libsel4vm, x86: interrupt model improvements#86
libsel4vm, x86: interrupt model improvements#86andybui01 wants to merge 5 commits intoseL4:masterfrom
Conversation
andybui01
force-pushed
the
andyb/x86_vmm
branch
from
8758323 to
31258f8
Compare
andybui01
changed the title
|
@andybui01 glad to see someone else has worked on this and released a version for the community! I noticed you mentioned DornerWorks in the PR, and you're right that we did some work on this topic, but our code can't be released yet :( What i've noticed is that MSIs don't need the IOAPIC for functionality, which is good, because adding IOAPIC support is a pain, due to needed to compile your own DSDT which needs to match the definitions in the |
Writing to a pci ioport currently double adds the offset to the register. This was not picked up before as all writes were done to a register with an offset of 0, so adding 0 again had no effect. Signed-off-by: Andy Bui <andy.bui@unsw.edu.au>
The ACPI tables currently report addresses of follow on tables in terms of the VMM's virtual address. This needs to be set to the guests' physical address in order for the guest to interpret the tables correctly. Signed-off-by: Andy Bui <andy.bui@unsw.edu.au>
The current x86 apic emulation is very barebones and is not used. This commit makes the apic the primary interrupt controller, backed up by the i8259 PIC. Signed-off-by: Andy Bui <andy.bui@unsw.edu.au>
This commit adds MSI support to the x86 VMM. It allows MSIs to be set up by the guest, and includes code that intercepts MSI setup so the VMM can ensure the guest OS is not programming the MSI to be sent to somewhere it shouldn't be sent (e.g. another core or vm). Signed-off-by: Andy Bui <andy.bui@unsw.edu.au>
Every apic has a timer and guest OS' generally expect the existence of the apic timer if there is an apic. This commit attempts to add apic timer emulation. It works sometimes (TM), but there is currently a bug where the timer interrupt arrives in the VMM but does not make it to the guest (even though it does before). This behaviour is also non-deterministic. Signed-off-by: Andy Bui <andy.bui@unsw.edu.au>
abrandnewusername
force-pushed
the
andyb/x86_vmm
branch
from
31258f8 to
79243d0
Compare
|
@andybui01 can you explain the non-deterministic behavior you're seeing with the lapic timer? I pulled this into a code base that is running real-time Linux, and the timer seemed to initialize properly, but then stop triggering during the boot process. I'm unsure of whether that is Linux stopping the timer for some reason, or the behavior you noticed. |
Hi, thanks for having a look. It's been a while since I worked on this, I remember it would work sometimes and then a fresh boot would freeze things up. I never investigated beyond this. |
|
I looked into this a bit, and on my boot, Linux is switching to oneshot mode, and the time server is never set to trigger an event. |
|
@andybui01 quick question - where was the reference for the lapic timer? KVM? |
Yep, KVM. |
2 participants
This PR attempts to do the following:
An example CAmkES app configuration can be found here: https://github.com/andybui01/camkes-vm-examples/blob/andyb/supermicro_msi/apps/x86/supermicro/supermicro.camkes
A working manifest for the supermicro board can also be found here: https://github.com/andybui01/camkes-vm-examples-manifest/blob/andyb/supermicro_msi/master.xml