Re: [PATCH v2 12/15] KVM: MMU: allow locklessly access shadow page table out of vcpu thread

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Oct 10, 2013 at 10:16:46PM +0300, Gleb Natapov wrote:
> On Thu, Oct 10, 2013 at 01:42:22PM -0300, Marcelo Tosatti wrote:
> > On Thu, Oct 10, 2013 at 03:08:45PM +0300, Gleb Natapov wrote:
> > > On Wed, Oct 09, 2013 at 10:47:10PM -0300, Marcelo Tosatti wrote:
> > > > > >> Gleb has a idea that uses RCU_DESTORY to protect the shadow page table
> > > > > >> and encodes the page-level into the spte (since we need to check if the spte
> > > > > >> is the last-spte. ).  How about this?
> > > > > > 
> > > > > > Pointer please? Why is DESTROY_SLAB_RCU any safer than call_rcu with
> > > > > > regards to limitation? (maybe it is).
> > > > > 
> > > > > For my experience, freeing shadow page and allocing shadow page are balanced,
> > > > > we can check it by (make -j12 on a guest with 4 vcpus and):
> > > > > 
> > > > > # echo > trace
> > > > > [root@eric-desktop tracing]# cat trace > ~/log | sleep 3
> > > > > [root@eric-desktop tracing]# cat ~/log | grep new | wc -l
> > > > > 10816
> > > > > [root@eric-desktop tracing]# cat ~/log | grep prepare | wc -l
> > > > > 10656
> > > > > [root@eric-desktop tracing]# cat set_event
> > > > > kvmmmu:kvm_mmu_get_page
> > > > > kvmmmu:kvm_mmu_prepare_zap_page
> > > > > 
> > > > > alloc VS. free = 10816 : 10656
> > > > > 
> > > > > So that, mostly all allocing and freeing are done in the slab's
> > > > > cache and the slab frees shdadow pages very slowly, there is no rcu issue.
> > > > 
> > > > A more detailed test case would be:
> > > > 
> > > > - cpu0-vcpu0 releasing pages as fast as possible
> > > > - cpu1 executing get_dirty_log
> > > > 
> > > > Think of a very large guest.
> > > > 
> > > The number of shadow pages allocated from slab will be bounded by
> > > n_max_mmu_pages, 
> > 
> > Correct, but that limit is not suitable (maximum number of mmu pages
> > should be larger than number of mmu pages freeable in a rcu grace
> > period).
> > 
> I am not sure I understand what you mean here. What I was sating is that if
> we change code to allocate sp->spt from slab, this slab will never have
> more then n_max_mmu_pages objects in it.

n_max_mmu_pages is not a suitable limit to throttle freeing of pages via
RCU (its too large). If the free memory watermarks are smaller than 
n_max_mmu_pages for all guests, OOM is possible.

> > > and, in addition, page released to slab is immediately
> > > available for allocation, no need to wait for grace period. 
> > 
> > See SLAB_DESTROY_BY_RCU comment at include/linux/slab.h.
> > 
> This comment is exactly what I was referring to in the code you quoted. Do
> you see anything problematic in what comment describes?

"This delays freeing the SLAB page by a grace period, it does _NOT_
delay object freeing." The page is not available for allocation.