systemd

Random Seeds

systemd can help in a number of ways with providing reliable, high quality random numbers from early boot on.

Linux Kernel Entropy Pool

Today’s computer systems require random number generators for numerous cryptographic and other purposes. On Linux systems, the kernel’s entropy pool is typically used as high-quality source of random numbers. The kernel’s entropy pool combines various entropy inputs together, mixes them and provides an API to userspace as well as to internal kernel subsystems to retrieve it. This entropy pool needs to be initialized with a minimal level of entropy before it can provide high quality, cryptographic random numbers to applications. Until the entropy pool is fully initialized application requests for high-quality random numbers cannot be fulfilled.

The Linux kernel provides three relevant userspace APIs to request random data from the kernel’s entropy pool:

(Strictly speaking there are more APIs, for example /dev/random, but these should not be used by almost any application and hence aren’t mentioned here.)

Note that the time it takes to initialize the random pool may differ between systems. If local hardware random number generators are available, initialization is likely quick, but particularly in embedded and virtualized environments available entropy is small and thus random pool initialization might take a long time (up to tens of minutes!).

Modern hardware tends to come with a number of hardware random number generators (hwrng), that may be used to relatively quickly fill up the entropy pool. Specifically:

If neither of these are available (in fact, even if they are), Linux generates entropy from various non-hwrng sources in various subsystems, all of which ultimately are rooted in IRQ noise, a very “slow” source of entropy, in particular in virtualized environments.

systemd’s Use of Random Numbers

systemd is responsible for bringing up the OS. It generally runs as the first userspace process the kernel invokes. Because of that it runs at a time where the entropy pool is typically not yet initialized, and thus requests to acquire random bytes will either be delayed, will fail or result in a noisy kernel log message (see above).

Various other components run during early boot that require random bytes. For example, initial RAM disks nowadays communicate with encrypted networks or access encrypted storage which might need random numbers. systemd itself requires random numbers as well, including for the following uses:

Note that these cases generally do not require a cryptographic-grade random number generator, as most of these utilize random numbers to minimize risk of collision and not to generate secret key material. However, they usually do require “medium-grade” random data. For example: systemd’s hash-maps are reseeded if they grow beyond certain thresholds (and thus collisions are more likely). This means they are generally fine with low-quality (even constant) random numbers initially as long as they get better with time, so that collision attacks are eventually thwarted as better, non-guessable seeds are acquired.

Keeping systemd's Demand on the Kernel Entropy Pool Minimal

Since most of systemd’s own use of random numbers do not require cryptographic-grade RNGs, it tries to avoid reading entropy from the kernel entropy pool if possible. If it succeeds this has the benefit that there’s no need to delay the early boot process until entropy is available, and noisy kernel log messages about early reading from /dev/urandom are avoided too. Specifically:

  1. When generating Type 4 UUIDs, systemd tries to use Intel’s and AMD’s RDRAND CPU opcode directly, if available. While some doubt the quality and trustworthiness of the entropy provided by these opcodes, they should be good enough for generating UUIDs, if not key material (though, as mentioned, today’s big distributions opted to trust it for that too, now, see above — but we are not going to make that decision for you, and for anything key material related will only use the kernel’s entropy pool). If RDRAND is not available or doesn’t work, it will use synchronous getrandom() as fallback, and /dev/urandom on old kernels where that system call doesn’t exist yet. This means on non-Intel/AMD systems UUID generation will block on kernel entropy initialization.

  2. For seeding hash tables, and all the other similar purposes systemd first tries RDRAND, and if that’s not available will try to use asynchronous getrandom() (if the kernel doesn’t support this system call, /dev/urandom is used). This may fail too in case the pool is not initialized yet, in which case it will fall back to glibc’s internal rand() calls, i.e. weak pseudo-random numbers. This should make sure we use good random bytes if we can, but neither delay boot nor trigger noisy kernel log messages during early boot for these use-cases.

systemd’s Support for Filling the Kernel Entropy Pool

systemd has various provisions to ensure the kernel entropy is filled during boot, in order to ensure the entropy pool is filled up quickly.

  1. When systemd’s PID 1 detects it runs in a virtualized environment providing the virtio-rng interface it will load the necessary kernel modules to make use of it during earliest boot, if possible — much earlier than regular kernel module loading done by systemd-udevd.service. This should ensure that in VM environments the entropy pool is quickly filled, even before systemd invokes the first service process — as long as the VM environment provides virtualized RNG hardware (and VM environments really should!).

  2. The systemd-random-seed.service system service will load a random seed from /var/lib/systemd/random-seed into the kernel entropy pool. By default it does not credit entropy for it though, since the seed is — more often than not — not reset when ‘golden’ master images of an OS are created, and thus replicated into every installation. If OS image builders carefully reset the random seed file before generating the image it should be safe to credit entropy, which can be enabled by setting the $SYSTEMD_RANDOM_SEED_CREDIT environment variable for the service to 1 (or even force, see man page). Note however, that this service typically runs relatively late during early boot: long after the initial RAM disk (initrd) completed, and after the /var/ file system became writable. This is usually too late for many applications, it is hence not advised to rely exclusively on this functionality to seed the kernel’s entropy pool. Also note that this service synchronously waits until the kernel’s entropy pool is initialized before completing start-up. It may thus be used by other services as synchronization point to order against, if they require an initialized entropy pool to operate correctly.

  3. The systemd-boot EFI boot loader included in systemd is able to maintain and provide a random seed stored in the EFI System Partition (ESP) to the booted OS, which allows booting up with a fully initialized entropy pool from earliest boot on. During installation of the boot loader (or when invoking bootctl random-seed) a seed file with an initial seed is placed in a file /loader/random-seed in the ESP. In addition, an identically sized randomized EFI variable called the the ‘system token’ is set, which is written to the machine’s firmware NVRAM. During boot, when systemd-boot finds both the random seed file and the system token they are combined and hashed with SHA256 (in counter mode, to generate sufficient data), to generate a new random seed file to store in the ESP as well as a random seed to pass to the OS kernel. The new random seed file for the ESP is then written to the ESP, ensuring this is completed before the OS is invoked. Very early during initialization PID 1 will read the random seed provided in the EFI variable and credit it fully to the kernel’s entropy pool.

    This mechanism is able to safely provide an initialized entropy pool already in the initrd and guarantees that different seeds are passed from the boot loader to the OS on every boot (in a way that does not allow regeneration of an old seed file from a new seed file). Moreover, when an OS image is replicated between multiple images and the random seed is not reset, this will still result in different random seeds being passed to the OS, as the per-machine ‘system token’ is specific to the physical host, and not included in OS disk images. If the ‘system token’ is properly initialized and kept sufficiently secret it should not be possible to regenerate the entropy pool of different machines, even if this seed is the only source of entropy.

    Note that the writes to the ESP needed to maintain the random seed should be minimal. The size of the random seed file is directly derived from the Linux kernel’s entropy pool size, which defaults to 512 bytes. This means updating the random seed in the ESP should be doable safely with a single sector write (since hard-disk sectors typically happen to be 512 bytes long, too), which should be safe even with FAT file system drivers built into low-quality EFI firmwares.

    As a special restriction: in virtualized environments PID 1 will refrain from using this mechanism, for safety reasons. This is because on VM environments the EFI variable space and the disk space is generally not maintained physically separate (for example, qemu in EFI mode stores the variables in the ESP itself). The robustness towards sloppy OS image generation is the main purpose of maintaining the ‘system token’ however, and if the EFI variable storage is not kept physically separate from the OS image there’s no point in it. That said, OS builders that know that they are not going to replicate the built image on multiple systems may opt to turn off the ‘system token’ concept by setting random-seed-mode always in the ESP’s /loader/loader.conf file. If done, systemd-boot will use the random seed file even if no system token is found in EFI variables.

With the three mechanisms described above it should be possible to provide early-boot entropy in most cases. Specifically:

  1. On EFI systems, systemd-boot’s random seed logic should make sure good entropy is available during earliest boot — as long as systemd-boot is used as boot loader, and outside of virtualized environments.

  2. On virtualized systems, the early virtio-rng hookup should ensure entropy is available early on — as long as the VM environment provides virtualized RNG devices, which they really should all do in 2019. Complain to your hosting provider if they don’t.

  3. On Intel/AMD systems systemd’s own reliance on the kernel entropy pool is minimal (as RDRAND is used on those for UUID generation). This only works if the CPU has RDRAND of course, which most physical CPUs do (but I hear many virtualized CPUs do not. Pity.)

  4. In all other cases, systemd-random-seed.service will help a bit, but — as mentioned — is too late to help with early boot.

This primarily leaves two kind of systems in the cold:

  1. Some embedded systems. Many embedded chipsets have hwrng functionality these days. Consider using them while crediting entropy. (i.e. rng_core.default_quality=1000 on the kernel command line is your friend). Or accept that the system might take a bit longer to boot. Alternatively, consider implementing a solution similar to systemd-boot’s random seed concept in your platform’s boot loader.

  2. Virtualized environments that lack both virtio-rng and RDRAND. Tough luck. Talk to your hosting provider, and ask them to fix this.

  3. Also note: if you deploy an image without any random seed and/or without installing any ‘system token’ in an EFI variable, as described above, this means that on the first boot no seed can be passed to the OS either. However, as the boot completes (with entropy acquired elsewhere), systemd will automatically install both a random seed in the GPT and a ‘system token’ in the EFI variable space, so that any future boots will have entropy from earliest boot on — all provided systemd-boot is used.

Frequently Asked Questions

  1. Why don’t you just use getrandom()? That’s all you need!

    Did you read any of the above? getrandom() is hooked to the kernel entropy pool, and during early boot it’s not going to be filled yet, very likely. We do use it in many cases, but not in all. Please read the above again!

  2. Why don’t you use getentropy()? That’s all you need!

    Same story. That call is just a different name for getrandom() with flags set to zero, and some additional limitations, and thus it also needs the kernel’s entropy pool to be initialized, which is the whole problem we are trying to address here.

  3. Why don’t you generate your UUIDs with uuidd? That’s all you need!

    First of all, that’s a system service, i.e. something that runs as “payload” of systemd, long after systemd is already up and hence can’t provide us UUIDs during earliest boot yet. Don’t forget: to assign the invocation UUID for the uuidd.service start we already need a UUID that the service is supposed to provide us. More importantly though, uuidd needs state/a random seed/a MAC address/host ID to operate, all of which are not available during early boot.

  4. Why don’t you generate your UUIDs with /proc/sys/kernel/random/uuid? That’s all you need!

    This is just a different, more limited interface to /dev/urandom. It gains us nothing.

  5. Why don’t you use rngd, haveged, egd? That’s all you need!

    Like uuidd above these are system services, hence come too late for our use-case. In addition much of what rngd provides appears to be equivalent to CONFIG_RANDOM_TRUST_CPU=y or rng_core.default_quality=1000, except being more complex and involving userspace. These services partly measure system behavior (such as scheduling effects) which the kernel either already feeds into its pool anyway (and thus shouldn’t be fed into it a second time, crediting entropy for it a second time) or is at least something the kernel could much better do on its own. Hence, if what these daemons do is still desirable today, this would be much better implemented in kernel (which would be very welcome of course, but wouldn’t really help us here in our specific problem, see above).

  6. Why don’t you use arc4random()? That’s all you need!

    This doesn’t solve the issue, since it requires a nonce to start from, and it gets that from getrandom(), and thus we have to wait for random pool initialization the same way as calling getrandom() directly. arc4random() is nothing more than optimization, in fact it implements similar algorithms that the kernel entropy pool implements anyway, hence besides being able to provide random bytes with higher throughput there’s little it gets us over just using getrandom(). Also, it’s not supported by glibc. And as long as that’s the case we are not keen on using it, as we’d have to maintain that on our own, and we don’t want to maintain our own cryptographic primitives if we don’t have to. Since systemd’s uses are not performance relevant (besides the pool initialization delay, which this doesn’t solve), there’s hence little benefit for us to call these functions. That said, if glibc learns these APIs one day, we’ll certainly make use of them where appropriate.

  7. This is boring: NetBSD had boot loader entropy seed support since ages!

    Yes, NetBSD has that, and the above is inspired by that (note though: this article is about a lot more than that). NetBSD’s support is not really safe, since it neither updates the random seed before using it, nor has any safeguards against replicating the same disk image with its random seed on multiple machines (which the ‘system token’ mentioned above is supposed to address). This means reuse of the same random seed by the boot loader is much more likely.

  8. Why does PID 1 upload the boot loader provided random seed into kernel instead of kernel doing that on its own?

    That’s a good question. Ideally the kernel would do that on its own, and we wouldn’t have to involve userspace in this.

  9. What about non-EFI?

    The boot loader random seed logic described above uses EFI variables to pass the seed from the boot loader to the OS. Other systems might have similar functionality though, and it shouldn’t be too hard to implement something similar for them. Ideally, we’d have an official way to pass such a seed as part of the struct boot_params from the boot loader to the kernel, but this is currently not available.

  10. I use a different boot loader than systemd-boot, I’d like to use boot loader random seeds too!

    Well, consider just switching to systemd-boot, it’s worth it. See systemd-boot(7) for an introduction why. That said, any boot loader can re-implement the logic described above, and can pass a random seed that systemd as PID 1 will then upload into the kernel’s entropy pool. For details see the Boot Loader Interface documentation.

  11. Why not pass the boot loader random seed via kernel command line instead of as EFI variable?

    The kernel command line is accessible to unprivileged processes via /proc/cmdline. It’s not desirable if unprivileged processes can use this information to possibly gain too much information about the current state of the kernel’s entropy pool.

  12. Why doesn’t systemd-boot rewrite the ‘system token’ too each time when updating the random seed file stored in the ESP?

    The system token is stored as persistent EFI variable, i.e. in some form of NVRAM. These memory chips tend be of low quality in many machines, and hence we shouldn’t write them too often. Writing them once during installation should generally be OK, but rewriting them on every single boot would probably wear the chip out too much, and we shouldn’t risk that.