mbox series

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

Message ID 20230916-nolibc-initramfs-v1-0-4416ecedca6d@weissschuh.net
Headers
Series selftests/nolibc: avoid spurious kernel relinks |

Message

Thomas Weißschuh Sept. 16, 2023, 7:13 a.m. UTC 
  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.

Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
---
Thomas Weißschuh (3):
      kbuild: add toplevel target for usr/gen_init_cpio
      selftests/nolibc: don't embed initramfs into kernel image
      selftests/nolibc: drop target "rerun"

 Makefile                                |  4 +++
 tools/testing/selftests/nolibc/Makefile | 50 +++++++++++++++------------------
 2 files changed, 27 insertions(+), 27 deletions(-)
---
base-commit: 3f79a57865b33f49fdae6655510bd27c8e6610e0
change-id: 20230916-nolibc-initramfs-4fd00eac3256

Best regards,

Comments

Willy Tarreau Sept. 17, 2023, 3:22 a.m. UTC | #1 
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).

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.

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 ;-)

Cheers,
Willy
Thomas Weißschuh Sept. 17, 2023, 2:44 p.m. UTC | #2 
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 :-)