RE: [PATCH V2 1/2] x86/efi: Unmap EFI boot services code/data regions from efi_pgd

From: Thomas Gleixner
Date: Mon Oct 29 2018 - 14:11:15 EST


On Mon, 29 Oct 2018, Prakhya, Sai Praneeth wrote:

> Hi Thomas and Peter,
> [off the mailing list because I wasn't sure if it's a good idea to spam others with my questions]

In general it's good to do on list because others can learn from the
answers as well.

> > > > + retval = __change_page_attr_set_clr(&cpa, 0);
> > > > + __flush_tlb_all();
> > >
> > > So this looks like you copied it from kernel_map_pages_in_pgd() which
> > > has been discussed before to be not sufficient, but it can't be
> > > changed right now due to locking issues.
> >
> > Managed to confuse myself. The place which cannot be changed is a different
> > one, but still for your call site __flush_tlb_all() might not be sufficient.
> Since the mappings are changed, I thought a flush_tlb() might be needed.
> But, (to be honest), I don't completely understand the x86/mm code. So, could you
> please elaborate the issue more for my better understanding?
> So some questions I have are,
> 1. What did Peter Z mean here? "How is that not a TLB invalidation bug ?"

__flush_tlb_all() flushes only the mapping on the current CPU. So in a SMP
environment it's not sufficient.

> 2. I assumed kernel_map_pages_in_pgd() to be a good code but you said that
> it has some locking issues. So, could you please elaborate more on that.

I corrected myself. It's the other function which cannot use the proper
cpa_flush.*() functions.

> Or, could you please provide me some pointers in the source code that I
> can take a look at so that I could understand the issue much better.

I'll have a second look at the whole thing and reply on list.