Re: [PATCH] x86/kexec: fix incorrect end address passed to kernel_ident_mapping_init()

[Yuntao Wang]

On Thu, 21 Dec 2023 14:03:00 -0800, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:

> On Thu, 21 Dec 2023 18:17:02 +0800 Yuntao Wang <ytcoode@xxxxxxxxx> wrote:
> 
> > kernel_ident_mapping_init() takes an exclusive memory range [pstart, pend)
> > where pend is not included in the range, while res represents an inclusive
> > memory range [start, end] where end is considered part of the range.
> > 
> > Therefore, passing res->end directly to kernel_ident_mapping_init() is
> > incorrect, the correct end address should be `res->end + 1`.
> > 
> > ...
> >
> > --- a/arch/x86/kernel/machine_kexec_64.c
> > +++ b/arch/x86/kernel/machine_kexec_64.c
> > @@ -44,7 +44,7 @@ static int mem_region_callback(struct resource *res, void *arg)
> >  	struct init_pgtable_data *data = arg;
> >  
> >  	return kernel_ident_mapping_init(data->info, data->level4p,
> > -					 res->start, res->end);
> > +					 res->start, res->end + 1);
> >  }
> 
> Thanks.  When fixing a bug, please always full describe the
> userspace-visible effects of that bug.
> 
> In this case I'm assuming "none", due to kernel_ident_mapping_init()'s
> upward rounding of the address?

Passing [start, end] rather than [start, end+1) to kernel_ident_mapping_init()
may result in the identity mapping for the end address not being set up.

For example, when res->start is equal to res->end, kernel_ident_mapping_init()
will not establish any identity mapping. Similarly, when the value of res->end
is a multiple of 2M and the page table maps 2M pages, kernel_ident_mapping_init()
will also not set up identity mapping for res->end.