On Wed, 9 Sep 2020 at 11:17, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
>
> (+ Atish, Palmer)
>
> On Fri, 4 Sep 2020 at 18:50, Heinrich Schuchardt <xypron.glpk@xxxxxx> wrote:
> >
> > In the memory map the regions with the lowest addresses may be of type
> > EFI_RESERVED_TYPE. The reserved areas may be discontinuous relative to the
> > rest of the memory. So for calculating the maximum loading address for the
> > device tree and the initial ramdisk image these reserved areas should not
> > be taken into account.
> >
> > Signed-off-by: Heinrich Schuchardt <xypron.glpk@xxxxxx>
> > ---
> > drivers/firmware/efi/libstub/efi-stub.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/firmware/efi/libstub/efi-stub.c b/drivers/firmware/efi/libstub/efi-stub.c
> > index c2484bf75c5d..13058ac75765 100644
> > --- a/drivers/firmware/efi/libstub/efi-stub.c
> > +++ b/drivers/firmware/efi/libstub/efi-stub.c
> > @@ -106,7 +106,8 @@ static unsigned long get_dram_base(void)
> > map.map_end = map.map + map_size;
> >
> > for_each_efi_memory_desc_in_map(&map, md) {
> > - if (md->attribute & EFI_MEMORY_WB) {
> > + if (md->attribute & EFI_MEMORY_WB &&
> > + md->type != EFI_RESERVED_TYPE) {
> > if (membase > md->phys_addr)
> > membase = md->phys_addr;
> > }
> > --
> > 2.28.0
> >
>
> This is not the right fix - on RPi2, for instance, which has some
> reserved memory at the base of DRAM, this change will result in the
> first 16 MB of memory to be wasted.
>
In the EFI memmap provided to the kernel efi stub it will be 2
regions. First is EFI_RESERVED and second is EFI_CONVENTIONAL_MEMORY.
Even if they follow each other.
And for_each_efi_memory_desc_in_map will just return the second one.
Do not see where the problem is here.
> What I would prefer to do is get rid of get_dram_base() entirely -
> arm64 does not use its return value in the first place, and for ARM,
> the only reason we need it is so that we can place the uncompressed
> kernel image as low in memory as possible, and there are probably
> better ways to do that. RISC-V just started using it too, but only
> passes it from handle_kernel_image() to efi_relocate_kernel(), and
> afaict, passing 0x0 there instead would not cause any problems.
For prior 5.8 kernels there was limitation for maximum address to
unpack the kernel. As I understand that was copy-pasted from x86 code,
and now is missing in 5.9. That is why the suggestion was to point
dram_base to the region where it's possible to allocate. I.e. I assume
that
patch was created not to the latest kernel. Removing the upper
allocation limit should work here.
Maxim.
