(Possible) panic from BootServices->AllocatePages() returning the zero address

[kukrimate]

The out parameter of BootServices->AllocatePages() is an EFI_PHYSICAL_ADDRESS, not a pointer.

The uefi-rs wrapper boot::allocate_pages() converts this into a NonNull<u8> and panics if that fails.

However there is nothing in the standard that says physical address 0 is not a valid result, and X86 has a tendency of mapping physical memory at address 0. This can cause valid programs to panic on allocation.

Most UEFI implementations allocate from top-down, so this only really happens in practice just before running out memory, but there is nothing in the spec that says they need to. Code assuming AllocatePages() always returns non-0 addresses can panic on random allocations.

Some other UEFI apps (e.g. GRUB) do have checks to retry allocation if address 0 was yielded.

[nicholasbishop]

Thanks for reporting this, I didn't realize that was allowed. I put up #1558 to address this -- does that fix match your understanding of the problem?

Does allocate_pool have the same issue? From a quick skim of the spec I didn't see any guarantee about not returning null there either.

Just curious -- did you file this bug because you observed a failure on some real firmware, or was it just something you were aware could happen?

[kukrimate]

Your workaround in #1558 is basically the same logic as in GRUB2, so I believe that should be correct.

I also think BootServices->AllocatePool() does not suffer from the same issue:

• Spec says the out parameter is a "pointer to the allocated buffer" for that function, and it seems to be specified in terms of C, which also does not allow accessing a buffer at NULL.
• I have also not seen the workaround for AllocatePool in any existing code, people always seem to assume it returns a usable pointer.
• But there is no explicit language against it in the spec beyond the pointer language, so maybe rules lawyering is possible somehow?

How I saw it?

• Noticed the workaround in GRUB2 many times while working on it.
• I also remember seeing it on some whacky firmware a long time ago.
• But yesterday, in OVMF, I only managed to reproduce in a 10 year old version, with a fixed address allocation at address 0.

EDIT: also I accidentally clicked the close and comment button, instead of the close one, sorry.

[phip1611]

Hey, thanks for reporting this, @kukrimate. In the future, please avoid closing the issue before the PR is merged. While unlikely, we want to prevent cases where a PR isn't merged, but we mistakenly assume the issue was resolved.

[kukrimate]

Closing was an accidental click, apologies.