Re: [RFC KVM 19/27] kvm/isolation: initialize the KVM page table with core mappings

[Alexandre Chartre]

On 5/13/19 6:00 PM, Andy Lutomirski wrote:
> On Mon, May 13, 2019 at 8:50 AM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> > > +     /*
> > > +      * Copy the mapping for all the kernel text. We copy at the PMD
> > > +      * level since the PUD is shared with the module mapping space.
> > > +      */
> > > +     rv = kvm_copy_mapping((void *)__START_KERNEL_map, KERNEL_IMAGE_SIZE,
> > > +          PGT_LEVEL_PMD);
> > > +     if (rv)
> > > +             goto out_uninit_page_table;
> >
> > Could you double-check this?  We (I) have had some repeated confusion
> > with the PTI code and kernel text vs. kernel data vs. __init.
> > KERNEL_IMAGE_SIZE looks to be 512MB which is quite a bit bigger than
> > kernel text.
> >
> > > +     /*
> > > +      * Copy the mapping for cpu_entry_area and %esp fixup stacks
> > > +      * (this is based on the PTI userland address space, but probably
> > > +      * not needed because the KVM address space is not directly
> > > +      * enterered from userspace). They can both be copied at the P4D
> > > +      * level since they each have a dedicated P4D entry.
> > > +      */
> > > +     rv = kvm_copy_mapping((void *)CPU_ENTRY_AREA_PER_CPU, P4D_SIZE,
> > > +          PGT_LEVEL_P4D);
> > > +     if (rv)
> > > +             goto out_uninit_page_table;
> >
> > cpu_entry_area is used for more than just entry from userspace.  The gdt
> > mapping, for instance, is needed everywhere.  You might want to go look
> > at 'struct cpu_entry_area' in some more detail.
> >
> > > +#ifdef CONFIG_X86_ESPFIX64
> > > +     rv = kvm_copy_mapping((void *)ESPFIX_BASE_ADDR, P4D_SIZE,
> > > +          PGT_LEVEL_P4D);
> > > +     if (rv)
> > > +             goto out_uninit_page_table;
> > > +#endif
> >
> > Why are these mappings *needed*?  I thought we only actually used these
> > fixup stacks for some crazy iret-to-userspace handling.  We're certainly
> > not doing that from KVM context.
> >
> > Am I forgetting something?
> >
> > > +#ifdef CONFIG_VMAP_STACK
> > > +     /*
> > > +      * Interrupt stacks are vmap'ed with guard pages, so we need to
> > > +      * copy mappings.
> > > +      */
> > > +     for_each_possible_cpu(cpu) {
> > > +             stack = per_cpu(hardirq_stack_ptr, cpu);
> > > +             pr_debug("IRQ Stack %px\n", stack);
> > > +             if (!stack)
> > > +                     continue;
> > > +             rv = kvm_copy_ptes(stack - IRQ_STACK_SIZE, IRQ_STACK_SIZE);
> > > +             if (rv)
> > > +                     goto out_uninit_page_table;
> > > +     }
> > > +
> > > +#endif
> >
> > I seem to remember that the KVM VMENTRY/VMEXIT context is very special.
> >   Interrupts (and even NMIs?) are disabled.  Would it be feasible to do
> > the switching in there so that we never even *get* interrupts in the KVM
> > context?
>
> That would be nicer.
>
> Looking at this code, it occurs to me that mapping the IRQ stacks
> seems questionable.  As it stands, this series switches to a normal
> CR3 in some C code somewhere moderately deep in the APIC IRQ code.  By
> that time, I think you may have executed traceable code, and, if that
> happens, you lose.  i hate to say this, but any shenanigans like this
> patch does might need to happen in the entry code *before* even
> switching to the IRQ stack.  Or perhaps shortly thereafter.
>
> We've talked about moving context tracking to C.  If we go that route,
> then this KVM context mess could go there, too -- we'd have a
> low-level C wrapper for each entry that would deal with getting us
> ready to run normal C code.
>
> (We need to do something about terminology.  This kvm_mm thing isn't
> an mm in the normal sense.  An mm has normal kernel mappings and
> varying user mappings.  For example, the PTI "userspace" page tables
> aren't an mm.  And we really don't want a situation where the vmalloc
> fault code runs with the "kvm_mm" mm active -- it will totally
> malfunction.)

One of my next step is to try to put the KVM page table in the PTI userspace
page tables, and not switch CR3 on KVM_RUN ioctl. That way, we will run with
a regular mm (but using the userspace page table). Then interrupt would switch
CR3 to kernel page table (like paranoid idtentry currently do it).

alex.