Re: locking question: do_mmap(), do_munmap()

Alexander Viro (viro@math.psu.edu)
Mon, 11 Oct 1999 17:37:40 -0400 (EDT)


On Mon, 11 Oct 1999, Manfred Spraul wrote:

> Alexander Viro wrote:
> >
> > On Mon, 11 Oct 1999, Stephen C. Tweedie wrote:
> >
> > > Hi,
> > >
> > > On Sun, 10 Oct 1999 15:03:45 -0400 (EDT), Alexander Viro
> > > <viro@math.psu.edu> said:
> > >
> > > > Hold on. In swap_out_mm() you have to protect find_vma() (OK, it doesn't
> > > > block, but we'll have to take care of mm->mmap_cache) _and_ you'll have to
> > > > protect vma from destruction all way down to try_to_swap_out(). And to
> > > > vma->swapout(). Which can sleep, so spinlocks are out of question
> > > > here.
> > >
> > > No, spinlocks would be ideal. The vma swapout codes _have_ to be
> > > prepared for the vma to be destroyed as soon as we sleep. In fact, the
> > > entire mm may disappear if the process happens to exit. Once we know
> > > which page to write where, the swapout operation becomes a per-page
> > > operation, not per-vma.
> >
> > Aha, so you propose to drop it in ->swapout(), right? (after get_file() in
> > filemap_write_page()... Ouch. Probably we'ld better lambda-expand the call
> > in filemap_swapout() - the thing is called from other places too)...
>
> What about something like a rw-semaphore which protects the vma list:
> vma-list modifiers [ie merge_segments(), insert_vm_struct() and
> do_munmap()] grab it exclusive, swapper grabs it "shared, starve
> exclusive".
> All other vma-list readers are protected by mm->mmap_sem.
>
> This should not dead-lock, and no changes are required in
> vm_ops->swapout().

What does it buy you over the simple semaphore here? Do you really see a
contention scenario?

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/