| Message ID | 20240415035731.103097-1-vaibhav@linux.ibm.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v2] KVM: PPC: Book3S HV nestedv2: Cancel pending DEC exception | expand |
Vaibhav Jain <vaibhav@linux.ibm.com> writes: > This reverts commit 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not > cancel pending decrementer exception") [1] which prevented canceling a > pending HDEC exception for nestedv2 KVM guests. It was done to avoid > overhead of a H_GUEST_GET_STATE hcall to read the 'DEC expiry TB' register > which was higher compared to handling extra decrementer exceptions. > > However recent benchmarks indicate that overhead of not handling 'DECR' > expiry for Nested KVM Guest(L2) is higher and results in much larger exits > to Pseries Host(L1) as indicated by the Unixbench-arithoh bench[2] Any reason you chose that benchmark? At least on my system it seems to compile to an infinite loop incrementing a single register. Presumably the change is still good, but a more well known benchmark would be good, even if it's just stress-ng, at least that's a bit more standard. cheers > Metric | Current upstream | Revert [1] | Difference % > ======================================================================== > arithoh-count (10) | 3244831634 | 3403089673 | +04.88% > kvm_hv:kvm_guest_exit | 513558 | 152441 | -70.32% > probe:kvmppc_gsb_recv | 28060 | 28110 | +00.18% > > N=1 > > As indicated by the data above that reverting [1] results in substantial > reduction in number of L2->L1 exits with only slight increase in number of > H_GUEST_GET_STATE hcalls to read the value of 'DEC expiry TB'. This results > in an overall ~4% improvement of arithoh[2] throughput. > > [1] commit 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not cancel pending decrementer exception") > [2] https://github.com/kdlucas/byte-unixbench/ > > Fixes: 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not cancel pending decrementer exception") > Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com> > > --- > Changelog: > Since v1: https://lore.kernel.org/all/20240313072625.76804-1-vaibhav@linux.ibm.com > * Updated/Corrected patch title and description > * Included data on test benchmark results for Unixbench-arithoh bench. > --- > arch/powerpc/kvm/book3s_hv.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c > index 8e86eb577eb8..692a7c6f5fd9 100644 > --- a/arch/powerpc/kvm/book3s_hv.c > +++ b/arch/powerpc/kvm/book3s_hv.c > @@ -4857,7 +4857,7 @@ int kvmhv_run_single_vcpu(struct kvm_vcpu *vcpu, u64 time_limit, > * entering a nested guest in which case the decrementer is now owned > * by L2 and the L1 decrementer is provided in hdec_expires > */ > - if (!kvmhv_is_nestedv2() && kvmppc_core_pending_dec(vcpu) && > + if (kvmppc_core_pending_dec(vcpu) && > ((tb < kvmppc_dec_expires_host_tb(vcpu)) || > (trap == BOOK3S_INTERRUPT_SYSCALL && > kvmppc_get_gpr(vcpu, 3) == H_ENTER_NESTED))) > -- > 2.44.0
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c index 8e86eb577eb8..692a7c6f5fd9 100644 --- a/arch/powerpc/kvm/book3s_hv.c +++ b/arch/powerpc/kvm/book3s_hv.c @@ -4857,7 +4857,7 @@ int kvmhv_run_single_vcpu(struct kvm_vcpu *vcpu, u64 time_limit, * entering a nested guest in which case the decrementer is now owned * by L2 and the L1 decrementer is provided in hdec_expires */ - if (!kvmhv_is_nestedv2() && kvmppc_core_pending_dec(vcpu) && + if (kvmppc_core_pending_dec(vcpu) && ((tb < kvmppc_dec_expires_host_tb(vcpu)) || (trap == BOOK3S_INTERRUPT_SYSCALL && kvmppc_get_gpr(vcpu, 3) == H_ENTER_NESTED)))
This reverts commit 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not cancel pending decrementer exception") [1] which prevented canceling a pending HDEC exception for nestedv2 KVM guests. It was done to avoid overhead of a H_GUEST_GET_STATE hcall to read the 'DEC expiry TB' register which was higher compared to handling extra decrementer exceptions. However recent benchmarks indicate that overhead of not handling 'DECR' expiry for Nested KVM Guest(L2) is higher and results in much larger exits to Pseries Host(L1) as indicated by the Unixbench-arithoh bench[2] Metric | Current upstream | Revert [1] | Difference % ======================================================================== arithoh-count (10) | 3244831634 | 3403089673 | +04.88% kvm_hv:kvm_guest_exit | 513558 | 152441 | -70.32% probe:kvmppc_gsb_recv | 28060 | 28110 | +00.18% N=1 As indicated by the data above that reverting [1] results in substantial reduction in number of L2->L1 exits with only slight increase in number of H_GUEST_GET_STATE hcalls to read the value of 'DEC expiry TB'. This results in an overall ~4% improvement of arithoh[2] throughput. [1] commit 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not cancel pending decrementer exception") [2] https://github.com/kdlucas/byte-unixbench/ Fixes: 180c6b072bf3 ("KVM: PPC: Book3S HV nestedv2: Do not cancel pending decrementer exception") Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com> --- Changelog: Since v1: https://lore.kernel.org/all/20240313072625.76804-1-vaibhav@linux.ibm.com * Updated/Corrected patch title and description * Included data on test benchmark results for Unixbench-arithoh bench. --- arch/powerpc/kvm/book3s_hv.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)