Whenever possible it is best to install standard cvmfs from native OS packages, but sometimes that is not an option. This package is for mounting cvmfs as an unprivileged user, without the cvmfs package being installed by a system administrator. The package can do this in 4 different ways:
- On systems where only fusermount is available, the
mountrepoandumountrepocommands can be used to mount cvmfs repositories in the user's own file space. That path can then be bindmounted at /cvmfs by a container manager such as singularity. - On systems where fusermount is available and unprivileged user
namespaces are enabled, but unprivileged namespace fuse mounts are not
available (in particular RHEL <=7.7 with
sysctl user.max_user_namespaces> 0), thecvmfsexeccommand can mount cvmfs repositories, map them into /cvmfs, and unmount them when it exits. singularity may even be run unprivileged from cvmfs from within cvmfsexec (it has to run unprivileged because setuid-root does not work inside a user namespace). The main disadvantage is that if the processes are hard-killed (kill -9), mountpoints are left behind and are difficult to clean up. - On systems where unprivileged namespace fuse mounts are available
(newer kernels >= 4.18 as on RHEL8 or >= 3.10.0-1127 as on RHEL 7.8),
the
cvmfsexeccommand can entirely manage mounting and unmounting of cvmfs repositories in the namespace, so if they get killed everything gets cleanly unmounted. fusermount is not needed in this case. - On systems that have no fusermount nor unprivileged user namespace
fuse mounts but do have a setuid installation of singularity >= 3.4,
an entirely separate command in this package
singcvmfscan mount cvmfs repositories inside a container using thesingularity --fusemountfeature. With singularity >= 3.6 and RHEL >= 7.8 and unprivileged user namespaces enabled this can also be used with unprivileged singularity.
Operating systems currently supported by this package are Red Hat Enterprise Linux (versions 7, 8, and 9) and its derivatives (CentOS, Scientific Linux, Rocky Linux, Alma Linux) and SUSE Linux Enterprise (version 15) and its derivatives (openSUSE Leap). All of those support the x86_64 architecture, and RHEL8 also supports ppc64le and aarch64.
All of the ways this package supports unprivileged cvmfs make use of a
copy of the cvmfs software. The cvmfs software and configuration are
expected to be in a dist subdirectory under where the scripts are. The
easiest way to create the dist directory is to use makedist. It takes a
parameter of osg, egi, or default to download the latest cvmfs and
configuration rpm from one of those three sources. Note: egi does not
currently provide rpms for RHEL8 or 9. Additionally, specifying none
will download from the default source but exclude the
cvmfs-config-default package.
By default a distribution for cvmfsexec and mountrepo/umountrepo is
created. To instead make a distribution for singcvmfs, add the -s
makedist option. By default the distribution made will match the host
operating system it runs on, but another distribution can be selected
with the -m option. See the makedist usage for supported machine types.
To customize any cvmfs configuration settings, put them in
dist/etc/cvmfs/default.local. In particular you may want to set
CVMFS_HTTP_PROXY, although the default is to use WLCG Web Proxy Auto
Discovery. You may also want to set CVMFS_QUOTA_LIMIT, otherwise the
default is 4000 MB. The default CVMFS_CACHE_BASE for the cache
shared between the cvmfs repository is under the dist directory,
dist/var/lib/cvmfs, unless the -m option is used to add an e2fs
filesystem (details below).
Make sure that the cache does not get shared between multiple machines.
For the cases where cvmfsexec or singcvmfs can be used, you can also
make a self-contained distribution in a single file that contains both
the command and all supporting files. This makes it easy to share the
distribution with other users or distribute it to many machines.
After running makedist and making any customizations you want, use this to make a cvmfsexec distribution:
makedist -o /tmp/cvmfsexec
or this to make a singcvmfs distribution:
makedist -s -o /tmp/singcvmfs
Executing a cvmfsexec file that is created in that way leaves behind a .cvmfsexec directory in the directory where it is run from, and running a singcvmfs file leaves behind a .singcvmfs directory.
The cvmfsexec command requires unprivileged user namespaces. On RHEL8/9 unprivileged user namepaces (and user namespace fuse mounts) are available by default, but on RHEL7 they need to be enabled by setting a sysctl parameter as detailed in the OSG unprivileged singularity instructions. In addition cvmfsexec requires fusermount on kernels older than those that come with RHEL7.8.
To execute a command in an environment where cvmfs repositories are
mounted at /cvmfs and automatically unmounted upon exit, use
cvmfsexec repository.name ... -- [command] where the default command
is $SHELL. It will automatically mount the configuration repository if
one is defined.
Unless disabled with the -N option,
inside the command you can mount additional repositories by using
$CVMFSMOUNT repository.name. Since the mounts have to happen outside
the user namespace, it actually sends a message to the original process
to mount, and makes the current process wait until completion.
Repositories that are already mounted are ignored. You can also unmount
repositories from within the command with $CVMFSUMOUNT repository.name.
If you want to use this feature and also
invoke additional processes within the original process that are
not trustworthy, such as user payloads that are invoked with
singularity --contain, then close the $CVMFSEXEC_CMDFD file descriptor
for those processes. This can be done in bash with
exec {CVMFSEXEC_CMDFD}>&-.
Note that setuid-root programs do not work inside an unprivileged user namepace, so if you use singularity it has to be run unprivileged.
Cache considerations: by default cvmfsexec starts a cache manager
process for all the cvmfs repositories it mounts, which means only one
cvmfsexec process can share a cache on a single machine. The cvmfs
configuration could be set to use a different path for a cache for
different invocations (the default is dist/var/lib/cvmfs), or it could
be set to use cvmfs alien cache mode which doesn't use a cache manager,
but the best approach is to start cvmfsexec from a pilot process and run
only one pilot per machine. If possible the cache should be on local
disk, because otherwise the many file accesses can overwhelm a shared
filesystem's metadata server. If there is no local disk the next best
option is to mount a filesystem separately for each worker
node from a big enough file on the shared filesystem, or alternatively
a RAM disk on the local node if there is sufficient RAM.
The -m option (described in the next section) can mount that separate
scratch filesystem for you in the cvmfsexec namespace.
Otherwise, if the cache directory needs to be changed that can be done
by setting CVMFS_CACHE_BASE in dist/etc/cvmfs/default.local.
If there is no local disk available, cvmfsexec can mount a scratch
ext2/3/4 filesystem with the -m option, using the fuse2fs command.
The filesystem will appear in the cvmfsexec namespace at /e2fs.
If you have not set CVMFS_CACHE_BASE in dist/etc/cvmfs/default.local
then this filesystem will automatically be used for cvmfs cache, and
it can also be used as scratch workspace for jobs.
Make sure that there is a unique file for each running copy of cvmfsexec. Create the file with commands like this:
truncate -s 600G scratch.img
mkfs.ext4 -F -O ^has_journal scratch.img
Choose a count of the number of gigabytes you want. The default cache
size is 4000 megabytes, and the recommendation is to reserve an
additional 1000 megabytes plus 20%, so make it at least 6 gigabytes
and add any additional space you want to use as scratch workspace.
It's a sparse file so there's little penalty for creating it bigger if
the space is never used.
Then start cvmfsexec with -m. For example, to mount only the
cvmfs configuration repository and run a shell do
cvmfsexec -m scratch.img --
Then check out /e2fs.
A caveat on older kernels (for example RHEL7.7 and older) is that a
kill -9 of all the processes will not clean up the mounts, and they
have to be separately unmounted later with umountrepo or fusermount -u.
On kernels >= 4.18 (for example RHEL8) or >= 3.10.0-1127 (for example on
RHEL7.8) the operation changes to do fuse mounts only inside of
unprivileged user namespaces, which always completely cleans up mounts
even with kill -9. This also normally uses a pid namespace to ensure
that all fuse processes are always cleaned up when the command exits
(the exception is when running under docker and kubernetes default
configurations that "mask" /proc).
$CVMFSMOUNT/$CVMFSUMOUNT still send a request to a parent process to mount/umount but it's not the original process, it's an intermediate process that has fakeroot access in the user namespace.
When not using cvmfsexec, but with fusermount available use
mountrepo repository.name to mount a repository. Note that the osg
configuration requires "config-osg.opensciencegrid.org" to be mounted
first, and the egi configuration requires "config-egi.egi.eu".
If you are using a container system, bind mount $PWD/dist/cvmfs into the container as /cvmfs.
To unmount all repositories, use umountrepo -a, or to unmount an
individual repository use umountrepo repository.name. Make sure that
all the processes do not get killed or the repositories will remain
mounted but inaccessible.
Cache considerations for this mode are the same as with the cvmfsexec command.
Syslog messages from cvmfs go in the log subdirectory alongside
dist. A separate log file is created for each repository. cvmfs
debugging logs can also be enabled in the usual way, by setting
CVMFS_DEBUGLOG in the cvmfs configuration.
Docker supports unprivileged user namespaces including unprivileged fuse
mounts on the kernels that support it, without using the --privileged
option or adding capabilities. The following set of docker options is
sufficient:
--security-opt seccomp=unconfined --security-opt systempaths=unconfined --device=/dev/fuse
If you have no need for running any setuid executables in docker then you can improve security further by adding:
--security-opt no-new-privileges
Singularity always has the equivalent protection enabled for the containers it runs.
When a privileged setuid installation of singularity >= 3.4 is
available, the singcvmfs command can be used to mount cvmfs
repositories inside a container. With singularity >= 3.6 and
RHEL >= 7.8 or a kernel >= 4.18
with unprivileged user namespaces enabled
this can also be used with an
unprivileged non-setuid singularity installation.
The command line interface is different than cvmfsexec because it is
designed for ease of use by end users on a laptop/desktop and as a
drop-in replacement for singularity when it executes containers.
Put cvmfs repositories to mount comma-separated in a
SINGCVMFS_REPOSITORIES environment variable. If a configuration
repository is needed it will be automatically mounted. Then you can use
singcvmfs exactly like singularity with one of its exec, instance, run,
or shell commands (note: it cannot read an image from cvmfs). For example,
once you have made a singcvmfs distribution
the following should work (replace centos:7 with a container matching
the host's operating system):
$ export SINGCVMFS_REPOSITORIES="grid.cern.ch,atlas.cern.ch"
$ singcvmfs -s exec -cip docker://centos:7 bash
Singularity> ls /cvmfs
atlas.cern.ch config-osg.opensciencegrid.org grid.cern.ch
Singularity> ls /cvmfs/atlas.cern.ch
repo
Singularity> exit
# or using singularity instances:
$ export SINGCVMFS_REPOSITORIES="grid.cern.ch,atlas.cern.ch"
$ singcvmfs -s instance start docker://centos:7 myexampleinstance
$ singcvmfs -s run instance://myexampleinstance ls /cvmfs
atlas.cern.ch config-osg.opensciencegrid.org grid.cern.ch
$ singcvmfs -s run instance://myexampleinstance ls /cvmfs/atlas.cern.ch
repo
$ singcvmfs -s instance stop myexampleinstance
The first time you run the above it will take a long time as singularity downloads the image from dockerhub and cvmfs fills its cache, but running it again should be very fast.
Alternatively, to make it easier to execute repeatedly interactively
from the command line, you can put the singularity container path in a
SINGCVMFS_IMAGE environment variable and leave out the singularity
command. The image cannot come from cvmfs, but it can come from docker,
shub, a local image file, or a local "sandbox" unpacked image directory.
Then the usage is singcvmfs [command] where the default command is
$SHELL. For example:
$ export SINGCVMFS_REPOSITORIES="grid.cern.ch,atlas.cern.ch"
$ export SINGCVMFS_IMAGE="docker://centos:7"
$ singcvmfs ls /cvmfs
atlas.cern.ch config-osg.opensciencegrid.org grid.cern.ch
$ singcvmfs ls /cvmfs/atlas.cern.ch
repo
There are other optional environment variables that may be set.
Run singcvmfs -h for more details.
Caveat: singcvmfs works by bind-mounting all of the files from the cvmfs distribution into the container, including the fuse3 libraries. It expects other base system libraries to be available inside the container, so if the container uses a different base OS distribution it will likely not work unless compatible libraries are in the container.
Cache considerations: by default singcvmfs starts a cache manager for all the cvmfs repositories it mounts, which means that caches cannot be shared between different invocations of singcvmfs on the same machine. This tends to be more of a problem than with the cvmfsexec command because it is more common to run many payload jobs on a machine with singularity than it is to run many pilots. You can select a different cache directory for each invocation by setting SINGCVMFS_CACHEDIR. Alternatively it's possible to use the cvmfs alien cache feature to share the cache, but you would have to make sure that the cache doesn't grow too big and that it gets cleaned up at some point. Make sure that caches are not on shared filesystems because they're likely to do too many metadata operations.
There is also a SINGCVMFS_CACHEIMAGE variable which can be set to an
ext3 filesystem image for the cache. If set, it must point to a
file with such a filesystem including a directory that is writable by
the user or a shared directory within it that is writable by the user.
The directory within the filesystem can be changed by setting
SINGCVMFS_CACHEDIR but defaults to the root directory, /.
If running as an unprivileged user, this can directory can be created
using the mkfs.ext3 -d option. Unfortunately the version of
mkfs.ext3 on RHEL7 is too old, but it can be compiled from
source.
Then to make the image file these commands should work:
$ truncate -s 6G scratch.img
$ mkdir -p tmp/shared
$ mkfs.ext3 -F -O ^has_journal -d tmp scratch.img