vmm svm: Refactor per-vCPU data.

[jhb]

Summary

• Allocate VMCBs separately to avoid excessive padding in struct svm_vcpu.

• Allocate APIC pages dynamically directly in struct vlapic.

• Move vm_mtrr into struct svm_vcpu rather than using a separate parallel array.

[corvink]

sys/amd64/vmm/amd/svm.c

616

I'm not sure if we should check if the vmcb is page aligned or not. Same for other structures of your patch series. malloc docs say:

IMPLEMENTATION NOTES
     The memory	allocator allocates memory in chunks that have size a power of
     two for requests up to the	size of	a page of memory.  For larger re-
     quests, one or more pages is allocated.  While it should not be relied
     upon, this	information may	be useful for optimizing the efficiency	of
     memory use.

I think it's unlikely that this behaviour changes. Nevertheless, it would be safer to add an assertion.

[markj]

sys/amd64/vmm/amd/svm.c

569

Presumably it doesn't matter if the softc isn't page-aligned?

616

Indeed, or use malloc_aligned().

2406–2407

Same here.

[jhb]

sys/amd64/vmm/amd/svm.c

569

I don't think it does. I think this was only for the VMCB.

616

I can use malloc_aligned(). I do think many many places in the kernel assume that malloc(PAGE_SIZE) returns an aligned page though.

[jhb]

• Use malloc_aligned

[corvink]

sys/amd64/vmm/amd/svm.c

616

Thanks for changing it. IMHO, we should avoid not well defined behaviour just because many other places assume it. Additionally, malloc_aligned documents the alignment requirement.