Re: [kvm-devel] [RFC 7/8]KVM: swap out guest pages

["Shaohua Li" <shaoh.li@gmail.com>]

2007/7/24, Avi Kivity <avi@qumranet.com>:
> Shaohua Li wrote:
> > Make KVM guest pages be allocated dynamically and able to be swaped out.
> >
> > One issue: all inodes returned from anon_inode_getfd are shared,
> > if one module changes field of the inode, other moduels might break.
> > Should we introduce a new API to not share inode?
> >
> > Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> > ---
> >
> > +static int kvm_set_page_dirty(struct page *page)
> > +{
> > +     if (!PageDirty(page))
> > +             SetPageDirty(page);
> > +     return 0;
> > +}
> > +
> > +static int kvm_writepage(struct page *page, struct writeback_control *wbc)
> > +{
> > +     struct address_space *mapping = page->mapping;
> > +     struct kvm *kvm = address_space_to_kvm(mapping);
> > +     int ret = 0;
> > +
> > +     /*
> > +      * gfn_to_page is called with kvm->lock hold, which might invoke page
> > +      * reclaim. So the .writepage should check if we already hold the lock
> > +      * to avoid deadlock.
> > +      */
> > +     if (!mutex_trylock(&kvm->lock)) {
> > +             set_page_dirty(page);
> > +             return AOP_WRITEPAGE_ACTIVATE;
> > +     }
> > +
> > +     /*
> > +      * We just zap vcpu 0's page table. For a SMP guest, we should zap all
> > +      * vcpus'. It's better shadow page table is per-vm.
> > +      */
> > +     if (PagePrivate(page))
> > +             kvm_mmu_zap_pagetbl(&kvm->vcpus[0], page->index);
> > +
> > +     ret = kvm_move_to_swap(page);
> > +     if (ret) {
> > +             set_page_dirty(page);
> > +             goto out;
> > +     }
> > +     unlock_page(page);
> > +out:
> > +     mutex_unlock(&kvm->lock);
> > +
> > +     return ret;
> > +}
> > +
> >
>
> Perhaps we can use this as a base for userspace-allocated memory.  We
> still have a kvm inode and address_space; but instead of calling
> kvm_move_to_swap(), we use the memory slot and virtual address offset to
> locate the underlying address_space and call that ->writepage().
>
> So:
>   kvm_writepage() removes any shadow page table references
>   the underlying ->writepage() does the work of paging to the underlying
> store
So write to a file, right? Yes, it can avoid use move to swap, and
should be feasible.

> We need to figure out how to avoid the underlying ->writepage() from not
> within the context of kvm_writepage().  Maybe have a page flag
> signifying layered address spaces?
>
> [it probably violates fifteen different mm assumptions; I need to study
> that code]
>
> An alternative would be to have kvm set a page flag signifying it has
> references to the page when it installs it in a shadow pte.  The mm
> would notice the flag and call kvm to clear it below proceeding with
> normal ->writepage().
This page_private flag's job, I think.

Thanks,
Shaohua