Re: [RFC PATCH 0/4 v2] Define killable version for access_remote_vm() and use it in fs/proc

[Yang Shi]

On 3/6/18 1:41 PM, Andrew Morton wrote:
> On Tue, 6 Mar 2018 13:17:37 -0800 Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote:
>
> > It just mitigates the hung task warning, can't resolve the mmap_sem
> > scalability issue. Furthermore, waiting on pure uninterruptible state
> > for reading /proc sounds unnecessary. It doesn't wait for I/O completion.
> OK.

Since we already had down_read_killable() APIs available, IMHO, giving 
application a chance to abort at some circumstances sounds not bad.

> > > Where the heck are we holding mmap_sem for so long?  Can that be fixed?
> > The mmap_sem is held for unmapping a large map which has every single
> > page mapped. This is not a issue in real production code. Just found it
> > by running vm-scalability on a machine with ~600GB memory.
> >
> > AFAIK, I don't see any easy fix for the mmap_sem scalability issue. I
> > saw range locking patches (https://lwn.net/Articles/723648/) were
> > floating around. But, it may not help too much on the case that a large
> > map with every single page mapped.
> Well it sounds fairly simple to mitigate?  Simplistically: don't unmap
> 600G in a single hit; do it 1G at a time, dropping mmap_sem each time.
> A smarter version might only come up for air if there are mmap_sem
> waiters and if it has already done some work.  I don't think we have
> any particular atomicity requirements when unmapping?

I'm not quite sure. But, the existing applications may assume munmap is 
atomic?

Thanks,
Yang

>