Had some discussion with Gao, Chao.
On 6/26/2024 5:09 AM, Edgecombe, Rick P wrote:
On Tue, 2024-06-25 at 14:54 +0800, Binbin Wu wrote:TDX GHCI doesn't have such restriction.
Can we add a comment or something to that check, on why KVM doesn't handle it?+ gpa = vcpu->mmio_fragments[0].gpa;Since MMIO cross page boundary is not allowed according to the input
+ size = vcpu->mmio_fragments[0].len;
checks from TDVMCALL, these mmio_fragments[] is not needed.
Just use vcpu->run->mmio.phys_addr and vcpu->run->mmio.len?
Is it documented somewhere in the TDX ABI that it is not expected to be
supported?
According to the reply from Isaku in the below link, I think current restriction is due to software implementation for simplicity.
https://lore.kernel.org/kvm/20240419173423.GD3596705@xxxxxxxxxxxxxxxxxxxxx/
+ /* Disallow MMIO crossing page boundary for simplicity. */
+ if (((gpa + size - 1) ^ gpa) & PAGE_MASK)
goto error;
According to https://lore.kernel.org/all/165550567214.4207.3700499203810719676.tip-bot2@tip-bot2/,
for Linux as TDX guest, it rejects EPT violation #VEs that split pages based on the reason "MMIO accesses are supposed to be naturally aligned and therefore never cross page boundaries" to handle the load_unaligned_zeropad() case.
I am not sure "MMIO accesses are supposed to be naturally aligned" is true for all other OS as TDX guest, though.
Any suggestion?