On 09/03/20 07:57, Thomas Gleixner wrote: > Thomas Gleixner <tglx@xxxxxxxxxxxxx> writes: >> Thomas Gleixner <tglx@xxxxxxxxxxxxx> writes: >>> Andy Lutomirski <luto@xxxxxxxxxx> writes: >>>> On Sat, Mar 7, 2020 at 7:47 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: >>>>> If MCE, NMI trigger a real pagefault then the #PF injection needs to >>>>> clear apf_reason and set the correct CR2. When that #PF returns then the >>>>> old CR2 and apf_reason need to be restored. >>>> >>> The host does not care about the IRET. It solely has to check whether >>> apf_reason is 0 or not. That way it knows that the guest has read CR2 >>> and apf_reason. > > Some hours or sleep and not staring at this meess later and while > reading the leaves of my morning tea: > > guest side: > > nmi()/mce() ... > > stash_crs(); > > + stash_and_clear_apf_reason(); > > .... > > + restore_apf_reason(); > > restore_cr2(); > > Too obvious, isn't it? Yes, this works but Andy was not happy about adding more save-and-restore to NMIs. If you do not want to do that, I'm okay with disabling async page fault support for now. Storing the page fault reason in memory was not a good idea. Better options would be to co-opt the page fault error code (e.g. store the reason in bits 31:16, mark bits 15:0 with the invalid error code RSVD=1/P=0), or to use the virtualization exception area. Paolo
