Make sure to zero out memory before calling madvise to increase robustness

Avi pointed out that it's not entirely safe to rely on madvise zeroing out
memory.  So let's do it explicitly before calling madvise.

Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>

vl.c

diff --git a/vl.c b/vl.c
index 60a00e1..1c077b4 100644
--- a/vl.c
+++ b/vl.c
@@ -3358,13 +3358,13 @@
         
         if (flags & RAM_SAVE_FLAG_COMPRESS) {
             uint8_t ch = qemu_get_byte(f);
-#if defined(__linux__)
+            memset(qemu_get_ram_ptr(addr), ch, TARGET_PAGE_SIZE);
+#ifndef _WIN32
             if (ch == 0 &&
                 (!kvm_enabled() || kvm_has_sync_mmu())) {
                 madvise(qemu_get_ram_ptr(addr), TARGET_PAGE_SIZE, MADV_DONTNEED);
-            } else
+            }
 #endif
-            memset(qemu_get_ram_ptr(addr), ch, TARGET_PAGE_SIZE);
         } else if (flags & RAM_SAVE_FLAG_PAGE)
             qemu_get_buffer(f, qemu_get_ram_ptr(addr), TARGET_PAGE_SIZE);
     } while (!(flags & RAM_SAVE_FLAG_EOS));