Re: [PATCH 2/2] mm: set PG_dma_pinned on get_user_pages*()

[John Hubbard]

On 06/20/2018 05:08 AM, Jan Kara wrote:
> On Tue 19-06-18 11:11:48, John Hubbard wrote:
>> On 06/19/2018 03:41 AM, Jan Kara wrote:
>>> On Tue 19-06-18 02:02:55, Matthew Wilcox wrote:
>>>> On Tue, Jun 19, 2018 at 10:29:49AM +0200, Jan Kara wrote:
[...]
>>> I'm also still pondering the idea of inserting a "virtual" VMA into vma
>>> interval tree in the inode - as the GUP references are IMHO closest to an
>>> mlocked mapping - and that would achieve all the functionality we need as
>>> well. I just didn't have time to experiment with it.
>>
>> How would this work? Would it have the same virtual address range? And how
>> does it avoid the problems we've been discussing? Sorry to be a bit slow
>> here. :)
> 
> The range covered by the virtual mapping would be the one sent to
> get_user_pages() to get page references. And then we would need to teach
> page_mkclean() to check for these virtual VMAs and block / skip / report
> (different situations would need different behavior) such page. But this
> second part is the same regardless how we identify a page that is pinned by
> get_user_pages().


OK. That neatly avoids the need a new page flag, I think. But of course it is 
somewhat more extensive to implement. Sounds like something to keep in mind,
in case it has better tradeoffs than the direction I'm heading so far.

 
>>> And then there's the aspect that both these approaches are a bit too
>>> heavyweight for some get_user_pages_fast() users (e.g. direct IO) - Al Viro
>>> had an idea to use page lock for that path but e.g. fs/direct-io.c would have
>>> problems due to lock ordering constraints (filesystem ->get_block would
>>> suddently get called with the page lock held). But we can probably leave
>>> performance optimizations for phase two.
>>
>>  
>> So I assume that phase one would be to apply this approach only to
>> get_user_pages_longterm. (Please let me know if that's wrong.)
> 
> No, I meant phase 1 would be to apply this to all get_user_pages() flavors.
> Then phase 2 is to try to find a way to make get_user_pages_fast() fast
> again. And then in parallel to that, we also need to find a way for
> get_user_pages_longterm() to signal to the user pinned pages must be
> released soon. Because after phase 1 pinned pages will block page
> writeback and such system won't oops but will become unusable
> sooner rather than later. And again this problem needs to be solved
> regardless of a mechanism of identifying pinned pages.
> 

OK, thanks, that does help. I had the priorities of these get_user_pages*()
changes all scrambled, but between your and Dan's explanation, I finally 
understand the preferred ordering of this work.