Re: [PATCH v2] fix setup_efi_pci()

[Matt Fleming <matt.fleming-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>]

On Fri, 2013-01-25 at 07:41 +0000, Jan Beulich wrote:
> >>> On 24.01.13 at 19:20, Matt Fleming <matt.fleming-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> > On Fri, 2013-01-18 at 12:35 +0000, Jan Beulich wrote:
> >> This fixes two issues:
> >> - wrong memory type used for allocation intended to persist post-boot
> > 
> > [...]
> > 
> >> @@ -311,7 +311,7 @@ static efi_status_t setup_efi_pci(struct
> >>  		size = pci->romsize + sizeof(*rom);
> >>  
> >>  		status = efi_call_phys3(sys_table->boottime->allocate_pool,
> >> -				EFI_LOADER_DATA, size, &rom);
> >> +					EFI_RUNTIME_SERVICES_DATA, size, &rom);
> >>  
> >>  		if (status != EFI_SUCCESS)
> >>  			continue;
> > 
> > I'm curious why you made this change. No one should be stealing that
> > region of memory because that's all handled in parse_e820_ext() - it's
> > marked as off limits wrt memory for the kernel's use. And the firmware
> > certainly shouldn't start touching it.
> 
> According to my reading of do_add_efi_memmap(), EFI_LOADER_DATA
> gets treated the same as EFI_CONVENTIONAL_MEMORY, i.e. gets
> marked as available for allocation. That's also how I'd expect things
> to be. I don't think parse_e820_ext() matters here at all.

Oh, I totally misread the flow of execution here. parse_setup_data()
seems to be missing a SETUP_PCI case.

Matthew, what guarantees do we have that the memory we've allocated for
the PCI ROMs isn't going to be reused? I suspect what's needed is for
parse_setup_data() to reserve the SETUP_PCI regions, just like you would
for any other key data structure.

It's the kernel's responsibility to protect any EFI_LOADER_DATA regions
that it needs, we shouldn't have to trick ourselves by using the
EFI_RUNTIME_SERVICES_DATA pool for allocations.