Vipin Sharma <vipinsh@xxxxxxxxxx> writes: > Hi Vitaly, Yury, Sunil, Tianyu Hi Vipin! > > Before I work on a patch series and send it out to the KVM mailing > list, I wanted to check with you a potential Windows VM optimization > and see if you have worked on it or if you know about some obvious > known blockers regarding this feature. > > Hypervisor Top-Level Functional Specification v6.0b mentions a hypercall: > > HvExtCallGetBootZeroedMemory > Call Code = 0x8002 > > This hypercall can be used by Windows guest to know which pages are > already zeroed and then guest can avoid zeroing them again during the > boot, resulting in Windows VM faster boot time and less memory usage. > > KVM currently doesn't implement this feature. I am thinking of > implementing it, here is a rough code flow: > 1. KVM will set bit 20 in EBX of CPUID leaf 0x40000003 to let the > Windows guest know that it can use the extended hypercall interface. > 2. Guest during the boot will use hypercall HvExtCallQueryCapabilities > (Call Code = 0x8001) to see which extended calls are available. > 3. KVM will respond to guest that the hypercall > HvExtCallGetBootZeroedMemory is available. > 4. Guest will issue the hypercall HvExtCallGetBootZeroedMemory to know > which pages are zeroed. > 5. KVM or userspace VMM will respond with GPA and page count to guest. I think it's VMM's responsibility. How would KVM know if the memory allocated to the guest was zeroed or not? The easiest solution would be to just pass through this hypercall to the VMM and let it respond. Alternatively, we can probably add a flag to KVM_SET_USER_MEMORY_REGION to either indicate that the memory is zeroed or to actually ask KVM to zero it. This way we will have the required information in KVM. I'm not sure if it's worth it, Windows probably calls HvExtCallGetBootZeroedMemory just once upon boot so handling it in the VMM is totally fine. > 6. Guest will skip zeroing these pages, resulting in faster boot and > less memory utilization of guest. > > This seems like a very easy win for KVM to increase Windows guest boot > performance but I am not sure if I am overlooking something. If you > are aware of any potential side effects of enabling these hypercalls > or some other issue I am not thinking about please let me know, > otherwise, I can start working on this feature and send RFC patches to > the mailing list. I dug through my git archives and found that I've actually tried HvExtCallQueryCapabilities back in 2018 but for some reason Windows versions I was testing didn't use it (hope it wasn't some silly mistake like forgotten CPUID bit on my part :-) so I put it aside and never got back to it. Thanks for picking this up! -- Vitaly
