[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] PLEASE REVERT URGENTLY: Re: [PATCH v5 2/3] x86/boot: add acpi rsdp address to setup_header

On Sun, Nov 11, 2018 at 10:49:39AM -0800, H. Peter Anvin wrote:
> On 11/10/18 1:03 AM, Juergen Gross wrote:
> > 
> > How would that help? The garabge data written could have the correct
> > terminal sentinel value by chance.
> > 
> > That's why I re-used an existing field in setup_header (the version) to
> > let grub tell the kernel which part of setup_header was written by grub.
> > 
> > That's the only way I could find to let the kernel distinguish between
> > garbage and actual data.
> There is plenty of space *before* the setup_header part of struct boot_params
> too -- look a the various __pad fields, especially (in your case), __pad3[16]
> and __pad4[116] would suit the bill just fine.
> >> It would be enormously helpful if you could find out any more details about
> >> exactly what they are doing to break things.
> > 
> > That's easy:
> > 
> > The memory layout is:
> > 
> > 0x1f1 bytes of data, including the sentinel, the setup_header, and then
> > more data.
> > 
> > grub did read the kernel's setup_header in the correct size into its
> > buffer (which contains random garbage before that), intializes the first
> > 0x1f1 including the sentinel byte, and then writes back the buffer, but
> > using a too large length for that.
> Are you kidding me... it really overwrites it with completely random data, and
> not simply overspilling contents of the file?
> In that case it might not be possible (or desirable) to use those N bytes
> following the setup_heaader, or we need to a bigger sentinel than one byte
> (probability being what it is, 256^n gets to be a pretty big number for any n,
> very quickly drowning in the noise compared to other potential sources of boot
> failures, and most likely less fatal than most.)
> How big is this garbage dump?  I'm going to brave a guess it is 512 bytes.  In
> that case this is hardly a big deal: the E820 map begins at 0x290, and the
> setup_header maximum goes to 0x280, so it is only 15 bytes lost.  If it is
> worse than that, we would risk losing __pad8[48] and __pad9[276], and
> especially the latter would be painful. In those case perhaps we should use
> 0x281..0x290 as a 15-byte sentinel; that is going to be virtually foolproof.
> I'm also thinking that it might be desirable to add a flags field (__pad2
> would be ideal) to struct boot_params; it would let us recycle some of the
> obsolete fields (hd0_info, hd1_info, sys_desc_table, olpc_ofw_header, ...) and
> perhaps be able to add some more robustness against these sort of things. This
> would be the right way to do what your version feedback mechanism would do.
> The reason why the feedback mechanism is fundamentally broken is that it only
> gives the boot loader a way to assert that it supports a certain version of
> the protocol, but it doesn't say *which* bootloader does such an assert -- and
> therefore it is still wide open to implementation error.
> We do, in fact, already have a feedback mechanism: the bootloader ID and
> bootloader version. One way we could deal with this problem is to bump the
> bootloader version reported by Grub, and add a quirk to the kernel that if the
> bootloader ID is Grub (7) and the version is less than a certain number, zero
> those fields. In fact, the more I think about it, this is what we should do.
> That being said, Grub really needs to be kept honest.  They keep making both
> severe design (like refusing to use the BIOS and UEFI entry points provided by
> the kernel by default) and implementation errors, apparently without
> meaningful oversight. I really ask that the people more closely involved with
> Grub try to keep a closer eye on their code as it applies to Linux.

Cc-ing GRUB and Daniel Kiper (maintainer of GRUB).

Could folks please please CC Daniel Kiper on any of these patches in the future?

>       -hpa

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.