On 09/24/2014 05:28 AM, zhang bo wrote: > The patch > http://libvirt.org/git/?p=libvirt.git;a=commitdiff;h=de0aeafe9ce3eb414c8b... > removes invtsc flag in the host-model CPU. > > I'm wondering, will it be better to pass args "migratable=no/yes" to qemu, and let qemu complete the > remaining work? As that qemu has checked whether it's necessary to use invtsc or not. The 'migratable' property is only for -cpu host (<cpu mode='host-passthrough'> in libvirt). For mode='host-model', libvirt detects the model and features of the host CPU and passes it as -cpu <model>,+feat,+feat2,... so we can't leave that to QEMU. > ---------- > "invtsc is available only if using: -cpu host,migratable=no,+invtsc." > http://git.qemu.org/?p=qemu.git;a=commit;h=120eee7d1fdb2eba15766cfff7b9bc... > ---------- > > There's another problem, if we do not pass "migratable=no" to qemu. > Consider if we set host mode to pass-through > --------- > <cpu mode='host-passthrough'> > </cpu> > --------- > then the vm->def->cpu->features contains invtsc. however, qemu will automatically remove this cpu flag > as that "migration=no" is not passed to it. thus, the guest will not start up. This problem is in fact > caused by the patch: > http://libvirt.org/git/?p=libvirt.git;a=commit;h=fba6bc47cbcabbe08d422796..., > it forbids guest domain to start up if the host has INVTSC while the guest(qemu) does not. Regardless of QEMU support for invtsc, I'm only able to start the domain, restore or migration fails. As far as I know, only 'invtsc' is the problematic feature, because it both a) can appear in the host CPU (so libvirt assumes -cpu host will add it) b) is checked by qemuProcessVerifyGuestCPU (and libvirt complains when it's not there) For other features, we only add them to qemu command line and let qemu filter out the unsupported ones. > ------------- > for (i = 0; def->cpu && i < def->cpu->nfeatures; i++) { > virCPUFeatureDefPtr feature = &def->cpu->features[i]; > > if (feature->policy != VIR_CPU_FEATURE_REQUIRE) > continue; > > if (STREQ(feature->name, "invtsc") && > !cpuHasFeature(guestcpu, feature->name)) { > virReportError(VIR_ERR_CONFIG_UNSUPPORTED, "%s", > _("host doesn't support invariant TSC")); > goto cleanup; > } > } > break; > -------------- > > > In conclusion: > 1 Will it better to pass args "migratable=yes/no" to qemu rather than doing the mask-invtsc job in libvirt? > 2 If the guest has "pass-through" cpu mode, then it's unable to start up, because qemu removes invtsc, and > vm->def->cpu->features has it. It seems a BUG. > I think the simplest fix for host-passthrough would be to apply the same filter host-model has. But since using invtsc with host-passthrough requires both +invtsc and migratable=no, so we'd need to either add a 'migratable' option to host-passthrough (this would skip the filter and add migratable=on), or allow fine-tuning the features for host-passthrough too. Jan

On 09/25/2014 04:31 AM, zhang bo wrote: > On 2014/9/24 19:49, Ján Tomko wrote: >> I think the simplest fix for host-passthrough would be to apply the same >> filter host-model has. >> >> But since using invtsc with host-passthrough requires both +invtsc and >> migratable=no, so we'd need to either add a 'migratable' option to >> host-passthrough (this would skip the filter and add migratable=on), or allow >> fine-tuning the features for host-passthrough too. >> >> Jan >> > > Additional to the 2 suggestions, will that be OK to remove the codes in qemuProcessVerifyGuestCPU that checks whether the vm->def has > invtsc flag while qemu doesn't? > > - if (STREQ(feature->name, "invtsc") && > - !cpuHasFeature(guestcpu, feature->name)) { > - virReportError(VIR_ERR_CONFIG_UNSUPPORTED, "%s", > - _("host doesn't support invariant TSC")); > - goto cleanup; > - } Without this check, the feature would be quietly discarded by QEMU if the host kernel or host CPU does not support this feature. I think it's better to leave "invtsc" out when we're generating the cpu definition for host-passthrough, as we do for host-model. Jan > > Removing these codes, plus with the solution that "add 'migratable' option to host-passthrough", it seems the problem would > be gone, and invtsc would not be so 'distinctive' in libvirt any more. >