On 6/10/26 16:57, Eva Winterschön wrote:
Hi Mario,
I have perhaps too many nvidia GPUs, from Maxwell generation to Kepler, Pascal,
Turing, Ampere, and GB10 series, with consumer RTX in the mix but mostly
datacenter grade SKUs from those series. Some are in systems, some are in
boxes, some will never be used, some have passive cooling, some have active,
etc. Otherwise to say, generations of GPUs are important for testing
compatibility and operational consistency. I can potentially run the same tests
if there's more detailed information.
Back when I had time for anything other than wasting my attention span on
cleaning up the security messes from May's Github CVE, I was running a lot of
test-bench and workload profiling on 14.x with bhyve passthrough for nvidia as
well as CUDA support (without virtualization involved). While reading these
threads I wonder about the following:
1. What does your test harness look like?
2. What is your kernel config?
3. What version and config of everything is in use on the OS (hardware specs,
uefi with/without csm enabled, x2acpi opt-out, pcie lane settings, aspm
settings, bhyve global settings, qemu settings, vm specific settings, pciconf
output, uefi loader settings/flags, loader.conf settings, rc.conf settings, pkg
versions, driver versions, firmware version for the MB and BMC, sysctl.conf
settings, firmware version on the GPU, model of GPU, nvidia driver version, etc)
4. Are you tracking and keeping syslog entries from the start to end of the
test runs?
5. Are you using additional process debugging for bhyve, kernel debug symbol
support, using drace + flamegraphs etc.
6. Which method of patching is used, what exactly was observed every time a
test iteration occurred? how idempotent and repeatable are your tests?
There's a lot of work to get from "doesn't work" to "it works", including the necessary
definition of "works" (are we talking, the VM can see the GPU, or the VM can use the GPU, the VM
can see and use the GPU and also execute CUDA workloads on the GPU).
Thanks,
Eva
On Wed, Jun 10, 2026, at 02:18, Mario Marietto wrote:
--> Does it occur on stock 15? .
Error happened with stock 15 upgraded until 15p9.
---> Nvidia GPU passthrough should work on stock 15, so Please test it
without patching any files.
I asked Claude to create the patches because it didn't work with stock
15. With the patches created by Claude that I have applied it worked
again.
Mario.
On Wed, Jun 10, 2026 at 10:59 AM Corvin Köhne <[email protected]> wrote:
On Tue, 2026-06-09 at 08:40 +0200, Mario Marietto wrote:
---
https://github.com/freebsd/freebsd-src/commit/effcd0ded0ef1c92bef29c1e9bbf51
7c97bb45b8
oh ok. That's not the bug I was talking about. You gave no reply about that
(bhyve: bootrom_alloc: vm_mmap_mapseg: Invalid argument on FreeBSD 15*-
RELEASE).
Does it occur on stock 15? Nvidia GPU passthrough should work on stock 15, so
please test it without patching any files.
--
Best regards,
Corvin
--
Mario.
HI!
As a starting point, you could get relevant informations with:
# pciconf -lbcevV
# grep vmm /boot/loader.conf
# grep ppt /boot/loader.conf
# acpidump -t | grep DMAR
# kldstat
# grep -iE '(vmm|ppt|APIC|DMAR|IOMMU|mapp)' /var/log/messages
Regards,
Stephan
