Re: [PATCH v11 4/9] mm: introduce memfd_secret system call to create "secret" memory areas

From: Mike Rapoport
Date: Tue Nov 24 2020 - 11:49:49 EST


On Tue, Nov 24, 2020 at 10:59:48AM +0000, Catalin Marinas wrote:
> Hi Mike,
>
> On Tue, Nov 24, 2020 at 11:25:51AM +0200, Mike Rapoport wrote:
> > +static vm_fault_t secretmem_fault(struct vm_fault *vmf)
> > +{
> > + struct address_space *mapping = vmf->vma->vm_file->f_mapping;

...

> > +
> > + err = set_direct_map_invalid_noflush(page, 1);
> > + if (err)
> > + goto err_del_page_cache;
>
> On arm64, set_direct_map_default_noflush() returns 0 if !rodata_full but
> no pgtable changes happen since the linear map can be a mix of small and
> huge pages. The arm64 implementation doesn't break large mappings. I
> presume we don't want to tell the user that the designated memory is
> "secret" but the kernel silently ignored it.
>
> We could change the arm64 set_direct_map* to return an error, however, I
> think it would be pretty unexpected for the user to get a fault when
> trying to access it. It may be better to return a -ENOSYS or something
> on the actual syscall if the fault-in wouldn't be allowed later.
>
> Alternatively, we could make the linear map always use pages on arm64,
> irrespective of other config or cmdline options (maybe not justified
> unless we have clear memsecret users). Yet another idea is to get
> set_direct_map* to break pmd/pud mappings into pte but that's not always
> possible without a stop_machine() and potentially disabling the MMU.

My preference would be to check at secretmem initialization if
set_direct_map_*() actually do anything and then return an error from
the syscall if they are essentially nop.

I'll update the patches with something like this in v12.

> --
> Catalin

--
Sincerely yours,
Mike.