[PATCH v1 3/5] KVM: selftests: Dirty host pages in dirty_log_test

[Gavin Shan]

Hi Drew,

On 8/22/22 4:29 PM, Gavin Shan wrote:
> On 8/19/22 3:28 PM, Andrew Jones wrote:
>> On Fri, Aug 19, 2022 at 08:55:59AM +0800, Gavin Shan wrote:
>>> It's assumed that 1024 host pages, instead of guest pages, are dirtied
>>> in each iteration in guest_code(). The current implementation misses
>>> the case of mismatched page sizes in host and guest. For example,
>>> ARM64 could have 64KB page size in guest, but 4KB page size in host.
>>> (TEST_PAGES_PER_LOOP / 16), instead of TEST_PAGES_PER_LOOP, host pages
>>> are dirtied in every iteration.
>>>
>>> Fix the issue by touching all sub-pages when we have mismatched
>>> page sizes in host and guest.
>>
>> I'll let the dirty-log test authors decide what's best to do for this
>> test, but I'd think we should let the guest continue dirtying its
>> pages without knowledge of the host pages. Then, adjust the host test
>> code to assert all sub-pages, other than the ones it expects the guest
>> to have written, remain untouched.
>>
> 
> I don't think what is clarified in the change log is correct. The current
> implementation already had the logic to handle the mismatched page sizes
> in vm_dirty_log_verify() where 'step' is used for it by fetching value
> from vm_num_host_pages(mode, 1). Please ignore this patch for now, as
> explained below.
> 
> The issue I have is the 'dirty_log_test' hangs when I have 4KB host page size
> and 64KB guest page size. It seems the vcpu doesn't exit due to full ring
> buffer state or kick-off. I will have more investigations to figure out the
> root cause.
> 

[...]

Please ignore this PATCH[3/5], I think this should be fixed by selecting
correct dirty ring count and the fix will be folded to PATCH[5/5] in next
revision.

In dirty_log_test, we have 1GB memory for guest to write and make them
dirty. When we have mismatch page sizes on host and guest, which is either
4kb-host-64kb-guest or 64kb-host-4kb-guest apart from 16kb case, 16384 host
pages are dirtied in each iteration. The default dirty ring count is 65536.
So the vcpu never exit due to full-dirty-ring-buffer state. This leads the
guest's code keep running and the dirty log isn't collected by the main
thread.

     #define TEST_DIRTY_RING_COUNT           65536

     dirty_pages_per_iteration = (0x40000000 / 0x10000)
                               = 0x4000
                               = 16384

Thanks,
Gavin