From mboxrd@z Thu Jan 1 00:00:00 1970 From: Nicholas Piggin Date: Mon, 22 Mar 2021 18:22:31 +0000 Subject: Re: [PATCH v3 19/41] KVM: PPC: Book3S HV P9: Stop handling hcalls in real-mode in the P9 path Message-Id: <1616436906.owrt3o4wh1.astroid@bobo.none> List-Id: References: <20210305150638.2675513-1-npiggin@gmail.com> <20210305150638.2675513-20-npiggin@gmail.com> <1616417941.ksskhyvg3t.astroid@bobo.none> In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable To: Alexey Kardashevskiy , =?iso-8859-1?q?C=E9dric?= Le Goater , kvm-ppc@vger.kernel.org Cc: linuxppc-dev@lists.ozlabs.org Excerpts from C=C3=A9dric Le Goater's message of March 23, 2021 2:01 am: > On 3/22/21 2:15 PM, Nicholas Piggin wrote: >> Excerpts from Alexey Kardashevskiy's message of March 22, 2021 5:30 pm: >>> >>> >>> On 06/03/2021 02:06, Nicholas Piggin wrote: >>>> In the interest of minimising the amount of code that is run in>>> "re= al-mode", don't handle hcalls in real mode in the P9 path. >>>> >>>> POWER8 and earlier are much more expensive to exit from HV real mode >>>> and switch to host mode, because on those processors HV interrupts get >>>> to the hypervisor with the MMU off, and the other threads in the core >>>> need to be pulled out of the guest, and SLBs all need to be saved, >>>> ERATs invalidated, and host SLB reloaded before the MMU is re-enabled >>>> in host mode. Hash guests also require a lot of hcalls to run. The >>>> XICS interrupt controller requires hcalls to run. >>>> >>>> By contrast, POWER9 has independent thread switching, and in radix mode >>>> the hypervisor is already in a host virtual memory mode when the HV >>>> interrupt is taken. Radix + xive guests don't need hcalls to handle >>>> interrupts or manage translations. >=20 > Do we need to handle the host-is-a-P9-without-xive case ? I'm not sure really. Is there an intention for OPAL to be able to=20 provide a fallback layer in the worst case? Maybe microwatt grows HV capability before XIVE? >=20 >>>> So it's much less important to handle hcalls in real mode in P9. >>> >>> So acde25726bc6034b (which added if(kvm_is_radix(vcpu->kvm))return=20 >>> H_TOO_HARD) can be reverted, pretty much? >>=20 >> Yes. Although that calls attention to the fact I missed doing >> a P9 h_random handler in this patch. I'll fix that, then I think >> acde2572 could be reverted entirely. >>=20 >> [...] >>=20 >>>> } else { >>>> kvmppc_xive_push_vcpu(vcpu); >>>> trap =3D kvmhv_load_hv_regs_and_go(vcpu, time_limit, lpcr); >>>> - kvmppc_xive_pull_vcpu(vcpu); >>>> + /* H_CEDE has to be handled now, not later */ >>>> + /* XICS hcalls must be handled before xive is pulled */ >>>> + if (trap =3D=3D BOOK3S_INTERRUPT_SYSCALL && >>>> + !(vcpu->arch.shregs.msr & MSR_PR)) { >>>> + unsigned long req =3D kvmppc_get_gpr(vcpu, 3); >>>> =20 >>>> + if (req =3D=3D H_CEDE) { >>>> + kvmppc_cede(vcpu); >>>> + kvmppc_xive_cede_vcpu(vcpu); /* may un-cede */ >>>> + kvmppc_set_gpr(vcpu, 3, 0); >>>> + trap =3D 0; >>>> + } >>>> + if (req =3D=3D H_EOI || req =3D=3D H_CPPR || >>> >>> else if (req =3D=3D H_EOI ... ? >>=20 >> Hummm, sure. >=20 > you could integrate the H_CEDE in the switch statement below. Below is in a different file just for the emulation calls. >>=20 >> [...] >>=20 >>>> +void kvmppc_xive_cede_vcpu(struct kvm_vcpu *vcpu) >>>> +{ >>>> + void __iomem *esc_vaddr =3D (void __iomem *)vcpu->arch.xive_esc_vadd= r; >>>> + >>>> + if (!esc_vaddr) >>>> + return; >>>> + >>>> + /* we are using XIVE with single escalation */ >>>> + >>>> + if (vcpu->arch.xive_esc_on) { >>>> + /* >>>> + * If we still have a pending escalation, abort the cede, >>>> + * and we must set PQ to 10 rather than 00 so that we don't >>>> + * potentially end up with two entries for the escalation >>>> + * interrupt in the XIVE interrupt queue. In that case >>>> + * we also don't want to set xive_esc_on to 1 here in >>>> + * case we race with xive_esc_irq(). >>>> + */ >>>> + vcpu->arch.ceded =3D 0; >>>> + /* >>>> + * The escalation interrupts are special as we don't EOI them. >>>> + * There is no need to use the load-after-store ordering offset >>>> + * to set PQ to 10 as we won't use StoreEOI. >>>> + */ >>>> + __raw_readq(esc_vaddr + XIVE_ESB_SET_PQ_10); >>>> + } else { >>>> + vcpu->arch.xive_esc_on =3D true; >>>> + mb(); >>>> + __raw_readq(esc_vaddr + XIVE_ESB_SET_PQ_00); >>>> + } >>>> + mb(); >>> >>> >>> Uff. Thanks for cut-n-pasting the comments, helped a lot to match this = c=20 >>> to that asm! >>=20 >> Glad it helped. >>>> +} >=20 > I had to do the PowerNV models in QEMU to start understanding that stuff = ...=20 >=20 >>>> +EXPORT_SYMBOL_GPL(kvmppc_xive_cede_vcpu); >>>> + >>>> /* >>>> * This is a simple trigger for a generic XIVE IRQ. This must >>>> * only be called for interrupts that support a trigger page >>>> @@ -2106,6 +2140,32 @@ static int kvmppc_xive_create(struct kvm_device= *dev, u32 type) >>>> return 0; >>>> } >>>> =20 >>>> +int kvmppc_xive_xics_hcall(struct kvm_vcpu *vcpu, u32 req) >>>> +{ >>>> + struct kvmppc_vcore *vc =3D vcpu->arch.vcore; >>> >>> >>> Can a XIVE enabled guest issue these hcalls? Don't we want if=20 >>> (!kvmppc_xics_enabled(vcpu)) and >>> if (xics_on_xive()) here, as kvmppc_rm_h_xirr() have? Some of these=20 >>> hcalls do write to XIVE registers but some seem to change=20 >>> kvmppc_xive_vcpu. Thanks, >>=20 >> Yes I think you're right, good catch. I'm not completely sure about all = >> the xive and xics modes but a guest certainly can make any kind of hcall= =20 >> it likes and we have to sanity check it. >=20 > Yes.=20 >=20 >> We want to take the hcall here (in replacement of the real mode hcalls) >> with the same condition. So it would be: >>=20 >> if (!kvmppc_xics_enabled(vcpu)) >> return H_TOO_HARD; >=20 > Yes. >=20 > This test covers the case in which a vCPU does XICS hcalls without QEMU=20 > having connected the vCPU to a XICS ICP. The ICP is the KVM XICS device=20 > on P8 or XICS-on-XIVE on P9. It catches QEMU errors when the interrupt=20 > mode is negotiated, we don't want the OS to do XICS hcalls after having=20 > negotiated the XIVE interrupt mode.=20 Okay. > It's different for the XIVE hcalls (when running under XICS) because they= =20 > are all handled in QEMU.=20 XIVE guest hcalls running on XICS host? >> if (!xics_on_xive()) >> return H_TOO_HARD; >=20 > I understand that this code is only called on P9 and with translation on. Yes. > On P9, we could have xics_on_xive() =3D=3D 0 if XIVE is disabled at compi= le=20 > time or with "xive=3Doff" at boot time. But guests should be supported.=20 > I don't see a reason to restrict the support even if these scenarios=20 > are rather unusual if not very rare. >=20 > on P10, it's the same but since we don't have the XICS emulation layer=20 > in OPAL, the host will be pretty useless. We don't care. >=20 > Since we are trying to handle hcalls, this is L0 and it can not be called= =20 > for nested guests, which would be another case of xics_on_xive() =3D=3D 0= .=20 > We don't care either. Okay so no xics_on_xive() test. I'll change that. Thanks, Nick