On Fri, Mar 20, 2020 at 04:30:07PM +0100, Pavel Hrdina wrote: > On Fri, Mar 20, 2020 at 01:40:10PM +0000, Daniel P. Berrangé wrote: > > This simple series allows apps to choose how cgroups are managed by > > libvirt, between cgroupfs and machined, or disabled entirely. > > I'm not so sure about this series. The situation with cgroups and > systemd is a bit more complex then the current code handles. > > There is an existing issue where we are violating the delegation rules > described by cgroups and systemd. > > Currently the "cgroupfs" approach is used only on non-systemd hosts and > we should keep it that way. Libvirt is not allowed to mangle with > cgroups owned by systemd so IMHO the warning in configuration file is > not enough because without delegation cgroups will not work properly in > libvirt. That isn't the case currently AFAICT from current code. Before this series, the virCgroupNewMachine method will first try systemd and then fallback to directly cgroupfs. This fallback can happen on a systemd host, when machined is not installed, as machined is an optional component. If we need to mandate use of systemd on systemd hosts, then our existing code is broken and needs fixing. I'm happy todo such a fix, and then adjust this series to take account of it. Essentially we'd allow apps to specify 'cgroupfs' or 'machined' but we'd enforce they only make safe choices. ie on a systemd host we'd only allow 'none' or 'machined' On a non-systemd host, or on a systemd host where we've been delegated a subtree, we'd only allow 'none' or 'cgroupfs' > One example is that without delegation the VM cgroups would not have any > controllers enabled by default. > > Having an option to completely disable cgroups is OK but that's the only > thing I would expose for now. > > When using machined the current topology looks like this (only the files > that we use are listed): > > /sys/fs/cgroup/machine.slice/ > └── machine-qemu\x2d1\x2dcentos8.scope > ├── emulator > │ ├── cgroup.controllers > │ ├── cgroup.threads > │ ├── cgroup.type > │ ├── cpuset.cpus > │ └── cpuset.mems > ├── vcpu0 > │ ├── cgroup.controllers > │ ├── cgroup.threads > │ ├── cgroup.type > │ ├── cpuset.cpus > │ └── cpuset.mems > ├── vcpu1 > │ ├── cgroup.controllers > │ ├── cgroup.threads > │ ├── cgroup.type > │ ├── cpuset.cpus > │ └── cpuset.mems > ├── vcpu2 > │ ├── cgroup.controllers > │ ├── cgroup.threads > │ ├── cgroup.type > │ ├── cpuset.cpus > │ └── cpuset.mems > ├── vcpu3 > │ ├── cgroup.controllers > │ ├── cgroup.threads > │ ├── cgroup.type > │ ├── cpuset.cpus > │ └── cpuset.mems > ├── cgroup.controllers > ├── cgroup.procs > ├── cgroup.subtree_control > ├── cgroup.type > ├── cpu.max > ├── cpu.stat > ├── cpu.weight > ├── io.bfq.weight > ├── io.max > ├── io.stat > ├── io.weight > ├── memory.high > ├── memory.max > ├── memory.stat > └── memory.swap.max > > which is incorrect. The group "machine-qemu\x2d1\x2dcentos8.scope" is > marked as delegated which means we can create new sub-cgroups there and > that we can access only these files directly in that group: > > cgroup.subtree_control - to enable cgroup controllers for our > sub-cgroups > > cgroup.procs - to move processes around into our sub-cgroups > > Other files are off limit and we should not touch them. In addition > systemd may create its own sub-cgroups and it would fail to do so > because there is another limitation from kernel that processes can live > only in the leaves except for the root cgroup. > > Exactly that is happening right now and it was discovered by user > reporting a BUG that systemctl daemon-reloaed changed the values in > cpu.shares because it owns that file and it was our fault to change it. > > Currently we use these cgroups files: > > cgroup.type > > - to enable threaded mode that is required for vcpu, emulator > and iothread pinning. We should not do that in the delegated > cgroup because we may break systemd, this has to be moved to > a sub-cgroup together with the pinning. > > io.weight / io.bfq.weight, cpu.weight > > - These cannot be moved into the sub-cgroup because it would > lose the effect and would be completely pointless to set it. > All the *.weight files works in a way that the resources > available to parent are distributed using the weight to the > children. If we would move it to the sub-cgroup there would > be only single child all the time. Therefore we have to use > D-Bus to talk to systemd to have these values configured > directly for the delegated cgroup. > > io.max, memory.max, memory.high, memory.swap.max, cpu.max > > - All of these can be safely set in the sub-cgroup that we will > have to create as they are absolute limits and they are not > affected by siblings. > > cpuset.mems, cpuset.cpus > > - These are used together with the threaded model and can also > be safely set in the sub-cgroup. > > Based on all of the above this is the new topology that has to be > created by libvirt if systemd is present: > > /sys/fs/cgroup/machine.slice/ > └── machine-qemu\x2d1\x2dcentos8.scope > ├── libvirt-vm (or something else) So IIUC, we can allow 'cgroupfs' on a systemd host, provided that the path for the <partition> points to a sub-tree that is delegated. > │ ├── emulator > │ │ ├── cgroup.controllers > │ │ ├── cgroup.threads > │ │ ├── cgroup.type > │ │ ├── cpuset.cpus > │ │ └── cpuset.mems > │ ├── vcpu0 > │ │ ├── cgroup.controllers > │ │ ├── cgroup.threads > │ │ ├── cgroup.type > │ │ ├── cpuset.cpus > │ │ └── cpuset.mems > │ ├── vcpu1 > │ │ ├── cgroup.controllers > │ │ ├── cgroup.threads > │ │ ├── cgroup.type > │ │ ├── cpuset.cpus > │ │ └── cpuset.mems > │ ├── vcpu2 > │ │ ├── cgroup.controllers > │ │ ├── cgroup.threads > │ │ ├── cgroup.type > │ │ ├── cpuset.cpus > │ │ └── cpuset.mems > │ ├── vcpu3 > │ │ ├── cgroup.controllers > │ │ ├── cgroup.threads > │ │ ├── cgroup.type > │ │ ├── cpuset.cpus > │ │ └── cpuset.mems > │ ├── cgroup.controllers > │ ├── cgroup.procs > │ ├── cgroup.subtree_control > │ ├── cgroup.type > │ ├── cpu.max > │ ├── cpu.stat > │ ├── io.max > │ ├── io.stat > │ ├── memory.high > │ ├── memory.max > │ ├── memory.stat > │ └── memory.swap.max > ├── cgroup.procs > ├── cgroup.subtree_control > ├── cpu.weight (only via systemd using D-Bus) > ├── io.bfq.weight (only via systemd using D-Bus) > └── io.weight (only via systemd using D-Bus) > > Pavel Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
