Can EFI memory descriptors overlap? [PING v1]

Can EFI memory descriptors overlap? [PING v1]

[Martin Fernandez]

Hi, I'm working on the rework of
https://lore.kernel.org/linux-efi/CAMj1kXEqjHVRNV131=o_aO3TX+58xhYzgfaVDrd0RngAqqtrVg@mail.gmail.com/T/
and I wanted to know if the memory map given by EFI have any guarantee
that the regions don't overlap.

I couldn't find anything in the UEFI spec, but there is a piece of
code in the kernel that seems to "doesn't care" and assumes that the
regions given by EFI are sorted and don't overlap.

https://elixir.bootlin.com/linux/latest/source/arch/x86/platform/efi/efi.c#L502

static void __init efi_merge_regions(void)
{
    efi_memory_desc_t *md, *prev_md = NULL;

    for_each_efi_memory_desc(md) {
        u64 prev_size;

        if (!prev_md) {
            prev_md = md;
            continue;
        }

        if (prev_md->type != md->type ||
            prev_md->attribute != md->attribute) {
            prev_md = md;
            continue;
        }

        prev_size = prev_md->num_pages << EFI_PAGE_SHIFT;

        if (md->phys_addr == (prev_md->phys_addr + prev_size)) {
            prev_md->num_pages += md->num_pages;
            md->type = EFI_RESERVED_TYPE;
            md->attribute = 0;
            continue;
        }
        prev_md = md;
    }
}

Is this a bug or in practice EFI always gives a "nice" memory map? I
think that if there is no overlap then the sorting (EFI doesn't give
me a sorted memory map on my machine) is a kind of easy problem to
solve, but if there is overlap involved then we need to decide for the
type and the attributes of those overlaps, which can be tricky.

I'm working on x86 and with the master branch of the kernel.

Martin.

Re: Can EFI memory descriptors overlap? [PING v1]

[Ard Biesheuvel]

On Mon, 18 Oct 2021 at 20:49, Martin Fernandez
<martin.fernandez@eclypsium.com> wrote:
>
> Hi, I'm working on the rework of
> https://lore.kernel.org/linux-efi/CAMj1kXEqjHVRNV131=o_aO3TX+58xhYzgfaVDrd0RngAqqtrVg@mail.gmail.com/T/
> and I wanted to know if the memory map given by EFI have any guarantee
> that the regions don't overlap.
>

Please don't ask people to refer to postings that are legally dubious.
I understand that this may sound exaggerated to you, but it is the
reality of open source.

> I couldn't find anything in the UEFI spec, but there is a piece of
> code in the kernel that seems to "doesn't care" and assumes that the
> regions given by EFI are sorted and don't overlap.
>

In my experience with EDK2 based UEFI firmware, the entries covering
system DRAM are usually sorted, with additional MMIO entries appearing
unordered. This is because the UEFI memory map is constructed from the
GCD memory map, which is seeded with the DRAM regions at boot, and
allocations are created by splitting them up.

> https://elixir.bootlin.com/linux/latest/source/arch/x86/platform/efi/efi.c#L502
>
> static void __init efi_merge_regions(void)
> {
>     efi_memory_desc_t *md, *prev_md = NULL;
>
>     for_each_efi_memory_desc(md) {
>         u64 prev_size;
>
>         if (!prev_md) {
>             prev_md = md;
>             continue;
>         }
>
>         if (prev_md->type != md->type ||
>             prev_md->attribute != md->attribute) {
>             prev_md = md;
>             continue;
>         }
>
>         prev_size = prev_md->num_pages << EFI_PAGE_SHIFT;
>
>         if (md->phys_addr == (prev_md->phys_addr + prev_size)) {
>             prev_md->num_pages += md->num_pages;
>             md->type = EFI_RESERVED_TYPE;
>             md->attribute = 0;
>             continue;
>         }
>         prev_md = md;
>     }
> }
>
> Is this a bug or in practice EFI always gives a "nice" memory map?

I agree that this practice has no basis in the spec.

> I
> think that if there is no overlap then the sorting (EFI doesn't give
> me a sorted memory map on my machine) is a kind of easy problem to
> solve, but if there is overlap involved then we need to decide for the
> type and the attributes of those overlaps, which can be tricky.
>
> I'm working on x86 and with the master branch of the kernel.

Overlap is another matter, though. Overlap means that the EFI memory
map is ambiguous, and contains more than one description of the same
memory region. Since the spec does not describe this case, I will pull
the 'undefined behavior' card, and propose that we pretend that those
can never occur.

-- 
Ard.