On Fri, Feb 22, 2019 at 12:28:33PM +0100, Cédric Le Goater wrote: > When migration of a VM is initiated, a first copy of the RAM is > transferred to the destination before the VM is stopped, but there is > no guarantee that the EQ pages in which the event notification are > queued have not been modified. > > To make sure migration will capture a consistent memory state, the > XIVE device should perform a XIVE quiesce sequence to stop the flow of > event notifications and stabilize the EQs. This is the purpose of the > KVM_DEV_XIVE_EQ_SYNC control which will also marks the EQ pages dirty > to force their transfer. > > Signed-off-by: Cédric Le Goater > --- > arch/powerpc/include/uapi/asm/kvm.h | 1 + > arch/powerpc/kvm/book3s_xive_native.c | 67 ++++++++++++++++++++++ > Documentation/virtual/kvm/devices/xive.txt | 29 ++++++++++ > 3 files changed, 97 insertions(+) > > diff --git a/arch/powerpc/include/uapi/asm/kvm.h b/arch/powerpc/include/uapi/asm/kvm.h > index 289c504b7c1d..cd78ad1020fe 100644 > --- a/arch/powerpc/include/uapi/asm/kvm.h > +++ b/arch/powerpc/include/uapi/asm/kvm.h > @@ -678,6 +678,7 @@ struct kvm_ppc_cpu_char { > /* POWER9 XIVE Native Interrupt Controller */ > #define KVM_DEV_XIVE_GRP_CTRL 1 > #define KVM_DEV_XIVE_RESET 1 > +#define KVM_DEV_XIVE_EQ_SYNC 2 > #define KVM_DEV_XIVE_GRP_SOURCE 2 /* 64-bit source attributes */ > #define KVM_DEV_XIVE_GRP_SOURCE_CONFIG 3 /* 64-bit source attributes */ > #define KVM_DEV_XIVE_GRP_EQ_CONFIG 4 /* 64-bit eq attributes */ > diff --git a/arch/powerpc/kvm/book3s_xive_native.c b/arch/powerpc/kvm/book3s_xive_native.c > index dd2a9d411fe7..3debc876d5a0 100644 > --- a/arch/powerpc/kvm/book3s_xive_native.c > +++ b/arch/powerpc/kvm/book3s_xive_native.c > @@ -640,6 +640,70 @@ static int kvmppc_xive_reset(struct kvmppc_xive *xive) > return 0; > } > > +static void kvmppc_xive_native_sync_sources(struct kvmppc_xive_src_block *sb) > +{ > + int j; > + > + for (j = 0; j < KVMPPC_XICS_IRQ_PER_ICS; j++) { > + struct kvmppc_xive_irq_state *state = &sb->irq_state[j]; > + struct xive_irq_data *xd; > + u32 hw_num; > + > + if (!state->valid) > + continue; > + if (state->act_priority == MASKED) Is this correct? If you masked an irq, then immediately did a sync, couldn't there still be some of the irqs in flight? I thought the reason we needed a sync was that masking and other such operations _didn't_ implicitly synchronize. > + continue; > + > + arch_spin_lock(&sb->lock); > + kvmppc_xive_select_irq(state, &hw_num, &xd); > + xive_native_sync_source(hw_num); > + xive_native_sync_queue(hw_num); > + arch_spin_unlock(&sb->lock); > + } > +} > + > +static int kvmppc_xive_native_vcpu_eq_sync(struct kvm_vcpu *vcpu) > +{ > + struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu; > + unsigned int prio; > + > + if (!xc) > + return -ENOENT; > + > + for (prio = 0; prio < KVMPPC_XIVE_Q_COUNT; prio++) { > + struct xive_q *q = &xc->queues[prio]; > + > + if (!q->qpage) > + continue; > + > + /* Mark EQ page dirty for migration */ > + mark_page_dirty(vcpu->kvm, gpa_to_gfn(q->guest_qpage)); > + } > + return 0; > +} > + > +static int kvmppc_xive_native_eq_sync(struct kvmppc_xive *xive) > +{ > + struct kvm *kvm = xive->kvm; > + struct kvm_vcpu *vcpu; > + unsigned int i; > + > + pr_devel("%s\n", __func__); > + > + for (i = 0; i <= xive->max_sbid; i++) { > + if (xive->src_blocks[i]) > + kvmppc_xive_native_sync_sources(xive->src_blocks[i]); > + } > + > + mutex_lock(&kvm->lock); > + kvm_for_each_vcpu(i, vcpu, kvm) { > + kvmppc_xive_native_vcpu_eq_sync(vcpu); > + } > + mutex_unlock(&kvm->lock); > + > + return 0; > +} > + > static int kvmppc_xive_native_set_attr(struct kvm_device *dev, > struct kvm_device_attr *attr) > { > @@ -650,6 +714,8 @@ static int kvmppc_xive_native_set_attr(struct kvm_device *dev, > switch (attr->attr) { > case KVM_DEV_XIVE_RESET: > return kvmppc_xive_reset(xive); > + case KVM_DEV_XIVE_EQ_SYNC: > + return kvmppc_xive_native_eq_sync(xive); > } > break; > case KVM_DEV_XIVE_GRP_SOURCE: > @@ -688,6 +754,7 @@ static int kvmppc_xive_native_has_attr(struct kvm_device *dev, > case KVM_DEV_XIVE_GRP_CTRL: > switch (attr->attr) { > case KVM_DEV_XIVE_RESET: > + case KVM_DEV_XIVE_EQ_SYNC: > return 0; > } > break; > diff --git a/Documentation/virtual/kvm/devices/xive.txt b/Documentation/virtual/kvm/devices/xive.txt > index 267634eae9e0..a26be635cff9 100644 > --- a/Documentation/virtual/kvm/devices/xive.txt > +++ b/Documentation/virtual/kvm/devices/xive.txt > @@ -23,6 +23,12 @@ the legacy interrupt mode, referred as XICS (POWER7/8). > queues. To be used by kexec and kdump. > Errors: none > > + 1.2 KVM_DEV_XIVE_EQ_SYNC (write only) > + Sync all the sources and queues and mark the EQ pages dirty. This > + to make sure that a consistent memory state is captured when > + migrating the VM. > + Errors: none > + > 2. KVM_DEV_XIVE_GRP_SOURCE (write only) > Initializes a new source in the XIVE device and mask it. > Attributes: > @@ -95,3 +101,26 @@ the legacy interrupt mode, referred as XICS (POWER7/8). > -ENOENT: Unknown source number > -EINVAL: Not initialized source number, invalid priority or > invalid CPU number. > + > +* Migration: > + > + Saving the state of a VM using the XIVE native exploitation mode > + should follow a specific sequence. When the VM is stopped : > + > + 1. Mask all sources (PQ=01) to stop the flow of events. > + > + 2. Sync the XIVE device with the KVM control KVM_DEV_XIVE_EQ_SYNC to > + flush any in-flight event notification and to stabilize the EQs. At > + this stage, the EQ pages are marked dirty to make sure they are > + transferred in the migration sequence. > + > + 3. Capture the state of the source targeting, the EQs configuration > + and the state of thread interrupt context registers. > + > + Restore is similar : > + > + 1. Restore the EQ configuration. As targeting depends on it. > + 2. Restore targeting > + 3. Restore the thread interrupt contexts > + 4. Restore the source states > + 5. Let the vCPU run -- David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson