After installing RC2 of VirtualBox 5.0, my FreeNAS Virtual Machine no longer boots. The VM gets dropped into the Guru Meditation state quite rapidly after the FreeBSD kernel boots. Last lines (from the kernel) that are visible on the console revolve around determining available memory.
VBox.log and config file of the VM have been attached in archive. Log file says:
This is a bug in VirtualBox, though it shows a strange difference between FreeNAS and FreeBSD 9.3 from which is derived. FreeBSD 9.3 does not use the CLFLUSH instruction when it detects that it's being virtualized (it's useless in a VM anyway), but your FreeNAS setup does use it (see CLFLUSH listed among the CPU features), even though it detected that it's virtualized (see HV in Features2). It's not clear to me why.
Did the same problem not exist in RC1 or did you never try?
I can confirm this problem for FreeBSD 10.2-PRERELEASE amd64 1001519. Reverting to RC1 fixes it. The crash occurs as soon as X11 is started, the kernel boots fine.
ctheis wrote:I can confirm this problem for FreeBSD 10.2-PRERELEASE amd64 1001519.
Are you saying that that version also enables CLFLUSH use even though it's running in a VM?
FWIW, the problem seems to also somehow depend on the specific CPU used. Some are affected and some are not even when CLFLUSH is used, we're not really sure why.
Thanks! This is indeed the same comment as the OP's problem (CLFLUSH instruction going wrong). Once again it's unclear why the kernel is using it at all. Turning on the "minimal" paravirtualization option might make a difference.
At any rate, the CLFLUSH guru meditation will be fixed in the next build.
Using RC3 I could not start the machine from the GUI. Another thread mentioned to use build 101470 and with that build, I no longer run into the original problem (FreeNAS not wanting to start).
Thanks! This is indeed the same comment as the OP's problem (CLFLUSH instruction going wrong). Once again it's unclear why the kernel is using it at all. Turning on the "minimal" paravirtualization option might make a difference.
At any rate, the CLFLUSH guru meditation will be fixed in the next build.
Thanks for the confirmation. The problem was elusive enough that only the people who saw in in the first place could verify the fix. It only seemed to happen on certain CPUs.
It would still be interesting to know why the FreeBSD kernel thought it needed to use CLFLUSH at all even when running in a VM... but it's not that important.