On Tue, Jun 19, 2012 at 6:41 AM, Andrea Arcangeli <aarcange at redhat.com> wrote: > On Tue, Jun 19, 2012 at 12:32:06PM +0300, Avi Kivity wrote: >> On 06/19/2012 01:20 AM, Christoffer Dall wrote: >> > On Mon, Jun 18, 2012 at 9:45 AM, Avi Kivity <avi at redhat.com> wrote: >> >> On 06/15/2012 10:09 PM, Christoffer Dall wrote: >> >>> From: Christoffer Dall <cdall at cs.columbia.edu> >> >>> >> >>> Handles the guest faults in KVM by mapping in corresponding user pages >> >>> in the 2nd stage page tables. >> >>> >> >>> Introduces new ARM-specific kernel memory types, PAGE_KVM_GUEST and >> >>> pgprot_guest variables used to map 2nd stage memory for KVM guests. >> >>> >> >>> Leverages MMU notifiers on KVM/ARM by supporting the kvm_unmap_hva() operation, >> >>> where we remove the HVA from the 2nd stage translation. All other KVM MMU >> >>> notifierhooks are NOPs. >> >> >> >> I think you must at least support change_pte (possibly by unmapping). >> >> Andrea? >> >> >> > hmmm, at least for KSM support we need to support change_pte (are >> > there other callers for this type of memory?) >> > >> > It's not trivial I guess, since we would need to support COW and >> > thereby stage-2 permission faults... Marc, right? >> >> As I mentioned, you can support change_pte by unmapping. ?This will >> cause ksm to be ineffective (pages will only be shared if the guest >> doesn't touch them at all), but it's enough to get started. > > The main reason change_pte initially was required for KSM to be > effective was because gup_fast was called with write=1 > unconditionally. change_pte was also responsible to set the spte > readonly. But that should have been fixed now on x86, so KSM should be > effective even despite lack of change_pte on x86. > > If the KVM page fault is calling gfn_to_pfn_async(write=0/1) depending > if the vmexit was caused by a write or read access (instead of > gfn_to_pfn which still has the unconditional write=1), and in turn > it's forced to sete the spte readonly after calling > gfn_to_pfn_async(write=0), change_pte is still useful but it's only a > worthwhile optimization to avoid a spte read fault after every KSM > page merged, it's not strictly required for KSM effectiveness anymore. > > In short if ARM does the right thing with regard of KVM read faults > passed to gup_fast(write=0) and setting the spte readonly, all should > work good with KSM (even if not as optimal as with change_pte). ah, we don't do things right, we use gfn_to_pfn() flat out and will always break the COW :) I guess now, when change_pte is a nop, it's outright incorrect if anyone runs KSM. This has just been added to my todo-list. -Christoffer
