Re: [PATCH RESEND 00/62] initrd: remove classic initrd support

On 9/24/25 11:17, Alexander Patrakov wrote:

quoted

Therefore is it really initrd you are removing or just some corner case
? If it is really initrd, then how does QEMU still work with that
-initrd parameter ?

The QEMU -initrd parameter is a misnomer. It can be used to pass an
initrd or an initramfs, and the kernel automatically figures out what
it is.

It's not a misnomer, initrams has always been able to make use of the 
existing initrd loading mechanism to read images externally supplied by 
the bootloader. It's what grub calls it too. I documented it in the 
"External initramfs images" section of 
https://kernel.org/doc/Documentation/filesystems/ramfs-rootfs-initramfs.txt 
back in 2005. The mechanism itself is 30 years old 
(Documentation/initrd.txt was written by Werner Almsberger in linux 
1.3.73 from March 7, 1996, ala 
https://github.com/mpe/linux-fullhistory/commit/afc106342783 ).

Since initrd contents could always be in a bunch of different 
autodetected formats (and optionally compressed just like the kernel), 
initramfs just hooked in to the staircase and said "if the format is 
cpio, call this function to handle it". The patch series proposes 
removing all the other formats, but not otherwise changing the existing 
external image loader mechanism. (Personally I think removing the 
architecture-specific hacks but leaving the generic support under init/ 
would probably have made more sense as a first step.)

The bootloader hands off an initrd image, initramfs is the boot-time 
cpio extraction plumbing that's _init tagged and gets freed, and rootfs 
is the persistent mounted instance of ramfs or tmpfs that's always there 
and is analogous to the init task (PID 1) except for the mount tree. 
(And is often overmounted so it's not visible, but it's still there. And 
is NOT SPECIAL: overmounts aren't a new concept, nor is hiding them in 
things like "df".)

There's a REASON my documentation file was called 
ramfs-rootfs-initramfs.txt: the naming's always been a bit... layered. 
(And yes, I have always spelled initmpfs with only one t.)

Rob