2017-04-07 14:24+0200, Radim Krčmář:
> 2017-04-07 12:55+0200, Christian Borntraeger:
>> On 04/06/2017 10:20 PM, Radim Krčmář wrote:
>>> static inline bool kvm_check_request(int req, struct kvm_vcpu *vcpu)
>>> {
>>> - if (test_bit(req, &vcpu->requests)) {
>>> - clear_bit(req, &vcpu->requests);
>>> + if (kvm_test_request(req, vcpu)) {
>>> + kvm_clear_request(req, vcpu);
>>
>> This looks fine. I am just asking myself why we do not use
>> test_and_clear_bit? Do we expect gcc to merge all test bits as
>> a fast path? This does not seem to work as far as I can tell and
>> almost everybody does a fast path like in
>
> test_and_clear_bit() is a slower operation even if the test is false (at
> least on x86), because it needs to be fully atomic.
>
>> arch/s390/kvm/kvm-s390.c:
>> if (!vcpu->requests)
>> return 0;
>>
>> arch/x86/kvm/x86.c:
>> if (vcpu->requests) {
>
> We'll mostly have only one request set, so splitting the test_and_clear
> improves the performance of many subsequent tests_and_clear()s even if
> the compiler doesn't optimize.
>
> GCC couldn't even optimize if we used test_and_clear_bit(), because that
> instruction adds barriers, but the forward check for vcpu->requests is
> there because we do not trust the optimizer to do it for us and it would
> make a big difference.
Ugh, I started thinking that bitops are not atomic because I looked at
wrong boot/bitops.h by mistake. The compiler cannot merge test_bit()s,
but the speed difference holds.
