Re: linux-next: manual merge of the akpm-current tree with the kvms390 tree

[John Hubbard <jhubbard@nvidia.com>]

On 2/26/20 7:11 PM, Stephen Rothwell wrote:
> Hi all,
> 
> Today's linux-next merge of the akpm-current tree got a conflict in:
> 
>   mm/gup.c
> 
> between commit:
> 
>   732b80e677b8 ("mm/gup/writeback: add callbacks for inaccessible pages")
> 
> from the kvms390 tree and commit:
> 
>   9947ea2c1e60 ("mm/gup: track FOLL_PIN pages")
> 
> from the akpm-current tree.
> 
> I fixed it up (see below - maybe not optimally) and can carry the fix as
> necessary. This is now fixed as far as linux-next is concerned, but any
> non trivial conflicts should be mentioned to your upstream maintainer
> when your tree is submitted for merging.  You may also want to consider
> cooperating with the maintainer of the conflicting tree to minimise any
> particularly complex conflicts.
> 

Yes. Changes to mm/gup.c really should normally go through linux-mm and 
Andrew's tree, if at all possible. This would have been caught, and figured out
on linux-mm, had that been done--instead of leaving the linux-next maintainer
trying to guess at how to resolve the conflict.

+Cc David Hildenbrand, who I see looked at the kvms390 proposed patch a bit.
Maybe he has some opinions, especially about my questions below.

The fix-up below may (or may not) need some changes:


diff --cc mm/gup.c
index 354bcfbd844b,f589299b0d4a..000000000000
--- a/mm/gup.c
+++ b/mm/gup.c
@@@ -269,18 -470,11 +468,19 @@@ retry
  		goto retry;
  	}
  
+ 	/* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
+ 	if (unlikely(!try_grab_page(page, flags))) {
+ 		page = ERR_PTR(-ENOMEM);
+ 		goto out;
+ 	}
 +	if (flags & FOLL_GET) {


If I'm reading the diff correctly, I believe that line should *maybe* be changed to:

	if (flags & (FOLL_GET | FOLL_PIN)) {

...because each of those flags has a similar effect: pinned pages for DMA or RDMA
use. So either flag will require a call to arch_make_page_accessible()...except that
I'm not sure that's what you want. Would the absence of a call to 
arch_make_page_accessible() cause things like pin_user_pages() to not work correctly?
Seems like it would, to me.

(I'm pretty unhappy that we have to ask this at the linux-next level.)

Also below...


- 		if (unlikely(!try_get_page(page))) {
- 			page = ERR_PTR(-ENOMEM);
- 			goto out;
- 		}
 +		ret = arch_make_page_accessible(page);
 +		if (ret) {
 +			put_page(page);


put_page() only works with FOLL_GET. So if we do allow to get here via either FOLL_GET or
FOLL_PIN, the we need to do an unpin_user_page(), like this:

		if (flags & FOLL_PIN)
			unpin_user_page(page);
		else
			put_page(page);



 +			page = ERR_PTR(ret);
 +			goto out;
 +		}
 +	}
  	if (flags & FOLL_TOUCH) {
  		if ((flags & FOLL_WRITE) &&
  		    !pte_dirty(pte) && !PageDirty(page))

thanks,
-- 
John Hubbard
NVIDIA