[Intel-gfx] [PATCH 1/5] drm/i915/userptr: Beware recursive lock_page()

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-09-11 12:31:32)
> 
> On 09/09/2019 14:52, Chris Wilson wrote:
> > Quoting Lionel Landwerlin (2019-07-26 14:38:40)
> >> On 17/07/2019 21:09, Tvrtko Ursulin wrote:
> >>>
> >>> On 17/07/2019 15:06, Chris Wilson wrote:
> >>>> Quoting Tvrtko Ursulin (2019-07-17 14:46:15)
> >>>>>
> >>>>> On 17/07/2019 14:35, Chris Wilson wrote:
> >>>>>> Quoting Tvrtko Ursulin (2019-07-17 14:23:55)
> >>>>>>>
> >>>>>>> On 17/07/2019 14:17, Chris Wilson wrote:
> >>>>>>>> Quoting Tvrtko Ursulin (2019-07-17 14:09:00)
> >>>>>>>>>
> >>>>>>>>> On 16/07/2019 16:37, Chris Wilson wrote:
> >>>>>>>>>> Quoting Tvrtko Ursulin (2019-07-16 16:25:22)
> >>>>>>>>>>>
> >>>>>>>>>>> On 16/07/2019 13:49, Chris Wilson wrote:
> >>>>>>>>>>>> Following a try_to_unmap() we may want to remove the userptr
> >>>>>>>>>>>> and so call
> >>>>>>>>>>>> put_pages(). However, try_to_unmap() acquires the page lock
> >>>>>>>>>>>> and so we
> >>>>>>>>>>>> must avoid recursively locking the pages ourselves -- which
> >>>>>>>>>>>> means that
> >>>>>>>>>>>> we cannot safely acquire the lock around set_page_dirty().
> >>>>>>>>>>>> Since we
> >>>>>>>>>>>> can't be sure of the lock, we have to risk skip dirtying the
> >>>>>>>>>>>> page, or
> >>>>>>>>>>>> else risk calling set_page_dirty() without a lock and so risk fs
> >>>>>>>>>>>> corruption.
> >>>>>>>>>>>
> >>>>>>>>>>> So if trylock randomly fail we get data corruption in whatever
> >>>>>>>>>>> data set
> >>>>>>>>>>> application is working on, which is what the original patch
> >>>>>>>>>>> was trying
> >>>>>>>>>>> to avoid? Are we able to detect the backing store type so at
> >>>>>>>>>>> least we
> >>>>>>>>>>> don't risk skipping set_page_dirty with anonymous/shmemfs?
> >>>>>>>>>>
> >>>>>>>>>> page->mapping???
> >>>>>>>>>
> >>>>>>>>> Would page->mapping work? What is it telling us?
> >>>>>>>>
> >>>>>>>> It basically tells us if there is a fs around; anything that is
> >>>>>>>> the most
> >>>>>>>> basic of malloc (even tmpfs/shmemfs has page->mapping).
> >>>>>>>
> >>>>>>> Normal malloc so anonymous pages? Or you meant everything _apart_
> >>>>>>> from
> >>>>>>> the most basic malloc?
> >>>>>>
> >>>>>> Aye missed the not.
> >>>>>>
> >>>>>>>>>> We still have the issue that if there is a mapping we should be
> >>>>>>>>>> taking
> >>>>>>>>>> the lock, and we may have both a mapping and be inside
> >>>>>>>>>> try_to_unmap().
> >>>>>>>>>
> >>>>>>>>> Is this a problem? On a path with mappings we trylock and so
> >>>>>>>>> solve the
> >>>>>>>>> set_dirty_locked and recursive deadlock issues, and with no
> >>>>>>>>> mappings
> >>>>>>>>> with always dirty the page and avoid data corruption.
> >>>>>>>>
> >>>>>>>> The problem as I see it is !page->mapping are likely an
> >>>>>>>> insignificant
> >>>>>>>> minority of userptr; as I think even memfd are essentially
> >>>>>>>> shmemfs (or
> >>>>>>>> hugetlbfs) and so have mappings.
> >>>>>>>
> >>>>>>> Better then nothing, no? If easy to do..
> >>>>>>
> >>>>>> Actually, I erring on the opposite side. Peeking at mm/ internals does
> >>>>>> not bode confidence and feels indefensible. I'd much rather throw my
> >>>>>> hands up and say "this is the best we can do with the API provided,
> >>>>>> please tell us what we should have done." To which the answer is
> >>>>>> probably to not have used gup in the first place :|
> >>>>>
> >>>>> """
> >>>>> /*
> >>>>>    * set_page_dirty() is racy if the caller has no reference against
> >>>>>    * page->mapping->host, and if the page is unlocked. This is
> >>>>> because another
> >>>>>    * CPU could truncate the page off the mapping and then free the
> >>>>> mapping.
> >>>>>    *
> >>>>>    * Usually, the page _is_ locked, or the caller is a user-space
> >>>>> process which
> >>>>>    * holds a reference on the inode by having an open file.
> >>>>>    *
> >>>>>    * In other cases, the page should be locked before running
> >>>>> set_page_dirty().
> >>>>>    */
> >>>>> int set_page_dirty_lock(struct page *page)
> >>>>> """
> >>>>>
> >>>>> Could we hold a reference to page->mapping->host while having pages
> >>>>> and then would be okay to call plain set_page_dirty?
> >>>>
> >>>> We would then be hitting the warnings in ext4 for unlocked pages again.
> >>>
> >>> Ah true..
> >>>
> >>>> Essentially the argument is whether or not that warn is valid, to
> >>>> which I
> >>>> think requires inner knowledge of vfs + ext4. To hold a reference on the
> >>>> host would require us tracking page->mapping (reasonable since we
> >>>> already hooked into mmu and so will get an invalidate + fresh gup on
> >>>> any changes), plus iterating over all to acquire the extra reference if
> >>>> applicable -- and I have no idea what the side-effects of that would be.
> >>>> Could well be positive side-effects. Just feels like wandering even
> >>>> further off the beaten path without a map. Good news hmm is just around
> >>>> the corner (which will probably prohibit this use-case) :|
> >>>
> >>> ... can we reach out to someone more knowledgeable in mm matters to
> >>> recommend us what to do?
> >>>
> >>> Regards,
> >>>
> >>> Tvrtko
> >>
> >>
> >> Just a reminder to not let this slip.
> >> We run into userptr bugs in CI quite regularly.
> > 
> > Remind away. Revert or trylock, there doesn't seem to be a good answer.
> 
> Rock and a hard place. Data corruption for userptr users (with either 
> trylock or no lock) or a deadlock (with the lock). I honestly can't 
> decide what is worse. Tiny preference to deadlock rather than silent 
> corruption. Misguided? Don't know really..

The deadlock is pretty easy to hit as soon as the system is under
mempressure and it tries to free pages as we do the userptr gup...
(Hah, easy in theory, but not in CI.)
-Chris