On Thu, 27 Feb 2020 at 16:54, Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote: > > On Thu, Feb 27, 2020 at 04:21:32PM +0100, Ard Biesheuvel wrote: > > On Thu, 27 Feb 2020 at 16:16, Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote: > > > > > > On Thu, Feb 27, 2020 at 09:12:29AM +0100, Ingo Molnar wrote: > > > > > > > > * Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote: > > > > > > > > > Commit ef5a7b5eb13e ("efi/x86: Remove GDT setup from efi_main") > > > > > introduced GDT setup into the 32-bit kernel's startup_32, and reloads > > > > > the GDTR after relocating the kernel for paranoia's sake. > > > > > > > > > > Commit 32d009137a56 ("x86/boot: Reload GDTR after copying to the end of > > > > > the buffer") introduced a similar GDTR reload in the 64-bit kernel. > > > > > > > > > > The GDTR is adjusted by init_size - _end, however this may not be the > > > > > correct offset to apply if the kernel was loaded at a misaligned address > > > > > or below LOAD_PHYSICAL_ADDR, as in that case the decompression buffer > > > > > has an additional offset from the original load address. > > > > > > > > > > This should never happen for a conformant bootloader, but we're being > > > > > paranoid anyway, so just store the new GDT address in there instead of > > > > > adding any offsets, which is simpler as well. > > > > > > > > > > Signed-off-by: Arvind Sankar <nivedita@xxxxxxxxxxxx> > > > > > Fixes: ef5a7b5eb13e ("efi/x86: Remove GDT setup from efi_main") > > > > > Fixes: 32d009137a56 ("x86/boot: Reload GDTR after copying to the end of the buffer") > > > > > > > > Have you or anyone else observed this condition practice, or have a > > > > suspicion that this could happen - or is this a mostly theoretical > > > > concern? > > > > > > > > Thanks, > > > > > > > > Ingo > > > > > > Right now it's a theoretical concern. > > > > > > I'm working on another patch, to tell the EFI firmware PE loader what > > > the kernel's preferred address is, so that we can avoid having to > > > relocate the kernel in the EFI stub in most cases (ie if the PE loader > > > manages to load us at that address). With those changes, the required > > > adjustment won't be init_size - _end any more, and while fixing it up > > > there, I noticed that it could already be the case that the required > > > adjustment is different. > > > > > > > Do you mean setting the image address in the PE/COFF header to the > > preferred address? > > Yep. I'm doing that and then making a few adjustments to the PE entry > code and head_* so that it can decompress starting at image_base instead > of startup_32. Interesting. I am going to rip most of the EFI handover protocol stuff out of OVMF, since it is mostly unnecessary, and having the PE/COFF loader put the image in the correct place right away is a nice complimentary improvement to that. (Note that the OVMF implementation of the EFI handover protocol does not currently honor the preferred address from the setup header anyway)