On 02/08/2023 11.28, Thomas Huth wrote:
The code already tries to do some endianness handling here, but currently fails badly: - While it already swaps the data when logging errors / tracing, it fails to byteswap the value before e.g. accessing entry->irte.present - entry->irte.source_id is swapped with le32_to_cpu(), though this is a 16-bit value - The whole union is apparently supposed to be swapped via the 64-bit data[2] array, but the struct is a mixture between 32 bit values (the first 8 bytes) and 64 bit values (the second 8 bytes), so this cannot work as expected.Fix it by converting the struct to two proper 64-bit bitfields, and by swapping the values only once for everybody right after reading the data from memory. Signed-off-by: Thomas Huth <th...@redhat.com> --- Note: There are more endianness issues in the code, I haven't figured out all of them yet, Linux fails to boot in the guest when I use this device on a s390x host. But I wanted to publish this patch now already since this should also fix the "issue" with the Clang ms_struct packing that we recently discussed on the mailing list.
I just found the all remaining issues (I hope). I can now run a Linux guest with -device intel-iommu and it works without crashing now. I'll send out the patches shorty, after cleaning them up a little bit.
Thomas
