Re: [PATCH v4 0/1] Use HMM for ODP v4
From: Jason Gunthorpe
Date: Wed May 22 2019 - 18:46:09 EST
On Wed, May 22, 2019 at 05:49:18PM -0400, Jerome Glisse wrote:
> > > > So why is mm suddenly guarenteed valid? It was a bug report that
> > > > triggered the race the mmget_not_zero is fixing, so I need a better
> > > > explanation why it is now safe. From what I see the hmm_range_fault
> > > > is doing stuff like find_vma without an active mmget??
> > >
> > > So the mm struct can not go away as long as we hold a reference on
> > > the hmm struct and we hold a reference on it through both hmm_mirror
> > > and hmm_range struct. So struct mm can not go away and thus it is
> > > safe to try to take its mmap_sem.
> >
> > This was always true here, though, so long as the umem_odp exists the
> > the mm has a grab on it. But a grab is not a get..
> >
> > The point here was the old code needed an mmget() in order to do
> > get_user_pages_remote()
> >
> > If hmm does not need an external mmget() then fine, we delete this
> > stuff and rely on hmm.
> >
> > But I don't think that is true as we have:
> >
> > CPU 0 CPU1
> > mmput()
> > __mmput()
> > exit_mmap()
> > down_read(&mm->mmap_sem);
> > hmm_range_dma_map(range, device,..
> > ret = hmm_range_fault(range, block);
> > if (hmm->mm == NULL || hmm->dead)
> > mmu_notifier_release()
> > hmm->dead = true
> > vma = find_vma(hmm->mm, start);
> > .. rb traversal .. while (vma) remove_vma()
> >
> > *goes boom*
> >
> > I think this is violating the basic constraint of the mm by acting on
> > a mm's VMA's without holding a mmget() to prevent concurrent
> > destruction.
> >
> > In other words, mmput() destruction does not respect the mmap_sem - so
> > holding the mmap sem alone is not enough locking.
> >
> > The unlucked hmm->dead simply can't save this. Frankly every time I
> > look a struct with 'dead' in it, I find races like this.
> >
> > Thus we should put the mmget_notzero back in.
>
> So for some reason i thought exit_mmap() was setting the mm_rb
> to empty node and flushing vmacache so that find_vma() would
> fail.
It would still be racy without locks.
> Note that right before find_vma() there is also range->valid
> check which will also intercept mm release.
There is no locking on range->valid so it is just moves the race
around. You can't solve races with unlocked/non-atomic variables.
> Anyway the easy fix is to get ref on mm user in range_register.
Yes a mmget_not_zero inside range_register would be fine.
How do you want to handle that patch?
> > I saw some other funky looking stuff in hmm as well..
> >
> > > Hence it is safe to take mmap_sem and it is safe to call in hmm, if
> > > mm have been kill it will return EFAULT and this will propagate to
> > > RDMA.
> >
> > > As per_mm i removed the per_mm->mm = NULL from release so that it is
> > > always safe to use that field even in face of racing mm "killing".
> >
> > Yes, that certainly wasn't good.
> >
> > > > > - * An array of the pages included in the on-demand paging umem.
> > > > > - * Indices of pages that are currently not mapped into the device will
> > > > > - * contain NULL.
> > > > > + * An array of the pages included in the on-demand paging umem. Indices
> > > > > + * of pages that are currently not mapped into the device will contain
> > > > > + * 0.
> > > > > */
> > > > > - struct page **page_list;
> > > > > + uint64_t *pfns;
> > > >
> > > > Are these actually pfns, or are they mangled with some shift? (what is range->pfn_shift?)
> > >
> > > They are not pfns they have flags (hence range->pfn_shift) at the
> > > bottoms i just do not have a better name for this.
> >
> > I think you need to have a better name then
>
> Suggestion ? i have no idea for a better name, it has pfn value
> in it.
pfn_flags?
Jason