Date:   Thu, 13 Apr 2023 14:41:43 +0200
From:   David Hildenbrand <david@...hat.com>
To:     David Howells <dhowells@...hat.com>
Cc:     "Teterevkov, Ivan" <Ivan.Teterevkov@....com>,
        "linux-mm@...ck.org" <linux-mm@...ck.org>,
        "jhubbard@...dia.com" <jhubbard@...dia.com>,
        "jack@...e.cz" <jack@...e.cz>,
        "rppt@...ux.ibm.com" <rppt@...ux.ibm.com>,
        "jglisse@...hat.com" <jglisse@...hat.com>,
        "ira.weiny@...el.com" <ira.weiny@...el.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Christoph Hellwig <hch@....de>,
        Matthew Wilcox <willy@...radead.org>
Subject: Re: find_get_page() VS pin_user_pages()

On 12.04.23 10:41, David Howells wrote:
> David Hildenbrand <david@...hat.com> wrote:
> 
>> I suspect that find_get_page() is not the kind of interface you want to use
>> for the purpose you describe. find_get_page() is a wrapper around
>> pagecache_get_page() and seems more like a helper for implementing an fs
>> (looking at the users and the fact that it only considers pages that are in
>> the pagecache).
> 
> Btw, at some point we're going to need public functions to get extra pins on
> pages.  vmsplice() should be pinning the pages it pushes into a pipe - so all
> pages in a pipe should probably be pinned - and anyone who splices a page out
> of a pipe and retains it (skbuffs spring strongly to mind) should also get a
> pin on the page.

As discussed, vmsplice() is a bit special, because it has 
longterm-pinning semantics: we'd want to migrate the page out of 
ZONE_MOVABLE/MIGRATE_CMA/... because the page might remain pinned in the 
pipe possibly forever, controlled by user space. 
pin_user_pages(FOLL_LONGTERM) would do the right thing, but we might 
ahve to be careful with extra pins.


I guess it depends on what we want to achieve. Let's discuss what would 
happen when we want to pin some page (and not going via pin_user_page()) 
that's definitely not an anon page -- so let's assume a pagecache page:

(a) Short-term pinning when already pinned (extra pins): easy.
(b) Short-term pinning when not pinned yet: should be fairly easy
     (pin_user_pages() doesn't do anything special for pagecache pages
      either).
(c) Long-term pinning when already long-term pinned (extra long-term
     pinnings): easy
(d) Long-term pinning when already short-term pinned: problematic,
     because we might have to migrate the page first, but it's already
     pinned ... and if we obtained the page via pin_user_page() from a
     MAP_PRIVATE VMA, we'd  have to do another
     pin_user_page(FOLL_LONGTERM) that would properly break COW and give
     us an anon page ...
(e) Long-term pinning when not pinned yet: fairly easy, but we might
     have to migrate the page first (like FOLL_LONGTERM would).


Regarding anon pages, we should pin only via pin_user_page(), so the 
"not pinned" case does not apply. Replicating pins -- (a) and (c) -- is 
usually easy, but (d) is similarly problematic.

Focusing again on !anon pages: if it's just "get another short-term pin 
on an already pinned page", it's easy (and I recall John H. had 
patches). If it's "get a long-term pin on an already pinned page", it 
can be problematic.

Any pages that will never have to be migrated when long-term pinning 
(just some allocated kernel page without MOVABLE semantics) are super 
easy to pin, and to add extra pins to.

> 
> So should all pages held by an skbuff be pinned rather than ref'd?  I have a
> patch to use the bottom two bits of an skb frag's page pointer to keep track
> of whether the page it points to is ref'd, pinned or neither, but if we can
> make it pin/not-pin them, I only need one bit for that.

It might possibly be the right thing. But ref'd vs. pinned really only 
makes a difference to (a) pages mapped into user space or (b) pages in 
the pageache. Of course, in any case, long-term semantics have to be 
respected if the page to pin might have been allocated with MOVABLE 
semantics.

-- 
Thanks,

David / dhildenb

Powered by blists - more mailing lists