On Wed, 23 Dec 2020, JaromÃr DoleÄ~Mek wrote:
TL;DR ZFS + Xen doesn't work together under NetBSD 9.1. You need to either use NetBSD-current, or use something other than ZFS. Longer version: Xen under NetBSD up to 9.x had a limitation of 32KB MAXBSIZE related to xbd(4), the Xen kernels had an option override to force the lower limit. This limitation was removed in NetBSD-current. ZFS is loaded as a module, and is compiled with default MAXBSIZE of 64kb, which doesn't work with the Xen kernel's 32KB.
But overriding as per http://mail-index.netbsd.org/port-xen/2020/09/11/msg009793.html works. What I don't get is why seeing as xen modules are distinct from non-xen (/stand/amd64/9.1/modules/zfs/zfs.kmod vs /stand/amd64-xen/9.1/modules/zfs/zfs.kmod), the xen versions cannot be built with this override automatically. The patch below would also impact non-Xen kernels.
I've not yet managed to follow through the build process to work out how amd64-xen kernel modules are built, I expect someone (hopefluly reading this list!) will know off the top of their head.
--- src/sys/modules/zfs/Makefile.zfsmod.DIST 2019-02-05 04:54:36.000000000 -0500
+++ src/sys/modules/zfs/Makefile.zfsmod 2020-09-10 18:35:34.683493656 -0400
@@ -135,5 +135,6 @@
CPPFLAGS+= -I${ZFSDIR}/dist/uts/common
CPPFLAGS+= -D_PROPLIB_ZFS_CONFLICT
+CPPFLAGS+= -DMAXPHYS=32768
NEED_ISYSTEM= # duplicate headers, depends on -isystem to build
Le mer. 23 déc. 2020 à 11:04, Matthias Petermann <mp%petermann-it.de@localhost> a écrit :Hello Brian, Am 22.12.2020 um 18:51 schrieb Brian Buhrow:hello. Just for giggles, what if you reduce the vcpu count to 1 to make sure you're not running into some sort of crazy race condition with the domu or a constraint in the dom0? If it works, then you can explore why a count of 8 doesn't.I did that - but same effect here. Anyway, during my tests I observed that one attempt with the ZVOL renders the system in state in which even the file backed volume will panic. After this, restarting xencommons doesn't help but a complete system reboot will do. Kind regards Matthias