Re: RFC x86/boot/64: BOOT_PGT_SIZE definition for compressed kernel

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Sat, Oct 24, 2020 at 08:41:58PM -0400, Arvind Sankar wrote:
> Hi, I think the definition of BOOT_PGT_SIZE in
> arch/x86/include/asm/boot.h is insufficient, especially after
>   ca0e22d4f011 ("x86/boot/compressed/64: Always switch to own page table")
> 
> Currently, it allocates 6 pages if KASLR is disabled, and either 17 or
> 19 pages depending on X86_VERBOSE_BOOTUP if KASLR is enabled.
> 
> - The X86_VERBOSE_BOOTUP test shouldn't be done: that only disables
>   debug messages, but warnings/errors are always output to VGA memory,
>   so the two extra pages for mapping video RAM are always needed.
> 
> - The calculation wasn't updated for X86_5LEVEL, which requires at least
>   one more page for the P4D level, and in theory could require two extra
>   pages for each of the 4 mappings (compressed kernel, output kernel,
>   boot_params and command line), though that would require a system with
>   truly ginormous amounts of RAM.

Or sparse physical memory map. I hacked QEMU before for testing 5-level
paging:

https://gist.github.com/kiryl/d45eb54110944ff95e544972d8bdac1d

> - If KASLR is disabled, there are only 6 pages, but now that we're
>   always setting up our own page table, we need 1+(2+2)*3 (one PGD, and
>   two PUD and two PMD pages for kernel, boot_params and command line),
>   and 2 more pages for the video RAM, and more for 5-level. Even for
>   !RELOCATABLE, 13 pages might be needed.

The comment for BOOT_PGT_SIZE has to be updated.

BTW, what happens if we underestimate BOOT_PGT_SIZE? Do we overwrite
something?

> - SEV-ES needs one more page because it needs to do a PTE-level mapping
>   for the GHCB page.
> 
> - The static calculation is also busted because
>   boot/compressed/{kaslr.c,acpi.c} can scan the setup data, EFI
>   configuration tables and the EFI memmap, and none of these are
>   accounted for. They used to be scanned while still on the
>   firmware/bootloader page tables, but now our page tables have to cover
>   them as well. Trying to add up the worst case for all of these, and
>   anything else the compressed kernel might potentially access seems
>   like a lost cause.
> 
> We could do something similar to what the main kernel does with
> early_dynamic_pgts: map the compressed kernel at a fixed virtual
> address (in negative address space, say); recycle all the other mappings
> until we're done with decompression, and then map the output,
> boot_params and command line. The number of pages needed for this can be
> statically calculated, for 4-level paging we'd need 2 pages for the
> fixed mapping, 12 pages for the other three, and one PGD page.

Recycling idea look promising to me, but it would require handling #PF in
decompression code, right? It is considerable complication of the code.

-- 
 Kirill A. Shutemov