Out-of-sync L1 shadows.

Hello,

This set of patches let the guest write directly to L1 pagetables, thus making them temporarily out of sync (OOS) with their shadows. L2 and higher level page tables are still always emulated.
This let us avoids most of the pagetable write emulations, resulting in a generic speedup, especially in Windows guests.

The policy implemented in the fourth patch (that put OOS a pagetable whenever the guest tries to write to it), has proven to be a win for guest with up to 4 vcpus.

Patches:

- oos-base.patch: implement the basic mechanisms to unsync and resync pages at appropriate times;

- oos-optimize-1.oos_fixups.patch: OOS writable mappings reverse map to avoid shadow brute-force search on resyncs;

- oos-optimize-2.snapshots.patch: add snapshots of guest pages to allow faster revalidation of OOS pages on resync;

- oos-policy.always.patch: always unsync pages on guest writes;

Benchmarks results:

Results for a Windows Server 2003 DDK Build (build -cPzM 3) with warm cache (third run) on a Core2 6420 @ 2.13GhZ.

             current   OOS        Diff
32bit 1vcpu  05:27.00  04:38.00   -14.98%
PAE 1vcpu    06:49.00  05:05.00   -25.43%
64bit 1vcpu  06:56.00  05:35.00   -19.47%

32bit 2vcpu  03:31.00  03:06.00   -11.85%
PAE 2vcpu    04:07.00  03:17.00   -20.24%
64bit 2vcpu  04:05.00  03:47.00   -7.35%

Thanks,
Gianluca