On Tue, Oct 06, 2020 at 04:50:44PM +0200, Vitaly Kuznetsov wrote: > Vivek Goyal <vgoyal@xxxxxxxxxx> writes: > > > On Tue, Oct 06, 2020 at 04:05:16PM +0200, Vitaly Kuznetsov wrote: > >> Vivek Goyal <vgoyal@xxxxxxxxxx> writes: > >> > >> > A. Just exit to user space with -EFAULT (using kvm request) and don't > >> > wait for the accessing task to run on vcpu again. > >> > >> What if we also save the required information (RIP, GFN, ...) in the > >> guest along with the APF token > > > > Can you elaborate a bit more on this. You mean save GFN on stack before > > it starts waiting for PAGE_READY event? > > When PAGE_NOT_PRESENT event is injected as #PF (for now) in the guest > kernel gets all the registers of the userspace process (except for CR2 > which is replaced with a token). In case it is not trivial to extract > accessed GFN from this data we can extend the shared APF structure and > add it there, KVM has it when it queues APF. > > > > >> so in case of -EFAULT we can just 'crash' > >> the guest and the required information can easily be obtained from > >> kdump? This will solve the debugging problem even for TDX/SEV-ES (if > >> kdump is possible there). > > > > Just saving additional info in guest will not help because there might > > be many tasks waiting and you don't know which GFN is problematic one. > > But KVM knows which token caused the -EFAULT when we exit to userspace > (and we can pass this information to it) so to debug the situation you > take this token and then explore the kdump searching for what's > associated with this exact token. So you will have to report token (along with -EFAULT) to user space. So this is basically the 3rd proposal which is extension of kvm API and will report say HVA/GFN also to user space along with -EFAULT. Thanks Vivek
