Re: [RFC PATCH v2 2/2] KVM: x86: Not wr-protect huge page with init_all_set dirty log

From: Keqian Zhu
Date: Wed Apr 28 2021 - 06:51:37 EST




On 2021/4/28 0:33, Ben Gardon wrote:
> On Mon, Apr 26, 2021 at 10:04 PM Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>>
>> Hi Ben,
>>
>> Sorry for the delay reply!
>>
>> On 2021/4/21 0:30, Ben Gardon wrote:
>>> On Tue, Apr 20, 2021 at 12:49 AM Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>>>>
>>>> Hi Ben,
>>>>
>>>> On 2021/4/20 3:20, Ben Gardon wrote:
>>>>> On Fri, Apr 16, 2021 at 1:25 AM Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>>>>>>
>>>>>> Currently during start dirty logging, if we're with init-all-set,
>>>>>> we write protect huge pages and leave normal pages untouched, for
>>>>>> that we can enable dirty logging for these pages lazily.
>>>>>>
>>>>>> Actually enable dirty logging lazily for huge pages is feasible
>>>>>> too, which not only reduces the time of start dirty logging, also
>>>>>> greatly reduces side-effect on guest when there is high dirty rate.
>>>>>>
>>>>>> Signed-off-by: Keqian Zhu <zhukeqian1@xxxxxxxxxx>
>>>>>> ---
>>>>>> arch/x86/kvm/mmu/mmu.c | 48 ++++++++++++++++++++++++++++++++++++++----
>>>>>> arch/x86/kvm/x86.c | 37 +++++++++-----------------------
>>>>>> 2 files changed, 54 insertions(+), 31 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
>>>>>> index 2ce5bc2ea46d..98fa25172b9a 100644
>>>>>> --- a/arch/x86/kvm/mmu/mmu.c
>>>>>> +++ b/arch/x86/kvm/mmu/mmu.c
>>>>>> @@ -1188,8 +1188,7 @@ static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
>>>>>> * @gfn_offset: start of the BITS_PER_LONG pages we care about
>>>>>> * @mask: indicates which pages we should protect
>>>>>> *
>>>>>> - * Used when we do not need to care about huge page mappings: e.g. during dirty
>>>>>> - * logging we do not have any such mappings.
>>>>>> + * Used when we do not need to care about huge page mappings.
>>>>>> */
>>>>>> static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
>>>>>> struct kvm_memory_slot *slot,
>>>>>> @@ -1246,13 +1245,54 @@ static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
>>>>>> * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
>>>>>> * enable dirty logging for them.
>>>>>> *
>>>>>> - * Used when we do not need to care about huge page mappings: e.g. during dirty
>>>>>> - * logging we do not have any such mappings.
>>>>>> + * We need to care about huge page mappings: e.g. during dirty logging we may
>>>>>> + * have any such mappings.
>>>>>> */
>>>>>> void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
>>>>>> struct kvm_memory_slot *slot,
>>>>>> gfn_t gfn_offset, unsigned long mask)
>>>>>> {
>>>>>> + gfn_t start, end;
>>>>>> +
>>>>>> + /*
>>>>>> + * Huge pages are NOT write protected when we start dirty log with
>>>>>> + * init-all-set, so we must write protect them at here.
>>>>>> + *
>>>>>> + * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn
>>>>>> + * of memslot has no such restriction, so the range can cross two large
>>>>>> + * pages.
>>>>>> + */
>>>>>> + if (kvm_dirty_log_manual_protect_and_init_set(kvm)) {
>>>>>> + start = slot->base_gfn + gfn_offset + __ffs(mask);
>>>>>> + end = slot->base_gfn + gfn_offset + __fls(mask);
>>>>>> + kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M);
>>>>>> +
>>>>>> + /* Cross two large pages? */
>>>>>> + if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) !=
>>>>>> + ALIGN(end << PAGE_SHIFT, PMD_SIZE))
>>>>>> + kvm_mmu_slot_gfn_write_protect(kvm, slot, end,
>>>>>> + PG_LEVEL_2M);
>>>>>> + }
>>>>>> +
>>>>>> + /*
>>>>>> + * RFC:
>>>>>> + *
>>>>>> + * 1. I don't return early when kvm_mmu_slot_gfn_write_protect() returns
>>>>>> + * true, because I am not very clear about the relationship between
>>>>>> + * legacy mmu and tdp mmu. AFAICS, the code logic is NOT an if/else
>>>>>> + * manner.
>>>>>> + *
>>>>>> + * The kvm_mmu_slot_gfn_write_protect() returns true when we hit a
>>>>>> + * writable large page mapping in legacy mmu mapping or tdp mmu mapping.
>>>>>> + * Do we still have normal mapping in that case? (e.g. We have large
>>>>>> + * mapping in legacy mmu and normal mapping in tdp mmu).
>>>>>
>>>>> Right, we can't return early because the two MMUs could map the page
>>>>> in different ways, but each MMU could also map the page in multiple
>>>>> ways independently.
>>>>> For example, if the legacy MMU was being used and we were running a
>>>>> nested VM, a page could be mapped 2M in EPT01 and 4K in EPT02, so we'd
>>>>> still need kvm_mmu_slot_gfn_write_protect calls for both levels.
>>>>> I don't think there's a case where we can return early here with the
>>>>> information that the first calls to kvm_mmu_slot_gfn_write_protect
>>>>> access.
>>>> Thanks for the detailed explanation.
>>>>
>>>>>
>>>>>> + *
>>>>>> + * 2. kvm_mmu_slot_gfn_write_protect() doesn't tell us whether the large
>>>>>> + * page mapping exist. If it exists but is clean, we can return early.
>>>>>> + * However, we have to do invasive change.
>>>>>
>>>>> What do you mean by invasive change?
>>>> We need the kvm_mmu_slot_gfn_write_protect to report whether all mapping are large
>>>> and clean, so we can return early. However it's not a part of semantics of this function.
>>>>
>>>> If this is the final code, compared to old code, we have an extra gfn_write_protect(),
>>>> I don't whether it's acceptable?
>>>
>>> Ah, I see. Please correct me if I'm wrong, but I think that in order
>>> to check that the only mappings on the GFN range are large, we'd still
>>> have to go over the rmap for the 4k mappings, at least for the legacy
>>> MMU. In that case, we're doing about as much work as the extra
>>> gfn_write_protect and I don't think that we'd get any efficiency gain
>>> for the change in semantics.
>>>
>>> Likewise for the TDP MMU, if the GFN range is mapped both large and
>>> 4k, it would have to be in different TDP structures, so the efficiency
>>> gains would again not be very big.
>> I am not familiar with the MMU virtualization of x86 arch, but I think
>> you are right.
>>
>>>
>>> I'm really just guessing about those performance characteristics
>>> though. It would definitely help to have some performance data to back
>>> all this up. Even just a few runs of the dirty_log_perf_test (in
>>> selftests) could provide some interesting results, and I'd be happy to
>>> help review any improvements you might make to that test.
>>>
>>> Regardless, I'd be inclined to keep this change as simple as possible
>>> for now and the early return optimization could happen in a follow-up
>>> patch. I think the extra gfn_write_protect is acceptable, especially
>>> if you can show that it doesn't cause a big hit in performance when
>>> running the dirty_log_perf_test with 4k and 2m backing memory.
>> I tested it using dirty_log_perf_test, the result shows that performance
>> of clear_dirty_log different within 2%.
>
> I think there are a couple obstacles which make the stock
> dirty_log_perf_test less useful for measuring this optimization.
>
> 1. Variance between runs
> With only 16 vCPUs and whatever the associated default guest memory
> size is, random system events and daemons introduce a lot of variance,
> at least in my testing. I usually try to run the biggest VM I can to
> smooth that out, but even with a 96 vCPU VM, a 2% difference is often
> not statistically significant. CPU pinning for the vCPU threads would
> help a lot to reduce variance. I don't remember if anyone has
> implemented this yet.
Yes, this makes sense.

>
> 2. The guest dirty pattern
> By default, each guest vCPU will dirty it's entire partition of guest
> memory on each iteration. This means that instead of amortizing out
> the cost of write-protecting and splitting large pages, we simply move
> the burden later in the process. I see you didn't include the time for
> each iteration below, but I would expect this patch to move some of
> the time from "Enabling dirty logging time" and "Dirtying memory time"
> for pass 1 to "Clear dirty log time" and "Dirtying memory time" for
> pass 2. I wouldn't expect the total time over 5 iterations to change
> for this test.
If we have large page mapping and are with this optimization, the "Enabling dirty logging time"
and the first round "Dirtying memory time" will be greatly reduced.

However, I don't think other times (dirty_memory except first round, get_log, clear_log) are
expected to change compared to w/o optimization. Because after the first round "Dirtying memory",
all mappings have been split to normal mappings, so the situation is same as w/o this optimization.

Maybe I miss something?

>
> It would probably also serve us well to have some kind of "hot" subset
> of memory for each vCPU, since some of the benefit of lazy large page
> splitting depend on that access pattern.
>
> 3. Lockstep dirtying and dirty log collection
> While this test is currently great for timing dirty logging
> operations, it's not great for trickier analysis, especially
> reductions to guest degradation. In order to measure that we'd need to
> change the test to collect the dirty log as quickly as possible,
> independent of what the guest is doing and then also record how much
> "progress" the guest is able to make while all that is happening.
Yes, make sense.

Does the "dirty log collection" contains "dirty log clear"? As I understand, the dirty log
collection is very fast, just some memory copy. But for "dirty log clear", we should modify mappings
and perform TLBI, the time is much longer.

>
> I'd be happy to help review any improvements to the test which you
> feel like making.
Thanks, Ben. emm... I feel very sorry that perhaps I don't have enough time to do this, many works are queued...
On the other hand, I think the "Dirtying memory time" of first round can show us the optimization.

>
>>
>> *Without this patch*
>>
>> ./dirty_log_perf_test -i 5 -v 16 -s anonymous
>>
>> Testing guest mode: PA-bits:ANY, VA-bits:48, 4K pages
>> guest physical test memory offset: 0xffbfffff000
>> Populate memory time: 3.105203579s
>> Enabling dirty logging time: 0.000323444s
>> [...]
>> Get dirty log over 5 iterations took 0.000595033s. (Avg 0.000119006s/iteration)
>> Clear dirty log over 5 iterations took 0.713212922s. (Avg 0.142642584s/iteration)
>>
>> ./dirty_log_perf_test -i 5 -v 16 -s anonymous_hugetlb
>>
>> Testing guest mode: PA-bits:ANY, VA-bits:48, 4K pages
>> guest physical test memory offset: 0xffbfffff000
>> Populate memory time: 3.922764235s
>> Enabling dirty logging time: 0.000316473s
>> [...]
>> Get dirty log over 5 iterations took 0.000485459s. (Avg 0.000097091s/iteration)
>> Clear dirty log over 5 iterations took 0.603749670s. (Avg 0.120749934s/iteration)
>>
>>
>> *With this patch*
>>
>> ./dirty_log_perf_test -i 5 -v 16 -s anonymous
>>
>> Testing guest mode: PA-bits:ANY, VA-bits:48, 4K pages
>> guest physical test memory offset: 0xffbfffff000
>> Populate memory time: 3.244515198s
>> Enabling dirty logging time: 0.000280207s
>> [...]
>> Get dirty log over 5 iterations took 0.000484953s. (Avg 0.000096990s/iteration)
>> Clear dirty log over 5 iterations took 0.727620114s. (Avg 0.145524022s/iteration)
>>
>> ./dirty_log_perf_test -i 5 -v 16 -s anonymous_hugetlb
>>
>> Testing guest mode: PA-bits:ANY, VA-bits:48, 4K pages
>> guest physical test memory offset: 0xffbfffff000
>> Populate memory time: 3.244294061s
>> Enabling dirty logging time: 0.000273590s
>> [...]
>> Get dirty log over 5 iterations took 0.000474244s. (Avg 0.000094848s/iteration)
>> Clear dirty log over 5 iterations took 0.600593672s. (Avg 0.120118734s/iteration)
>>
>>
>> I faced a problem that there is no huge page mapping when test with
>> "-s anonymous_hugetlb", both for TDP enabled or disabled.
>
> Do you mean that even before dirty logging was enabled, KVM didn't
> create any large mappings? That's odd. I would assume the backing
> memory allocation would just fail if there aren't enough hugepages
> available.
It's odd indeed. I can see there are large mapping when I do normal migration, but I
don't see large mapping when run this test.

I have proofed the time of "clear dirty log" is not effected, what about send a
formal patch?

Thanks,
Keqian