LXCFS is a small FUSE filesystem written with the intention of making Linux
containers feel more like a virtual machine. It started as a side-project of
LXC but is useable by any runtime.
LXCFS will take care that the information provided by crucial files in procfs
such as:
/proc/cpuinfo
/proc/diskstats
/proc/meminfo
/proc/stat
/proc/swaps
/proc/uptime
/proc/slabinfo
/sys/devices/system/cpu
/sys/devices/system/cpu/online
are container aware such that the values displayed (e.g. in /proc/uptime)
really reflect how long the container is running and not how long the host is
running.
Prior to the implementation of cgroup namespaces by Serge Hallyn LXCFS also
provided a container aware cgroupfs tree. It took care that the container
only had access to cgroups underneath it's own cgroups and thus provided
additional safety. For systems without support for cgroup namespaces LXCFS
will still provide this feature but it is mostly considered deprecated.
LXCFS is split into a shared library (a libtool module, to be precise)
liblxcfs and a simple binary lxcfs. When upgrading to a newer version of
LXCFS the lxcfs binary will not be restarted. Instead it will detect that
a new version of the shared library is available and will reload it using
dlclose(3) and dlopen(3). This design was chosen so that the fuse main loop
that LXCFS uses will not need to be restarted. If it were then all containers
using LXCFS would need to be restarted since they would otherwise be left
with broken fuse mounts.
To force a reload of the shared library at the next possible instance simply
send SIGUSR1 to the pid of the running LXCFS process. This can be as simple
as doing:
rm /usr/lib64/lxcfs/liblxcfs.so # MUST to delete the old library file first
cp liblxcfs.so /usr/lib64/lxcfs/liblxcfs.so # to place new library file
kill -s USR1 $(pidof lxcfs) # reload
To achieve smooth upgrades through shared library reloads LXCFS also relies
on the fact that when dlclose(3) drops the last reference to the shared
library destructors are run and when dlopen(3) is called constructors are
run. While this is true for glibc it is not true for musl (See the section
Unloading libraries.).
So users of LXCFS on musl are advised to restart LXCFS completely and all
containers making use of it.
In order to build LXCFS install fuse and the fuse development headers according
to your distro. LXCFS prefers fuse3 but does work with new enough fuse2
versions:
git clone git://github.com/lxc/lxcfs
cd lxcfs
meson setup -Dinit-script=systemd --prefix=/usr build/
meson compile -C build/
sudo meson install -C build/
The recommended command to run lxcfs is:
sudo mkdir -p /var/lib/lxcfs
sudo lxcfs /var/lib/lxcfs
A container runtime wishing to use LXCFS should then bind mount the
approriate files into the correct places on container startup.
In order to use lxcfs with systemd-based containers, you can either use
LXC 1.1 in which case it should work automatically, or otherwise, copy
the lxc.mount.hook and lxc.reboot.hook files (once built) from this tree to
/usr/share/lxcfs, make sure it is executable, then add the
following lines to your container configuration:
lxc.mount.auto = cgroup:mixed
lxc.autodev = 1
lxc.kmsg = 0
lxc.include = /usr/share/lxc/config/common.conf.d/00-lxcfs.conf
docker run -it -m 256m --memory-swap 256m \
-v /var/lib/lxcfs/proc/cpuinfo:/proc/cpuinfo:rw \
-v /var/lib/lxcfs/proc/diskstats:/proc/diskstats:rw \
-v /var/lib/lxcfs/proc/meminfo:/proc/meminfo:rw \
-v /var/lib/lxcfs/proc/stat:/proc/stat:rw \
-v /var/lib/lxcfs/proc/swaps:/proc/swaps:rw \
-v /var/lib/lxcfs/proc/uptime:/proc/uptime:rw \
-v /var/lib/lxcfs/proc/slabinfo:/proc/slabinfo:rw \
ubuntu:18.04 /bin/bash
In a system with swap enabled, the parameter "-u" can be used to set all values in "meminfo" that refer to the swap to 0.
sudo lxcfs -u /var/lib/lxcfs
If you noticed LXCFS not showing any SWAP in your container despite having SWAP on your system, please read this section carefully and look for instructions on how to enable SWAP accounting for your distribution.
Swap cgroup handling on Linux is very confusing and there just isn't a perfect way for LXCFS to handle it.
Terminology used below:
- RAM refers to
memory.usage_in_bytesandmemory.limit_in_bytes - RAM+SWAP refers to
memory.memsw.usage_in_bytesandmemory.memsw.limit_in_bytes
The main issues are:
-
SWAP accounting is often opt-in and, requiring a special kernel boot time option (
swapaccount=1) and/or special kernel build options (CONFIG_MEMCG_SWAP). -
Both a RAM limit and a RAM+SWAP limit can be set. The delta however isn't the available SWAP space as the kernel is still free to SWAP as much of the RAM as it feels like. This makes it impossible to render a SWAP device size as using the delta between RAM and RAM+SWAP for that wouldn't account for the kernel swapping more pages, leading to swap usage exceeding swap total.
-
It's impossible to disable SWAP in a given container. The closest that can be done is setting swappiness down to 0 which severly limits the risk of swapping pages but doesn't eliminate it.
As a result, LXCFS had to make some compromise which go as follow:
-
When SWAP accounting isn't enabled, no SWAP space is reported at all. This is simply because there is no way to know the SWAP consumption. The container may very much be using some SWAP though, there's just no way to know how much of it and showing a SWAP device would require some kind of SWAP usage to be reported. Showing the host value would be completely wrong, showing a 0 value would be equallty wrong.
-
Because SWAP usage for a given container can exceed the delta between RAM and RAM+SWAP, the SWAP size is always reported to be the smaller of the RAM+SWAP limit or the host SWAP device itself. This ensures that at no point SWAP usage will be allowed to exceed the SWAP size.
-
If the swappiness is set to 0 and there is no SWAP usage, no SWAP is reported. However if there is SWAP usage, then a SWAP device of the size of the usage (100% full) is reported. This provides adequate reporting of the memory consumption while preventing applications from assuming more SWAP is available.