Slashdot Mirror
Discover the Anatomy of initrd
IdaAshley writes "The Linux initial RAM disk (initrd) is a temporary root file system that is mounted during system boot to support the two-state boot process. It contains various executables and drivers that permit the real root file system to be mounted, after which the initrd RAM disk is unmounted and its memory freed. In this article explore the initial RAM disk for Linux 2.6, including its creation and use in the Linux kernel. In many embedded Linux systems, the initrd is the final root file system."
This looks to be a pretty good tutorial.
As well as mkinitrd, there are some cool tools coming along that help build an initrd.
Here is one I have used, and although it's very early in it's development cycle.
http://yaird.alioth.debian.org/
Are there any more that are actually easy to use?
Signature v3.0, now with 42% less memory usage.
Is it just me, or does maintaining initrd seem like more trouble than it is worth? I guess it is good for generic distributions which want to support every hardware config under the sun and keep the kernel small, but if you know the (set of) hardware you are going to be running a given kernel on, you might as well just blow that initrd crap away. Whenever I recompile a kernel, initrd is the first thing to go. All you need in the kernel to make initrd irrelevent is disk and filesystem (and maybe software RAID modules) support. You can keep everything else as modules.
-matthew
"THERE IS NO JUSTICE, THERE IS ONLY ME." -Death
You're still a coward.
"THERE IS NO JUSTICE, THERE IS ONLY ME." -Death
I thought that initramfs was the hotness? Why would we want to use initrd over initramfs (or vice versa, for that matter)?
Ewige Blumenkraft.
Stick to our techie roots.
We want more politics, thinly disguised corporate press releases, and pseudoscience articles, please!
- None can love freedom heartily, but good men; the rest love not freedom, but license. -- John Milton
There are definitely times where it is required.
Take, for example, booting onto a root on a SAN--specifically a Coraid Ethernet SAN.
During boot, you must bring up the ethernet interface, load the AoE module, probe for the SAN, wait for it to quiesce, then enable clustering, join the cluster, join the fence domain, allow the cluster to quiesce, load the DLM, enable CLVM, probe and activate everything, then mount your root GFS (or OCFS2 if you can get that working).
This would not be possible without a good initrd.
Also, not so obviously, distros aren't the only people who deploy a single kernel on lots of hardware. When you have 300 old Dells, 150 old gateways, 75 custom built boxen, and a handful of laptops, maintaining a single kernel and initrd beats maintaining ten of them. Not your use case, but definitely important.
I think Mauve has the most RAM. --PHB (Dilbert Comic)
We use both in our port of Linux to proprietary Cray architectures.
/, /dev, and /dev/console.
Initramfs is synthesized from the stock page and dentry caches. The nice thing about
initramfs is that RAM gets released to the dynamic pool when an inode is deleted. So
it's a good choice for files that need to be present at boot time but not kept around,
like kernel modules. The bad thing: each file occupies at least one page, and that's
64KiB here. An initramfs with lots of little files can waste beaucoup bytes.
RAM "disks" are devices that occupy statically allocated memory. Storage doesn't get
reclaimed, but no storage is wasted by rounding up file sizes. An initrd is an image
of a RAM disk, possibly compressed.
Note that a cpio archive, possibly compressed, that's loaded in place of an initrd image
gets automatically
expanded and copied into the initramfs.
Confusingly, the kernel also has its own compressed cpio archive within its own text containing
the initial content of the initramfs. At a minimum, it holds
If you're using an architecture with small pages (read: x86), use initramfs. Otherwise,
you may need a clever hybrid solution, such as populating initramfs from a compressed cpio
archive and then moving some of its content into a RAM disk.
"Skill shows through where genius wears thin." -Wittgenstein || Religion: uniting aviation and architecture.
I don't know much about this, but am curious.
Why, once the system is booted cleanly, can't it save its state in initrd, and just load the state the next time without going through the boot process? Would that result in fast booting?
* I had a very small problem getting it to work the first time, but that problem has been fixed since then.
The difference between spam and poop is that you don't have to dig through septic tanks looking for real food. -- Me
Why? I may be missing something, but the only time I've ever modified my initrd is when rebuilding the kernel.
Unfortunatly, the article starts well by explaining initramfs (using a compressed cpio rather than a block device image), but then leads the reader down the initrd path.
Initramfs is intended to superceed initrd for several good reasons.
For one, it's more efficient. Files in an initramfs are a special case of filesystem where the file is loaded into the generic file cache but has no backing storage (that is, it doesn't live on a block device). OTOH, an initrd is a simple image of a block device. It has a fixed number of blocks defined at creation time. On top of that it has filesystem formatting.
For one thing, access to the file takes a much simpler path through the code in the case of initramfs. At a minimum, an initrd requires one more filesystem to be built in to the kernel. Initramfs itself is just a thin layer over the generic filesystem code, so is much smaller and simpler.
An initramfs image will be exactly the size that is needed, never more or less. There is no need to guess what size block device will be needed to exactly hold everything you need in advance. There is no need to go with anything larger 'just in case'.
As for memory usage, once an initramfs is mounted, if you rm a file, that memory is immediatly freed (assuming nothing is holding the file open). With initrd, you can rm a file but it won't help you, it just means you now have more wasted blocks held in RAM.
Initramfs can grow as necessary (memory permitting). For example if the boot process fails, you can log WHY to a file and drop to a root shell for debugging. An initrd can be writable, but can never be any larger than it was created to be. If you MIGHT need to write, you will have to build the initrd large enough to cover the worst case anyone might encounter in debugging and hope that's not too big for machines with less RAM to accomodate.
Once you get a swap partition mounted from somewhere, your initramfs can be swapped out as needed, so can grow larger than RAM if necessary. With an initrd, it's stuck right where it is until you unmount it.
It seems to me that the reasons to keep the old initrd support around are down to support for legacy uses and as long as we have the code and it's not too much trouble, why not?
So, in the shell script from the article, instead of creating /tmp/ramdisk.img and mounting it with -oloop, just create a directory and copy (or hardlink) what you want more or less as the script does. Then make a compressed cpio archive of it as the initramfs image. Possibly like cd initrd; find . |cpio -H newc -o |gzip -9 >../initrd.img.gz