7. ch-image

Build and manage images; completely unprivileged.

7.1. Synopsis

$ ch-image [...] build [-t TAG] [-f DOCKERFILE] [...] CONTEXT
$ ch-image [...] build-cache [...]
$ ch-image [...] delete IMAGE_GLOB [IMAGE_GLOB ...]
$ ch-image [...] gestalt [SELECTOR]
$ ch-image [...] import PATH IMAGE_REF
$ ch-image [...] list [-l] [IMAGE_REF]
$ ch-image [...] pull [...] IMAGE_REF [DEST_REF]
$ ch-image [...] push [--image DIR] IMAGE_REF [DEST_REF]
$ ch-image [...] reset
$ ch-image [...] undelete IMAGE_REF
$ ch-image { --help | --version | --dependencies }

7.2. Description

ch-image is a tool for building and manipulating container images, but not running them (for that you want ch-run). It is completely unprivileged, with no setuid/setgid/setcap helpers. Many operations can use caching for speed. The action to take is specified by a sub-command.

Options that print brief information and then exit:

-h, --help

Print help and exit successfully. If specified before the sub-command, print general help and list of sub-commands; if after the sub-command, print help specific to that sub-command.


Report dependency problems on standard output, if any, and exit. If all is well, there is no output and the exit is successful; in case of problems, the exit is unsuccessful.


Print version number and exit successfully.

Common options placed before or after the sub-command:

-a, --arch ARCH

Use ARCH for architecture-aware registry operations. (See section “Architecture” below for details.)


Download all files when pulling, even if they are already in builder storage. Note that ch-image pull will always retrieve the most up-to-date image; this option is mostly for debugging.


Authenticate with the remote repository, then (if successful) make all subsequent requests in authenticated mode. For most subcommands, the default is to never authenticate, i.e., make all requests anonymously. The exception is push, which implies --auth.


Set a PDB breakpoint at line number LINE of module named MODULE (typically the filename with .py removed, or __main__ for ch-image itself). That is, a PDB debugger shell will open before executing the specified line.

This is accomplished by re-parsing the module, injecting import pdb; pdb.set_trace() into the parse tree, re-compiling the tree, and replacing the module’s code with the result. This has various gotchas, including (1) module-level code in the target module is executed twice, (2) the option is parsed with bespoke early code so command line argument parsing itself can be debugged, (3) breakpoints on function definition will trigger while the module is being re-executed, not when the function is called (break on the first line of the function body instead), and (4) other weirdness we haven’t yet characterized.


Enable build cache. Default if a sufficiently new Git is available. See section Build cache for details.

--cache-large SIZE

Set the cache’s large file threshold to SIZE MiB, or 0 for no large files, which is the default. Values greater than zero can speed up many builds but can also cause performance degradation. Experimental. See section Large file threshold for details.


Add a stack trace to fatal error hints. This can also be done by setting the environment variable CH_IMAGE_DEBUG.


Disable build cache. Default if a sufficiently new Git is not available. This option turns off the cache completely; if you want to re-execute a Dockerfile and store the new results in cache, use --rebuild instead.


Disable storage directory locking. This lets you run as many concurrent ch-image instances as you want against the same storage directory, which risks corruption but may be OK for some workloads.


Enforce default handling of xattrs, i.e. do not save them in the build cache or restore them on rebuild. This is the default, but the option is provided to override the $CH_XATTRS environment variable.


Re-prompt the user every time a registry password is needed.


Dump profile to files /tmp/chofile.p (cProfile dump format) and /tmp/chofile.txt (text summary). You can convert the former to a PDF call graph with gprof2dot -f pstats /tmp/chofile.p | dot -Tpdf -o /tmp/chofile.pdf. This excludes time spend in subprocesses. Profile data should still be written on fatal errors, but not if the program crashes.

-q, --quiet

Be quieter; can be repeated. Incompatible with -v and suppresses --debug regardless of option order. See the FAQ entry on verbosity for details.


Execute all instructions, even if they are build cache hits, except for FROM which is retrieved from cache on hit.

-s, --storage DIR

Set the storage directory (see below for important details).


Don’t verify TLS certificates of the repository. (Do not use this option unless you understand the risks.)

-v, --verbose

Print extra chatter; can be repeated. See the FAQ entry on verbosity for details.


Save xattrs and ACLs in the build cache, and restore them when rebuilding from the cache.

7.3. Architecture

Charliecloud provides the option --arch ARCH to specify the architecture for architecture-aware registry operations. The argument ARCH can be: (1) yolo, to bypass architecture-aware code and use the registry’s default architecture; (2) host, to use the host’s architecture, obtained with the equivalent of uname -m (default if --arch not specified); or (3) an architecture name. If the specified architecture is not available, the error message will list which ones are.


  1. ch-image is limited to one image per image reference in builder storage at a time, regardless of architecture. For example, if you say ch-image pull --arch=foo baz and then ch-image pull --arch=bar baz, builder storage will contain one image called “baz”, with architecture “bar”.

  2. Images’ default architecture is usually amd64, so this is usually what you get with --arch=yolo. Similarly, if a registry image is architecture-unaware, it will still be pulled with --arch=amd64 and --arch=host on x86-64 hosts (other host architectures must specify --arch=yolo to pull architecture-unaware images).

  3. uname -m and image registries often use different names for the same architecture. For example, what uname -m reports as “x86_64” is known to registries as “amd64”. --arch=host should translate if needed, but it’s useful to know this is happening. Directly specified architecture names are passed to the registry without translation.

  4. Registries treat architecture as a pair of items, architecture and sometimes variant (e.g., “arm” and “v7”). Charliecloud treats architecture as a simple string and converts to/from the registry view transparently.

7.4. Authentication

Charliecloud does not have configuration files; thus, it has no separate login subcommand to store secrets. Instead, Charliecloud will prompt for a username and password when authentication is needed. Note that some repositories refer to the secret as something other than a “password”; e.g., GitLab calls it a “personal access token (PAT)”, Quay calls it an “application token”, and nVidia NGC calls it an “API token”.

For non-interactive authentication, you can use environment variables CH_IMAGE_USERNAME and CH_IMAGE_PASSWORD. Only do this if you fully understand the implications for your specific use case, because it is difficult to securely store secrets in environment variables.

By default for most subcommands, all registry access is anonymous. To instead use authenticated access for everything, specify --auth or set the environment variable $CH_IMAGE_AUTH=yes. The exception is push, which always runs in authenticated mode. Even for pulling public images, it can be useful to authenticate for registries that have per-user rate limits, such as Docker Hub. (Older versions of Charliecloud started with anonymous access, then tried to upgrade to authenticated if it seemed necessary. However, this turned out to be brittle; see issue #1318.)

The username and password are remembered for the life of the process and silently re-offered to the registry if needed. One case when this happens is on push to a private registry: many registries will first offer a read-only token when ch-image checks if something exists, then re-authenticate when upgrading the token to read-write for upload. If your site uses one-time passwords such as provided by a security device, you can specify --password-many to provide a new secret each time.

These values are not saved persistently, e.g. in a file. Note that we do use normal Python variables for this information, without pinning them into physical RAM with mlock(2) or any other special treatment, so we cannot guarantee they will never reach non-volatile storage.

Technical details

Most registries use something called Bearer authentication, where the client (e.g., Charliecloud) includes a token in the headers of every HTTP request.

The authorization dance is different from the typical UNIX approach, where there is a separate login sequence before any content requests are made. The client starts by simply making the HTTP request it wants (e.g., to GET an image manifest), and if the registry doesn’t like the client’s token (or if there is no token because the client doesn’t have one yet), it replies with HTTP 401 Unauthorized, but crucially it also provides instructions in the response header on how to get a token. The client then follows those instructions, obtains a token, re-tries the request, and (hopefully) all is well. This approach also allows a client to upgrade a token if needed, e.g. when transitioning from asking if a layer exists to uploading its content.

The distinction between Charliecloud’s anonymous mode and authenticated modes is that it will only ask for anonymous tokens in anonymous mode and authenticated tokens in authenticated mode. That is, anonymous mode does involve an authentication procedure to obtain a token, but this “authentication” is done anonymously. (Yes, it’s confusing.)

Registries also often reply HTTP 401 when an image does not exist, rather than the seemingly more correct HTTP 404 Not Found. This is to avoid information leakage about the existence of images the client is not allowed to pull, and it’s why Charliecloud never says an image simply does not exist.

7.5. Storage directory

ch-image maintains state using normal files and directories located in its storage directory; contents include various caches and temporary images used for building.

In descending order of priority, this directory is located at:

-s, --storage DIR

Command line option.


Environment variable. The path must be absolute, because the variable is likely set in a very different context than when it’s used, which seems error-prone on what a relative path is relative to.


Default. (Previously, the default was /var/tmp/$USER/ch-image. If a valid storage directory is found at the old default path, ch-image tries to move it to the new default path.)

Unlike many container implementations, there is no notion of storage drivers, graph drivers, etc., to select and/or configure.

The storage directory can reside on any single filesystem (i.e., it cannot be split across multiple filesystems). However, it contains lots of small files and metadata traffic can be intense. For example, the Charliecloud test suite uses approximately 400,000 files and directories in the storage directory as of this writing. Place it on a filesystem appropriate for this; tmpfs’es such as /var/tmp are a good choice if you have enough RAM (/tmp is not recommended because ch-run bind-mounts it into containers by default).

While you can currently poke around in the storage directory and find unpacked images runnable with ch-run, this is not a supported use case. The supported workflow uses ch-convert to obtain a packed image; see the tutorial for details.

The storage directory format changes on no particular schedule. ch-image is normally able to upgrade directories produced by a given Charliecloud version up to one year after that version’s release. Upgrades outside this window and downgrades are not supported. In these cases, ch-image will refuse to run until you delete and re-initialize the storage directory with ch-image reset.


Network filesystems, especially Lustre, are typically bad choices for the storage directory. This is a site-specific question and your local support will likely have strong opinions.

7.6. Build cache

7.6.1. Overview

Subcommands that create images, such as build and pull, can use a build cache to speed repeated operations. That is, an image is created by starting from the empty image and executing a sequence of instructions, largely Dockerfile instructions but also some others like “pull” and “import”. Some instructions are expensive to execute (e.g., RUN wget http://slow.example.com/bigfile or transferring data billed by the byte), so it’s often cheaper to retrieve their results from cache instead.

The build cache uses a relatively new Git under the hood; see the installation instructions for version requirements. Charliecloud implements workarounds for Git’s various storage limitations, so things like file metadata and Git repositories within the image should work. Important exception: No files named .git* or other Git metadata are permitted in the image’s root directory.

Extended attributes (xattrs) are ignored by the build cache by default. Cache support for xattrs belonging to unprivileged xattr namespaces (e.g. user) can be enabled by specifying the --xattrs option or by setting the CH_XATTRS environment variable. If CH_XATTRS is set, you override it with --no-xattrs. Note that extended attributes in privileged xattr namespaces (e.g. :code:‘trusted‘) cannot be read by :code:‘ch-image‘ and will always be lost without warning.

The cache has three modes: enabled, disabled, and a hybrid mode called rebuild where the cache is fully enabled for FROM instructions, but all other operations re-execute and re-cache their results. The purpose of rebuild is to do a clean rebuild of a Dockerfile atop a known-good base image.

Enabled mode is selected with --cache or setting $CH_IMAGE_CACHE to enabled, disabled mode with --no-cache or disabled, and rebuild mode with --rebuild or rebuild. The default mode is enabled if an appropriate Git is installed, otherwise disabled.

7.6.2. Compared to other implementations


This section is a lightly edited excerpt from our paper “Charliecloud’s layer-free, Git-based container build cache”.

Existing tools such as Docker and Podman implement their build cache with a layered (union) filesystem such as OverlayFS or FUSE-OverlayFS and tar archives to represent the content of each layer; this approach is standardized by OCI. The layered cache works, but it has drawbacks in three critical areas:

  1. Diff format. The tar format is poorly standardized and not designed for diffs. Notably, tar cannot represent file deletion. The workaround used for OCI layers is specially named whiteout files, which means the tar archives cannot be unpacked by standard UNIX tools and require special container-specific processing.

  2. Cache overhead. Each time a Dockerfile instruction is started, a new overlay filesystem is mounted atop the existing layer stack. File metadata operations in the instruction then start at the top layer and descend the stack until the layer containing the desired file is reached. The cost of these operations is therefore proportional to the number of layers, i.e., the number of instructions between the empty root image and the instruction being executed. This results in a best practice of large, complex instructions to minimize their number, which can conflict with simpler, more numerous instructions the user might prefer.

  3. De-duplication. Identical files on layers with an ancestry relationship (i.e., instruction A precedes B in a build) are stored only once. However, identical files on layers without this relationship are stored multiple times. For example, if instructions B and B’ both follow A — perhaps because B was modified and the image rebuilt — then any files created by both B and B’ will be stored twice.

    Also, similar files are never de-duplicated, regardless of ancestry. For example, if instruction A creates a file and subsequently instruction B modifies a single bit in that file, both versions are stored in their entirety.

Our Git-based cache addresses the three drawbacks: (1) Git is purpose-built to store changing directory trees, (2) cache overhead is imposed only at instruction commit time, and (3) Git de-duplicates both identical and similar files. Also, it is based on an extremely widely used tool that enjoys development support from well-resourced actors, in particular on scaling (e.g., Microsoft’s large-repository accelerator Scalar was recently merged into Git).

In addition to these structural advantages, performance experiments reported in our paper above show that the Git-based approach is as good as (and sometimes better than) overlay-based caches. On build time, the two approaches are broadly similar, with one or the other being faster depending on context. Both had performance problems on NFS. Notably, however, the Git-based cache was much faster for a 129-instruction Dockerfile. On disk usage, the winner depended on the condition. For example, we saw the layered cache storing large sibling layers redundantly; on the other hand, the Git-based cache has some obvious redundancies as well, and one must compact it for full de-duplication benefit. However, Git’s de-duplication was quite effective in some conditions and we suspect will prove even better in more realistic scenarios.

That is, we believe our results show that the Git-based build cache is highly competitive with the layered approach, with no obvious inferiority so far and hints that it may be superior on important dimensions. We have ongoing work to explore these questions in more detail.

7.6.3. De-duplication and garbage collection

Charliecloud’s build cache takes advantage of Git’s file de-duplication features. This operates across the entire build cache, i.e., files are de-duplicated no matter where in the cache they are found or the relationship between their container images. Files are de-duplicated at different times depending on whether they are identical or merely similar.

Identical files are de-duplicated at git add time; in ch-image build terms, that’s upon committing a successful instruction. That is, it’s impossible to store two files with the same content in the build cache. If you try — say with RUN yum install -y foo in one Dockerfile and RUN yum install -y foo bar in another, which are different instructions but both install RPM foo’s files — the content is stored once and each copy gets its own metadata and a pointer to the content, much like filesystem hard links.

Similar files, however, are only de-duplicated during Git’s garbage collection process. When files are initially added to a Git repository (with git add), they are stored inside the repository as (possibly compressed) individual files, called objects in Git jargon. Upon garbage collection, which happens both automatically when certain parameters are met and explicitly with git gc, these files are archived and (re-)compressed together into a single file called a packfile. Also, existing packfiles may be re-written into the new one.

During this process, similar files are identified, and each set of similar files is stored as one base file plus diffs to recover the others. (Similarity detection seems to be based primarily on file size.) This delta process is agnostic to alignment, which is an advantage over alignment-sensitive block-level de-duplicating filesystems. Exception: “Large” files are not compressed or de-duplicated. We use the Git default threshold of 512 MiB (as of this writing).

Charliecloud runs Git garbage collection at two different times. First, a lighter-weight garbage pass runs automatically when the number of loose files (objects) grows beyond a limit. This limit is in flux as we learn more about build cache performance, but it’s quite a bit higher than the Git default. This garbage runs in the background and can continue after the build completes; you may see Git processes using a lot of CPU.

An important limitation of the automatic garbage is that large packfiles (again, this is in flux, but it’s several GiB) will not be re-packed, limiting the scope of similar file detection. To address this, a heavier garbage collection can be run manually with ch-image build-cache --gc. This will re-pack (and re-write) the entire build cache, de-duplicating all similar files. In both cases, garbage uses all available cores.

git build-cache prints the specific garbage collection parameters in use, and -v can be added for more detail.

7.6.4. Large file threshold

Because Git uses content-addressed storage, upon commit, it must read in full all files modified by an instruction. This I/O cost can be a significant fraction of build time for some images. To mitigate this, regular files larger than the experimental large file threshold are stored outside the Git repository, somewhat like Git Large File Storage.

ch-image copies large files in and out of images at each instruction commit. It tries to do this with a fast metadata-only copy-on-write operation called “reflink”, but that is only supported with the right Python version, Linux kernel version, and filesystem. If unsupported, Charliecloud falls back to an expensive standard copy, which is likely slower than letting Git deal with the files. See File copy performance for details.

Every version of a large file is stored verbatim and uncompressed (e.g., a large file with a one-byte change will be stored in full twice), so Git’s de-duplication does not apply. However, on filesystems with reflink support, files can share extents (e.g., each of the two files will have its own extent containing the changed byte, but the rest of the extents will remain shared). This provides de-duplication between large files images that share ancestry. Also, unused large files are deleted by ch-image build-cache --gc.

A final caveat: Large files in any image with the same path, mode, size, and mtime (to nanosecond precision if possible) are considered identical, even if their content is not actually identical (e.g., touch(1) shenanigans can corrupt an image).

Option --cache-large sets the threshold in MiB; if not set, environment variable CH_IMAGE_CACHE_LARGE is used; if that is not set either, the default value 0 indicates that no files are considered large.

(Note that Git has an unrelated setting called core.bigFileThreshold.)

7.6.5. Example

Suppose we have this Dockerfile:

$ cat a.df
FROM alpine:3.17
RUN echo foo
RUN echo bar

On our first build, we get:

$ ch-image build -t foo -f a.df .
  1. FROM alpine:3.17
[ ... pull chatter omitted ... ]
  2. RUN echo foo
copying image ...
  3. RUN echo bar
grown in 3 instructions: foo

Note the dot after each instruction’s line number. This means that the instruction was executed. You can also see this by the output of the two echo commands.

But on our second build, we get:

$ ch-image build -t foo -f a.df .
  1* FROM alpine:3.17
  2* RUN echo foo
  3* RUN echo bar
copying image ...
grown in 3 instructions: foo

Here, instead of being executed, each instruction’s results were retrieved from cache. (Charliecloud uses lazy retrieval; nothing is actually retrieved until the end, as seen by the “copying image” message.) Cache hit for each instruction is indicated by an asterisk (*) after the line number. Even for such a small and short Dockerfile, this build is noticeably faster than the first.

We can also try a second, slightly different Dockerfile. Note that the first three instructions are the same, but the third is different:

$ cat c.df
FROM alpine:3.17
RUN echo foo
RUN echo qux
$ ch-image build -t c -f c.df .
  1* FROM alpine:3.17
  2* RUN echo foo
  3. RUN echo qux
copying image ...
grown in 3 instructions: c

Here, the first two instructions are hits from the first Dockerfile, but the third is a miss, so Charliecloud retrieves that state and continues building.

We can also inspect the cache:

$ ch-image build-cache --tree
*  (c) RUN echo qux
| *  (a) RUN echo bar
*  RUN echo foo
*  (alpine+3.9) PULL alpine:3.17
*  (root) ROOT

named images:     4
state IDs:        5
commits:          5
files:          317
disk used:        3 MiB

Here there are four named images: a and c that we built, the base image alpine:3.17 (written as alpine+3.9 because colon is not allowed in Git branch names), and the empty base of everything root. Also note how a and c diverge after the last common instruction RUN echo foo.

7.7. build

Build an image from a Dockerfile and put it in the storage directory.

7.7.1. Synopsis

$ ch-image [...] build [-t TAG] [-f DOCKERFILE] [...] CONTEXT

7.7.2. Description

See below for differences with other Dockerfile interpreters. Charliecloud supports an extended instruction (RSYNC), a few other instructions behave slightly differently, and a few are ignored.

Note that FROM implicitly pulls the base image if needed, so you may want to read about the pull subcommand below as well.

Required argument:


Path to context directory. This is the root of COPY instructions in the Dockerfile. If a single hyphen (-) is specified: (a) read the Dockerfile from standard input, (b) specifying --file is an error, and (c) there is no context, so COPY will fail. (See --file for how to provide the Dockerfile on standard input while also having a context.)


-b, --bind SRC[:DST]

For RUN instructions only, bind-mount SRC at guest DST. The default destination if not specified is to use the same path as the host; i.e., the default is equivalent to --bind=SRC:SRC. If DST does not exist, try to create it as an empty directory, though images do have ten directories /mnt/[0-9] already available as mount points. Can be repeated.

Note: See documentation for ch-run --bind for important caveats and gotchas.

Note: Other instructions that modify the image filesystem, e.g. COPY, can only access host files from the context directory, regardless of this option.

--build-arg KEY[=VALUE]

Set build-time variable KEY defined by ARG instruction to VALUE. If VALUE not specified, use the value of environment variable KEY.

-f, --file DOCKERFILE

Use DOCKERFILE instead of CONTEXT/Dockerfile. If a single hyphen (-) is specified, read the Dockerfile from standard input; like docker build, the context directory is still available in this case.


Use unprivileged build with root emulation mode MODE, which can be fakeroot, seccomp (the default), or none. See section “Privilege model” below for details on what this does and when you might need it.


If command CMD is found in a RUN instruction, add the comma-separated ARGs to it. For example, --force-cmd=foo,-a,--bar=baz would transform RUN foo -c into RUN foo -a --bar=baz -c. This is intended to suppress validation that defeats --force=seccomp and implies that option. Can be repeated. If specified, replaces (does not extend) the default suppression options. Literal commas can be escaped with backslash; importantly however, backslash will need to be protected from the shell also. Section “Privilege model” below explains why you might need this.

-n, --dry-run

Don’t actually execute any Dockerfile instructions.


Stop after parsing the Dockerfile.

-t, --tag TAG

Name of image to create. If not specified, infer the name:

  1. If Dockerfile named Dockerfile with an extension: use the extension with invalid characters stripped, e.g. Dockerfile.@FOO.barfoo.bar.

  2. If Dockerfile has extension df or dockerfile: use the basename with the same transformation, e.g. baz.@QUX.dockerfile -> baz.qux.

  3. If context directory is not /: use its name, i.e. the last component of the absolute path to the context directory, with the same transformation,

  4. Otherwise (context directory is /): use root.

If no colon present in the name, append :latest.

Uses ch-run -w -u0 -g0 --no-passwd --unsafe to execute RUN instructions.

7.7.3. Privilege model Overview

ch-image is a fully unprivileged image builder. It does not use any setuid or setcap helper programs, and it does not use configuration files /etc/subuid or /etc/subgid. This contrasts with the “rootless” or “fakeroot” modes of some competing builders, which do require privileged supporting code or utilities.

Without root emulation, this approach does confuse programs that expect to have real root privileges, most notably distribution package installers. This subsection describes why that happens and what you can do about it.

ch-image executes all instructions as the normal user who invokes it. For RUN, this is accomplished with ch-run arguments including -w --uid=0 --gid=0. That is, your host EUID and EGID are both mapped to zero inside the container, and only one UID (zero) and GID (zero) are available inside the container. Under this arrangement, processes running in the container for each RUN appear to be running as root, but many privileged system calls will fail without the root emulation methods described below. This affects any fully unprivileged container build, not just Charliecloud.

The most common time to see this is installing packages. For example, here is RPM failing to chown(2) a file, which makes the package update fail:

  Updating   : 1:dbus-1.10.24-13.el7_6.x86_64                            2/4
Error unpacking rpm package 1:dbus-1.10.24-13.el7_6.x86_64
error: unpacking of archive failed on file /usr/libexec/dbus-1/dbus-daemon-launch-helper;5cffd726: cpio: chown
  Cleanup    : 1:dbus-libs-1.10.24-12.el7.x86_64                         3/4
error: dbus-1:1.10.24-13.el7_6.x86_64: install failed

This one is (ironically) apt-get failing to drop privileges:

E: setgroups 65534 failed - setgroups (1: Operation not permitted)
E: setegid 65534 failed - setegid (22: Invalid argument)
E: seteuid 100 failed - seteuid (22: Invalid argument)
E: setgroups 0 failed - setgroups (1: Operation not permitted)

Charliecloud provides two different mechanisms to avoid these problems. Both involve lying to the containerized process about privileged system calls, but at very different levels of complexity. Root emulation mode fakeroot

This mode uses fakeroot(1) to maintain an elaborate web of deceit that is internally consistent. This program intercepts both privileged system calls (e.g., setuid(2)) as well as other system calls whose return values depend on those calls (e.g., getuid(2)), faking success for privileged system calls (perhaps making no system call at all) and altering return values to be consistent with earlier fake success. Charliecloud automatically installs the fakeroot(1) program inside the container and then wraps RUN instructions having known privilege needs with it. Thus, this mode is only available for certain distributions.

The advantage of this mode is its consistency; e.g., careful programs that check the new UID after attempting to change it will not notice anything amiss. Its disadvantage is complexity: detailed knowledge and procedures for multiple Linux distributions.

This mode has three basic steps:

  1. After FROM, analyze the image to see what distribution it contains, which determines the specific workarounds.

  2. Before the user command in the first RUN instruction where the injection seems needed, install fakeroot(1) in the image, if one is not already installed, as well as any other necessary initialization commands. For example, we turn off the apt sandbox (for Debian Buster) and configure EPEL but leave it disabled (for CentOS/RHEL).

  3. Prepend fakeroot to RUN instructions that seem to need it, e.g. ones that contain apt, apt-get, dpkg for Debian derivatives and dnf, rpm, or yum for RPM-based distributions.

RUN instructions that do not seem to need modification are unaffected by this mode.

The details are specific to each distribution. ch-image analyzes image content (e.g., grepping /etc/debian_version) to select a configuration; see lib/force.py for details. ch-image prints exactly what it is doing.


Because of fakeroot mode’s complexity, we plan to remove it if seccomp mode performs well enough. If you have a situation where fakeroot mode works and seccomp does not, please let us know. Root emulation mode seccomp (default)

This mode uses the kernel’s seccomp(2) system call filtering to intercept certain privileged system calls, do absolutely nothing, and return success to the program.

Some system calls are quashed regardless of their arguments: capset(2); chown(2) and friends; kexec_load(2) (used to validate the filter itself); ; and setuid(2), setgid(2), and setgroups(2) along with the other system calls that change user or group. mknod(2) and mknodat(2) are quashed if they try to create a device file (e.g., creating FIFOs works normally).

The advantages of this approach is that it’s much simpler, it’s faster, it’s completely agnostic to libc, and it’s mostly agnostic to distribution. The disadvantage is that it’s a very lazy liar; even the most cursory consistency checks will fail, e.g., getuid(2) after setuid(2).

While this mode does not provide consistency, it does offer a hook to help prevent programs asking for consistency. For example, apt-get -o APT::Sandbox::User=root will prevent apt-get from attempting to drop privileges, which it verifies, exiting with failure if the correct IDs are not found (which they won’t be under this approach). This can be expressed with --force-cmd=apt-get,-o,APT::Sandbox::User=root, though this particular case is built-in and does not need to be specified. The full default configuration, which is applied regardless of the image distribution, can be examined in the source file force.py. If any --force-cmd are specified, this replaces (rather than extends) the default configuration.

Note that because the substitutions are a simple regex with no knowledge of shell syntax, they can cause unwanted modifications. For example, RUN apt-get install -y apt-get will be run as /bin/sh -c "apt-get -o APT::Sandbox::User=root install -y apt-get -o APT::Sandbox::User=root". One workaround is to add escape syntax transparent to the shell; e.g., RUN apt-get install -y apt-get.

This mode executes all RUN instructions with the seccomp(2) filter and has no knowledge of which instructions actually used the intercepted system calls. Therefore, the printed “instructions modified” number is only a count of instructions with a hook applied as described above. RUN logging

In terminal output, image metadata, and the build cache, the RUN instruction is always logged as RUN.S, RUN.F, or RUN.N. The letter appended to the instruction reflects the root emulation mode used during the build in which the instruction was executed. RUN.S indicates seccomp, RUN.F indicates fakeroot, and RUN.N indicates that neither form of root emulation was used (--force=none).

7.7.4. Compatibility and behavior differences

ch-image is an independent implementation and shares no code with other Dockerfile interpreters. It uses a formal Dockerfile parsing grammar developed from the Dockerfile reference documentation and miscellaneous other sources, which you can examine in the source code.

We believe this independence is valuable for several reasons. First, it helps the community examine Dockerfile syntax and semantics critically, think rigorously about what is really needed, and build a more robust standard. Second, it yields disjoint sets of bugs (note that Podman, Buildah, and Docker all share the same Dockerfile parser). Third, because it is a much smaller code base, it illustrates how Dockerfiles work more clearly. Finally, it allows straightforward extensions if needed to support scientific computing.

ch-image tries hard to be compatible with Docker and other interpreters, though as an independent implementation, it is not bug-compatible.

The following subsections describe differences from the Dockerfile reference that we expect to be approximately permanent. For not-yet-implemented features and bugs in this area, see related issues on GitHub.

None of these are set in stone. We are very interested in feedback on our assessments and open questions. This helps us prioritize new features and revise our thinking about what is needed for HPC containers. Context directory

The context directory is bind-mounted into the build, rather than copied like Docker. Thus, the size of the context is immaterial, and the build reads directly from storage like any other local process would (i.e., it is reasonable use / for the context). However, you still can’t access anything outside the context directory. Variable substitution

Variable substitution happens for all instructions, not just the ones listed in the Dockerfile reference.

ARG and ENV cause cache misses upon definition, in contrast with Docker where these variables miss upon use, except for certain cache-excluded variables that never cause misses, listed below.

Note that ARG and ENV have different syntax despite very similar semantics.

ch-image passes the following proxy environment variables in to the build. Changes to these variables do not cause a cache miss. They do not require an ARG instruction, as documented in the Dockerfile reference. Unlike Docker, they are available if the same-named environment variable is defined; --build-arg is not required.


In addition to those listed in the Dockerfile reference, these environment variables are passed through in the same way:


Finally, these variables are also pre-defined but are unrelated to the host environment:

TAR_OPTIONS=--no-same-owner ARG

Variables set with ARG are available anywhere in the Dockerfile, unlike Docker, where they only work in FROM instructions, and possibly in other ARG before the first FROM. FROM

The FROM instruction accepts option --arg=NAME=VALUE, which serves the same purpose as the ARG instruction. It can be repeated. LABEL

The LABEL instruction accepts key=value pairs to add metadata for an image. Unlike Docker, multiline values are not supported; see issue #1512. Can be repeated. COPY


The behavior described here matches Docker’s now-deprecated legacy builder. Docker’s new builder, BuildKit, has different behavior in some cases, which we have not characterized.

Especially for people used to UNIX cp(1), the semantics of the Dockerfile COPY instruction can be confusing.

Most notably, when a source of the copy is a directory, the contents of that directory, not the directory itself, are copied. This is documented, but it’s a real gotcha because that’s not what cp(1) does, and it means that many things you can do in one cp(1) command require multiple COPY instructions.

Also, the reference documentation is incomplete. In our experience, Docker also behaves as follows; ch-image does the same in an attempt to be bug-compatible.

  1. You can use absolute paths in the source; the root is the context directory.

  2. Destination directories are created if they don’t exist in the following situations:

    1. If the destination path ends in slash. (Documented.)

    2. If the number of sources is greater than 1, either by wildcard or explicitly, regardless of whether the destination ends in slash. (Not documented.)

    3. If there is a single source and it is a directory. (Not documented.)

  3. Symbolic links behave differently depending on how deep in the copied tree they are. (Not documented.)

    1. Symlinks at the top level — i.e., named as the destination or the source, either explicitly or by wildcards — are dereferenced. They are followed, and whatever they point to is used as the destination or source, respectively.

    2. Symlinks at deeper levels are not dereferenced, i.e., the symlink itself is copied.

  4. If a directory appears at the same path in source and destination, and is at the 2nd level or deeper, the source directory’s metadata (e.g., permissions) are copied to the destination directory. (Not documented.)

  5. If an object (a) appears in both the source and destination, (b) is at the 2nd level or deeper, and (c) is different file types in source and destination, the source object will overwrite the destination object. (Not documented.)

We expect the following differences to be permanent:

  • Wildcards use Python glob semantics, not the Go semantics.

  • COPY --chown is ignored, because it doesn’t make sense in an unprivileged build. Features we do not plan to support

  • Parser directives are not supported. We have not identified a need for any of them.

  • EXPOSE: Charliecloud does not use the network namespace, so containerized processes can simply listen on a host port like other unprivileged processes.

  • HEALTHCHECK: This instruction’s main use case is monitoring server processes rather than applications. Also, it requires a container supervisor daemon, which we have no plans to add.

  • MAINTAINER is deprecated.

  • STOPSIGNAL requires a container supervisor daemon process, which we have no plans to add.

  • USER does not make sense for unprivileged builds.

  • VOLUME: Charliecloud has good support for bind mounts; we anticipate that it will continue to focus on that and will not introduce the volume management features that Docker has.

7.7.5. RSYNC (Dockerfile extension)


This instruction is experimental and may change or be removed. Overview

Copying files is often simple but has numerous difficult corner cases, e.g. when dealing with symbolic or hard links. The standard instruction COPY deals with many of these corner cases differently from other UNIX utilities, lacks complete documentation, and behaves inconsistently between different Dockerfile interpreters (e.g., Docker’s legacy builder vs. BuildKit), as detailed above. On the other hand, rsync(1) is an extremely capable, widely used file copy tool, with detailed options to specify behavior and 25 years of history dealing with weirdness.

RSYNC (also spelled NSYNC) is a Charliecloud extension that gives copying behavior identical to rsync(1). In fact, Charliecloud’s current implementation literally calls the host’s rsync(1) to do the copy, though this may change in the future. There is no list form of RSYNC.

The two key usage challenges are trailing slashes on paths and symlink handling. In particular, the default symlink handling seemed reasonable to us, but you may want something different. See the arguments and examples below. Importantly, COPY is not any less fraught, and you have no choice about what to do with symlinks. Arguments

RSYNC takes the same arguments as rsync(1), so refer to its man page for a detailed explanation of all the options (with possible emphasis on its symlink options). Sources are relative to the context directory even if they look absolute with a leading slash. Any globbed sources are processed by ch-image(1) using Python rules, i.e., rsync(1) sees the expanded sources with no wildcards. Relative destinations are relative to the image’s current working directory, while absolute destinations refer to the image’s root.

For arguments that read input from a file (e.g. --exclude-from or --files-from), relative paths are relative to the context directory, absolute paths refer to the image root, and - (standard input) is an error.

For example,

RSYNC --foo src1 src2 dst

is translated to (the equivalent of):

$ mkdir -p /foo
$ rsync -@=-1 -AHSXpr --info=progress2 -l --safe-links \
        --foo /context/src1 /context/src2 /storage/imgroot/foo/dst2

Note the extensive default arguments to rsync(1). RSYNC takes a single instruction option beginning with + (plus) that is shorthand for a group of rsync(1) options. This single option is one of:


Preserves metadata and directory structure. Symlinks are skipped with a warning. Equivalent to all of:

  • -@=-1: use nanosecond precision when comparing timestamps.

  • -A: preserve ACLs.

  • -H: preserve hard link groups.

  • -S: preserve file sparseness when possible.

  • -X: preserve xattrs in user.* namespace.

  • -p: preserve permissions.

  • -r: recurse into directories.

  • --info=progress2 (only if stderr is a terminal): show progress meter (note subtleties in interpretation).

+l (default)

Like +u, but silently skips “unsafe” symlinks whose target is outside the top-of-transfer directory. Preserves:

  • Metadata.

  • Directory structure.

  • Symlinks, if a link’s target is within the “top-of-transfer directory”. This is not the context directory and often not the source either. Also, this creates broken symlinks if the target is not within the source but is within the top-of-transfer. See examples below.

Equivalent to the rsync(1) options listed for +m plus --links (copy symlinks as symlinks unless otherwise specified) and --safe-links (silently skip unsafe symlinks).


Like +l, but replaces with their target “unsafe” symlinks whose target is outside the top-of-transfer directory, and thus can copy data outside the context directory into the image. Preserves:

  • Metadata.

  • Directory structure.

  • Symlinks, if a link’s target is within the “top-of-transfer directory”. This is not the context directory and often not the source either. Also, this creates broken symlinks if the target is not within the source but is within the top-of-transfer. See examples below.

Equivalent to the rsync(1) options listed for +m plus --links (copy symlinks as symlinks unless otherwise specified) and --copy-unsafe-links (copy the target of unsafe symlinks).


No default arguments. Directories will not be descended, no metadata will be preserved, and both hard and symbolic links will be ignored, except as otherwise specified by rsync(1) options starting with a hyphen. (Note that -a/--archive is discouraged because it omits some metadata and handles symlinks inappropriately for containers.)


rsync(1) supports a configuration file ~/.popt that alters its command line processing. Currently, this configuration is respected for RSYNC arguments, but that may change without notice. Disallowed rsync(1) features

A small number of rsync(1) features are actively disallowed:

  1. rsync: and ssh: transports are an error. Charliecloud needs access to the entire input to compute cache hit or miss, and these transports make that impossible. It is possible these will become available in the future (please let us know if that is your use case!). For now, the workaround is to install rsync(1) in the image and use it in a RUN instruction, though only the instruction text will be considered for the cache.

  2. Option arguments must be delimited with = (equals). For example, to set the block size to 4 MiB, you must say --block-size=4M or -B=4M. -B4M will be interpreted as the three arguments -B, -4, and -M; --block-size 4M will be interpreted as --block-size with no argument and a copy source named 4M. This is so Charliecloud can process rsync(1) options without knowing which ones take an argument.

  3. Invalid rsync(1) options:


    Running rsync(1) in daemon mode does not make sense for container build.

    -n, --dry-run

    This makes the copy a no-op, and Charliecloud may want to use it internally in the future.


    This would let the instruction alter the context directory.

Note that there are likely other flags that don’t make sense and/or cause undesirable behavior. We have not characterized this problem. Build cache

The instruction is a cache hit if the metadata of all source files is unchanged (specifically: filename, file type and permissions, xattrs, size, and last modified time). Unlike Docker, Charliecloud does not use file contents. This has two implications. First, it is possible to fool the cache by manually restoring the last-modified time. Second, RSYNC is I/O-intensive even when it hits, because it must stat(2) every source file before checking the cache. However, this is still less I/O than reading the file content too.

Notably, Charliecloud’s cache ignores rsync(1)’s own internal notion of whether anything would be transferred (e.g., rsync -ni). This may change in the future. Examples and tutorial

All of these examples use the same input, whose content will be introduced gradually, using edited output of ls -oghR (which is like ls -lhR but omits user and group). Examples assume a umask of 0007. The Dockerfile instructions listed also assume a preceding:

FROM alpine:3.17
RUN mkdir /dst

i.e., a simple base image containing a top-level directory dst.

Many additional examples are available in the source code in the file test/build/50_rsync.bats.

We begin by copying regular files. The context directory ctx contains, in part, two directories containing one regular file each. Note that one of these files (file-basic1) and one of the directories (basic1) have strange permissions.

drwx---r-x 2  60 Oct 11 13:20 basic1
drwxrwx--- 2  60 Oct 11 13:20 basic2

-rw----r-- 1 12 Oct 11 13:20 file-basic1

-rw-rw---- 1 12 Oct 11 13:20 file-basic2

The simplest form of RSYNC is to copy a single file into a specified directory:

RSYNC /basic1/file-basic1 /dst

resulting in:

$ ls -oghR dst
-rw----r-- 1 12 Oct 11 13:26 file-basic1

Note that file-basic1’s metadata — here its odd permissions — are preserved. 1 is the number of hard links to the file, and 12 is the file size.

One can also rename the destination by specifying a new file name, and with +z, not copy metadata (from here on the ls command is omitted for brevity):

RSYNC +z /basic1/file-basic1 /dst/file-basic1_nom
-rw------- 1 12 Sep 21 15:51 file-basic1_nom

A trailing slash on the destination creates a new directory and places the source file within:

RSYNC /basic1/file-basic1 /dst/new/
drwxrwx--- 1 22 Oct 11 13:26 new

-rw----r-- 1 12 Oct 11 13:26 file-basic1

With multiple source files, the destination trailing slash is optional:

RSYNC /basic1/file-basic1 /basic2/file-basic2 /dst/newB
drwxrwx--- 1 44 Oct 11 13:26 newB

-rw----r-- 1 12 Oct 11 13:26 file-basic1
-rw-rw---- 1 12 Oct 11 13:26 file-basic2

For directory sources, the presence or absence of a trailing slash is highly significant. Without one, the directory itself is placed in the destination (recall that this would rename a source file):

RSYNC /basic1 /dst/basic1_new
drwxrwx--- 1 12 Oct 11 13:28 basic1_new

drwx---r-x 1 22 Oct 11 13:28 basic1

-rw----r-- 1 12 Oct 11 13:28 file-basic1

A source trailing slash means copy the contents of a directory rather than the directory itself. Importantly, however, the directory’s metadata is copied to the destination directory.

RSYNC /basic1/ /dst/basic1_renamed
drwx---r-x 1 22 Oct 11 13:28 basic1_renamed

-rw----r-- 1 12 Oct 11 13:28 file-basic1

One gotcha is that RSYNC +z is a no-op if the source is a directory:

RSYNC +z /basic1 /dst/basic1_newC

At least -r is needed with +z in this case:

RSYNC +z -r /basic1/ /dst/basic1_newD
drwx------ 1 22 Oct 11 13:28 basic1_newD

-rw------- 1 12 Oct 11 13:28 file-basic1

Multiple source directories can be specified, including with wildcards. This example also illustrates that copies files are by default merged with content already existing in the image.

RUN mkdir /dst/dstC && echo file-dstC > /dst/dstC/file-dstC
RSYNC /basic* /dst/dstC
drwxrwx--- 1 42 Oct 11 13:33 dstC

drwx---r-x 1 22 Oct 11 13:33 basic1
drwxrwx--- 1 22 Oct 11 13:33 basic2
-rw-rw---- 1 10 Oct 11 13:33 file-dstC

-rw----r-- 1 12 Oct 11 13:33 file-basic1

-rw-rw---- 1 12 Oct 11 13:33 file-basic2

Trailing slashes can be specified independently for each source:

RUN mkdir /dst/dstF && echo file-dstF > /dst/dstF/file-dstF
RSYNC /basic1 /basic2/ /dst/dstF
drwxrwx--- 1 52 Oct 11 13:33 dstF

drwx---r-x 1 22 Oct 11 13:33 basic1
-rw-rw---- 1 12 Oct 11 13:33 file-basic2
-rw-rw---- 1 10 Oct 11 13:33 file-dstF

-rw----r-- 1 12 Oct 11 13:33 file-basic1

Bare / (i.e., the entire context directory) is considered to have a trailing slash:

RSYNC / /dst
drwx---r-x 1  22 Oct 11 13:33 basic1
drwxrwx--- 1  22 Oct 11 13:33 basic2

-rw----r-- 1 12 Oct 11 13:33 file-basic1

-rw-rw---- 1 12 Oct 11 13:33 file-basic2

To replace (rather than merge with) existing content, use --delete. Note also that wildcards can be combined with trailing slashes and that the directory gets the metadata of the first slashed directory.

RUN mkdir /dst/dstG && echo file-dstG > /dst/dstG/file-dstG
RSYNC --delete /basic*/ /dst/dstG
drwx---r-x 1 44 Oct 11 14:00 dstG

-rw----r-- 1 12 Oct 11 14:00 file-basic1
-rw-rw---- 1 12 Oct 11 14:00 file-basic2

Symbolic links in the source(s) add significant complexity. Like rsync(1), RSYNC can do one of three things with a given symlink:

  1. Ignore it, silently or with a warning.

  2. Preserve it: copy as a symlink, with the same target.

  3. Dereference it: copy the target instead.

These actions are selected independently for safe symlinks and unsafe symlinks. Safe symlinks are those which point to a target within the top of transfer, which is the deepest directory in the source path with a trailing slash. For example, /foo/bar’s top-of-transfer is /foo (regardless of whether bar is a directory or file), while /foo/bar/’s top-of-transfer is /foo/bar.

For the symlink examples, the context contains two sub-directories with a variety of symlinks, as well as a sibling file and directory outside the context. All of these links are valid on the host. In this listing, the absolute path to the parent of the context directory is replaced with /....

drwxrwx--- 9 200 Oct 11 14:00 ctx
drwxrwx--- 2  60 Oct 11 14:00 dir-out
-rw-rw---- 1   9 Oct 11 14:00 file-out

drwxrwx--- 3 320 Oct 11 14:00 sym1

lrwxrwxrwx 1 13 Oct 11 14:00 dir-out_rel -> ../../dir-out
drwxrwx--- 2 60 Oct 11 14:00 dir-sym1
lrwxrwxrwx 1  8 Oct 11 14:00 dir-sym1_direct -> dir-sym1
lrwxrwxrwx 1 10 Oct 11 14:00 dir-top_rel -> ../dir-top
lrwxrwxrwx 1 47 Oct 11 14:00 file-out_abs -> /.../file-out
lrwxrwxrwx 1 14 Oct 11 14:00 file-out_rel -> ../../file-out
-rw-rw---- 1 10 Oct 11 14:00 file-sym1
lrwxrwxrwx 1 57 Oct 11 14:00 file-sym1_abs -> /.../ctx/sym1/file-sym1
lrwxrwxrwx 1  9 Oct 11 14:00 file-sym1_direct -> file-sym1
lrwxrwxrwx 1 17 Oct 11 14:00 file-sym1_upover -> ../sym1/file-sym1
lrwxrwxrwx 1 51 Oct 11 14:00 file-top_abs -> /.../ctx/file-top
lrwxrwxrwx 1 11 Oct 11 14:00 file-top_rel -> ../file-top

-rw-rw---- 1 14 Oct 11 14:00 dir-sym1.file

-rw-rw---- 1 13 Oct 11 14:00 dir-out.file

By default, safe symlinks are preserved while unsafe symlinks are silently ignored:

RSYNC /sym1 /dst
drwxrwx--- 1 206 Oct 11 17:10 sym1

drwxrwx--- 1 26 Oct 11 17:10 dir-sym1
lrwxrwxrwx 1  8 Oct 11 17:10 dir-sym1_direct -> dir-sym1
lrwxrwxrwx 1 10 Oct 11 17:10 dir-top_rel -> ../dir-top
-rw-rw---- 1 10 Oct 11 17:10 file-sym1
lrwxrwxrwx 1  9 Oct 11 17:10 file-sym1_direct -> file-sym1
lrwxrwxrwx 1 17 Oct 11 17:10 file-sym1_upover -> ../sym1/file-sym1
lrwxrwxrwx 1 17 Oct 11 17:10 file-sym2_upover -> ../sym2/file-sym2
lrwxrwxrwx 1 11 Oct 11 17:10 file-top_rel -> ../file-top

-rw-rw---- 1 14 Oct 11 17:10 dir-sym1.file

The source files have four rough fates:

  1. Regular files and directories (file-sym1 and dir-sym1). These are copied into the image unchanged, including metadata.

  2. Safe symlinks, now broken. This is one of the gotchas of RSYNC’s top-of-transfer directory (here host path ./ctx, image path /) differing from the source directory (./ctx/sym1, /sym1), because the latter lacks a trailing slash. dir-top_rel, file-sym2_upover, and file-top_rel all ascend only as high as ./ctx (host path, / image) before re-descending. This is within the top-of-transfer, so the symlinks are safe and thus copied unchanged, but their targets were not included in the copy.

  3. Safe symlinks, still valid.

    1. dir-sym1_direct and file-sym1_direct point directly to files in the same directory.

    2. dir-sym1_upover and file-sym1_upover point to files in the same directory, but by first ascending into their parent — within the top-of-transfer, so they are safe — and then re-descending. If sym1 were renamed during the copy, these links would break.

  4. Unsafe symlinks, which are ignored by the copy and do not appear in the image.

    1. Absolute symlinks are always unsafe (*_abs).

    2. dir-out_rel and file-out_rel are relative symlinks that ascend above the top-of-transfer, in this case to targets outside the context, and are thus unsafe.

The top-of-transfer can be changed to sym1 with a trailing slash. This also adds sym1 to the destination so the resulting directory structure is the same.

RSYNC /sym1/ /dst/sym1
drwxrwx--- 1 96 Oct 11 17:10 sym1

drwxrwx--- 1 26 Oct 11 17:10 dir-sym1
lrwxrwxrwx 1  8 Oct 11 17:10 dir-sym1_direct -> dir-sym1
-rw-rw---- 1 10 Oct 11 17:10 file-sym1
lrwxrwxrwx 1  9 Oct 11 17:10 file-sym1_direct -> file-sym1

-rw-rw---- 1 14 Oct 11 17:10 dir-sym1.file

*_upover and *-out_rel are now unsafe and replaced with their targets.

Another common use case is to follow unsafe symlinks and copy their targets in place of the links. This is accomplished with +u:

RSYNC +u /sym1/ /dst/sym1
drwxrwx--- 1 352 Oct 11 17:10 sym1

drwxrwx--- 1 24 Oct 11 17:10 dir-out_rel
drwxrwx--- 1 26 Oct 11 17:10 dir-sym1
lrwxrwxrwx 1  8 Oct 11 17:10 dir-sym1_direct -> dir-sym1
drwxrwx--- 1 24 Oct 11 17:10 dir-top_rel
-rw-rw---- 1  9 Oct 11 17:10 file-out_abs
-rw-rw---- 1  9 Oct 11 17:10 file-out_rel
-rw-rw---- 1 10 Oct 11 17:10 file-sym1
-rw-rw---- 1 10 Oct 11 17:10 file-sym1_abs
lrwxrwxrwx 1  9 Oct 11 17:10 file-sym1_direct -> file-sym1
-rw-rw---- 1 10 Oct 11 17:10 file-sym1_upover
-rw-rw---- 1 10 Oct 11 17:10 file-sym2_abs
-rw-rw---- 1 10 Oct 11 17:10 file-sym2_upover
-rw-rw---- 1  9 Oct 11 17:10 file-top_abs
-rw-rw---- 1  9 Oct 11 17:10 file-top_rel

-rw-rw---- 1 13 Oct 11 17:10 dir-out.file

-rw-rw---- 1 14 Oct 11 17:10 dir-sym1.file

-rw-rw---- 1 13 Oct 11 17:10 dir-top.file

Now all the unsafe symlinks noted above are present in the image, but they have changed to the normal files and directories pointed to.


This feature lets you copy files outside the context into the image, unlike other container builders where COPY can never access anything outside the context.

The sources themselves, if symlinks, do not get special treatment:

RSYNC /sym1/file-sym1_direct /sym1/file-sym1_upover /dst
lrwxrwxrwx 1 9 Oct 11 17:10 file-sym1_direct -> file-sym1

Note that file-sym1_upover does not appear in the image, despite being named explicitly in the instruction, because it is an unsafe symlink.

If the destination is a symlink to a file, and the source is a file, the link is replaced and the target is unchanged. (If the source is a directory, that is an error.)

RUN touch /dst/file-dst && ln -s file-dst /dst/file-dst_direct
RSYNC /file-top /dst/file-dst_direct
-rw-rw---- 1 0 Oct 11 17:42 file-dst
-rw-rw---- 1 9 Oct 11 17:42 file-dst_direct

If the destination is a symlink to a directory, the link is followed:

RUN mkdir /dst/dir-dst && ln -s dir-dst /dst/dir-dst_direct
RSYNC /file-top /dst/dir-dst_direct
drwxrwx--- 1 16 Oct 11 17:50 dir-dst
lrwxrwxrwx 1  7 Oct 11 17:50 dir-dst_direct -> dir-dst

-rw-rw---- 1 9 Oct 11 17:50 file-top

7.7.6. Examples

Build image bar using ./foo/bar/Dockerfile and context directory ./foo/bar:

$ ch-image build -t bar -f ./foo/bar/Dockerfile ./foo/bar
grown in 4 instructions: bar

Same, but infer the image name and Dockerfile from the context directory path:

$ ch-image build ./foo/bar
grown in 4 instructions: bar

Build using humongous vendor compilers you want to bind-mount instead of installing into the image:

$ ch-image build --bind /opt/bigvendor:/opt .
$ cat Dockerfile
FROM centos:7

RUN /opt/bin/cc hello.c
#COPY /opt/lib/*.so /usr/local/lib   # fail: COPY doesn’t bind mount
RUN cp /opt/lib/*.so /usr/local/lib  # possible workaround
RUN ldconfig

7.8. build-cache

$ ch-image [...] build-cache [...]

Print basic information about the cache. If -v is given, also print some Git statistics and the Git repository configuration.

If any of the following options are given, do the corresponding operation before printing. Multiple options can be given, in which case they happen in this order.


Create a DOT export of the tree named ./build-cache.dot and a PDF rendering ./build-cache.pdf. Requires graphviz and git2dot.


Run Git garbage collection on the cache, including full de-duplication of similar files. This will immediately remove all cache entries not currently reachable from a named branch (which is likely to cause corruption if the build cache is being accessed concurrently by another process). The operation can take a long time on large caches.


Clear and re-initialize the build cache.


Print a text tree of the cache using Git’s git log --graph feature. If -v is also given, the tree has more detail.

7.9. delete

$ ch-image [...] delete IMAGE_GLOB [IMAGE_GLOB ... ]

Delete the image(s) described by each IMAGE_GLOB from the storage directory (including all build stages).

IMAGE_GLOB can be either a plain image reference or an image reference with glob characters to match multiple images. For example, ch-image delete 'foo*' will delete all images whose names start with foo. Multiple images and/or globs can also be given in a single command line.

Importantly, this sub-command does not also remove the image from the build cache. Therefore, it can be used to reduce the size of the storage directory, trading off the time needed to retrieve an image from cache.


Glob characters must be quoted or otherwise protected from the shell, which also desires to interpret them and will do so incorrectly.

7.10. gestalt

$ ch-image [...] gestalt [SELECTOR]

Provide information about the configuration and available features of ch-image. End users generally will not need this; it is intended for testing and debugging.

SELECTOR is one of:

  • bucache. Exit successfully if the build cache is available, unsuccessfully with an error message otherwise. With -v, also print version information about dependencies.

  • bucache-dot. Exit successfully if build cache DOT trees can be written, unsuccessfully with an error message otherwise. With -v, also print version information about dependencies.

  • python-path. Print the path to the Python interpreter in use and exit successfully.

  • storage-path. Print the storage directory path and exit successfully.

7.11. list

Print information about images. If no argument given, list the images in builder storage.

7.11.1. Synopsis

$ ch-image [...] list [-l] [IMAGE_REF]

7.11.2. Description

Optional argument:

-l, --long

Use long format (name, last change timestamp) when listing images.

-u, --undeletable

List images that can be undeleted. Can also be spelled --undeleteable.


Print details of what’s known about IMAGE_REF, both locally and in the remote registry, if any.

7.11.3. Examples

List images in builder storage:

$ ch-image list
alpine:3.17 (amd64)
alpine:latest (amd64)
debian:buster (amd64)

Print details about Debian Buster image:

$ ch-image list debian:buster
details of image:    debian:buster
in local storage:    no
full remote ref:     registry-1.docker.io:443/library/debian:buster
available remotely:  yes
remote arch-aware:   yes
host architecture:   amd64
archs available:     386       bae2738ed83
                     amd64     98285d32477
                     arm/v7    97247fd4822
                     arm64/v8  122a0342878

For remotely available images like Debian Buster, the associated digest is listed beside each available architecture. Importantly, this feature does not provide the hash of the local image, which is only calculated on push.

7.12. import

$ ch-image [...] import PATH IMAGE_REF

Copy the image at PATH into builder storage with name IMAGE_REF. PATH can be:

  • an image directory

  • a tarball with no top-level directory (a.k.a. a “tarbomb”)

  • a standard tarball with one top-level directory

If the imported image contains Charliecloud metadata, that will be imported unchanged, i.e., images exported from ch-image builder storage will be functionally identical when re-imported.


Descendant images (i.e., FROM the imported IMAGE_REF) are linked using IMAGE_REF only. If a new image is imported under a new IMAGE_REF, all instructions descending from that IMAGE_REF will still hit, even if the new image is different.

7.13. pull

Pull the image described by the image reference IMAGE_REF from a repository to the local filesystem.

7.13.1. Synopsis

$ ch-image [...] pull [...] IMAGE_REF [DEST_REF]

See the FAQ for the gory details on specifying image references.

7.13.2. Description



If specified, use this as the destination image reference, rather than IMAGE_REF. This lets you pull an image with a complicated reference while storing it locally with a simpler one.


--last-layer N

Unpack only N layers, leaving an incomplete image. This option is intended for debugging.


Parse IMAGE_REF, print a parse report, and exit successfully without talking to the internet or touching the storage directory.

This script does a fair amount of validation and fixing of the layer tarballs before flattening in order to support unprivileged use despite image problems we frequently see in the wild. For example, device files are ignored, and file and directory permissions are increased to a minimum of rwx------ and rw------- respectively. Note, however, that symlinks pointing outside the image are permitted, because they are not resolved until runtime within a container.

The following metadata in the pulled image is retained; all other metadata is currently ignored. (If you have a need for additional metadata, please let us know!)

  • Current working directory set with WORKDIR is effective in downstream Dockerfiles.

  • Environment variables set with ENV are effective in downstream Dockerfiles and also written to /ch/environment for use in ch-run --set-env.

  • Mount point directories specified with VOLUME are created in the image if they don’t exist, but no other action is taken.

Note that some images (e.g., those with a “version 1 manifest”) do not contain metadata. A warning is printed in this case.

7.13.3. Examples

Download the Debian Buster image matching the host’s architecture and place it in the storage directory:

$ uname -m
pulling image:    debian:buster
requesting arch:  arm64/v8
manifest list: downloading
manifest: downloading
config: downloading
layer 1/1: c54d940: downloading
flattening image
layer 1/1: c54d940: listing
validating tarball members
resolving whiteouts
layer 1/1: c54d940: extracting
image arch: arm64

Same, specifying the architecture explicitly:

$ ch-image --arch=arm/v7 pull debian:buster
pulling image:    debian:buster
requesting arch:  arm/v7
manifest list: downloading
manifest: downloading
config: downloading
layer 1/1: 8947560: downloading
flattening image
layer 1/1: 8947560: listing
validating tarball members
resolving whiteouts
layer 1/1: 8947560: extracting
image arch: arm (may not match host arm64/v8)

7.14. push

Push the image described by the image reference IMAGE_REF from the local filesystem to a repository.

7.14.1. Synopsis

$ ch-image [...] push [--image DIR] IMAGE_REF [DEST_REF]

See the FAQ for the gory details on specifying image references.

7.14.2. Description



If specified, use this as the destination image reference, rather than IMAGE_REF. This lets you push to a repository without permanently adding a tag to the image.


--image DIR

Use the unpacked image located at DIR rather than an image in the storage directory named IMAGE_REF.

Because Charliecloud is fully unprivileged, the owner and group of files in its images are not meaningful in the broader ecosystem. Thus, when pushed, everything in the image is flattened to user:group root:root. Also, setuid/setgid bits are removed, to avoid surprises if the image is pulled by a privileged container implementation.

7.14.3. Examples

Push a local image to the registry example.com:5000 at path /foo/bar with tag latest. Note that in this form, the local image must be named to match that remote reference.

$ ch-image push example.com:5000/foo/bar:latest
pushing image:   example.com:5000/foo/bar:latest
layer 1/1: gathering
layer 1/1: preparing
preparing metadata
starting upload
layer 1/1: a1664c4: checking if already in repository
layer 1/1: a1664c4: not present, uploading
config: 89315a2: checking if already in repository
config: 89315a2: not present, uploading
manifest: uploading
cleaning up

Same, except use local image alpine:3.17. In this form, the local image name does not have to match the destination reference.

$ ch-image push alpine:3.17 example.com:5000/foo/bar:latest
pushing image:   alpine:3.17
destination:     example.com:5000/foo/bar:latest
layer 1/1: gathering
layer 1/1: preparing
preparing metadata
starting upload
layer 1/1: a1664c4: checking if already in repository
layer 1/1: a1664c4: not present, uploading
config: 89315a2: checking if already in repository
config: 89315a2: not present, uploading
manifest: uploading
cleaning up

Same, except use unpacked image located at /var/tmp/image rather than an image in ch-image storage. (Also, the sole layer is already present in the remote registry, so we don’t upload it again.)

$ ch-image push --image /var/tmp/image example.com:5000/foo/bar:latest
pushing image:   example.com:5000/foo/bar:latest
image path:      /var/tmp/image
layer 1/1: gathering
layer 1/1: preparing
preparing metadata
starting upload
layer 1/1: 892e38d: checking if already in repository
layer 1/1: 892e38d: already present
config: 546f447: checking if already in repository
config: 546f447: not present, uploading
manifest: uploading
cleaning up

7.15. reset

$ ch-image [...] reset

Delete all images and cache from ch-image builder storage.

7.16. undelete

$ ch-image [...] undelete IMAGE_REF

If IMAGE_REF has been deleted but is in the build cache, recover it from the cache. Only available when the cache is enabled, and will not overwrite IMAGE_REF if it exists.

7.17. Environment variables


Username and password for registry authentication. See important caveats in section “Authentication” above.


If set, append log chatter to this file, rather than standard error. This is useful for debugging situations where standard error is consumed or lost.

Also sets verbose mode if not already set (equivalent to --verbose).


If set, prepend PID and timestamp to logged chatter.


If set, save xattrs in the build cache and restore them when rebuilding from the cache (equivalent to --xattrs).