Re: [PATCH] mm: thp: use down_read_trylock in khugepaged to avoid long block

[Kirill A. Shutemov]

On Mon, Dec 18, 2017 at 09:41:19AM +0100, Michal Hocko wrote:
> On Sat 16-12-17 23:09:25, Kirill A. Shutemov wrote:
> > On Sat, Dec 16, 2017 at 12:45:25PM +0100, Michal Hocko wrote:
> > > On Sat 16-12-17 04:04:10, Yang Shi wrote:
> [...]
> > > > Shall we add "cond_resched()" in unmap_vmas(), i.e for every 100 vmas? It
> > > > may improve the responsiveness a little bit for non-preempt kernel, although
> > > > it still can't release the semaphore.
> > > 
> > > We already do, once per pmd (see zap_pmd_range).
> > 
> > It doesn't help. We would need to find a way to drop mmap_sem, if we're
> > holding it way too long. And doing it on per-vma count basis is not right
> > call. It won't address issue with single huge vma.
> 
> Absolutely agreed. I just wanted to point out that a new cond_resched is
> not really needed. One way to reduce the lock starvation is to use range
> locking.
> 
> > Do we have any instrumentation that would help detect starvation on a
> > rw_semaphore?
> 
> I am afraid we don't.

I guess we have enough info in mmu_gather to decide if we are doing munmap way
too long. Although, getting everything right would be tricky...

-- 
 Kirill A. Shutemov