Re: [PATCH 0/2] x86/mm: remove arch-specific PTE/PMD get-and-clear functions

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Wed, Jun 13, 2012 at 04:00:23PM +0100, David Vrabel wrote:
> On 13/06/12 15:04, Konrad Rzeszutek Wilk wrote:
> > On Wed, Jun 13, 2012 at 11:20:43AM +0100, David Vrabel wrote:
> >> This series removes the x86-specific implementation of
> >> ptep_get_and_clear() and pmdp_get_and_clear().
> >>
> >> The principal reason for this is it allows Xen paravitualized guests
> >> to batch the PTE clears which is a significant performance
> >> optimization of munmap() and mremap() -- the number of entries into
> >> the hypervisor is reduced by about a factor of about 30 (60 in 32-bit
> >> guests) for munmap().
> >>
> >> There may be minimal gains on native and KVM guests due to the removal
> >> of the locked xchg.
> > 
> > What about lguest?
> 
> As I note in the description of patch 1:
> 
> "There may be a performance regression with lguest guests as
> an optimization for avoiding calling pte_update() when doing a full
> teardown of an mm is removed."
> 
> I don't know how much this performance regression would be or if the
> performance of lguest guests is something people care about.
> 
> We could have an x86-specific ptep_get_and_clear_full() which looks like:
> 
> pte_t ptep_get_and_clear_full(
> 	struct mm_struct *mm, unsigned long addr, pte_t *ptep,
> 	int full)
> {
> 	pte_t pte = *ptep;
> 
> 	pte_clear(mm, address, ptep);
> 	if (!full)
> 		pte_update(mm, addr, ptep);
> 
> 	return pte;
> }
> 
> Which would have all the performance benefits of the proposed patch
> without the performance regression with lguest.

Lets rope Rusty in this since he is the maintainer of lguest.
> 
> David
> 
> >>
> >> Removal of arch-specific functions where generic ones are suitable
> >> seems to be a generally useful thing to me.
> >>
> >> The full reasoning for why this is safe is included in the commit
> >> message of patch 1 but to summarize.  The atomic get-and-clear does
> >> not guarantee that the latest dirty/accessed bits are returned as TLB
> >> as there is a still a window after the get-and-clear and before the
> >> TLB flush that the bits may be updated on other processors.  So, user
> >> space applications accessing pages that are being unmapped or remapped
> >> already have unpredictable behaviour.
> >>
> >> David