Re: Linux 5.0-rc1 (test results)

From: Linus Torvalds
Date: Mon Jan 07 2019 - 18:22:07 EST


On Mon, Jan 7, 2019 at 11:26 AM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
>
> Bisect points to commit 4cf58924951ef ("mm: treewide: remove unused address
> argument from pte_alloc functions"). Interesting - wasn't that supposed
> to be automatic ?
>
> csky does use the the removed address argument, so I won't even try to
> provide a patch. Copying csky maintainer instead.

Hmm. Interesting. The csky code seems to have some odd "poison pte
contents with ones if the address has the high bit set".

Which makes little or no sense. The "high bit set" case is for kernel
page tables, but that's exactly the "pte_alloc_one()" vs
"pte_alloc_one_kernel()" distinction.

So testing the address seems entirely wrong.

But there's other strangeness in there too. For example,
pte_alloc_one_kernel() will just write directly to the page. And
pte_alloc_one() will do a "kmap_atomic()" on the page it allocates,
except since it uses GFP_KERNEL, that's entirely pointless.

Is the alloc_pages() in pte_alloc_one() perhaps meant to use
GFP_HIGHUSER instead? Is this perhaps some copy-paste issue?

So I *think* the removal of the 'address' use in csky should be
simple, but yes, this needs a csky maintainer to look at.

Thanks,

Linus