Re: [PATCH RFC 0/3] selftests/nolibc: avoid spurious kernel relinks

[Thomas Weißschuh]

On 2023-09-17 05:22:19+0200, Willy Tarreau wrote:
> On Sat, Sep 16, 2023 at 09:13:26AM +0200, Thomas Weißschuh wrote:
> > Currently the nolibc testsuite embeds the test executable into a kernel
> > image with CONFIG_INITRAMFS_SOURCE.
> > This forces a full kernel relink everytime the test executable is
> > updated.
> > 
> > This relinking step dominates the test cycle.
> > It is slower than building and running the test in qemu together.
> > 
> > With a bit of Makefile-shuffling the relinking can be avoided.
> 
> That's pretty nice as indeed it still takes a while to relink it into
> the kernel. I agree that for running nolibc-test in qemu we don't need
> a unified image. However I've seldom used it on real hardware and I
> find it significantly more convenient to use as a single image in this
> case. Maybe we should just rename targets so that everything qemu-related
> just uses two distinct images while a "unified-image" (or anything else)
> still assembles the image into the kernel (BTW the help on the "kernel"
> target still mentions initramfs).

Sounds good, "unified-image" is a bit close to "unified kernel image"
(UKI) which is similar but different.

What about kernel-standalone?

> Note that we don't need to modify anything in the build system to create
> an initrd, I usually make them by hand using "cpio -o -H newc", we don't
> need anything else here.

I'd like to keep the build self-contained. But actually the kernel build
will always build a minimal initramfs if CONFIG_BLK_DEV_INITRD is set.
This config is needed for the nolibc testsuite in any case.

The fact that the initrd is always build means that usr/gen_init_cpio
is also always build.
So we can add "kernel" as a prerequisite to "initramfs.cpio" and
everything should work out without any buildsystem modifications.

> Regarding rerun, I'd rather keep it for the sole reason that I've used
> it to check for randomly failing errors (typically the timing-based
> ones). It's convenient to run the same image 100 times if needed. I'm
> not strongly attached to it, but if "make run" is slower, then we can
> keep it. However if you really want to drop it, it also needs to be
> dropped from the help message ;-)

Fine for me, let's keep it :-)