Re: [PATCH 4/6] x86/efi: Hoist page table switching code into efi_call_virt()

[Matt Fleming <matt-mF/unelCI9GS6iBeEJttW/XRex20P6io@public.gmane.org>]

On Fri, 13 Nov, at 08:59:43AM, Borislav Petkov wrote:
> On Thu, Nov 12, 2015 at 08:01:08PM +0000, Matt Fleming wrote:
> > > That PUSHF implicitly pushes on the stack pointed by %rsp. But(!) we
> > > have switched the pagetable (i.e., %cr3 has efi_scratch.efi_pgt) and
> > > we're pushing to the VA where the stack *was* but is not anymore.
> >  
> > All the kernel mappings will still exist in the page table we switch
> > to, so pushing to the stack should be fine.
> > 
> > The mappings have to exist so that the firmware can dereference
> > pointer arguments, e.g. when writing variable data to a kernel buffer.
> > 
> > Or have I misunderstood your point?
> > 
> > > Or maybe it is because you're copying all the PUDs. It is still not 100%
> > > clean, IMHO.
> ^^^^^^^^^^^^^^^^^
> 
> I think we're on the same page - you're copying the PUDs in
> efi_sync_low_kernel_mappings() so the stack should be there.

Correct.

> > > Can you do the prolog/epilog calls inside the IRQs-off section?
> > 
> > Not really because in the efi_enabled(EFI_OLD_MEMMAP) case we perform
> > kmalloc(), see commit 23a0d4e8fa6d ("efi: Disable interrupts around
> > EFI calls, not in the epilog/prolog calls").
> 
>         n_pgds = DIV_ROUND_UP((max_pfn << PAGE_SHIFT), PGDIR_SIZE);
>         save_pgd = kmalloc(n_pgds * sizeof(pgd_t), GFP_KERNEL);
> 
> This?
> 
> That n_pgds thing is of static size so you can pre-alloc it maybe even
> once during boot and reuse it all the time when EFI_OLD_MEMMAP is
> enabled.

Well, this code is only executed once during boot anyway, for
phys_efi_set_virtual_address_map().

FYI, I'm still planning on ripping out all the EFI_OLD_MEMMAP code, as
it's getting particularly crufty.