lkml.org 
[lkml]   [2019]   [Jan]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/2] mm: introduce put_user_page*(), placeholder versions
On Tue, Jan 15, 2019 at 09:07:59AM +0100, Jan Kara wrote:
> On Mon 14-01-19 12:21:25, Jerome Glisse wrote:
> > On Mon, Jan 14, 2019 at 03:54:47PM +0100, Jan Kara wrote:
> > > On Fri 11-01-19 19:06:08, John Hubbard wrote:
> > > > On 1/11/19 6:46 PM, Jerome Glisse wrote:
> > > > > On Fri, Jan 11, 2019 at 06:38:44PM -0800, John Hubbard wrote:
> > > > > [...]
> > > > >
> > > > >>>> The other idea that you and Dan (and maybe others) pointed out was a debug
> > > > >>>> option, which we'll certainly need in order to safely convert all the call
> > > > >>>> sites. (Mirror the mappings at a different kernel offset, so that put_page()
> > > > >>>> and put_user_page() can verify that the right call was made.) That will be
> > > > >>>> a separate patchset, as you recommended.
> > > > >>>>
> > > > >>>> I'll even go as far as recommending the page lock itself. I realize that this
> > > > >>>> adds overhead to gup(), but we *must* hold off page_mkclean(), and I believe
> > > > >>>> that this (below) has similar overhead to the notes above--but is *much* easier
> > > > >>>> to verify correct. (If the page lock is unacceptable due to being so widely used,
> > > > >>>> then I'd recommend using another page bit to do the same thing.)
> > > > >>>
> > > > >>> Please page lock is pointless and it will not work for GUP fast. The above
> > > > >>> scheme do work and is fine. I spend the day again thinking about all memory
> > > > >>> ordering and i do not see any issues.
> > > > >>>
> > > > >>
> > > > >> Why is it that page lock cannot be used for gup fast, btw?
> > > > >
> > > > > Well it can not happen within the preempt disable section. But after
> > > > > as a post pass before GUP_fast return and after reenabling preempt then
> > > > > it is fine like it would be for regular GUP. But locking page for GUP
> > > > > is also likely to slow down some workload (with direct-IO).
> > > > >
> > > >
> > > > Right, and so to crux of the matter: taking an uncontended page lock
> > > > involves pretty much the same set of operations that your approach does.
> > > > (If gup ends up contended with the page lock for other reasons than these
> > > > paths, that seems surprising.) I'd expect very similar performance.
> > > >
> > > > But the page lock approach leads to really dramatically simpler code (and
> > > > code reviews, let's not forget). Any objection to my going that
> > > > direction, and keeping this idea as a Plan B? I think the next step will
> > > > be, once again, to gather some performance metrics, so maybe that will
> > > > help us decide.
> > >
> > > FWIW I agree that using page lock for protecting page pinning (and thus
> > > avoid races with page_mkclean()) looks simpler to me as well and I'm not
> > > convinced there will be measurable difference to the more complex scheme
> > > with barriers Jerome suggests unless that page lock contended. Jerome is
> > > right that you cannot just do lock_page() in gup_fast() path. There you
> > > have to do trylock_page() and if that fails just bail out to the slow gup
> > > path.
> > >
> > > Regarding places other than page_mkclean() that need to check pinned state:
> > > Definitely page migration will want to check whether the page is pinned or
> > > not so that it can deal differently with short-term page references vs
> > > longer-term pins.
> > >
> > > Also there is one more idea I had how to record number of pins in the page:
> > >
> > > #define PAGE_PIN_BIAS 1024
> > >
> > > get_page_pin()
> > > atomic_add(&page->_refcount, PAGE_PIN_BIAS);
> > >
> > > put_page_pin();
> > > atomic_add(&page->_refcount, -PAGE_PIN_BIAS);
> > >
> > > page_pinned(page)
> > > (atomic_read(&page->_refcount) - page_mapcount(page)) > PAGE_PIN_BIAS
> > >
> > > This is pretty trivial scheme. It still gives us 22-bits for page pins
> > > which should be plenty (but we should check for that and bail with error if
> > > it would overflow). Also there will be no false negatives and false
> > > positives only if there are more than 1024 non-page-table references to the
> > > page which I expect to be rare (we might want to also subtract
> > > hpage_nr_pages() for radix tree references to avoid excessive false
> > > positives for huge pages although at this point I don't think they would
> > > matter). Thoughts?
> >
> > Racing PUP are as likely to cause issues:
> >
> > CPU0 | CPU1 | CPU2
> > | |
> > | PUP() |
> > page_pinned(page) | |
> > (page_count(page) - | |
> > page_mapcount(page)) | |
> > | | GUP()
> >
> > So here the refcount snap-shot does not include the second GUP and
> > we can have a false negative ie the page_pinned() will return false
> > because of the PUP happening just before on CPU1 despite the racing
> > GUP on CPU2 just after.
> >
> > I believe only either lock or memory ordering with barrier can
> > guarantee that we do not miss GUP ie no false negative. Still the
> > bias idea might be usefull as with it we should not need a flag.
>
> Right. We need similar synchronization (i.e., page lock or careful checks
> with memory barriers) if we want to get a reliable page pin information.
>
> > So to make the above safe it would still need the page write back
> > double check that i described so that GUP back-off if it raced with
> > page_mkclean,clear_page_dirty_for_io and the fs write page call back
> > which call test_set_page_writeback() (yes it is very unlikely but
> > might still happen).
>
> Agreed. So with page lock it would actually look like:
>
> get_page_pin()
> lock_page(page);
> wait_for_stable_page();
> atomic_add(&page->_refcount, PAGE_PIN_BIAS);
> unlock_page(page);
>
> And if we perform page_pinned() check under page lock, then if
> page_pinned() returned false, we are sure page is not and will not be
> pinned until we drop the page lock (and also until page writeback is
> completed if needed).
>

So i still can't see anything wrong with that idea, i had similar
one in the past and diss-missed and i can't remember why :( But
thinking over and over i do not see any issue beside refcount wrap
around. Which is something that can happens today thought i don't
think it can be use in an evil way and we can catch it and be
loud about it.

So i think the following would be bullet proof:


get_page_pin()
atomic_add(&page->_refcount, PAGE_PIN_BIAS);
smp_wmb();
if (PageWriteback(page)) {
// back off
atomic_add(&page->_refcount, -PAGE_PIN_BIAS);
// re-enable preempt if in fast
wait_on_page_writeback(page);
goto retry;
}

put_page_pin();
atomic_add(&page->_refcount, -PAGE_PIN_BIAS);

page_pinned(page)
(atomic_read(&page->_refcount) - page_mapcount(page)) > PAGE_PIN_BIAS

test_set_page_writeback()
...
wb = TestSetPageWriteback(page)
smp_mb();
if (page_pinned(page)) {
// report page as pinned to caller of test_set_page_writeback()
}
...

This is text book memory barrier. Either get_page_pin() see racing
test_set_page_writeback() or test_set_page_writeback() see racing GUP


An optimization for GUP:
get_page_pin()
pwp = PageWriteback(page);
smp_rmb();
waspinned = page_pinned(page);
if (!waspinned && pwp) {
// backoff
}

atomic_add(&page->_refcount, PAGE_PIN_BIAS);
smp_wmb();
if (PageWriteback(page)) {
// back off
atomic_add(&page->_refcount, -PAGE_PIN_BIAS);
// re-enable preempt if in fast
wait_on_page_writeback(page);
goto retry;
}

If page was not pin prior to this GUP than we can back off early.


Anyway i think this is better than mapcount. I started an analysis
of all places that were looking at mapcount a few of them would have
need an update if we were to increment mapcount with GUP.

I will go take a look at THP and hugetlbfs in respect to this just
to check for way to mitigate false positive.

Cheers,
Jérôme

\
 
 \ /
  Last update: 2019-01-15 18:16    [W:0.257 / U:5.288 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site