From fbb59a17eebbb238223e66940e8fe5b9f64a8d74 Mon Sep 17 00:00:00 2001 From: Mikhail Malyshev Date: Mon, 9 Nov 2026 18:22:30 +0000 Subject: [PATCH 1/2] Describe IO threads that follow the HOUSEKEEPING pool Under IO_PLACEMENT_HOUSEKEEPING a VM's non-vCPU threads ran on the CPUs reserved for EVE alone, with a fallback for a node that reserves none. EVE now runs them on the whole HOUSEKEEPING pool -- the CPUs reserved for EVE and every other CPU no workload holds, never a kernel-isolated one -- and moves them as other workloads take and release CPUs, so they spread over every shared CPU instead of piling onto EVE's. Say so in IO_PLACEMENT_HOUSEKEEPING and AppCPUAssignment.io_cpu_ids. CPU_POOL_KIND_HOUSEKEEPING now says why the pool is never empty: EVE never gives the CPUs reserved for it (the lowest eve_max_vcpus) to a workload of its own, so the pool always holds them, and housekeeping workloads share them with EVE. That also leaves IO_PLACEMENT_HOUSEKEEPING without a case to refuse, so CPU_PLACEMENT_ERROR_INSUFFICIENT no longer names one. Comments only: no field or value changes. Signed-off-by: Mikhail Malyshev --- proto/config/vm.proto | 12 +++++++----- proto/info/info.proto | 14 +++++++++----- 2 files changed, 16 insertions(+), 10 deletions(-) diff --git a/proto/config/vm.proto b/proto/config/vm.proto index 0e42f3f0..a70aaede 100644 --- a/proto/config/vm.proto +++ b/proto/config/vm.proto @@ -75,11 +75,13 @@ enum IoPlacement { IO_PLACEMENT_UNSPECIFIED = 0; // The threads stay on the workload's own CPUs. IO_PLACEMENT_DEDICATED = 1; - // The threads run on the CPUs reserved for EVE, off the workload's own - // CPUs, so device emulation does not take cycles from busy vCPUs; on a node - // that reserves none, on HOUSEKEEPING CPUs no workload is planned onto. A - // node with no such CPU refuses the workload (info - // CPU_PLACEMENT_ERROR_INSUFFICIENT). + // The threads run on the CPUs of the HOUSEKEEPING pool (info + // CPU_POOL_KIND_HOUSEKEEPING), off the workload's own CPUs, so device + // emulation does not take cycles from busy vCPUs: the CPUs reserved for EVE + // and every other CPU no workload holds, never a kernel-isolated one. They + // follow the pool as other workloads take and release CPUs. The pool always + // holds the CPUs reserved for EVE, so the threads always have somewhere to + // run. IO_PLACEMENT_HOUSEKEEPING = 2; } diff --git a/proto/info/info.proto b/proto/info/info.proto index 7e645d52..189c92e4 100644 --- a/proto/info/info.proto +++ b/proto/info/info.proto @@ -579,7 +579,11 @@ message StorageInfo { enum CPUPoolKind { CPU_POOL_KIND_UNSPECIFIED = 0; // The CPUs reserved for EVE plus every CPU outside the isolated set that - // no workload holds: where CPU_POLICY_HOUSEKEEPING workloads run. + // no workload holds: where CPU_POLICY_HOUSEKEEPING workloads and the + // non-vCPU threads of IO_PLACEMENT_HOUSEKEEPING ones run. EVE never gives + // the CPUs reserved for it (the lowest eve_max_vcpus) to a workload of its + // own; housekeeping workloads share them with EVE. So the pool always holds + // them, and the DEDICATED pool can grow to every CPU but those. CPU_POOL_KIND_HOUSEKEEPING = 1; // CPUs outside the isolated set held by CPU_POLICY_DEDICATED workloads, // including SMT siblings held idle by a one-thread-per-core request. @@ -1208,8 +1212,9 @@ message AppCPUAssignment { // its own threads runs there either. In the pool of cpu_ids, never free. repeated uint32 parked_cpu_ids = 2; // Where a VM's non-vCPU QEMU threads run under IO_PLACEMENT_HOUSEKEEPING: - // CPUs of the HOUSEKEEPING pool, shared with EVE and not held. Empty - // otherwise, when those threads share cpu_ids. + // the CPUs of the HOUSEKEEPING pool, shared and not held, which change as + // other workloads take and release CPUs. Empty otherwise, when those threads + // share cpu_ids. repeated uint32 io_cpu_ids = 3; } @@ -1284,8 +1289,7 @@ enum CPUPlacementError { // restarts at once. CPU_PLACEMENT_ERROR_NEEDS_REPACK = 1; // The plan cannot place it, or only by leaving out a running workload, so - // no repack makes room for it. Includes a full isolated pool, and an - // IO_PLACEMENT_HOUSEKEEPING on a node with no CPU to serve it. + // no repack makes room for it. Includes a full isolated pool. CPU_PLACEMENT_ERROR_INSUFFICIENT = 2; // A field the request cannot carry, a value out of range, or a combination // this EVE does not implement. From 60c64c752f11249540b37cdf7e1548d3f538ef95 Mon Sep 17 00:00:00 2001 From: Mikhail Malyshev Date: Mon, 9 Nov 2026 18:23:17 +0000 Subject: [PATCH 2/2] Generate golang and python code for the IO placement comments Regenerated with make proto after the comment changes to vm.proto and info.proto; only the comments in the generated Go code change. Signed-off-by: Mikhail Malyshev --- go/config/vm.pb.go | 12 +++++++----- go/info/info.pb.go | 14 +++++++++----- 2 files changed, 16 insertions(+), 10 deletions(-) diff --git a/go/config/vm.pb.go b/go/config/vm.pb.go index b6378c4b..6f752535 100644 --- a/go/config/vm.pb.go +++ b/go/config/vm.pb.go @@ -230,11 +230,13 @@ const ( IoPlacement_IO_PLACEMENT_UNSPECIFIED IoPlacement = 0 // The threads stay on the workload's own CPUs. IoPlacement_IO_PLACEMENT_DEDICATED IoPlacement = 1 - // The threads run on the CPUs reserved for EVE, off the workload's own - // CPUs, so device emulation does not take cycles from busy vCPUs; on a node - // that reserves none, on HOUSEKEEPING CPUs no workload is planned onto. A - // node with no such CPU refuses the workload (info - // CPU_PLACEMENT_ERROR_INSUFFICIENT). + // The threads run on the CPUs of the HOUSEKEEPING pool (info + // CPU_POOL_KIND_HOUSEKEEPING), off the workload's own CPUs, so device + // emulation does not take cycles from busy vCPUs: the CPUs reserved for EVE + // and every other CPU no workload holds, never a kernel-isolated one. They + // follow the pool as other workloads take and release CPUs. The pool always + // holds the CPUs reserved for EVE, so the threads always have somewhere to + // run. IoPlacement_IO_PLACEMENT_HOUSEKEEPING IoPlacement = 2 ) diff --git a/go/info/info.pb.go b/go/info/info.pb.go index 211b42a3..fc1f7922 100644 --- a/go/info/info.pb.go +++ b/go/info/info.pb.go @@ -1065,7 +1065,11 @@ type CPUPoolKind int32 const ( CPUPoolKind_CPU_POOL_KIND_UNSPECIFIED CPUPoolKind = 0 // The CPUs reserved for EVE plus every CPU outside the isolated set that - // no workload holds: where CPU_POLICY_HOUSEKEEPING workloads run. + // no workload holds: where CPU_POLICY_HOUSEKEEPING workloads and the + // non-vCPU threads of IO_PLACEMENT_HOUSEKEEPING ones run. EVE never gives + // the CPUs reserved for it (the lowest eve_max_vcpus) to a workload of its + // own; housekeeping workloads share them with EVE. So the pool always holds + // them, and the DEDICATED pool can grow to every CPU but those. CPUPoolKind_CPU_POOL_KIND_HOUSEKEEPING CPUPoolKind = 1 // CPUs outside the isolated set held by CPU_POLICY_DEDICATED workloads, // including SMT siblings held idle by a one-thread-per-core request. @@ -1994,8 +1998,7 @@ const ( // restarts at once. CPUPlacementError_CPU_PLACEMENT_ERROR_NEEDS_REPACK CPUPlacementError = 1 // The plan cannot place it, or only by leaving out a running workload, so - // no repack makes room for it. Includes a full isolated pool, and an - // IO_PLACEMENT_HOUSEKEEPING on a node with no CPU to serve it. + // no repack makes room for it. Includes a full isolated pool. CPUPlacementError_CPU_PLACEMENT_ERROR_INSUFFICIENT CPUPlacementError = 2 // A field the request cannot carry, a value out of range, or a combination // this EVE does not implement. @@ -6768,8 +6771,9 @@ type AppCPUAssignment struct { // its own threads runs there either. In the pool of cpu_ids, never free. ParkedCpuIds []uint32 `protobuf:"varint,2,rep,packed,name=parked_cpu_ids,json=parkedCpuIds,proto3" json:"parked_cpu_ids,omitempty"` // Where a VM's non-vCPU QEMU threads run under IO_PLACEMENT_HOUSEKEEPING: - // CPUs of the HOUSEKEEPING pool, shared with EVE and not held. Empty - // otherwise, when those threads share cpu_ids. + // the CPUs of the HOUSEKEEPING pool, shared and not held, which change as + // other workloads take and release CPUs. Empty otherwise, when those threads + // share cpu_ids. IoCpuIds []uint32 `protobuf:"varint,3,rep,packed,name=io_cpu_ids,json=ioCpuIds,proto3" json:"io_cpu_ids,omitempty"` }