Re: [RFC] KVM: arm64: support enabling dirty log graually in small chunks

[zhukeqian]

Hi Marc,

On 2020/3/10 21:16, Marc Zyngier wrote:
> On 2020-03-10 08:26, zhukeqian wrote:
>> Hi Marc,
>>
>> On 2020/3/9 19:45, Marc Zyngier wrote:
>>> Kegian,
> 
> [...]
> 
>>> Is there a userspace counterpart to it?
>>>
>> As this KVM/x86 related changes have not been merged to mainline
>> kernel, some little modification is needed on mainline Qemu.
> 
> Could you please point me to these changes?
I made some changes locally listed below.

However, Qemu can choose to enable KVM_DIRTY_LOG_INITIALLY_SET or not.
Here I made no judgement on dirty_log_manual_caps because I just want
to verify the optimization of this patch.

diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
index 439a4efe52..1611f644a4 100644
--- a/accel/kvm/kvm-all.c
+++ b/accel/kvm/kvm-all.c
@@ -2007,14 +2007,16 @@ static int kvm_init(MachineState *ms)
     s->coalesced_pio = s->coalesced_mmio &&
                        kvm_check_extension(s, KVM_CAP_COALESCED_PIO);

-    s->manual_dirty_log_protect =
+    uint64_t dirty_log_manual_caps =
         kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
-    if (s->manual_dirty_log_protect) {
-        ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0, 1);
+    if (dirty_log_manual_caps) {
+        ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0,
+                                dirty_log_manual_caps);
         if (ret) {
             warn_report("Trying to enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 "
                         "but failed.  Falling back to the legacy mode. ");
-            s->manual_dirty_log_protect = false;
+        } else {
+            s->manual_dirty_log_protect = true;
         }
     }

> 
>> As I tested this patch on a 128GB RAM Linux VM with no huge pages, the
>> time of enabling dirty log will decrease obviously.
> 
> I'm not sure how realistic that is. Not having huge pages tends to lead
> to pretty bad performance in general...
Sure, this has no effect on guests which are all of huge pages.

For my understanding, once a guest has normal pages (maybe are initialized
at beginning or dissloved from huge pages), it can benefit from this patch.
> 
> Thanks,
> 
>         M.
Pretty thanks for your review.

Thanks,
Keqian