[PATCH 2/4] PM: hibernate: improve robustness of mapping pages in the direct map

[Mike Rapoport]

On Thu, Oct 29, 2020 at 11:19:18PM +0000, Edgecombe, Rick P wrote:
> On Thu, 2020-10-29 at 09:54 +0200, Mike Rapoport wrote:
> > __kernel_map_pages() on arm64 will also bail out if rodata_full is
> > false:
> > void __kernel_map_pages(struct page *page, int numpages, int enable)
> > {
> >         if (!debug_pagealloc_enabled() && !rodata_full)
> >                 return;
> > 
> >         set_memory_valid((unsigned long)page_address(page), numpages,
> > enable);
> > }
> > 
> > So using set_direct_map() to map back pages removed from the direct
> > map
> > with __kernel_map_pages() seems safe to me.
> 
> Heh, one of us must have some simple boolean error in our head. I hope
> its not me! :) I'll try on more time.

Well, then it's me :)
You are right, I misread this and I could not understand why
!rodata_full bothers you.

> __kernel_map_pages() will bail out if rodata_full is false **AND**
> debug page alloc is off. So it will only bail under conditions where
> there could be nothing unmapped on the direct map.
> 
> Equivalent logic would be:
> 	if (!(debug_pagealloc_enabled() || rodata_full))
> 		return;
> 
> Or:
> 	if (debug_pagealloc_enabled() || rodata_full)
> 		set_memory_valid(blah)
> 
> So if either is on, the existing code will try to re-map. But the
> set_direct_map_()'s will only work if rodata_full is on. So switching
> hibernate to set_direct_map() will cause the remap to be missed for the
> debug page alloc case, with !rodata_full.
> 
> It also breaks normal debug page alloc usage with !rodata_full for
> similar reasons after patch 3. The pages would never get unmapped.

I've updated the patches, there should be no regression now.

-- 
Sincerely yours,
Mike.