Messages in this thread |  | | Date | Sat, 31 Dec 2022 19:21:06 -0800 | Subject | Re: [PATCH qemu] x86: don't let decompressed kernel image clobber setup_data | From | "H. Peter Anvin" <> |
| |
On 12/31/22 11:00, Borislav Petkov wrote: > On Sat, Dec 31, 2022 at 07:22:47PM +0100, Jason A. Donenfeld wrote: >> So with that understanding confirmed, I'm confused at your surprise that >> hpa's unrelated fix to the different issue didn't fix this issue. > > No surprise there - I used a qemu variant without your patch to prevent the > setup_data clobbering so hpa's fix can't help there. > >> But since the kernel doesn't do this now, and the 62MiB bug also seems >> to apply to existing kernels, for the purposes of QEMU for now, I think >> the v3 patch is probably best, since it'll handle existing kernels. > > Right, we can't fix all those guests which are out there. > >> Alternatively, setup_data could be relocated, the boot param protocol >> could be bumped, and then QEMU could conditionalized it's use of >> setup_data based on that protocol version. That'd work, but seems a bit >> more involved. > > I think this is at least worth considering because the kernel overwriting > setup_data after having been told where that setup_data is located is not really > nice. > > Well not explicitly at least - it gets the pointer to the first element and > something needs to traverse that list to know which addresses are live. But > still, that info is there so perhaps we need to take setup_data into > consideration too before decompressing... >
As far as the decompression itself goes, it should only a problem if we are using physical KASLR since otherwise the kernel has a guaranteed safe zone already allocated by the boot loader. However, if physical KASLR is in use, then the decompressor needs to know everything there is to know about the memory map.
However, there also seems to be some kind of interaction with AMD SEV-SNP.
The bug appears to have been introduced by:
b57feed2cc2622ae14b2fa62f19e973e5e0a60cf x86/compressed/64: Add identity mappings for setup_data entries https://lore.kernel.org/r/TYCPR01MB694815CD815E98945F63C99183B49@TYCPR01MB6948.jpnprd01.prod.outlook.com
... which was included in version 5.19, so it is relatively recent.
For a small amount of setup_data, the solution of just putting it next to the command line makes a lot of sense, and should be safe indefinitely.
-hpa
|  |