Limit what a jail can use
Memory, processes, CPU, disk, firewall states and network bandwidth limits for a jail with jailops config set. What each limit does when it is reached, and how to check that it applies.
Why a ceiling
One busy jail should not take the host with it.
A jail that leaks memory, forks without end, fills the disk or opens thousands of connections is a problem for every other jail on the host. A limit makes it its own problem: the jail slows down or fails an allocation, and everything else carries on.
Limits are keys of the jail. They are enforced by the kernel and the filesystem, not by a program that watches.
Setting a limit
One key, one value.
$ jailops config set web limits.memory 2G --project shop $ jailops config set web limits.cpu 2 --project shop $ jailops config set web limits.pids 500 --project shop $ jailops config set web limits.disk 20GiB --project shop
The same keys work in a profile and in a file, so a limit can be part of what you declare for a jail.
What each one does
Nothing is killed.
| Key | At the limit |
|---|---|
limits.memory | An allocation beyond it fails. The programs that asked get the error; nothing is killed. |
limits.pids | fork fails while the jail has that many processes. |
limits.cpu | The jail is slowed down to that many cores. |
limits.disk | Writes fail. The limit covers the root, the data, the volumes and the snapshots of the jail. |
limits.states | The jail's firewall rules create no new connections beyond the number. One jail cannot fill the host's table. |
A jail that sets no limits.states gets a default of 10000 per rule, which a busy web or DNS server does not reach. Memory counts what the programs reserve, so a busy database needs more than it uses when idle: start generously and lower it.
Bandwidth
In each direction.
The network device of a jail can be limited for what it sends and for what it receives, in bits per second. The limit applies to everything the jail sends or receives, whoever started the connection, and it survives a restart.
$ jailops config set web nic.eth0.egress 20M --project shop $ jailops config set web nic.eth0.ingress 50M --project shop
Checking it
The doctor says when a limit does not apply.
The memory, process and CPU limits need the kernel's resource accounting, which is a boot setting: kern.racct.enable=1 in /boot/loader.conf. Without it a jail with such a limit still starts, but the limit is not enforced. jailops doctor fails the check sysctl.kern.racct.enable and the jail shows the condition Limits as NotEnforced, so you see it before it matters.
The memory, process and CPU limits take effect at the next start of the jail. The disk limit applies at once.
When a jail reaches a limit, the host publishes an event you can watch: What the host tells you by itself.
More articles
Jail updates without cron
Replace the nightly script that upgrades every jail: automatic jail upgrades per image, with a health check, a rollback and a record of every run, what they leave alone, and how a new FreeBSD release comes in.
Read the article: Jail updates without cronJail search and shell
Ctrl+K, a few letters, Shift+Enter: find a jail by name, label or app and open its shell. Labels that filter the jail list, bash in new jails, and the jailops command next to every step.
Read the article: Jail search and shellGetting started with the JailOps CLI
Root on the host, the command line and nothing else: install JailOps, build an image, create a network and a jail, reach it by name, snapshot it, open a port and install an app.
Read the article: Getting started with the JailOps CLIJailOps on a netcup server, with IPv6 that stays up
A /64, a router pair and a FreeBSD default that drops their probes. How every jail gets its own IPv6 address on netcup, and the one sysctl that ends the random unreachable alerts.
Read the article: JailOps on a netcup server, with IPv6 that stays upOne small cloud server, many services, solid IPv6
Run your database, web tier and workers on one low-cost server, each jail with its own IPv6 address, and fix the alerts a shared /64 causes with one sysctl.
Read the article: One small cloud server, many services, solid IPv6Write your own app: memcached, from package to catalog
Turn a small FreeBSD package into a JailOps app: the definition line by line, the offline check, a local install, the five traps, and the pull request that puts it in the catalog.
Read the article: Write your own app: memcached, from package to catalogFrom catalog to running service in one command
Install nginx from the app catalog with one command, see each of its steps, and find the data, the configuration, the logs and the snapshot afterwards.
Read the article: From catalog to running service in one commandA firewall you can read
Policies, address sets and jails by name (jail:web), each with a comment, so the firewall can be reviewed by someone who did not write it.
Read the article: A firewall you can readPostgreSQL and git, closed except for who needs them
Close a database and a git server, then let in exactly the web jail and the office, with a comment on every entry.
Read the article: PostgreSQL and git, closed except for who needs themProjects are boundaries
A project is a boundary. Another project is an external host, so only addresses cross it, written out on both sides. Why that is simpler to review, and what a project does not isolate.
Read the article: Projects are boundariesThe life of a jail
A jail starts at boot only when you declared it with --autostart. Why JailOps does it this way, how the order is decided, and how to see it with jailops boot.
Read the article: The life of a jailTab completion for jailops
Press Tab and get the names of your jails, networks and firewall policies, even on another host. How to install and turn on completion in bash, zsh and fish.
Read the article: Tab completion for jailopsTest the firewall before you trust it
Ask the firewall whether a connection would pass, and why, before you try it. Read the answer, put it in a script, and know what it cannot tell you.
Read the article: Test the firewall before you trust itSee a connection before it happens
Pick a From and a To, a protocol and a port, press Run, and see the connection flow through the firewall: what passes, what stops it, and which rule decides.
Read the article: See a connection before it happensStop every jail cleanly when the host shuts down
A host that reboots should stop its databases before it stops the disks. How JailOps stops the jails in order, and how to start them in order again.
Read the article: Stop every jail cleanly when the host shuts downOne line in pf.conf
The host's firewall stays yours. JailOps asks for one line, and everything for the jails lives below it.
Read the article: One line in pf.confWhat the host tells you by itself
A pool that reports an error, an uplink that goes down, a jail that reaches a limit: they show up in the event stream by themselves.
Read the article: What the host tells you by itselfNetworks the host controls
No bridge, no DHCP. The host hands out the addresses, gives each jail a link of its own and is the router, so jails cannot see or impersonate each other.
Read the article: Networks the host controlsVolumes: data that outlives a jail
Keep the data in a volume and the jail disposable. One volume can be attached to several jails, has its own snapshots and its own backup.
Read the article: Volumes: data that outlives a jailBack up volumes once
One copy of a volume, however many jails use it. Leave volumes out when you want to, bring them back with the jail.
Read the article: Back up volumes onceA private Docker registry
A registry of your own with users in a file: one jail, docker-registry and docker-auth, htpasswd for the passwords, the images on a volume.
Read the article: A private Docker registryNames for jails
A jail is reached by its name inside its project: db.shop.internal, or just db. Create a jail and its name is there; delete it and the name is gone.
Read the article: Names for jails
Try it on one server.
Try the console from the screenshots, install on a clean FreeBSD 15.1 machine, or add JailOps to a host you already run.