Re: [PATCH] x86/sev: Apply RMP table fixups for kexec.

From: Kalra, Ashish
Date: Tue Apr 02 2024 - 17:17:59 EST



On 4/2/2024 4:00 PM, Kalra, Ashish wrote:

On 4/2/2024 3:21 PM, Borislav Petkov wrote:
On Tue, Apr 02, 2024 at 02:33:44PM -0500, Kalra, Ashish wrote:
And we can't do this in snp_rmptable_init() as e820_table_firmware can't be
fixed at that point and by that time this table has been mapped into sysfs
(/sys/firmware) which is used by kexec -c variant.
Well, you have to do something here because if snp_rmptable_init()
late-disables SNP, your RMP table fixups are moot and invalid.

Which means, your RMP table fixups need to happen at the *very* *late*
step after we know that SNP is enabled and won't get disabled anymore.

I.e., in snp_rmptable_init().

The main issue with doing that in snp_rmptable_init() is that there is no e820 API interfaces available to update the e820_table_kexec and e820_table_firmware and e820_table_firmware has already been exposed to sysfs.

The e820 API only exports e820__range_update() which *only* fixes e820_table.

The important point to note here is that in most cases BIOS would have reserved RMP table start and end aligned to 2M boundary and setup the e820 table which the BIOS passes to the kernel as such, so even if the kernel does not enable SNP or disables SNP later, these reservations will remain aligned as such. So what we are doing here in-kernel fixups is doing the same alignment fixups which the BIOS would have done. The summary here is that e820 table adjustments for RMP table done either by BIOS and/or kernel will exist/remain even if SNP is not enabled by the kernel.

Again, to reiterate here, RMP table memory is reserved by BIOS regardless of the kernel enabling SNP (and also passed on the e820 map to ensure that kernel does not map anything in that memory), so any adjustments/fixups on top of that reserved memory should not matter, after all we don't free this reserved RMP table memory if kernel does not enable SNP.

Thanks, Ashish