Re: [PATCH v3 17/35] mm/mmap: write-lock VMA before shrinking or expanding it

[Liam R. Howlett]

* Suren Baghdasaryan <surenb@xxxxxxxxxx> [230223 21:06]:
> On Thu, Feb 23, 2023 at 5:46 PM Liam R. Howlett <Liam.Howlett@xxxxxxxxxx> wrote:
> >
> > * Suren Baghdasaryan <surenb@xxxxxxxxxx> [230223 16:16]:
> > > On Thu, Feb 23, 2023 at 12:28 PM Liam R. Howlett
> > > <Liam.Howlett@xxxxxxxxxx> wrote:
> > > >
> > > >
> > > > Wait, I figured a better place to do this.
> > > >
> > > > init_multi_vma_prep() should vma_start_write() on any VMA that is passed
> > > > in.. that we we catch any modifications here & in vma_merge(), which I
> > > > think is missed in this patch set?
> > >
> > > Hmm. That looks like a good idea but in that case, why not do the
> > > locking inside vma_prepare() itself? From the description of that
> > > function it sounds like it was designed to acquire locks before VMA
> > > modifications, so would be the ideal location for doing that. WDYT?
> >
> > That might be even better.  I think it will result in even less code.
> 
> Yes.
> 
> >
> > There is also a vma_complete() which might work to call
> > vma_end_write_all() as well?
> 
> If there are other VMAs already locked before vma_prepare() then we
> would unlock them too. Safer to just let mmap_unlock do
> vma_end_write_all().
> 
> >
> > > The only concern is vma_adjust_trans_huge() being called before
> > > vma_prepare() but I *think* that's safe because
> > > vma_adjust_trans_huge() does its modifications after acquiring PTL
> > > lock, which page fault handlers also have to take. Does that sound
> > > right?
> >
> > I am not sure.  We are certainly safe the way it is, and the PTL has to
> > be safe for concurrent faults.. but this could alter the walk to a page
> > table while that walk is occurring and I don't think that happens today.
> >
> > It might be best to leave the locking order the way you have it, unless
> > someone can tell us it's safe?
> 
> Yes, I have the same feelings about changing this.
> 
> >
> > We could pass through the three extra variables that are needed to move
> > the vma_adjust_trans_huge() call within that function as well?  This
> > would have the added benefit of having all locking grouped in the one
> > location, but the argument list would be getting long, however we could
> > use the struct.
> 
> Any issues if I change the order to have vma_prepare() called always
> before vma_adjust_trans_huge()? That way the VMA will always be locked
> before vma_adjust_trans_huge() executes and we don't need any
> additional arguments.

I preserved the locking order from __vma_adjust() to ensure there was no
issues.

I am not sure but, looking through the page table information [1], it
seems that vma_adjust_trans_huge() uses the pmd lock, which is part of
the split page table lock.  According to the comment in rmap, it should
be fine to reverse the ordering here.

Instead of:

mmap_lock()
vma_adjust_trans_huge()
	pte_lock
	pte_unlock

vma_prepare()
	mapping->i_mmap_rwsem lock
	anon_vma->rwsem lock

<changes to tree/VMAs>

vma_complete()
	anon_vma->rwsem unlock
	mapping->i_mmap_rwsem unlock

mmap_unlock()

---------

We would have:

mmap_lock()
vma_prepare()
	mapping->i_mmap_rwsem lock
	anon_vma->rwsem lock

vma_adjust_trans_huge()
	pte_lock
	pte_unlock

<changes to tree/VMAs>

vma_complete()
	anon_vma->rwsem unlock
	mapping->i_mmap_rwsem unlock

mmap_unlock()


Essentially, increasing the nesting of the pte lock, but not violating
the ordering.

1. https://docs.kernel.org/mm/split_page_table_lock.html

> 
> >
> > remove & remove2 should be be detached in vma_prepare() or
> > vma_complete() as well?
> 
> They are marked detached in vma_complete() (see
> https://lore.kernel.org/all/20230216051750.3125598-25-surenb@xxxxxxxxxx/)
> and that should be enough. We should be safe as long as we mark them
> detached before unlocking the VMA.
> 

Right, Thanks.

...