Re: [PATCH v3 04/20] mm: VMA sequence count

Hi,

On 14/09/2017 02:31, Sergey Senozhatsky wrote:

Hi,

On (09/13/17 18:56), Laurent Dufour wrote:

quoted

Hi Sergey,

On 13/09/2017 13:53, Sergey Senozhatsky wrote:

quoted

Hi,

On (09/08/17 20:06), Laurent Dufour wrote:

[..]

quoted

quoted

ok, so what I got on my box is:

vm_munmap()  -> down_write_killable(&mm->mmap_sem)
 do_munmap()
  __split_vma()
   __vma_adjust()  -> write_seqcount_begin(&vma->vm_sequence)
                   -> write_seqcount_begin_nested(&next->vm_sequence, SINGLE_DEPTH_NESTING)

so this gives 3 dependencies  ->mmap_sem   ->   ->vm_seq
                              ->vm_seq     ->   ->vm_seq/1
                              ->mmap_sem   ->   ->vm_seq/1


SyS_mremap() -> down_write_killable(&current->mm->mmap_sem)
 move_vma()   -> write_seqcount_begin(&vma->vm_sequence)
              -> write_seqcount_begin_nested(&new_vma->vm_sequence, SINGLE_DEPTH_NESTING);
  move_page_tables()
   __pte_alloc()
    pte_alloc_one()
     __alloc_pages_nodemask()
      fs_reclaim_acquire()


I think here we have prepare_alloc_pages() call, that does

        -> fs_reclaim_acquire(gfp_mask)
        -> fs_reclaim_release(gfp_mask)

so that adds one more dependency  ->mmap_sem   ->   ->vm_seq    ->   fs_reclaim
                                  ->mmap_sem   ->   ->vm_seq/1  ->   fs_reclaim


now, under memory pressure we hit the slow path and perform direct
reclaim. direct reclaim is done under fs_reclaim lock, so we end up
with the following call chain

__alloc_pages_nodemask()
 __alloc_pages_slowpath()
  __perform_reclaim()       ->   fs_reclaim_acquire(gfp_mask);
   try_to_free_pages()
    shrink_node()
     shrink_active_list()
      rmap_walk_file()      ->   i_mmap_lock_read(mapping);


and this break the existing dependency. since we now take the leaf lock
(fs_reclaim) first and the the root lock (->mmap_sem).

Thanks for looking at this.
I'm sorry, I should have miss something.

no prob :)

quoted

My understanding is that there are 2 chains of locks:
 1. from __vma_adjust() mmap_sem -> i_mmap_rwsem -> vm_seq
 2. from move_vmap() mmap_sem -> vm_seq -> fs_reclaim
 2. from __alloc_pages_nodemask() fs_reclaim -> i_mmap_rwsem

yes, as far as lockdep warning suggests.

quoted

So the solution would be to have in __vma_adjust()
 mmap_sem -> vm_seq -> i_mmap_rwsem

But this will raised the following dependency from  unmap_mapping_range()
unmap_mapping_range() 		-> i_mmap_rwsem
 unmap_mapping_range_tree()
  unmap_mapping_range_vma()
   zap_page_range_single()
    unmap_single_vma()
     unmap_page_range()	 	-> vm_seq

And there is no way to get rid of it easily as in unmap_mapping_range()
there is no VMA identified yet.

That's being said I can't see any clear way to get lock dependency cleaned
here.
Furthermore, this is not clear to me how a deadlock could happen as vm_seq
is a sequence lock, and there is no way to get blocked here.

as far as I understand,
   seq locks can deadlock, technically. not on the write() side, but on
the read() side:

read_seqcount_begin()
 raw_read_seqcount_begin()
   __read_seqcount_begin()

and __read_seqcount_begin() spins for ever

   __read_seqcount_begin()
   {
    repeat:
     ret = READ_ONCE(s->sequence);
     if (unlikely(ret & 1)) {
         cpu_relax();
         goto repeat;
     }
     return ret;
   }


so if there are two CPUs, one doing write_seqcount() and the other one
doing read_seqcount() then what can happen is something like this

	CPU0					CPU1

						fs_reclaim_acquire()
	write_seqcount_begin()
	fs_reclaim_acquire()			read_seqcount_begin()
	write_seqcount_end()

CPU0 can't write_seqcount_end() because of fs_reclaim_acquire() from
CPU1, CPU1 can't read_seqcount_begin() because CPU0 did write_seqcount_begin()
and now waits for fs_reclaim_acquire(). makes sense?

Yes, this makes sense.

But in the case of this series, there is no call to
__read_seqcount_begin(), and the reader (the speculative page fault
handler), is just checking for (vm_seq & 1) and if this is true, simply
exit the speculative path without waiting.
So there is no deadlock possibility.

The bad case would be to have 2 concurrent threads calling
write_seqcount_begin() on the same VMA, leading a wrongly freed sequence
lock but this can't happen because of the mmap_sem holding for write in
such a case.

Cheers,
Laurent.