[Skiboot] Experiments with modules for payload kernel

Stewart Smith stewart at linux.vnet.ibm.com
Fri Jan 30 14:38:26 AEDT 2015


(while not strictly skiboot, of interest)

The payload that we load (a linux kernel with initramfs with petitboot)
currently has everything =Y rather than =m. This means that during boot
a lot of things are serialized, and on some systems it can take 20+
seconds to go from starting to boot the kernel to having the petitboot
UI show up (mainly spent detecting disks).

Another downside of a kernel without modules is that when booting in a
simulator (mambo), the calculations done for what to use to RAID
(altivec or blah) are really slow.

So, I tried building just about everything as modules and seeing if it
worked.

It does!

There's a bug I've found with pb-discover core dumping on one machine,
but on others, I've ended up getting petitboot UI up in about 0.6-1
second after control switches to linux.

So, my thoughts are we should seriously consider switching over to using
modules rather than building everything into the kernel. It slightly
parallelises startup and gets us to something the user can interact with
quickly - especially useful for environments where the first boot is
going to be a network boot (why wait for disks to finish being
detected?)

Obviously, more testing is needed, and we need a UI thing in petitboot
to say we're still scanning for disks to boot from.



More information about the Skiboot mailing list