Re: [PATCH v8 0/9] rwsem performance optimizations

[Davidlohr Bueso <davidlohr@hp.com>]

On Wed, 2013-10-09 at 20:14 -0700, Linus Torvalds wrote:
> On Wed, Oct 9, 2013 at 12:28 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > The workload that I got the report from was a virus scanner, it would
> > spawn nr_cpus threads and {mmap file, scan content, munmap} through your
> > filesystem.
> 
> So I suspect we could make the mmap_sem write area *much* smaller for
> the normal cases.
> 
> Look at do_mmap_pgoff(), for example: it is run entirely under
> mmap_sem, but 99% of what it does doesn't actually need the lock.
> 
> The part that really needs the lock is
> 
>         addr = get_unmapped_area(file, addr, len, pgoff, flags);
>         addr = mmap_region(file, addr, len, vm_flags, pgoff);
> 
> but we hold it over all the other stuff too.
> 

True. By looking at the callers, we're always doing:

down_write(&mm->mmap_sem);
do_mmap_pgoff()
...
up_write(&mm->mmap_sem);

That goes for shm, aio, and of course mmap_pgoff().

While I know you hate two level locking, one way to go about this is to
take the lock inside do_mmap_pgoff() after the initial checks (flags,
page align, etc.) and return with the lock held, leaving the caller to
unlock it. 

> In fact, even if we moved the mmap_sem down into do_mmap(), and moved
> code around a bit to only hold it over those functions, it would still
> cover unnecessarily much. For example, while merging is common, not
> merging is pretty common too, and we do that
> 
>         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
> 
> allocation under the lock. We could easily do things like preallocate
> it outside the lock.
> 

AFAICT there are also checks that should be done at the beginning of the
function, such as checking for MAP_LOCKED and VM_LOCKED flags before
calling get_unmapped_area().

Thanks,
Davidlohr