On Tue, Nov 05, 2024 at 04:57:55PM +0100, Phil Dennis-Jordan wrote: > This is a loose collection of patches against the x86 hvf accel. They > can be applied/pulled independently from one another. > > Patch 1 is a repost of a patch I've submitted a bunch of times already. > It wires up and enables x2APIC mode in conjunction with HVF - the > software APIC implementation in QEMU gained the feature earlier this > year but hvf wasn't included. > The change typically improves performance with modern SMP guest OSes by > a 2-digit percentage. (Exact values depend on workload.) > > Patch 2 fixes cases of undefined behaviour recently introduced by commit > 7cac7aa which made changes to HVF CPUID XSAVE functionality. > > Patch 3 fixes a minor one-off memory leak during hvf startup. > > Patch 4 ever so slightly improves APIC correctness under hvf: when > setting the APICBASE MSR, if the APIC deems the new value invalid, > we raise an exception (as per spec) rather than silently doing > nothing. This fixes a failing kvm-unit-tests test case. > > Patch 5 removes some unnecessary duplication and type-rechecking in > HVF's inner loop. (No need to cast the cpu state pointer to X86CPU > within, the hvf_vcp_exec function already does that once at the top.) > > Some of this work has been sponsored by Sauce Labs Inc. > > Phil Dennis-Jordan (5): > i386/hvf: Integrates x2APIC support with hvf accel > i386/hvf: Fix for UB in handling CPUID function 0xD > i386/hvf: Fixes startup memory leak (vmcs caps) > i386/hvf: Raise exception on error setting APICBASE > i386/hvf: Removes duplicate/shadowed variables in hvf_vcpu_exec >
To the series, Reviewed-by: Roman Bolshakov <rbolsha...@ddn.com> Tested-by: Roman Bolshakov <rbolsha...@ddn.com> I figured the issue with 24.04 guests, it was an issue on my side (too little memory provided to the guest). Paolo, please apply this if you have no objections. Regards, Roman
