zpool import blows up wired vm.kmem_size

[grarpamp]

10.3-release i386 2GB pentium4

vm.kmem_size=720000000
Needed to keep bumping kmem up to accomodate other zpools.
That's the upper limit that will boot on this box.

Then got rid of all but one zpool...

NAME      SIZE  ALLOC   FREE    CAP  DEDUP  HEALTH  ALTROOT
pool1     232G   228G  3.64G    98%  1.00x  ONLINE  -
pool1 with 6 zfs filesystems with 6+ million inodes.

When import attempt, wired mem will hover around 75M for a minute,
then jump in a few seconds to about 180M and sit for a few sec,
then jump straight to 700+M or so within a second, then the box just
sits there with some z-thread[s] stuck in 'kmem a' top state.
ARC was only in the low 10M's which sounds fine since no file reads.

zpool import -o readonly=on imports through to auto mounting ok.
Only read-write blows kmem up and never mounts or returns.
Obviously will copy data off it, and off i386.

I think it's having problems cleaning up
previous writes that hit kmem / powerfail
and rebooted.

But why the instantaneous jumps in kmem size though?

Is this the equivalent of fsck running out of userspace
ram on bigfs back in the day, but now zfs in kernel space?