[PATCH 2/5] kernel.h: Add non_block_start/end()

[Michal Hocko]

On Thu 15-08-19 17:13:23, Jason Gunthorpe wrote:
> On Thu, Aug 15, 2019 at 09:35:26PM +0200, Michal Hocko wrote:
> 
> > > The last detail is I'm still unclear what a GFP flags a blockable
> > > invalidate_range_start() should use. Is GFP_KERNEL OK?
> > 
> > I hope I will not make this muddy again ;)
> > invalidate_range_start in the blockable mode can use/depend on any sleepable
> > allocation allowed in the context it is called from. 
> 
> 'in the context is is called from' is the magic phrase, as
> invalidate_range_start is called while holding several different mm
> related locks. I know at least write mmap_sem and i_mmap_rwsem
> (write?)
> 
> Can GFP_KERNEL be called while holding those locks?

i_mmap_rwsem would be problematic because it is taken during the
reclaim.

> This is the question of indirect dependency on reclaim via locks you
> raised earlier.
> 
> > So in other words it is no different from any other function in the
> > kernel that calls into allocator. As the API is missing gfp context
> > then I hope it is not called from any restricted contexts (except
> > from the oom which we have !blockable for).
> 
> Yes, the callers are exactly my concern.
>  
> > > Lockdep has
> > > complained on that in past due to fs_reclaim - how do you know if it
> > > is a false positive?
> > 
> > I would have to see the specific lockdep splat.
> 
> See below. I found it when trying to understand why the registration
> of the mmu notififer was so oddly coded.
> 
> The situation was:
> 
>   down_write(&mm->mmap_sem);
>   mm_take_all_locks(mm);
>   kmalloc(GFP_KERNEL);  <--- lockdep warning

Ugh. mm_take_all_locks :/

> I understood Daniel said he saw this directly on a recent kernel when
> working with his lockdep patch?
> 
> Checking myself, on todays kernel I see a call chain:
> 
> shrink_all_memory
>   fs_reclaim_acquire(sc.gfp_mask);
>   [..]
>   do_try_to_free_pages
>    shrink_zones
>     shrink_node
>      shrink_node_memcg
>       shrink_list
>        shrink_active_list
>         page_referenced
>          rmap_walk
>           rmap_walk_file
>            i_mmap_lock_read
>             down_read(i_mmap_rwsem)
> 
> So it is possible that the down_read() above will block on
> i_mmap_rwsem being held in the caller of invalidate_range_start which
> is doing kmalloc(GPF_KERNEL).
> 
> Is this OK? The lockdep annotation says no..

It's not as per the above code patch which is easily possible because
mm_take_all_locks will lock all file vmas.

-- 
Michal Hocko
SUSE Labs