Re: failing regression tests (please fix them!)

[Antti Kantee <pooka%cs.hut.fi@localhost>]

On Wed Jun 02 2010 at 23:10:58 +0300, Andreas Gustafsson wrote:
> Antti Kantee wrote:
> > With a few hours of waiting you can easily produce it yourself (probably
> > required for debugging tests anyway) by running:
> > "anita test http://nyftp.netbsd.org/pub/NetBSD-daily/HEAD/<date>/i386/"
> 
> That won't actually let you debug the tests interactively, since it
> will just run the tests and then immediately kill the virtual machine.
> You probably want something like this:
> 
>    anita interact http://nyftp.netbsd.org/pub/NetBSD-daily/HEAD/<date>/i386/
>    <log in as root>
>    cd /usr/tests/lib/libevent
>    atf-run

Just to point out the obvious, you can also run "qemu -nographic wd0.img"
afterwards.

Even with interact, the problem might be that you don't have a test
compiled with -g.  I usually mount the image on the host, pax sources
over, compile with DESTDIR=/anita DBG=-g (it's significantly faster on
the host if you have the toolchain), make DESTDIR=/anita install, umount
and run qemu.  Also, need to "ln -s / /anita" for gdb in qemu to find the
sources (i'm sure there's a "better" way than a link to deal with that).

Speaking of which, the default image size used by anita is 384M.
I've locally patched it to be 1024M to fit some extra debugging crud.
Since it's a sparse file, the size isn't very critical, and one image
is more practical than two.

Speaking of speaking of which, I did some benchmarking on my system
(with my slow laptop disk and highly aged file system) and found a
non-sparse 384M file to be 2 seconds faster for "anita install" than a
fully allocated one.  Significantly larger disk images were slower in the
preallocated case (because of the time it takes to allocate the image).
I'm just mentioning this because I found the first result surprising.