KVM: SVM: Force flush caches before reboot CSV guest

Upstream: no

For memory encrypted guest, its pages' encrypt status will changed at
runtime. When user reboot the guest, the pages' encrypt status during
last boot were ignored. So during the boot flow of reboot, there may
be 2 versions of memory data lies in cache as follows:

                  +--------+        |              |
                  |        |        +--------------+  --+
                  |        |        |              |     \
                  |________|        |              |      \
cacheline for  -> |________| <-+    |              |       \
 pa1(c=0)         |        |    \   |______________|        \
                  |        |     \_ 64 bytes aligned <- pa1  \
                  |        |      _ |______________|           4K
                  |        |     /  |              |           page
cacheline for     |________|    /   |              |         /
 pa1(c=1)      -> |________| <-+    |              |        /
                  |        |        |              |       /
                  |        |        |              |      /
                  |        |        |              |     /
                  |        |        +--------------+  --+
                  |        |        |              |

If the older version cache was flushed after that of newer version, and
guest read the memory again, then it will get corrupted data and may
lead to crash.

In this change, for any memory encrypted guest, the cache is forcibly
flushed to memory before the next boot flow, which ensures that memory
access is up-to-date.

Signed-off-by: hanliyang <hanliyang@hygon.cn>
This commit is contained in:
hanliyang 2023-05-06 16:01:25 +08:00
parent c6d3aa44b4
commit 2fb5abf978
1 changed files with 5 additions and 2 deletions

View File

@ -1028,12 +1028,15 @@ static int csv_control_post_system_reset(struct kvm *kvm)
unsigned long i;
int ret;
if (!sev_es_guest(kvm))
if (!sev_guest(kvm))
return 0;
/* Flush both host and guest caches of VMSA */
/* Flush both host and guest caches before next boot flow */
wbinvd_on_all_cpus();
if (!sev_es_guest(kvm))
return 0;
kvm_for_each_vcpu(i, vcpu, kvm) {
struct vcpu_svm *svm = to_svm(vcpu);