails). It also uses RunContextWithExitCode2() so callers can detect a missing skopeo binary via errors.Is(err, exec.ErrNotFound), which is only a soft dependency of the Toolbx package, as it is not required for running native containers. Add CopyOverrideArch(), which uses 'skopeo copy --override-arch' to pull a specific architecture variant of a multi-arch image into Podman's local container storage. This is used instead of 'podman pull' because Podman does not support pulling a foreign architecture image into a locally addressable name. The way in which the cross-arch extension chooses the name for non-native images (and also containers) is described in the discussion at [1] [1] containers/podman#27780 (comment) containers#1784 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
…and names Add resolveArchitectureID(), which combines the --arch command-line flag with architecture detection from image tag suffixes (e.g., "fedora-toolbox:42-aarch64"). This detection applies only to images from distributions that Toolbx explicitly supports (see [1]), to avoid a false architecture approach on custom images where a dash-separated component might not represent an architecture, since there is no standard set regarding preserving architecture in the tag (see detailed explanation at [2]). When both sources specify an architecture, it validates that they do not conflict. Add resolveImageNameWithArchitectureSuffix(), which appends the OCI architecture name to supported distro image references when the target architecture differs from the host, to ensure the local Toolbx images naming convention [2]. Again, this applies only to supported distros. [1] https://containertoolbx.org/distros/ [2] containers/podman#27780 (comment) containers#1786 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Change resolveContainerAndImageNames() to accept an archID parameter. When the target architecture is non-native, and the container name was auto-generated (was not set by a user), append the architecture suffix to the container name (e.g., "fedora-toolbox-arm64") to distinguish it from native containers. Temporarily update the callers of resolveContainerAndImageNames() to pass in architecture.HostArchID to the updated signature, to maintain a default native behavior. Once implemented, the --arch argument in the callers will pass the actual architecture information. containers#1786 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Add the --arch flag to the 'create' command, allowing users to create Toolbx containers for architectures different from the host (e.g., 'toolbox create --arch arm64'). Utilize the architecture resolution pipeline in create() by using resolveArchitectureID() (added in [1]) to determine the target architecture from the --arch flag and image tags. Validate host support via IsArchSupportedOnCreation() (added in [2]), which checks for the required QEMU emulator and binfmt_misc registration. Pass architecture ID to resolveContainerAndImageNames() (updated in [1]) so that non-native containers get architecture-suffixed names. Update pullImage() to handle cross-architecture image pulling: when the target architecture is non-native, use skopeo.CopyOverrideArch() (added in [3]) instead of podman.Pull(), since Podman does not support pulling foreign architecture images into locally addressable names. The need for this is explained in a discussion in [4]. Add a 'toolbox-arch' label to created containers to record the target architecture in OCI format. Extract the image pull error formatting into createErrorImagePull() in utils.go to avoid duplication between the native and cross-arch pull paths. Update the createContainer() call in run.go to pass the default architecture config via GetArchConfigDefault(), maintaining the existing native-architecture behavior. [1] containers#1786 [2] containers#1783 [3] containers#1784 [4] containers/podman#27780 containers#1787 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Rework the image download prompt flow to support architecture verification before pulling non-native images. The new implementation ensures that the image inspection completes for the non-native creation path before it is pulled, so the image's architecture can be verified. The previous implementation used promptForDownloadError as a control flow mechanism between the first and second download prompts. Replace this with the pullImageDecision enum (pullNo, pullYes, pullUnknown) for clearer three-state signaling. Replaced getImageSizeFromRegistryAsync() with getImageFromRegistryAsync(), which now returns the full skopeo.Image struct instead of just the image size string. It calls skopeo.Inspect() (updated in [1]), making image metadata available throughout the download prompt flow for both size display and architecture verification in a single inspect call. Use Image.GetSizeHuman() (added in [1]) for image size display in the second download prompt, replacing the local size computation. Update showPromptForDownloadFirst() to return (pullImageDecision, *skopeo.Image, error). For non-native architectures, when the user confirms the download, the function now waits for the skopeo inspect to complete (with a spinner) before returning, ensuring that architecture verification can happen before the pull begins. Update pullImage() to verify the image architecture before pulling non-native images by calling VerifyArchitectureMatch() (added in [1]) to catch incompatible single-architecture images. Handle the case where the inspect returns nil (multi-arch manifest has no matching variant) with an explicit error. Detect a missing skopeo binary via exec.ErrNotFound, which is only a soft dependency of the Toolbx package, as it is not required for running non-native containers, and report it through createErrorSkopeoNotFound() added in utils.go. [1] containers#1784 containers#1787 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Add the --arch and --arch-emulator-path flags to the init-container command, passed from the create command when creating a cross-architecture container. The --arch flag defaults to the host architecture ID so that existing native containers continue to work without changes. When the container's architecture differs from the host, the init-container entry point configures QEMU emulation inside the container before any foreign-architecture binaries can run: 1. Validate QEMU emulation by running the 'true' command, which fails with ENOEXEC if the host's binfmt_misc registration is not working (detected via RunWithExitCode2() added in [1]), because it is necessary to have host emulation working to emulate the binfmt_misc registration in the following step. 2. Mount a fresh binfmt_misc filesystem inside the container via MountBinfmtMisc() (added in [2]) to create a sandboxed binfmt_misc registration with the C flag. 3. Validate architecture support via IsArchSupportedOnInitialization() (added in [3]), which verifies the QEMU interpreter at the host-mounted path under /run/host. 4. Register the QEMU interpreter with the C flag via RegisterBinfmtMisc() (added and explained in [2]) The binfmt_misc registration is performed inside the container rather than relying on the host's registration, as explained in [2]. Update showEntryPointLog() in run.go to propagate lines prefixed with 'Warning:' to stderr on the host, instead of treating them as errors. This is needed because the cross-architecture initialization may emit warnings that should be visible to the user but are not fatal. [1] containers#1780 [2] containers#1782 [3] containers#1783 containers#1788 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Add the --arch / -a flag to the enter command, allowing users to enter cross-architecture containers by specifying the target architecture (e.g., toolbox enter --arch arm64). Can be used with flags --distro and --release, just as for container creation. The flag value is resolved through resolveArchitectureID() (added in [1]) and passed to resolveContainerAndImageNames() (updated for cross-arch in [1]) so that it resolves to the architecture-suffixed container name. [1] containers#1786 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
Add the --arch / -a flag to the run command, allowing users to run commands inside cross-architecture containers by specifying the target architecture (e.g., toolbox run --arch arm64 uname -m). Can be used with flags --distro and --release, just as for container creation. The flag value is resolved through resolveArchitectureID() (added in [1]) and passed to resolveContainerAndImageNames() (updated for cross-arch in [1]) so that it resolves to the architecture-suffixed container name. [1] containers#1786 Signed-off-by: Dalibor Kricka <dalidalk@seznam.cz>
|
Build succeeded. ✔️ unit-test SUCCESS in 3m 04s |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is PR 9/10 in a series adding cross-architecture container support using QEMU and binfmt_misc.
Depends on: #1788
Please review #1788 first. The new commits in this PR are:
Summary