[PATCHv3 0/5] use __create_pgd_mapping() to implement idmap and unify codes

[Pingfan Liu]

On Tue, Jun 08, 2021 at 06:38:07PM +0100, Catalin Marinas wrote:
> On Tue, Jun 01, 2021 at 05:25:49PM +0800, Pingfan Liu wrote:
> > On Tue, Jun 1, 2021 at 3:50 AM Ard Biesheuvel <ardb at kernel.org> wrote:
> > > On Mon, 31 May 2021 at 10:46, Pingfan Liu <kernelfans at gmail.com> wrote:
> > > > v2 -> v3:
> > > >   -1. leave out the part of redefinition the CONFIG_PGTABLE_LEVEL,
> > > > concentrate on sharing __create_pgd_mapping() in head.S as the first
> > > > step.
> > > >   -2. make IDMAP_PGTABLE_LEVELS use the max value ([3/5])
> > > >
> > > > rfc -> v2:
> > > >   more debug and test
> > > >
> > > > *** Goal of this series ***
> > > >
> > > > __create_pgd_mapping() sets up the pgtable for mapping __va(paddr) ->
> > > > paddr under the MMU-on situation.  Since pgtable upper level holds the
> > > > paddr of the lower level, with a slight adaptation,
> > > > __create_pgd_mapping() can also set up the mapping under the MMU-off
> > > > situation. ([4/5])
> > > >
> > > > After that, both idmap_pg_dir and init_pg_dir can be created by
> > > > __create_pgd_mapping(). And the counterpart asm code can be simplified.
> > >
> > > I understand the desire to simplify the page table construction code
> > > in head.S, but up until now, we have been very careful to avoid
> > > calling into C code with the MMU off. There are simply too many
> > > assumptions in the compiler about the context the generated code will
> > > execute in: one example is unaligned access, which must be disabled
> > > for source files that may be called with the MMU off, as otherwise,
> > > the compiler is permitted to emit loads and stores that are not
> > > allowed on device memory (which is the default memory type used for
> > > all memory with the MMU off)
> >
> > You are right. These C routines happen to use "unsigned long", which
> > can exclude this unaligned case.
> > To make an guarantee, is "-mno-unaligned-access" good enough?
> > 

Unaligned-access is the main challedge, and the unaligned-access should be caused from the programer.
Compiler has enforced different default alignment, including global and local(stack) variable, function alignment.
Otherwise, the atomity can not be assumped on some base data type.

But some code may violate the alignment, which compiler has no method to prevent.
As Documentation/core-api/unaligned-memory-access.rst
  1. Casting variables to types of different lengths
  2. Pointer arithmetic followed by access to at least 2 bytes of data

So in order to prevent the unaligned-access, the involved codes should be checked carefully.

> > Besides unaligned-access, any further risk originating from compiler
> > assumption? (I think that the common optimization: reordering,
> > merging, reloading on this "device" memory has no bad effect)
> 

These should be not issue. Since MMU-off present a more strict memory model than MMU-on.
So if a code can run on a more relaxed one, it can also run on the strict one.

> There's also instrumentation that needs disabling (kasan, ubsan, kcov,
> gcov). You can look at arch/arm64/kvm/hyp/nvhe/Makefile for various
> flags added or filtered out, though the KVM hyp code runs with the MMU
> on. I'm not sure what other flags are needed to guarantee the generated
> code can run with the MMU off but we can always ask the toolchain folk.
> 

Thanks for pointing out these issues. I will check.

> However, I'm still not convinced about sharing __create_pgd_mapping()
> with the early head.S code. A better option would be a separate,
> stand-alone file where we have more control on what gets called or
> accessed (of course, if there's any value in going this route).
> 
> > > Do you have a killer use case for this feature? Or is it just a nice cleanup?
> >
> > Yes, in the omitted part in v2, I had planned to provide an unified
> > page table manipulation routine, and provide an create_idmap() API. So
> > there can be an handy interface to create a whole RAM addressable
> > idmapX where needed.
> 
> Where would this be needed?
> 

It was planned for Pavel's series: [PATCH v13 00/18] arm64: MMU enabled kexec relocation,
where Pavel uses linear mapping to copy the data. But that way should involve
copy of "EL2 vectors" to tackle the transition to EL2.  On the other hand, if
using a global addressable idmap, the logic of transition to EL2 keeps
untouched, and just enable MMU in SYM_CODE_START(arm64_relocate_new_kernel)

The other call site is in trans_pgd_idmap_page() 


Thanks,

Pingfan