Oddball small memory qemu images and networking

[Reinoud Zandijk <reinoud%NetBSD.org@localhost>]

Hi folks,

On Mon, May 18, 2020 at 09:41:17PM +0000, Andrew Doran wrote:
> Finally got around to trying this.  Having beaten on it for a while with
> real hardware I don't see any problem with swapping over NFS on 9.99.63.
> 
> On Sat, May 02, 2020 at 12:06:48PM +1000, Paul Ripke wrote:
> 
> > I have a qemu guest for experimenting with -current, 1 CPU & 64MiB RAM.
> 
> 64 megs, I'm surprised it makes to the login prompt.

I can get a prompt and compile some stuff locally in 48 MiB but no network
until sat 80 MiB.

As a challenge I tried a bog standard GENERIC 9.99.69/amd64 on qemu with small
amounts of memory with a `harddisc' including a swap space. I increased the
memory size in steps of 2 MiB.

With one CPU
40 MiB boots but fsck_ffs killed due to out of swap; swap is later?
42 MiB booted to multiuser, mount_ffs failing
44 MiB booted to multiuser but is bog slow due to swap, no dhcpcd
46 MiB booted to multiuser fine, usable, no dhcpcd
...
66 MiB booted to multiuser fine, dhcpcd sometimes works
...
80 MiB booted to multiuser fine, dhcpcd works reliable


With 2 CPUs :
44 MiB doesn't boot and creates a panic in pmap_get_physpage()
46 MiB booted to multiuser but slowish due to swapping
48 MiB booted to multiuser fine, more usable, no dhcpcd
...
54 MiB booted to multiuser fine, dhcpcd sporadically works
...
60 MiB booted to multiuser fine, dhcpcd sometimes works
...
70 MiB booted to multiuser fine, dhcpcd mostly works
...
80 MiB booted to multiuser fine, dhcpcd works reliable


Network buffers are thus limiting its useability in such odd small memory
systems even though swap is available; the kernel can't allocate memory for
SIOCAOFADDR when done manually. Its strange that it sometimes succeeds though!

Not useless though as without network, a 1 CPU machine can already compile on
a 48Mb machine on FFS.

Note that on very low memory (60MiB) I can compile programs hosted over NFS
(when dhcpcd works) though SIGINFO can print odd stuff like:

[ 174.2806880] load: 0.48  cmd: cc1 3164 [0x693f93fb] 0.00u 267344117007.14s
1% 9572k

When compiling on bigger, say 80MiB it prints normal times.

Hope this was of help,
Reinoud