Re: [PATCH 3/7] x86/mm/init: pass unconverted symbol addresses to free_init_pages()

[Hugh Dickins]

On Sat, 4 Aug 2018, Linus Torvalds wrote:
> On Fri, Aug 3, 2018 at 5:19 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> >
> > I thought that virt_to_page() only works on virtual addresses
> > in the direct map
> 
> You're right that virt_to_page() does not work on any _actual_ virtual
> mappings (ie no user pages, and no vmalloc() pages etc). It does not
> follow page tables at all.
> 
> And on 32-bit, it literally ends up doing (see __phys_addr_nodebug()) a simple
> 
>     #define __phys_addr_nodebug(x)  ((x) - PAGE_OFFSET)
> 
> However, on x86-64, we have *two* cases of direct mappings: we have
> the one at __START_KERNEL_map, and we have the one at PAGE_OFFSET.
> 
> And virt_to_page() handles both of those direct mappings.
> 
> Annoying? Yes. And it has caused bugs in the past. And I entirely
> forget why we needed it on x86-64.
> 
> [ Goes around and rummages ]
> 
> Oh, never mind, looking around reminded me why: we want to map the
> kernel text in the top 31 bits, so that we can use the faster
> -mcmodel=kernel because all symbols fit in sign-extended 32 bits.
> 
> Maybe there was some other reason too, but I think that's it.

Thanks a lot for writing that up. You shamed me into grepping
a little harder than I did yesterday, when all I could find were
"- PAGE_OFFSET" conversions (maybe I got lost in 32-bit-land).
I had missed __phys_addr_nodebug(), where the __START_KERNEL_map
alternative is handled.

Hugh