Re: [PATCH v3 3/8] KVM: Resolve memslot ID via a hash table instead of via a static array

["Maciej S. Szmigiero" <mail@maciej.szmigiero.name>]

On 21.05.2021 09:05, Maciej S. Szmigiero wrote:
> On 20.05.2021 00:31, Sean Christopherson wrote:
>> On Sun, May 16, 2021, Maciej S. Szmigiero wrote:
(..)
>>>           new_size = old_size;
>>>       slots = kvzalloc(new_size, GFP_KERNEL_ACCOUNT);
>>> -    if (likely(slots))
>>> -        memcpy(slots, old, old_size);
>>> +    if (unlikely(!slots))
>>> +        return NULL;
>>> +
>>> +    memcpy(slots, old, old_size);
>>> +
>>> +    hash_init(slots->id_hash);
>>> +    kvm_for_each_memslot(memslot, slots)
>>> +        hash_add(slots->id_hash, &memslot->id_node, memslot->id);
>>
>> What's the perf penalty if the number of memslots gets large?  I ask because the
>> lazy rmap allocation is adding multiple calls to kvm_dup_memslots().
> 
> I would expect the "move inactive" benchmark to be closest to measuring
> the performance of just a memslot array copy operation but the results
> suggest that the performance stays within ~10% window from 10 to 509
> memslots on the old code (it then climbs 13x for 32k case).
> 
> That suggests that something else is dominating this benchmark for these
> memslot counts (probably zapping of shadow pages).
> 
> At the same time, the tree-based memslots implementation is clearly
> faster in this benchmark, even for smaller memslot counts, so apparently
> copying of the memslot array has some performance impact, too.
> 
> Measuring just kvm_dup_memslots() performance would probably be done
> best by benchmarking KVM_MR_FLAGS_ONLY operation - will try to add this
> operation to my set of benchmarks and see how it performs with different
> memslot counts.

Update:
I've implemented a simple KVM_MR_FLAGS_ONLY benchmark, that repeatably
sets and unsets KVM_MEM_LOG_DIRTY_PAGES flag on a memslot with a single
page of memory in it. [1]

Since on the current code with higher memslot counts the "set flags"
operation spends a significant time in kvm_mmu_calculate_default_mmu_pages()
a second set of measurements was done with patch [2] applied.

In this case, the top functions in the perf trace are "memcpy" and
"clear_page" (called from kvm_set_memslot(), most likely from inlined
kvm_dup_memslots()).

For reference, a set of measurements with the whole patch series
(patches 1 - 8) applied was also done, as "new code".
In this case, SRCU-related functions dominate the perf trace.

32k memslots:
Current code:             0.00130s
Current code + patch [2]: 0.00104s (13x 4k result)
New code:                 0.0000144s

4k memslots:
Current code:             0.0000899s
Current code + patch [2]: 0.0000799s (+78% 2k result)
New code:                 0.0000144s

2k memslots:
Current code:             0.0000495s
Current code + patch [2]: 0.0000447s (+54% 509 result)
New code:                 0.0000143s

509 memslots:
Current code:             0.0000305s
Current code + patch [2]: 0.0000290s (+5% 100 result)
New code:                 0.0000141s

100 memslots:
Current code:             0.0000280s
Current code + patch [2]: 0.0000275s (same as for 10 slots)
New code:                 0.0000142s

10 memslots:
Current code:             0.0000272s
Current code + patch [2]: 0.0000272s
New code:                 0.0000141s

Thanks,
Maciej

[1]: The patch against memslot_perf_test.c is available here:
https://github.com/maciejsszmigiero/linux/commit/841e94898a55ff79af9d20a08205aa80808bd2a8

[2]: "[PATCH v3 1/8] KVM: x86: Cache total page count to avoid traversing the memslot array"