Experiment: zero-copy TCP send

(Yes, all EC n + k with k >= 2 users should upgrade as soon as possible)

CMakeLists.txt

@@ -2,6 +2,6 @@ cmake_minimum_required(VERSION 2.8.12)
 
 project(vitastor)
 
-set(VERSION "1.0.0")
+set(VERSION "1.1.0")
 
 add_subdirectory(src)

README-ru.md

@@ -50,6 +50,7 @@ Vitastor поддерживает QEMU-драйвер, протоколы NBD и
   - Параметры
     - [Общие](docs/config/common.ru.md)
     - [Сетевые](docs/config/network.ru.md)
+    - [Клиентский код](docs/config/client.en.md)
     - [Глобальные дисковые параметры](docs/config/layout-cluster.ru.md)
     - [Дисковые параметры OSD](docs/config/layout-osd.ru.md)
     - [Прочие параметры OSD](docs/config/osd.ru.md)

README.md

@@ -50,6 +50,7 @@ Read more details below in the documentation.
   - Parameter Reference
     - [Common](docs/config/common.en.md)
     - [Network](docs/config/network.en.md)
+    - [Client](docs/config/client.en.md)
     - [Global Disk Layout](docs/config/layout-cluster.en.md)
     - [OSD Disk Layout](docs/config/layout-osd.en.md)
     - [OSD Runtime Parameters](docs/config/osd.en.md)

csi/Makefile

@@ -1,4 +1,4 @@
-VERSION ?= v1.0.0
+VERSION ?= v1.1.0
 
 all: build push
 

csi/deploy/004-csi-nodeplugin.yaml

@@ -49,7 +49,7 @@ spec:
             capabilities:
               add: ["SYS_ADMIN"]
             allowPrivilegeEscalation: true
-          image: vitalif/vitastor-csi:v1.0.0
+          image: vitalif/vitastor-csi:v1.1.0
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/005-csi-provisioner-rbac.yaml

@@ -35,10 +35,13 @@ rules:
     verbs: ["get", "list", "watch"]
   - apiGroups: ["snapshot.storage.k8s.io"]
     resources: ["volumesnapshots"]
-    verbs: ["get", "list"]
+    verbs: ["get", "list", "patch"]
+  - apiGroups: ["snapshot.storage.k8s.io"]
+    resources: ["volumesnapshots/status"]
+    verbs: ["get", "list", "patch"]
   - apiGroups: ["snapshot.storage.k8s.io"]
     resources: ["volumesnapshotcontents"]
-    verbs: ["create", "get", "list", "watch", "update", "delete"]
+    verbs: ["create", "get", "list", "watch", "update", "delete", "patch"]
   - apiGroups: ["snapshot.storage.k8s.io"]
     resources: ["volumesnapshotclasses"]
     verbs: ["get", "list", "watch"]
@@ -53,7 +56,7 @@ rules:
     verbs: ["get", "list", "watch"]
   - apiGroups: ["snapshot.storage.k8s.io"]
     resources: ["volumesnapshotcontents/status"]
-    verbs: ["update"]
+    verbs: ["update", "patch"]
   - apiGroups: [""]
     resources: ["configmaps"]
     verbs: ["get"]

csi/deploy/007-csi-provisioner.yaml

@@ -23,6 +23,11 @@ metadata:
   name: csi-vitastor-provisioner
 spec:
   replicas: 3
+  strategy:
+    type: RollingUpdate
+    rollingUpdate:
+      maxUnavailable: 1
+      maxSurge: 0
   selector:
     matchLabels:
       app: csi-vitastor-provisioner
@@ -46,7 +51,7 @@ spec:
       priorityClassName: system-cluster-critical
       containers:
         - name: csi-provisioner
-          image: k8s.gcr.io/sig-storage/csi-provisioner:v2.2.0
+          image: k8s.gcr.io/sig-storage/csi-provisioner:v3.0.0
           args:
             - "--csi-address=$(ADDRESS)"
             - "--v=5"
@@ -116,7 +121,7 @@ spec:
             privileged: true
             capabilities:
               add: ["SYS_ADMIN"]
-          image: vitalif/vitastor-csi:v1.0.0
+          image: vitalif/vitastor-csi:v1.1.0
           args:
             - "--node=$(NODE_ID)"
             - "--endpoint=$(CSI_ENDPOINT)"

csi/deploy/009-storage-class.yaml

@@ -17,3 +17,4 @@ parameters:
   # multiple etcdUrls may be specified, delimited by comma
   #etcdUrl: "http://192.168.7.2:2379"
   #etcdPrefix: "/vitastor"
+allowVolumeExpansion: true

csi/deploy/example-snapshot-class.yaml

@@ -0,0 +1,7 @@
+apiVersion: snapshot.storage.k8s.io/v1
+kind: VolumeSnapshotClass
+metadata:
+  name: vitastor-snapclass
+driver: csi.vitastor.io
+deletionPolicy: Delete
+parameters:

csi/deploy/example-snapshot-clone.yaml

@@ -0,0 +1,16 @@
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: test-vitastor-clone
+spec:
+  storageClassName: vitastor
+  dataSource:
+    name: snap1
+    kind: VolumeSnapshot
+    apiGroup: snapshot.storage.k8s.io
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 10Gi

csi/deploy/example-snapshot.yaml

@@ -0,0 +1,8 @@
+apiVersion: snapshot.storage.k8s.io/v1
+kind: VolumeSnapshot
+metadata:
+  name: snap1
+spec:
+  volumeSnapshotClassName: vitastor-snapclass
+  source:
+    persistentVolumeClaimName: test-vitastor-pvc

csi/go.mod

@@ -9,6 +9,7 @@ require (
 	golang.org/x/net v0.0.0-20201202161906-c7110b5ffcbb
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
 	google.golang.org/grpc v1.33.1
+	google.golang.org/protobuf v1.24.0
 	k8s.io/klog v1.0.0
 	k8s.io/utils v0.0.0-20210305010621-2afb4311ab10
 )

csi/src/config.go

@@ -5,7 +5,7 @@ package vitastor
 
 const (
     vitastorCSIDriverName    = "csi.vitastor.io"
-    vitastorCSIDriverVersion = "1.0.0"
+    vitastorCSIDriverVersion = "1.1.0"
 )
 
 // Config struct fills the parameters of request or user input

csi/src/controllerserver.go

@@ -20,6 +20,7 @@ import (
 
     "google.golang.org/grpc/codes"
     "google.golang.org/grpc/status"
+    "google.golang.org/protobuf/types/known/timestamppb"
 
     "github.com/container-storage-interface/spec/lib/go/csi"
 )
@@ -45,6 +46,7 @@ type InodeConfig struct
     ParentPool uint64 `json:"parent_pool,omitempty"`
     ParentId uint64 `json:"parent_id,omitempty"`
     Readonly bool `json:"readonly,omitempty"`
+    CreateTs uint64 `json:"create_ts,omitempty"`
 }
 
 type ControllerServer struct
@@ -178,27 +180,43 @@ func (cs *ControllerServer) CreateVolume(ctx context.Context, req *csi.CreateVol
         return nil, status.Error(codes.InvalidArgument, "no etcdUrl in storage class configuration and no etcd_address in vitastor.conf")
     }
 
+    args := []string{ "create", volName, "-s", fmt.Sprintf("%v", volSize), "--pool", fmt.Sprintf("%v", poolId) }
+
+    // Support creation from snapshot
+    var src *csi.VolumeContentSource
+    if (req.VolumeContentSource.GetSnapshot() != nil)
+    {
+        snapId := req.VolumeContentSource.GetSnapshot().GetSnapshotId()
+        if (snapId != "")
+        {
+            snapVars := make(map[string]string)
+            err := json.Unmarshal([]byte(snapId), &snapVars)
+            if (err != nil)
+            {
+                return nil, status.Error(codes.Internal, "volume ID not in JSON format")
+            }
+            args = append(args, "--parent", snapVars["name"]+"@"+snapVars["snapshot"])
+            src = &csi.VolumeContentSource{
+                Type: &csi.VolumeContentSource_Snapshot{
+                    Snapshot: &csi.VolumeContentSource_SnapshotSource{
+                        SnapshotId: snapId,
+                    },
+                },
+            }
+        }
+    }
+
     // Create image using vitastor-cli
-    _, err := invokeCLI(ctxVars, []string{ "create", volName, "-s", fmt.Sprintf("%v", volSize), "--pool", fmt.Sprintf("%v", poolId) })
+    _, err := invokeCLI(ctxVars, args)
     if (err != nil)
     {
         if (strings.Index(err.Error(), "already exists") > 0)
         {
-            stat, err := invokeCLI(ctxVars, []string{ "ls", "--json", volName })
+            inodeCfg, err := invokeList(ctxVars, volName, true)
             if (err != nil)
             {
                 return nil, err
             }
-            var inodeCfg []InodeConfig
-            err = json.Unmarshal(stat, &inodeCfg)
-            if (err != nil)
-            {
-                return nil, status.Error(codes.Internal, "Invalid JSON in vitastor-cli ls: "+err.Error())
-            }
-            if (len(inodeCfg) == 0)
-            {
-                return nil, status.Error(codes.Internal, "vitastor-cli create said that image already exists, but ls can't find it")
-            }
             if (inodeCfg[0].Size < uint64(volSize))
             {
                 return nil, status.Error(codes.Internal, "image "+volName+" is already created, but size is less than expected")
@@ -217,6 +235,7 @@ func (cs *ControllerServer) CreateVolume(ctx context.Context, req *csi.CreateVol
             // Ugly, but VolumeContext isn't passed to DeleteVolume :-(
             VolumeId: string(volumeIdJson),
             CapacityBytes: volSize,
+            ContentSource: src,
         },
     }, nil
 }
@@ -230,15 +249,15 @@ func (cs *ControllerServer) DeleteVolume(ctx context.Context, req *csi.DeleteVol
         return nil, status.Error(codes.InvalidArgument, "request cannot be empty")
     }
 
-    ctxVars := make(map[string]string)
-    err := json.Unmarshal([]byte(req.VolumeId), &ctxVars)
+    volVars := make(map[string]string)
+    err := json.Unmarshal([]byte(req.VolumeId), &volVars)
     if (err != nil)
     {
         return nil, status.Error(codes.Internal, "volume ID not in JSON format")
     }
-    volName := ctxVars["name"]
+    volName := volVars["name"]
 
-    ctxVars, _, _ = GetConnectionParams(ctxVars)
+    ctxVars, _, _ := GetConnectionParams(volVars)
 
     _, err = invokeCLI(ctxVars, []string{ "rm", volName })
     if (err != nil)
@@ -344,6 +363,8 @@ func (cs *ControllerServer) ControllerGetCapabilities(ctx context.Context, req *
         csi.ControllerServiceCapability_RPC_LIST_VOLUMES,
         csi.ControllerServiceCapability_RPC_EXPAND_VOLUME,
         csi.ControllerServiceCapability_RPC_CREATE_DELETE_SNAPSHOT,
+        csi.ControllerServiceCapability_RPC_LIST_SNAPSHOTS,
+        // TODO: csi.ControllerServiceCapability_RPC_CLONE_VOLUME,
     } {
         controllerServerCapabilities = append(controllerServerCapabilities, functionControllerServerCapabilities(capability))
     }
@@ -353,28 +374,214 @@ func (cs *ControllerServer) ControllerGetCapabilities(ctx context.Context, req *
     }, nil
 }
 
+func invokeList(ctxVars map[string]string, pattern string, expectExist bool) ([]InodeConfig, error)
+{
+    stat, err := invokeCLI(ctxVars, []string{ "ls", "--json", pattern })
+    if (err != nil)
+    {
+        return nil, err
+    }
+    var inodeCfg []InodeConfig
+    err = json.Unmarshal(stat, &inodeCfg)
+    if (err != nil)
+    {
+        return nil, status.Error(codes.Internal, "Invalid JSON in vitastor-cli ls: "+err.Error())
+    }
+    if (expectExist && len(inodeCfg) == 0)
+    {
+        return nil, status.Error(codes.Internal, "Can't find expected image "+pattern+" via vitastor-cli ls")
+    }
+    return inodeCfg, nil
+}
+
 // CreateSnapshot create snapshot of an existing PV
 func (cs *ControllerServer) CreateSnapshot(ctx context.Context, req *csi.CreateSnapshotRequest) (*csi.CreateSnapshotResponse, error)
 {
-    return nil, status.Error(codes.Unimplemented, "")
+    klog.Infof("received controller create snapshot request %+v", protosanitizer.StripSecrets(req))
+    if (req == nil)
+    {
+        return nil, status.Errorf(codes.InvalidArgument, "request cannot be empty")
+    }
+    if (req.SourceVolumeId == "" || req.Name == "")
+    {
+        return nil, status.Error(codes.InvalidArgument, "source volume ID and snapshot name are required fields")
+    }
+
+    // snapshot name
+    snapName := req.Name
+
+    // req.VolumeId is an ugly json string in our case :)
+    ctxVars := make(map[string]string)
+    err := json.Unmarshal([]byte(req.SourceVolumeId), &ctxVars)
+    if (err != nil)
+    {
+        return nil, status.Error(codes.Internal, "volume ID not in JSON format")
+    }
+    volName := ctxVars["name"]
+
+    // Create image using vitastor-cli
+    _, err = invokeCLI(ctxVars, []string{ "create", "--snapshot", snapName, volName })
+    if (err != nil && strings.Index(err.Error(), "already exists") <= 0)
+    {
+        return nil, err
+    }
+
+    // Check created snapshot
+    inodeCfg, err := invokeList(ctxVars, volName+"@"+snapName, true)
+    if (err != nil)
+    {
+        return nil, err
+    }
+
+    // Use ugly JSON snapshot ID again, DeleteSnapshot doesn't have context :-(
+    ctxVars["snapshot"] = snapName
+    snapIdJson, _ := json.Marshal(ctxVars)
+    return &csi.CreateSnapshotResponse{
+        Snapshot: &csi.Snapshot{
+            SizeBytes: int64(inodeCfg[0].Size),
+            SnapshotId: string(snapIdJson),
+            SourceVolumeId: req.SourceVolumeId,
+            CreationTime: &timestamppb.Timestamp{ Seconds: int64(inodeCfg[0].CreateTs) },
+            ReadyToUse: true,
+        },
+    }, nil
 }
 
 // DeleteSnapshot delete provided snapshot of a PV
 func (cs *ControllerServer) DeleteSnapshot(ctx context.Context, req *csi.DeleteSnapshotRequest) (*csi.DeleteSnapshotResponse, error)
 {
-    return nil, status.Error(codes.Unimplemented, "")
+    klog.Infof("received controller delete snapshot request %+v", protosanitizer.StripSecrets(req))
+    if (req == nil)
+    {
+        return nil, status.Errorf(codes.InvalidArgument, "request cannot be empty")
+    }
+    if (req.SnapshotId == "")
+    {
+        return nil, status.Error(codes.InvalidArgument, "snapshot ID is a required field")
+    }
+
+    volVars := make(map[string]string)
+    err := json.Unmarshal([]byte(req.SnapshotId), &volVars)
+    if (err != nil)
+    {
+        return nil, status.Error(codes.Internal, "snapshot ID not in JSON format")
+    }
+    volName := volVars["name"]
+    snapName := volVars["snapshot"]
+
+    ctxVars, _, _ := GetConnectionParams(volVars)
+
+    _, err = invokeCLI(ctxVars, []string{ "rm", volName+"@"+snapName })
+    if (err != nil)
+    {
+        return nil, err
+    }
+
+    return &csi.DeleteSnapshotResponse{}, nil
 }
 
 // ListSnapshots list the snapshots of a PV
 func (cs *ControllerServer) ListSnapshots(ctx context.Context, req *csi.ListSnapshotsRequest) (*csi.ListSnapshotsResponse, error)
 {
-    return nil, status.Error(codes.Unimplemented, "")
+    klog.Infof("received controller list snapshots request %+v", protosanitizer.StripSecrets(req))
+    if (req == nil)
+    {
+        return nil, status.Error(codes.InvalidArgument, "request cannot be empty")
+    }
+
+    volVars := make(map[string]string)
+    err := json.Unmarshal([]byte(req.SourceVolumeId), &volVars)
+    if (err != nil)
+    {
+        return nil, status.Error(codes.Internal, "volume ID not in JSON format")
+    }
+    volName := volVars["name"]
+    ctxVars, _, _ := GetConnectionParams(volVars)
+
+    inodeCfg, err := invokeList(ctxVars, volName+"@*", false)
+    if (err != nil)
+    {
+        return nil, err
+    }
+
+    resp := &csi.ListSnapshotsResponse{}
+    for _, ino := range inodeCfg
+    {
+        snapName := ino.Name[len(volName)+1:]
+        if (len(req.StartingToken) > 0 && snapName < req.StartingToken)
+        {
+        }
+        else if (req.MaxEntries == 0 || len(resp.Entries) < int(req.MaxEntries))
+        {
+            volVars["snapshot"] = snapName
+            snapIdJson, _ := json.Marshal(volVars)
+            resp.Entries = append(resp.Entries, &csi.ListSnapshotsResponse_Entry{
+                Snapshot: &csi.Snapshot{
+                    SizeBytes: int64(ino.Size),
+                    SnapshotId: string(snapIdJson),
+                    SourceVolumeId: req.SourceVolumeId,
+                    CreationTime: &timestamppb.Timestamp{ Seconds: int64(ino.CreateTs) },
+                    ReadyToUse: true,
+                },
+            })
+        }
+        else
+        {
+            resp.NextToken = snapName
+            break
+        }
+    }
+
+    return resp, nil
 }
 
-// ControllerExpandVolume resizes a volume
+// ControllerExpandVolume increases the size of a volume
 func (cs *ControllerServer) ControllerExpandVolume(ctx context.Context, req *csi.ControllerExpandVolumeRequest) (*csi.ControllerExpandVolumeResponse, error)
 {
-    return nil, status.Error(codes.Unimplemented, "")
+    klog.Infof("received controller expand volume request %+v", protosanitizer.StripSecrets(req))
+    if (req == nil)
+    {
+        return nil, status.Error(codes.InvalidArgument, "request cannot be empty")
+    }
+    if (req.VolumeId == "" || req.CapacityRange == nil || req.CapacityRange.RequiredBytes == 0)
+    {
+        return nil, status.Error(codes.InvalidArgument, "VolumeId, CapacityRange and RequiredBytes are required fields")
+    }
+
+    volVars := make(map[string]string)
+    err := json.Unmarshal([]byte(req.VolumeId), &volVars)
+    if (err != nil)
+    {
+        return nil, status.Error(codes.Internal, "volume ID not in JSON format")
+    }
+    volName := volVars["name"]
+    ctxVars, _, _ := GetConnectionParams(volVars)
+
+    inodeCfg, err := invokeList(ctxVars, volName, true)
+    if (err != nil)
+    {
+        return nil, err
+    }
+
+    if (req.CapacityRange.RequiredBytes > 0 && inodeCfg[0].Size < uint64(req.CapacityRange.RequiredBytes))
+    {
+        sz := ((req.CapacityRange.RequiredBytes+4095)/4096)*4096
+        _, err := invokeCLI(ctxVars, []string{ "modify", "--inc_size", "1", "--resize", fmt.Sprintf("%d", sz), volName })
+        if (err != nil)
+        {
+            return nil, err
+        }
+        inodeCfg, err = invokeList(ctxVars, volName, true)
+        if (err != nil)
+        {
+            return nil, err
+        }
+    }
+
+    return &csi.ControllerExpandVolumeResponse{
+        CapacityBytes: int64(inodeCfg[0].Size),
+        NodeExpansionRequired: false,
+    }, nil
 }
 
 // ControllerGetVolume get volume info

csi/src/identityserver.go

@@ -49,6 +49,13 @@ func (is *IdentityServer) GetPluginCapabilities(ctx context.Context, req *csi.Ge
                     },
                 },
             },
+            {
+                Type: &csi.PluginCapability_VolumeExpansion_{
+                    VolumeExpansion: &csi.PluginCapability_VolumeExpansion{
+                        Type: csi.PluginCapability_VolumeExpansion_OFFLINE,
+                    },
+                },
+            },
         },
     }, nil
 }

csi/src/nodeserver.go

@@ -70,10 +70,10 @@ func (ns *NodeServer) NodePublishVolume(ctx context.Context, req *csi.NodePublis
     isBlock := req.GetVolumeCapability().GetBlock() != nil
 
     // Check that it's not already mounted
-    _, error := mount.IsNotMountPoint(ns.mounter, targetPath)
-    if (error != nil)
+    _, err := mount.IsNotMountPoint(ns.mounter, targetPath)
+    if (err != nil)
     {
-        if (os.IsNotExist(error))
+        if (os.IsNotExist(err))
         {
             if (isBlock)
             {
@@ -102,12 +102,12 @@ func (ns *NodeServer) NodePublishVolume(ctx context.Context, req *csi.NodePublis
         }
         else
         {
-            return nil, status.Error(codes.Internal, error.Error())
+            return nil, status.Error(codes.Internal, err.Error())
         }
     }
 
     ctxVars := make(map[string]string)
-    err := json.Unmarshal([]byte(req.VolumeId), &ctxVars)
+    err = json.Unmarshal([]byte(req.VolumeId), &ctxVars)
     if (err != nil)
     {
         return nil, status.Error(codes.Internal, "volume ID not in JSON format")
@@ -147,70 +147,74 @@ func (ns *NodeServer) NodePublishVolume(ctx context.Context, req *csi.NodePublis
     }
     devicePath := strings.TrimSpace(stdoutStr)
 
-    // Check existing format
     diskMounter := &mount.SafeFormatAndMount{Interface: ns.mounter, Exec: utilexec.New()}
-    existingFormat, err := diskMounter.GetDiskFormat(devicePath)
-    if (err != nil)
-    {
-        klog.Errorf("failed to get disk format for path %s, error: %v", err)
-        // unmap NBD device
-        unmapOut, unmapErr := exec.Command("/usr/bin/vitastor-nbd", "unmap", devicePath).CombinedOutput()
-        if (unmapErr != nil)
-        {
-            klog.Errorf("failed to unmap NBD device %s: %s, error: %v", devicePath, unmapOut, unmapErr)
-        }
-        return nil, err
-    }
-
-    // Format the device (ext4 or xfs)
-    fsType := req.GetVolumeCapability().GetMount().GetFsType()
-    opt := req.GetVolumeCapability().GetMount().GetMountFlags()
-    opt = append(opt, "_netdev")
-    if ((req.VolumeCapability.AccessMode.Mode == csi.VolumeCapability_AccessMode_MULTI_NODE_READER_ONLY ||
-        req.VolumeCapability.AccessMode.Mode == csi.VolumeCapability_AccessMode_SINGLE_NODE_READER_ONLY) &&
-        !Contains(opt, "ro"))
-    {
-        opt = append(opt, "ro")
-    }
-    if (fsType == "xfs")
-    {
-        opt = append(opt, "nouuid")
-    }
-    readOnly := Contains(opt, "ro")
-    if (existingFormat == "" && !readOnly)
-    {
-        args := []string{}
-        switch fsType
-        {
-            case "ext4":
-                args = []string{"-m0", "-Enodiscard,lazy_itable_init=1,lazy_journal_init=1", devicePath}
-            case "xfs":
-                args = []string{"-K", devicePath}
-        }
-        if (len(args) > 0)
-        {
-            cmdOut, cmdErr := diskMounter.Exec.Command("mkfs."+fsType, args...).CombinedOutput()
-            if (cmdErr != nil)
-            {
-                klog.Errorf("failed to run mkfs error: %v, output: %v", cmdErr, string(cmdOut))
-                // unmap NBD device
-                unmapOut, unmapErr := exec.Command("/usr/bin/vitastor-nbd", "unmap", devicePath).CombinedOutput()
-                if (unmapErr != nil)
-                {
-                    klog.Errorf("failed to unmap NBD device %s: %s, error: %v", devicePath, unmapOut, unmapErr)
-                }
-                return nil, status.Error(codes.Internal, cmdErr.Error())
-            }
-        }
-    }
     if (isBlock)
     {
-        opt = append(opt, "bind")
-        err = diskMounter.Mount(devicePath, targetPath, fsType, opt)
+        err = diskMounter.Mount(devicePath, targetPath, "", []string{"bind"})
     }
     else
     {
+        // Check existing format
+        existingFormat, err := diskMounter.GetDiskFormat(devicePath)
+        if (err != nil)
+        {
+            klog.Errorf("failed to get disk format for path %s, error: %v", err)
+            goto unmap
+        }
+
+        // Format the device (ext4 or xfs)
+        fsType := req.GetVolumeCapability().GetMount().GetFsType()
+        opt := req.GetVolumeCapability().GetMount().GetMountFlags()
+        opt = append(opt, "_netdev")
+        if ((req.VolumeCapability.AccessMode.Mode == csi.VolumeCapability_AccessMode_MULTI_NODE_READER_ONLY ||
+            req.VolumeCapability.AccessMode.Mode == csi.VolumeCapability_AccessMode_SINGLE_NODE_READER_ONLY) &&
+            !Contains(opt, "ro"))
+        {
+            opt = append(opt, "ro")
+        }
+        if (fsType == "xfs")
+        {
+            opt = append(opt, "nouuid")
+        }
+        readOnly := Contains(opt, "ro")
+        if (existingFormat == "" && !readOnly)
+        {
+            var cmdOut []byte
+            switch fsType
+            {
+                case "ext4":
+                    args := []string{"-m0", "-Enodiscard,lazy_itable_init=1,lazy_journal_init=1", devicePath}
+                    cmdOut, err = diskMounter.Exec.Command("mkfs.ext4", args...).CombinedOutput()
+                case "xfs":
+                    cmdOut, err = diskMounter.Exec.Command("mkfs.xfs", "-K", devicePath).CombinedOutput()
+            }
+            if (err != nil)
+            {
+                klog.Errorf("failed to run mkfs error: %v, output: %v", err, string(cmdOut))
+                goto unmap
+            }
+        }
+
         err = diskMounter.FormatAndMount(devicePath, targetPath, fsType, opt)
+
+        // Try to run online resize on mount.
+        // FIXME: Implement online resize. It requires online resize support in vitastor-nbd.
+        if (err == nil && existingFormat != "" && !readOnly)
+        {
+            var cmdOut []byte
+            switch (fsType)
+            {
+                case "ext4":
+                    cmdOut, err = diskMounter.Exec.Command("resize2fs", devicePath).CombinedOutput()
+                case "xfs":
+                    cmdOut, err = diskMounter.Exec.Command("xfs_growfs", devicePath).CombinedOutput()
+            }
+            if (err != nil)
+            {
+                klog.Errorf("failed to run resizefs error: %v, output: %v", err, string(cmdOut))
+                goto unmap
+            }
+        }
     }
     if (err != nil)
     {
@@ -218,15 +222,18 @@ func (ns *NodeServer) NodePublishVolume(ctx context.Context, req *csi.NodePublis
             "failed to mount device path (%s) to path (%s) for volume (%s) error: %s",
             devicePath, targetPath, volName, err,
         )
-        // unmap NBD device
-        unmapOut, unmapErr := exec.Command("/usr/bin/vitastor-nbd", "unmap", devicePath).CombinedOutput()
-        if (unmapErr != nil)
-        {
-            klog.Errorf("failed to unmap NBD device %s: %s, error: %v", devicePath, unmapOut, unmapErr)
-        }
-        return nil, status.Error(codes.Internal, err.Error())
+        goto unmap
     }
     return &csi.NodePublishVolumeResponse{}, nil
+
+unmap:
+    // unmap NBD device
+    unmapOut, unmapErr := exec.Command("/usr/bin/vitastor-nbd", "unmap", devicePath).CombinedOutput()
+    if (unmapErr != nil)
+    {
+        klog.Errorf("failed to unmap NBD device %s: %s, error: %v", devicePath, unmapOut, unmapErr)
+    }
+    return nil, status.Error(codes.Internal, err.Error())
 }
 
 // NodeUnpublishVolume unmounts the volume from the target path

debian/changelog

@@ -1,10 +1,10 @@
-vitastor (1.0.0-1) unstable; urgency=medium
+vitastor (1.1.0-1) unstable; urgency=medium
 
   * Bugfixes
 
  -- Vitaliy Filippov <vitalif@yourcmc.ru>  Fri, 03 Jun 2022 02:09:44 +0300
 
-vitastor (1.0.0-1) unstable; urgency=medium
+vitastor (1.1.0-1) unstable; urgency=medium
 
   * Implement NFS proxy
   * Add documentation

debian/control

@@ -2,7 +2,7 @@ Source: vitastor
 Section: admin
 Priority: optional
 Maintainer: Vitaliy Filippov <vitalif@yourcmc.ru>
-Build-Depends: debhelper, liburing-dev (>= 0.6), g++ (>= 8), libstdc++6 (>= 8), linux-libc-dev, libgoogle-perftools-dev, libjerasure-dev, libgf-complete-dev, libibverbs-dev, libisal-dev
+Build-Depends: debhelper, liburing-dev (>= 0.6), g++ (>= 8), libstdc++6 (>= 8), linux-libc-dev, libgoogle-perftools-dev, libjerasure-dev, libgf-complete-dev, libibverbs-dev, libisal-dev, cmake, pkg-config
 Standards-Version: 4.5.0
 Homepage: https://vitastor.io/
 Rules-Requires-Root: no

debian/patched-qemu.Dockerfile

@@ -54,7 +54,8 @@ RUN set -e; \
     quilt add block/vitastor.c; \
     cp /root/vitastor/src/qemu_driver.c block/vitastor.c; \
     quilt refresh; \
-    V=$(head -n1 debian/changelog | perl -pe 's/^.*\((.*?)(~bpo[\d\+]*)?\).*$/$1/')+vitastor3; \
+    V=$(head -n1 debian/changelog | perl -pe 's/5\.2\+dfsg-9/5.2+dfsg-11/; s/^.*\((.*?)(~bpo[\d\+]*)?\).*$/$1/')+vitastor4; \
+    if [ "$REL" = bullseye ]; then V=${V}bullseye; fi; \
     DEBEMAIL="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v $V 'Plug Vitastor block driver'; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \
     rm -rf /root/packages/qemu-$REL/qemu-*/

debian/vitastor.Dockerfile

@@ -35,8 +35,8 @@ RUN set -e -x; \
     mkdir -p /root/packages/vitastor-$REL; \
     rm -rf /root/packages/vitastor-$REL/*; \
     cd /root/packages/vitastor-$REL; \
-    cp -r /root/vitastor vitastor-1.0.0; \
-    cd vitastor-1.0.0; \
+    cp -r /root/vitastor vitastor-1.1.0; \
+    cd vitastor-1.1.0; \
     ln -s /root/fio-build/fio-*/ ./fio; \
     FIO=$(head -n1 fio/debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     ls /usr/include/linux/raw.h || cp ./debian/raw.h /usr/include/linux/raw.h; \
@@ -49,8 +49,8 @@ RUN set -e -x; \
     rm -rf a b; \
     echo "dep:fio=$FIO" > debian/fio_version; \
     cd /root/packages/vitastor-$REL; \
-    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_1.0.0.orig.tar.xz vitastor-1.0.0; \
-    cd vitastor-1.0.0; \
+    tar --sort=name --mtime='2020-01-01' --owner=0 --group=0 --exclude=debian -cJf vitastor_1.1.0.orig.tar.xz vitastor-1.1.0; \
+    cd vitastor-1.1.0; \
     V=$(head -n1 debian/changelog | perl -pe 's/^.*\((.*?)\).*$/$1/'); \
     DEBFULLNAME="Vitaliy Filippov <vitalif@yourcmc.ru>" dch -D $REL -v "$V""$REL" "Rebuild for $REL"; \
     DEB_BUILD_OPTIONS=nocheck dpkg-buildpackage --jobs=auto -sa; \

docs/config.en.md

@@ -33,6 +33,7 @@ In the future, additional configuration methods may be added:
 
 - [Common](config/common.en.md)
 - [Network](config/network.en.md)
+- [Client](config/client.en.md)
 - [Global Disk Layout](config/layout-cluster.en.md)
 - [OSD Disk Layout](config/layout-osd.en.md)
 - [OSD Runtime Parameters](config/osd.en.md)

docs/config.ru.md

@@ -36,6 +36,7 @@
 
 - [Общие](config/common.ru.md)
 - [Сеть](config/network.ru.md)
+- [Клиентский код](config/client.ru.md)
 - [Глобальные дисковые параметры](config/layout-cluster.ru.md)
 - [Дисковые параметры OSD](config/layout-osd.ru.md)
 - [Прочие параметры OSD](config/osd.ru.md)

docs/config/client.en.md

@@ -0,0 +1,103 @@
+[Documentation](../../README.md#documentation) → [Configuration](../config.en.md) → Client Parameters
+
+-----
+
+[Читать на русском](client.ru.md)
+
+# Client Parameters
+
+These parameters apply only to clients and affect their interaction with
+the cluster.
+
+- [client_max_dirty_bytes](#client_max_dirty_bytes)
+- [client_max_dirty_ops](#client_max_dirty_ops)
+- [client_enable_writeback](#client_enable_writeback)
+- [client_max_buffered_bytes](#client_max_buffered_bytes)
+- [client_max_buffered_ops](#client_max_buffered_ops)
+- [client_max_writeback_iodepth](#client_max_writeback_iodepth)
+
+## client_max_dirty_bytes
+
+- Type: integer
+- Default: 33554432
+- Can be changed online: yes
+
+Without [immediate_commit](layout-cluster.en.md#immediate_commit)=all this parameter sets the limit of "dirty"
+(not committed by fsync) data allowed by the client before forcing an
+additional fsync and committing the data. Also note that the client always
+holds a copy of uncommitted data in memory so this setting also affects
+RAM usage of clients.
+
+## client_max_dirty_ops
+
+- Type: integer
+- Default: 1024
+- Can be changed online: yes
+
+Same as client_max_dirty_bytes, but instead of total size, limits the number
+of uncommitted write operations.
+
+## client_enable_writeback
+
+- Type: boolean
+- Default: false
+- Can be changed online: yes
+
+This parameter enables client-side write buffering. This means that write
+requests are accumulated in memory for a short time before being sent to
+a Vitastor cluster which allows to send them in parallel and increase
+performance of some applications. Writes are buffered until client forces
+a flush with fsync() or until the amount of buffered writes exceeds the
+limit.
+
+Write buffering significantly increases performance of some applications,
+for example, CrystalDiskMark under Windows (LOL :-D), but also any other
+applications if they do writes in one of two non-optimal ways: either if
+they do a lot of small (4 kb or so) sequential writes, or if they do a lot
+of small random writes, but without any parallelism or asynchrony, and also
+without calling fsync().
+
+With write buffering enabled, you can expect around 22000 T1Q1 random write
+iops in QEMU more or less regardless of the quality of your SSDs, and this
+number is in fact bound by QEMU itself rather than Vitastor (check it
+yourself by adding a "driver=null-co" disk in QEMU). Without write
+buffering, the current record is 9900 iops, but the number is usually
+even lower with non-ideal hardware, for example, it may be 5000 iops.
+
+Even when this parameter is enabled, write buffering isn't enabled until
+the client explicitly allows it, because enabling it without the client
+being aware of the fact that his writes may be buffered may lead to data
+loss. Because of this, older versions of clients don't support write
+buffering at all, newer versions of the QEMU driver allow write buffering
+only if it's enabled in disk settings with `-blockdev cache.direct=false`,
+and newer versions of FIO only allow write buffering if you don't specify
+`-direct=1`. NBD and NFS drivers allow write buffering by default.
+
+You can overcome this restriction too with the `client_writeback_allowed`
+parameter, but you shouldn't do that unless you **really** know what you
+are doing.
+
+## client_max_buffered_bytes
+
+- Type: integer
+- Default: 33554432
+- Can be changed online: yes
+
+Maximum total size of buffered writes which triggers write-back when reached.
+
+## client_max_buffered_ops
+
+- Type: integer
+- Default: 1024
+- Can be changed online: yes
+
+Maximum number of buffered writes which triggers write-back when reached.
+Multiple consecutive modified data regions are counted as 1 write here.
+
+## client_max_writeback_iodepth
+
+- Type: integer
+- Default: 256
+- Can be changed online: yes
+
+Maximum number of parallel writes when flushing buffered data to the server.

docs/config/client.ru.md

@@ -0,0 +1,103 @@
+[Документация](../../README-ru.md#документация) → [Конфигурация](../config.ru.md) → Параметры клиентского кода
+
+-----
+
+[Read in English](client.en.md)
+
+# Параметры клиентского кода
+
+Данные параметры применяются только к клиентам Vitastor (QEMU, fio, NBD) и
+затрагивают логику их работы с кластером.
+
+- [client_max_dirty_bytes](#client_max_dirty_bytes)
+- [client_max_dirty_ops](#client_max_dirty_ops)
+- [client_enable_writeback](#client_enable_writeback)
+- [client_max_buffered_bytes](#client_max_buffered_bytes)
+- [client_max_buffered_ops](#client_max_buffered_ops)
+- [client_max_writeback_iodepth](#client_max_writeback_iodepth)
+
+## client_max_dirty_bytes
+
+- Тип: целое число
+- Значение по умолчанию: 33554432
+- Можно менять на лету: да
+
+При работе без [immediate_commit](layout-cluster.ru.md#immediate_commit)=all - это лимит объёма "грязных" (не
+зафиксированных fsync-ом) данных, при достижении которого клиент будет
+принудительно вызывать fsync и фиксировать данные. Также стоит иметь в виду,
+что в этом случае до момента fsync клиент хранит копию незафиксированных
+данных в памяти, то есть, настройка влияет на потребление памяти клиентами.
+
+## client_max_dirty_ops
+
+- Тип: целое число
+- Значение по умолчанию: 1024
+- Можно менять на лету: да
+
+Аналогично client_max_dirty_bytes, но ограничивает количество
+незафиксированных операций записи вместо их общего объёма.
+
+## client_enable_writeback
+
+- Тип: булево (да/нет)
+- Значение по умолчанию: false
+- Можно менять на лету: да
+
+Данный параметр разрешает включать буферизацию записи в памяти. Буферизация
+означает, что операции записи отправляются на кластер Vitastor не сразу, а
+могут небольшое время накапливаться в памяти и сбрасываться сразу пакетами,
+до тех пор, пока либо не будет превышен лимит неотправленных записей, либо
+пока клиент не вызовет fsync.
+
+Буферизация значительно повышает производительность некоторых приложений,
+например, CrystalDiskMark в Windows (ха-ха :-D), но также и любых других,
+которые пишут на диск неоптимально: либо последовательно, но мелкими блоками
+(например, по 4 кб), либо случайно, но без параллелизма и без fsync - то
+есть, например, отправляя 128 операций записи в разные места диска, но не
+все сразу с помощью асинхронного I/O, а по одной.
+
+В QEMU с буферизацией записи можно ожидать показателя примерно 22000
+операций случайной записи в секунду в 1 поток и с глубиной очереди 1 (T1Q1)
+без fsync, почти вне зависимости от того, насколько хороши ваши диски - эта
+цифра упирается в сам QEMU. Без буферизации рекорд пока что - 9900 операций
+в секунду, но на железе похуже может быть и поменьше, например, 5000 операций
+в секунду.
+
+При этом, даже если данный параметр включён, буферизация не включается, если
+явно не разрешена клиентом, т.к. если клиент не знает, что запросы записи
+буферизуются, это может приводить к потере данных. Поэтому в старых версиях
+клиентских драйверов буферизация записи не включается вообще, в новых
+версиях QEMU-драйвера включается, только если разрешена опцией диска
+`-blockdev cache.direct=false`, а в fio - только если нет опции `-direct=1`.
+В NBD и NFS драйверах буферизация записи разрешена по умолчанию.
+
+Можно обойти и это ограничение с помощью параметра `client_writeback_allowed`,
+но делать так не надо, если только вы не уверены в том, что делаете, на все
+100%. :-)
+
+## client_max_buffered_bytes
+
+- Тип: целое число
+- Значение по умолчанию: 33554432
+- Можно менять на лету: да
+
+Максимальный общий размер буферизованных записей, при достижении которого
+начинается процесс сброса данных на сервер.
+
+## client_max_buffered_ops
+
+- Тип: целое число
+- Значение по умолчанию: 1024
+- Можно менять на лету: да
+
+Максимальное количество буферизованных записей, при достижении которого
+начинается процесс сброса данных на сервер. При этом несколько
+последовательных изменённых областей здесь считаются 1 записью.
+
+## client_max_writeback_iodepth
+
+- Тип: целое число
+- Значение по умолчанию: 256
+- Можно менять на лету: да
+
+Максимальное число параллельных операций записи при сбросе буферов на сервер.

docs/config/layout-cluster.en.md

@@ -96,8 +96,9 @@ SSD cache or "media-cache" - for example, a lot of Seagate EXOS drives have
 it (they have internal SSD cache even though it's not stated in datasheets).
 
 Setting this parameter to "all" or "small" in OSD parameters requires enabling
-disable_journal_fsync and disable_meta_fsync, setting it to "all" also requires
-enabling disable_data_fsync.
+[disable_journal_fsync](layout-osd.en.yml#disable_journal_fsync) and
+[disable_meta_fsync](layout-osd.en.yml#disable_meta_fsync), setting it to
+"all" also requires enabling [disable_data_fsync](layout-osd.en.yml#disable_data_fsync).
 
 TLDR: For optimal performance, set immediate_commit to "all" if you only use
 SSDs with supercapacitor-based power loss protection (nonvolatile

docs/config/layout-cluster.ru.md

@@ -103,8 +103,9 @@ HDD-дисках с внутренним SSD или "медиа" кэшем - н
 указано в спецификациях).
 
 Указание "all" или "small" в настройках / командной строке OSD требует
-включения disable_journal_fsync и disable_meta_fsync, значение "all" также
-требует включения disable_data_fsync.
+включения [disable_journal_fsync](layout-osd.ru.yml#disable_journal_fsync) и
+[disable_meta_fsync](layout-osd.ru.yml#disable_meta_fsync), значение "all"
+также требует включения [disable_data_fsync](layout-osd.ru.yml#disable_data_fsync).
 
 Итого, вкратце: для оптимальной производительности установите
 immediate_commit в значение "all", если вы используете в кластере только SSD

docs/config/layout-osd.en.md

@@ -213,6 +213,6 @@ Thus, recommended setups are:
 3. Hybrid HDD+SSD: csum_block_size=4k + inmemory_metadata=false
 4. HDD-only, faster random read: csum_block_size=32k
 5. HDD-only, faster random write: csum_block_size=4k +
-   inmemory_metadata=false + cached_io_meta=true
+   inmemory_metadata=false + meta_io=cached
 
-See also [cached_io_meta](osd.en.md#cached_io_meta).
+See also [meta_io](osd.en.md#meta_io).

docs/config/layout-osd.ru.md

@@ -226,6 +226,6 @@ csum_block_size данных.
 3. Гибридные HDD+SSD: csum_block_size=4k + inmemory_metadata=false
 4. Только HDD, быстрее случайное чтение: csum_block_size=32k
 5. Только HDD, быстрее случайная запись: csum_block_size=4k +
-   inmemory_metadata=false + cached_io_meta=true
+   inmemory_metadata=false + meta_io=cached
 
-Смотрите также [cached_io_meta](osd.ru.md#cached_io_meta).
+Смотрите также [meta_io](osd.ru.md#meta_io).

docs/config/network.en.md

@@ -30,7 +30,6 @@ between clients, OSDs and etcd.
 - [etcd_slow_timeout](#etcd_slow_timeout)
 - [etcd_keepalive_timeout](#etcd_keepalive_timeout)
 - [etcd_ws_keepalive_timeout](#etcd_ws_keepalive_timeout)
-- [client_dirty_limit](#client_dirty_limit)
 
 ## tcp_header_buffer_size
 
@@ -240,17 +239,3 @@ etcd_report_interval to guarantee that keepalive actually works.
 
 etcd websocket ping interval required to keep the connection alive and
 detect disconnections quickly.
-
-## client_dirty_limit
-
-- Type: integer
-- Default: 33554432
-- Can be changed online: yes
-
-Without immediate_commit=all this parameter sets the limit of "dirty"
-(not committed by fsync) data allowed by the client before forcing an
-additional fsync and committing the data. Also note that the client always
-holds a copy of uncommitted data in memory so this setting also affects
-RAM usage of clients.
-
-This parameter doesn't affect OSDs themselves.

docs/config/network.ru.md

@@ -30,7 +30,6 @@
 - [etcd_slow_timeout](#etcd_slow_timeout)
 - [etcd_keepalive_timeout](#etcd_keepalive_timeout)
 - [etcd_ws_keepalive_timeout](#etcd_ws_keepalive_timeout)
-- [client_dirty_limit](#client_dirty_limit)
 
 ## tcp_header_buffer_size
 
@@ -251,17 +250,3 @@ etcd_report_interval, чтобы keepalive гарантированно рабо
 - Можно менять на лету: да
 
 Интервал проверки живости вебсокет-подключений к etcd.
-
-## client_dirty_limit
-
-- Тип: целое число
-- Значение по умолчанию: 33554432
-- Можно менять на лету: да
-
-При работе без immediate_commit=all - это лимит объёма "грязных" (не
-зафиксированных fsync-ом) данных, при достижении которого клиент будет
-принудительно вызывать fsync и фиксировать данные. Также стоит иметь в виду,
-что в этом случае до момента fsync клиент хранит копию незафиксированных
-данных в памяти, то есть, настройка влияет на потребление памяти клиентами.
-
-Параметр не влияет на сами OSD.

docs/config/osd.en.md

@@ -11,6 +11,7 @@ initialization and can be changed - either with an OSD restart or, for some of
 them, even without restarting by updating configuration in etcd.
 
 - [etcd_report_interval](#etcd_report_interval)
+- [etcd_stats_interval](#etcd_stats_interval)
 - [run_primary](#run_primary)
 - [osd_network](#osd_network)
 - [bind_address](#bind_address)
@@ -31,9 +32,9 @@ them, even without restarting by updating configuration in etcd.
 - [max_flusher_count](#max_flusher_count)
 - [inmemory_metadata](#inmemory_metadata)
 - [inmemory_journal](#inmemory_journal)
-- [cached_io_data](#cached_io_data)
-- [cached_io_meta](#cached_io_meta)
-- [cached_io_journal](#cached_io_journal)
+- [data_io](#data_io)
+- [meta_io](#meta_io)
+- [journal_io](#journal_io)
 - [journal_sector_buffer_count](#journal_sector_buffer_count)
 - [journal_no_same_sector_overwrites](#journal_no_same_sector_overwrites)
 - [throttle_small_writes](#throttle_small_writes)
@@ -56,11 +57,21 @@ them, even without restarting by updating configuration in etcd.
 - Type: seconds
 - Default: 5
 
-Interval at which OSDs report their state to etcd. Affects OSD lease time
+Interval at which OSDs report their liveness to etcd. Affects OSD lease time
 and thus the failover speed. Lease time is equal to this parameter value
 plus max_etcd_attempts * etcd_quick_timeout because it should be guaranteed
 that every OSD always refreshes its lease in time.
 
+## etcd_stats_interval
+
+- Type: seconds
+- Default: 30
+
+Interval at which OSDs report their statistics to etcd. Highly affects the
+imposed load on etcd, because statistics include a key for every OSD and
+for every PG. At the same time, low statistic intervals make `vitastor-cli`
+statistics more responsive.
+
 ## run_primary
 
 - Type: boolean
@@ -258,47 +269,59 @@ is typically very small because it's sufficient to have 16-32 MB journal
 for SSD OSDs. However, in theory it's possible that you'll want to turn it
 off for hybrid (HDD+SSD) OSDs with large journals on quick devices.
 
-## cached_io_data
+## data_io
 
-- Type: boolean
-- Default: false
+- Type: string
+- Default: direct
 
-Read and write *data* through Linux page cache, i.e. use a file descriptor
-opened with O_SYNC, but without O_DIRECT for I/O. May improve read
-performance for hot data and slower disks - HDDs and maybe SATA SSDs.
-Not recommended for desktop SSDs without capacitors because O_SYNC flushes
-disk cache on every write.
+I/O mode for *data*. One of "direct", "cached" or "directsync". Corresponds
+to O_DIRECT, O_SYNC and O_DIRECT|O_SYNC, respectively.
 
-## cached_io_meta
+Choose "cached" to use Linux page cache. This may improve read performance
+for hot data and slower disks - HDDs and maybe SATA SSDs - but will slightly
+decrease write performance for fast disks because page cache is an overhead
+itself.
 
-- Type: boolean
-- Default: false
+Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+(which requires disable_data_fsync) with drives having write-back cache
+which can't be turned off, for example, Intel Optane. Also note that *some*
+desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
+disable_data_fsync unsafe even with "directsync".
 
-Read and write *metadata* through Linux page cache. May improve read
-performance only if your drives are relatively slow (HDD, SATA SSD), and
-only if checksums are enabled and [inmemory_metadata](#inmemory_metadata)
-is disabled, because in this case metadata blocks are read from disk
-on every read request to verify checksums and caching them may reduce this
-extra read load.
+## meta_io
 
-Absolutely pointless to enable with enabled inmemory_metadata because all
-metadata is kept in memory anyway, and likely pointless without checksums,
-because in that case, metadata blocks are read from disk only during journal
+- Type: string
+- Default: direct
+
+I/O mode for *metadata*. One of "direct", "cached" or "directsync".
+
+"cached" may improve read performance, but only under the following conditions:
+1. your drives are relatively slow (HDD, SATA SSD), and
+2. checksums are enabled, and
+3. [inmemory_metadata](#inmemory_metadata) is disabled.
+Under all these conditions, metadata blocks are read from disk on every
+read request to verify checksums and caching them may reduce this extra
+read load. Without (3) metadata is never read from the disk after starting,
+and without (2) metadata blocks are read from disk only during journal
 flushing.
 
-If the same device is used for data and metadata, enabling [cached_io_data](#cached_io_data)
-also enables this parameter, given that it isn't turned off explicitly.
+"directsync" is the same as above.
 
-## cached_io_journal
+If the same device is used for data and metadata, meta_io by default is set
+to the same value as [data_io](#data_io).
 
-- Type: boolean
-- Default: false
+## journal_io
 
-Read and write *journal* through Linux page cache. May improve read
-performance if [inmemory_journal](#inmemory_journal) is turned off.
+- Type: string
+- Default: direct
 
-If the same device is used for metadata and journal, enabling [cached_io_meta](#cached_io_meta)
-also enables this parameter, given that it isn't turned off explicitly.
+I/O mode for *journal*. One of "direct", "cached" or "directsync".
+
+Here, "cached" may only improve read performance for recent writes and
+only if [inmemory_journal](#inmemory_journal) is turned off.
+
+If the same device is used for metadata and journal, journal_io by default
+is set to the same value as [meta_io](#meta_io).
 
 ## journal_sector_buffer_count
 

docs/config/osd.ru.md

@@ -12,6 +12,7 @@
 изменения конфигурации в etcd.
 
 - [etcd_report_interval](#etcd_report_interval)
+- [etcd_stats_interval](#etcd_stats_interval)
 - [run_primary](#run_primary)
 - [osd_network](#osd_network)
 - [bind_address](#bind_address)
@@ -32,9 +33,9 @@
 - [max_flusher_count](#max_flusher_count)
 - [inmemory_metadata](#inmemory_metadata)
 - [inmemory_journal](#inmemory_journal)
-- [cached_io_data](#cached_io_data)
-- [cached_io_meta](#cached_io_meta)
-- [cached_io_journal](#cached_io_journal)
+- [data_io](#data_io)
+- [meta_io](#meta_io)
+- [journal_io](#journal_io)
 - [journal_sector_buffer_count](#journal_sector_buffer_count)
 - [journal_no_same_sector_overwrites](#journal_no_same_sector_overwrites)
 - [throttle_small_writes](#throttle_small_writes)
@@ -57,11 +58,21 @@
 - Тип: секунды
 - Значение по умолчанию: 5
 
-Интервал, с которым OSD обновляет своё состояние в etcd. Значение параметра
-влияет на время резервации (lease) OSD и поэтому на скорость переключения
+Интервал, с которым OSD сообщает о том, что жив, в etcd. Значение параметра
+влияет на время резервации (lease) OSD и поэтому - на скорость переключения
 при падении OSD. Время lease равняется значению этого параметра плюс
 max_etcd_attempts * etcd_quick_timeout.
 
+## etcd_stats_interval
+
+- Тип: секунды
+- Значение по умолчанию: 30
+
+Интервал, с которым OSD обновляет свою статистику в etcd. Сильно влияет на
+создаваемую нагрузку на etcd, потому что статистика содержит по ключу на
+каждый OSD и на каждую PG. В то же время низкий интервал делает
+статистику, печатаемую `vitastor-cli`, отзывчивей.
+
 ## run_primary
 
 - Тип: булево (да/нет)
@@ -266,51 +277,62 @@ Flusher - это микро-поток (корутина), которая коп
 параметра может оказаться полезным для гибридных OSD (HDD+SSD) с большими
 журналами, расположенными на быстром по сравнению с HDD устройстве.
 
-## cached_io_data
+## data_io
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+- Тип: строка
+- Значение по умолчанию: direct
 
-Читать и записывать *данные* через системный кэш Linux (page cache), то есть,
-использовать для данных файловый дескриптор, открытый без флага O_DIRECT, но
-с флагом O_SYNC. Может улучшить скорость чтения для относительно медленных
-дисков - HDD и, возможно, SATA SSD. Не рекомендуется для потребительских
-SSD без конденсаторов, так как O_SYNC сбрасывает кэш диска при каждой записи.
+Режим ввода-вывода для *данных*. Одно из значений "direct", "cached" или
+"directsync", означающих O_DIRECT, O_SYNC и O_DIRECT|O_SYNC, соответственно.
 
-## cached_io_meta
+Выберите "cached", чтобы использовать системный кэш Linux (page cache) при
+чтении и записи. Это может улучшить скорость чтения горячих данных с
+относительно медленных дисков - HDD и, возможно, SATA SSD - но немного
+снижает производительность записи для быстрых дисков, так как кэш сам по
+себе тоже добавляет накладные расходы.
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+Выберите "directsync", если хотите задействовать
+[immediate_commit](layout-cluster.ru.md#immediate_commit) (требующий
+включенияd disable_data_fsync) на дисках с неотключаемым кэшем. Пример таких
+дисков - Intel Optane. При этом также стоит иметь в виду, что *некоторые*
+настольные SSD (например, HP EX950) игнорируют флаг O_SYNC, делая отключение
+fsync небезопасным даже с режимом "directsync".
 
-Читать и записывать *метаданные* через системный кэш Linux. Может улучшить
-скорость чтения, если у вас медленные диски, и только если контрольные суммы
-включены, а параметр [inmemory_metadata](#inmemory_metadata) отключён, так
-как в этом случае блоки метаданных читаются с диска при каждом запросе чтения
+## meta_io
+
+- Тип: строка
+- Значение по умолчанию: direct
+
+Режим ввода-вывода для *метаданных*. Одно из значений "direct", "cached" или
+"directsync".
+
+"cached" может улучшить скорость чтения, если:
+1. у вас медленные диски (HDD, SATA SSD)
+2. контрольные суммы включены
+3. параметр [inmemory_metadata](#inmemory_metadata) отключён.
+При этих условиях блоки метаданных читаются с диска при каждом запросе чтения
 для проверки контрольных сумм и их кэширование может снизить дополнительную
-нагрузку на диск.
+нагрузку на диск. Без (3) метаданные никогда не читаются с диска после
+запуска OSD, а без (2) блоки метаданных читаются только при сбросе журнала.
 
-Абсолютно бессмысленно включать данный параметр, если параметр
-inmemory_metadata включён (по умолчанию это так), и также вероятно
-бессмысленно включать его, если не включены контрольные суммы, так как в
-этом случае блоки метаданных читаются с диска только во время сброса
-журнала.
+Если одно и то же устройство используется для данных и метаданных, режим
+ввода-вывода метаданных по умолчанию устанавливается равным [data_io](#data_io).
 
-Если одно и то же устройство используется для данных и метаданных, включение
-[cached_io_data](#cached_io_data) также включает данный параметр, при
-условии, что он не отключён явным образом.
+## journal_io
 
-## cached_io_journal
+- Тип: строка
+- Значение по умолчанию: direct
 
-- Тип: булево (да/нет)
-- Значение по умолчанию: false
+Режим ввода-вывода для *журнала*. Одно из значений "direct", "cached" или
+"directsync".
 
-Читать и записывать *журнал* через системный кэш Linux. Может улучшить
-скорость чтения, если параметр [inmemory_journal](#inmemory_journal)
+Здесь "cached" может улучшить скорость чтения только недавно записанных
+данных и только если параметр [inmemory_journal](#inmemory_journal)
 отключён.
 
 Если одно и то же устройство используется для метаданных и журнала,
-включение [cached_io_meta](#cached_io_meta) также включает данный
-параметр, при условии, что он не отключён явным образом.
+режим ввода-вывода журнала по умолчанию устанавливается равным
+[meta_io](#meta_io).
 
 ## journal_sector_buffer_count
 

docs/config/pool.en.md

@@ -205,9 +205,8 @@ This parameter usually doesn't require to be changed.
 - Default: 131072
 
 Block size for this pool. The value from /vitastor/config/global is used when
-unspecified. If your cluster has OSDs with different block sizes then pool must
-be restricted by [osd_tags](#osd_tags) to only include OSDs with matching block
-size.
+unspecified. Only OSDs with matching block_size are used for each pool. If you
+want to further restrict OSDs for the pool, use [osd_tags](#osd_tags).
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#block_size).
 
@@ -216,10 +215,9 @@ Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-c
 - Type: integer
 - Default: 4096
 
-"Sector" size of virtual disks in this pool. The value from
-/vitastor/config/global is used when unspecified. Similar to block_size, the
-pool must be restricted by [osd_tags](#osd_tags) to only include OSDs with
-matching bitmap_granularity.
+"Sector" size of virtual disks in this pool. The value from /vitastor/config/global
+is used when unspecified. Similarly to block_size, only OSDs with matching
+bitmap_granularity are used for each pool.
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#bitmap_granularity).
 
@@ -229,10 +227,11 @@ Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-c
 - Default: none
 
 Immediate commit setting for this pool. The value from /vitastor/config/global
-is used when unspecified. Similar to block_size, the pool must be restricted by
-[osd_tags](#osd_tags) to only include OSDs with compatible immediate_commit.
-Compatible means that a pool with non-immediate commit will work with OSDs with
-immediate commit enabled, but not vice versa.
+is used when unspecified. Similarly to block_size, only OSDs with compatible
+bitmap_granularity are used for each pool. "Compatible" means that a pool with
+non-immediate commit will use OSDs with immediate commit enabled, but not vice
+versa. I.e., pools with "none" use all OSDs, pools with "small" only use OSDs
+with "all" or "small", and pools with "all" only use OSDs with "all".
 
 Read more about this parameter in [Cluster-Wide Disk Layout Parameters](layout-cluster.en.md#immediate_commit).
 

docs/config/pool.ru.md

@@ -208,8 +208,9 @@ PG в Vitastor эферемерны, то есть вы можете менят
 
 Размер блока для данного пула. Если не задан, используется значение из
 /vitastor/config/global. Если в вашем кластере есть OSD с разными размерами
-блока, пул должен быть ограничен только OSD, блок которых равен блоку пула,
-с помощью [osd_tags](#osd_tags).
+блока, пул будет использовать только OSD с размером блока, равным размеру блока
+пула. Если вы хотите сильнее ограничить набор используемых для пула OSD -
+используйте [osd_tags](#osd_tags).
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#block_size).
 
@@ -219,9 +220,8 @@ PG в Vitastor эферемерны, то есть вы можете менят
 - По умолчанию: 4096
 
 Размер "сектора" виртуальных дисков в данном пуле. Если не задан, используется
-значение из /vitastor/config/global. Аналогично block_size, пул должен быть
-ограничен OSD со значением bitmap_granularity, равным значению пула, с помощью
-[osd_tags](#osd_tags).
+значение из /vitastor/config/global. Аналогично block_size, каждый пул будет
+использовать только OSD с совпадающей с пулом настройкой bitmap_granularity.
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#bitmap_granularity).
 
@@ -231,11 +231,13 @@ PG в Vitastor эферемерны, то есть вы можете менят
 - По умолчанию: none
 
 Настройка мгновенного коммита для данного пула. Если не задана, используется
-значение из /vitastor/config/global. Аналогично block_size, пул должен быть
-ограничен OSD со значением bitmap_granularity, совместимым со значением пула, с
-помощью [osd_tags](#osd_tags). Совместимость означает, что пул с отключенным
-мгновенным коммитом может работать на OSD с включённым мгновенным коммитом, но
-не наоборот.
+значение из /vitastor/config/global. Аналогично block_size, каждый пул будет
+использовать только OSD с *совместимыми* настройками immediate_commit.
+"Совместимыми" означает, что пул с отключенным мгновенным коммитом будет
+использовать OSD с включённым мгновенным коммитом, но не наоборот. То есть,
+пул со значением "none" будет использовать все OSD, пул со "small" будет
+использовать OSD с "all" или "small", а пул с "all" будет использовать только
+OSD с "all".
 
 О самом параметре читайте в разделе [Дисковые параметры уровня кластера](layout-cluster.ru.md#immediate_commit).
 

docs/config/src/client.en.md

@@ -0,0 +1,4 @@
+# Client Parameters
+
+These parameters apply only to clients and affect their interaction with
+the cluster.

docs/config/src/client.ru.md

@@ -0,0 +1,4 @@
+# Параметры клиентского кода
+
+Данные параметры применяются только к клиентам Vitastor (QEMU, fio, NBD) и
+затрагивают логику их работы с кластером.

docs/config/src/client.yml

@@ -0,0 +1,124 @@
+- name: client_max_dirty_bytes
+  type: int
+  default: 33554432
+  online: true
+  info: |
+    Without [immediate_commit](layout-cluster.en.md#immediate_commit)=all this parameter sets the limit of "dirty"
+    (not committed by fsync) data allowed by the client before forcing an
+    additional fsync and committing the data. Also note that the client always
+    holds a copy of uncommitted data in memory so this setting also affects
+    RAM usage of clients.
+  info_ru: |
+    При работе без [immediate_commit](layout-cluster.ru.md#immediate_commit)=all - это лимит объёма "грязных" (не
+    зафиксированных fsync-ом) данных, при достижении которого клиент будет
+    принудительно вызывать fsync и фиксировать данные. Также стоит иметь в виду,
+    что в этом случае до момента fsync клиент хранит копию незафиксированных
+    данных в памяти, то есть, настройка влияет на потребление памяти клиентами.
+- name: client_max_dirty_ops
+  type: int
+  default: 1024
+  online: true
+  info: |
+    Same as client_max_dirty_bytes, but instead of total size, limits the number
+    of uncommitted write operations.
+  info_ru: |
+    Аналогично client_max_dirty_bytes, но ограничивает количество
+    незафиксированных операций записи вместо их общего объёма.
+- name: client_enable_writeback
+  type: bool
+  default: false
+  online: true
+  info: |
+    This parameter enables client-side write buffering. This means that write
+    requests are accumulated in memory for a short time before being sent to
+    a Vitastor cluster which allows to send them in parallel and increase
+    performance of some applications. Writes are buffered until client forces
+    a flush with fsync() or until the amount of buffered writes exceeds the
+    limit.
+
+    Write buffering significantly increases performance of some applications,
+    for example, CrystalDiskMark under Windows (LOL :-D), but also any other
+    applications if they do writes in one of two non-optimal ways: either if
+    they do a lot of small (4 kb or so) sequential writes, or if they do a lot
+    of small random writes, but without any parallelism or asynchrony, and also
+    without calling fsync().
+
+    With write buffering enabled, you can expect around 22000 T1Q1 random write
+    iops in QEMU more or less regardless of the quality of your SSDs, and this
+    number is in fact bound by QEMU itself rather than Vitastor (check it
+    yourself by adding a "driver=null-co" disk in QEMU). Without write
+    buffering, the current record is 9900 iops, but the number is usually
+    even lower with non-ideal hardware, for example, it may be 5000 iops.
+
+    Even when this parameter is enabled, write buffering isn't enabled until
+    the client explicitly allows it, because enabling it without the client
+    being aware of the fact that his writes may be buffered may lead to data
+    loss. Because of this, older versions of clients don't support write
+    buffering at all, newer versions of the QEMU driver allow write buffering
+    only if it's enabled in disk settings with `-blockdev cache.direct=false`,
+    and newer versions of FIO only allow write buffering if you don't specify
+    `-direct=1`. NBD and NFS drivers allow write buffering by default.
+
+    You can overcome this restriction too with the `client_writeback_allowed`
+    parameter, but you shouldn't do that unless you **really** know what you
+    are doing.
+  info_ru: |
+    Данный параметр разрешает включать буферизацию записи в памяти. Буферизация
+    означает, что операции записи отправляются на кластер Vitastor не сразу, а
+    могут небольшое время накапливаться в памяти и сбрасываться сразу пакетами,
+    до тех пор, пока либо не будет превышен лимит неотправленных записей, либо
+    пока клиент не вызовет fsync.
+
+    Буферизация значительно повышает производительность некоторых приложений,
+    например, CrystalDiskMark в Windows (ха-ха :-D), но также и любых других,
+    которые пишут на диск неоптимально: либо последовательно, но мелкими блоками
+    (например, по 4 кб), либо случайно, но без параллелизма и без fsync - то
+    есть, например, отправляя 128 операций записи в разные места диска, но не
+    все сразу с помощью асинхронного I/O, а по одной.
+
+    В QEMU с буферизацией записи можно ожидать показателя примерно 22000
+    операций случайной записи в секунду в 1 поток и с глубиной очереди 1 (T1Q1)
+    без fsync, почти вне зависимости от того, насколько хороши ваши диски - эта
+    цифра упирается в сам QEMU. Без буферизации рекорд пока что - 9900 операций
+    в секунду, но на железе похуже может быть и поменьше, например, 5000 операций
+    в секунду.
+
+    При этом, даже если данный параметр включён, буферизация не включается, если
+    явно не разрешена клиентом, т.к. если клиент не знает, что запросы записи
+    буферизуются, это может приводить к потере данных. Поэтому в старых версиях
+    клиентских драйверов буферизация записи не включается вообще, в новых
+    версиях QEMU-драйвера включается, только если разрешена опцией диска
+    `-blockdev cache.direct=false`, а в fio - только если нет опции `-direct=1`.
+    В NBD и NFS драйверах буферизация записи разрешена по умолчанию.
+
+    Можно обойти и это ограничение с помощью параметра `client_writeback_allowed`,
+    но делать так не надо, если только вы не уверены в том, что делаете, на все
+    100%. :-)
+- name: client_max_buffered_bytes
+  type: int
+  default: 33554432
+  online: true
+  info: |
+    Maximum total size of buffered writes which triggers write-back when reached.
+  info_ru: |
+    Максимальный общий размер буферизованных записей, при достижении которого
+    начинается процесс сброса данных на сервер.
+- name: client_max_buffered_ops
+  type: int
+  default: 1024
+  online: true
+  info: |
+    Maximum number of buffered writes which triggers write-back when reached.
+    Multiple consecutive modified data regions are counted as 1 write here.
+  info_ru: |
+    Максимальное количество буферизованных записей, при достижении которого
+    начинается процесс сброса данных на сервер. При этом несколько
+    последовательных изменённых областей здесь считаются 1 записью.
+- name: client_max_writeback_iodepth
+  type: int
+  default: 256
+  online: true
+  info: |
+    Maximum number of parallel writes when flushing buffered data to the server.
+  info_ru: |
+    Максимальное число параллельных операций записи при сбросе буферов на сервер.

docs/config/src/included.en.md

@@ -28,6 +28,8 @@
 
 {{../../config/network.en.md|indent=2}}
 
+{{../../config/client.en.md|indent=2}}
+
 {{../../config/layout-cluster.en.md|indent=2}}
 
 {{../../config/layout-osd.en.md|indent=2}}

docs/config/src/included.ru.md

@@ -28,6 +28,8 @@
 
 {{../../config/network.ru.md|indent=2}}
 
+{{../../config/client.ru.md|indent=2}}
+
 {{../../config/layout-cluster.ru.md|indent=2}}
 
 {{../../config/layout-osd.ru.md|indent=2}}

docs/config/src/layout-cluster.yml

@@ -87,8 +87,9 @@
     it (they have internal SSD cache even though it's not stated in datasheets).
 
     Setting this parameter to "all" or "small" in OSD parameters requires enabling
-    disable_journal_fsync and disable_meta_fsync, setting it to "all" also requires
-    enabling disable_data_fsync.
+    [disable_journal_fsync](layout-osd.en.yml#disable_journal_fsync) and
+    [disable_meta_fsync](layout-osd.en.yml#disable_meta_fsync), setting it to
+    "all" also requires enabling [disable_data_fsync](layout-osd.en.yml#disable_data_fsync).
 
     TLDR: For optimal performance, set immediate_commit to "all" if you only use
     SSDs with supercapacitor-based power loss protection (nonvolatile
@@ -140,8 +141,9 @@
     указано в спецификациях).
 
     Указание "all" или "small" в настройках / командной строке OSD требует
-    включения disable_journal_fsync и disable_meta_fsync, значение "all" также
-    требует включения disable_data_fsync.
+    включения [disable_journal_fsync](layout-osd.ru.yml#disable_journal_fsync) и
+    [disable_meta_fsync](layout-osd.ru.yml#disable_meta_fsync), значение "all"
+    также требует включения [disable_data_fsync](layout-osd.ru.yml#disable_data_fsync).
 
     Итого, вкратце: для оптимальной производительности установите
     immediate_commit в значение "all", если вы используете в кластере только SSD

docs/config/src/layout-osd.yml

@@ -244,9 +244,9 @@
     3. Hybrid HDD+SSD: csum_block_size=4k + inmemory_metadata=false
     4. HDD-only, faster random read: csum_block_size=32k
     5. HDD-only, faster random write: csum_block_size=4k +
-       inmemory_metadata=false + cached_io_meta=true
+       inmemory_metadata=false + meta_io=cached
 
-    See also [cached_io_meta](osd.en.md#cached_io_meta).
+    See also [meta_io](osd.en.md#meta_io).
   info_ru: |
     Размер блока расчёта контрольных сумм.
 
@@ -271,6 +271,6 @@
     3. Гибридные HDD+SSD: csum_block_size=4k + inmemory_metadata=false
     4. Только HDD, быстрее случайное чтение: csum_block_size=32k
     5. Только HDD, быстрее случайная запись: csum_block_size=4k +
-       inmemory_metadata=false + cached_io_meta=true
+       inmemory_metadata=false + meta_io=cached
 
-    Смотрите также [cached_io_meta](osd.ru.md#cached_io_meta).
+    Смотрите также [meta_io](osd.ru.md#meta_io).

docs/config/src/network.yml

@@ -30,6 +30,18 @@
     будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
     это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
     принципе, это может применяться для клиентов со старыми версиями ядра.
+- name: use_zerocopy_send
+  type: bool
+  default: false
+  info: |
+    If true, OSDs and clients will attempt to use TCP zero-copy send
+    (MSG_ZEROCOPY) for big buffers. It's recommended to raise net.ipv4.tcp_wmem
+    and net.core.wmem_max sysctls when using this mode.
+  info_ru: |
+    Если установлено в true, то OSD и клиенты будут стараться использовать
+    TCP-отправку без копирования (MSG_ZEROCOPY) для больших буферов данных.
+    Рекомендуется поднять значения sysctl net.ipv4.tcp_wmem и net.core.wmem_max
+    при использовании этого режима.
 - name: use_rdma
   type: bool
   default: true
@@ -259,23 +271,3 @@
     detect disconnections quickly.
   info_ru: |
     Интервал проверки живости вебсокет-подключений к etcd.
-- name: client_dirty_limit
-  type: int
-  default: 33554432
-  online: true
-  info: |
-    Without immediate_commit=all this parameter sets the limit of "dirty"
-    (not committed by fsync) data allowed by the client before forcing an
-    additional fsync and committing the data. Also note that the client always
-    holds a copy of uncommitted data in memory so this setting also affects
-    RAM usage of clients.
-
-    This parameter doesn't affect OSDs themselves.
-  info_ru: |
-    При работе без immediate_commit=all - это лимит объёма "грязных" (не
-    зафиксированных fsync-ом) данных, при достижении которого клиент будет
-    принудительно вызывать fsync и фиксировать данные. Также стоит иметь в виду,
-    что в этом случае до момента fsync клиент хранит копию незафиксированных
-    данных в памяти, то есть, настройка влияет на потребление памяти клиентами.
-
-    Параметр не влияет на сами OSD.

docs/config/src/osd.yml

@@ -2,15 +2,28 @@
   type: sec
   default: 5
   info: |
-    Interval at which OSDs report their state to etcd. Affects OSD lease time
+    Interval at which OSDs report their liveness to etcd. Affects OSD lease time
     and thus the failover speed. Lease time is equal to this parameter value
     plus max_etcd_attempts * etcd_quick_timeout because it should be guaranteed
     that every OSD always refreshes its lease in time.
   info_ru: |
-    Интервал, с которым OSD обновляет своё состояние в etcd. Значение параметра
-    влияет на время резервации (lease) OSD и поэтому на скорость переключения
+    Интервал, с которым OSD сообщает о том, что жив, в etcd. Значение параметра
+    влияет на время резервации (lease) OSD и поэтому - на скорость переключения
     при падении OSD. Время lease равняется значению этого параметра плюс
     max_etcd_attempts * etcd_quick_timeout.
+- name: etcd_stats_interval
+  type: sec
+  default: 30
+  info: |
+    Interval at which OSDs report their statistics to etcd. Highly affects the
+    imposed load on etcd, because statistics include a key for every OSD and
+    for every PG. At the same time, low statistic intervals make `vitastor-cli`
+    statistics more responsive.
+  info_ru: |
+    Интервал, с которым OSD обновляет свою статистику в etcd. Сильно влияет на
+    создаваемую нагрузку на etcd, потому что статистика содержит по ключу на
+    каждый OSD и на каждую PG. В то же время низкий интервал делает
+    статистику, печатаемую `vitastor-cli`, отзывчивей.
 - name: run_primary
   type: bool
   default: true
@@ -260,73 +273,96 @@
     достаточно 16- или 32-мегабайтного журнала. Однако в теории отключение
     параметра может оказаться полезным для гибридных OSD (HDD+SSD) с большими
     журналами, расположенными на быстром по сравнению с HDD устройстве.
-- name: cached_io_data
-  type: bool
-  default: false
+- name: data_io
+  type: string
+  default: direct
   info: |
-    Read and write *data* through Linux page cache, i.e. use a file descriptor
-    opened with O_SYNC, but without O_DIRECT for I/O. May improve read
-    performance for hot data and slower disks - HDDs and maybe SATA SSDs.
-    Not recommended for desktop SSDs without capacitors because O_SYNC flushes
-    disk cache on every write.
-  info_ru: |
-    Читать и записывать *данные* через системный кэш Linux (page cache), то есть,
-    использовать для данных файловый дескриптор, открытый без флага O_DIRECT, но
-    с флагом O_SYNC. Может улучшить скорость чтения для относительно медленных
-    дисков - HDD и, возможно, SATA SSD. Не рекомендуется для потребительских
-    SSD без конденсаторов, так как O_SYNC сбрасывает кэш диска при каждой записи.
-- name: cached_io_meta
-  type: bool
-  default: false
-  info: |
-    Read and write *metadata* through Linux page cache. May improve read
-    performance only if your drives are relatively slow (HDD, SATA SSD), and
-    only if checksums are enabled and [inmemory_metadata](#inmemory_metadata)
-    is disabled, because in this case metadata blocks are read from disk
-    on every read request to verify checksums and caching them may reduce this
-    extra read load.
+    I/O mode for *data*. One of "direct", "cached" or "directsync". Corresponds
+    to O_DIRECT, O_SYNC and O_DIRECT|O_SYNC, respectively.
 
-    Absolutely pointless to enable with enabled inmemory_metadata because all
-    metadata is kept in memory anyway, and likely pointless without checksums,
-    because in that case, metadata blocks are read from disk only during journal
+    Choose "cached" to use Linux page cache. This may improve read performance
+    for hot data and slower disks - HDDs and maybe SATA SSDs - but will slightly
+    decrease write performance for fast disks because page cache is an overhead
+    itself.
+
+    Choose "directsync" to use [immediate_commit](layout-cluster.ru.md#immediate_commit)
+    (which requires disable_data_fsync) with drives having write-back cache
+    which can't be turned off, for example, Intel Optane. Also note that *some*
+    desktop SSDs (for example, HP EX950) may ignore O_SYNC thus making
+    disable_data_fsync unsafe even with "directsync".
+  info_ru: |
+    Режим ввода-вывода для *данных*. Одно из значений "direct", "cached" или
+    "directsync", означающих O_DIRECT, O_SYNC и O_DIRECT|O_SYNC, соответственно.
+
+    Выберите "cached", чтобы использовать системный кэш Linux (page cache) при
+    чтении и записи. Это может улучшить скорость чтения горячих данных с
+    относительно медленных дисков - HDD и, возможно, SATA SSD - но немного
+    снижает производительность записи для быстрых дисков, так как кэш сам по
+    себе тоже добавляет накладные расходы.
+
+    Выберите "directsync", если хотите задействовать
+    [immediate_commit](layout-cluster.ru.md#immediate_commit) (требующий
+    включенияd disable_data_fsync) на дисках с неотключаемым кэшем. Пример таких
+    дисков - Intel Optane. При этом также стоит иметь в виду, что *некоторые*
+    настольные SSD (например, HP EX950) игнорируют флаг O_SYNC, делая отключение
+    fsync небезопасным даже с режимом "directsync".
+- name: meta_io
+  type: string
+  default: direct
+  info: |
+    I/O mode for *metadata*. One of "direct", "cached" or "directsync".
+
+    "cached" may improve read performance, but only under the following conditions:
+    1. your drives are relatively slow (HDD, SATA SSD), and
+    2. checksums are enabled, and
+    3. [inmemory_metadata](#inmemory_metadata) is disabled.
+    Under all these conditions, metadata blocks are read from disk on every
+    read request to verify checksums and caching them may reduce this extra
+    read load. Without (3) metadata is never read from the disk after starting,
+    and without (2) metadata blocks are read from disk only during journal
     flushing.
 
-    If the same device is used for data and metadata, enabling [cached_io_data](#cached_io_data)
-    also enables this parameter, given that it isn't turned off explicitly.
+    "directsync" is the same as above.
+
+    If the same device is used for data and metadata, meta_io by default is set
+    to the same value as [data_io](#data_io).
   info_ru: |
-    Читать и записывать *метаданные* через системный кэш Linux. Может улучшить
-    скорость чтения, если у вас медленные диски, и только если контрольные суммы
-    включены, а параметр [inmemory_metadata](#inmemory_metadata) отключён, так
-    как в этом случае блоки метаданных читаются с диска при каждом запросе чтения
+    Режим ввода-вывода для *метаданных*. Одно из значений "direct", "cached" или
+    "directsync".
+
+    "cached" может улучшить скорость чтения, если:
+    1. у вас медленные диски (HDD, SATA SSD)
+    2. контрольные суммы включены
+    3. параметр [inmemory_metadata](#inmemory_metadata) отключён.
+    При этих условиях блоки метаданных читаются с диска при каждом запросе чтения
     для проверки контрольных сумм и их кэширование может снизить дополнительную
-    нагрузку на диск.
+    нагрузку на диск. Без (3) метаданные никогда не читаются с диска после
+    запуска OSD, а без (2) блоки метаданных читаются только при сбросе журнала.
 
-    Абсолютно бессмысленно включать данный параметр, если параметр
-    inmemory_metadata включён (по умолчанию это так), и также вероятно
-    бессмысленно включать его, если не включены контрольные суммы, так как в
-    этом случае блоки метаданных читаются с диска только во время сброса
-    журнала.
-
-    Если одно и то же устройство используется для данных и метаданных, включение
-    [cached_io_data](#cached_io_data) также включает данный параметр, при
-    условии, что он не отключён явным образом.
-- name: cached_io_journal
-  type: bool
-  default: false
+    Если одно и то же устройство используется для данных и метаданных, режим
+    ввода-вывода метаданных по умолчанию устанавливается равным [data_io](#data_io).
+- name: journal_io
+  type: string
+  default: direct
   info: |
-    Read and write *journal* through Linux page cache. May improve read
-    performance if [inmemory_journal](#inmemory_journal) is turned off.
+    I/O mode for *journal*. One of "direct", "cached" or "directsync".
 
-    If the same device is used for metadata and journal, enabling [cached_io_meta](#cached_io_meta)
-    also enables this parameter, given that it isn't turned off explicitly.
+    Here, "cached" may only improve read performance for recent writes and
+    only if [inmemory_journal](#inmemory_journal) is turned off.
+
+    If the same device is used for metadata and journal, journal_io by default
+    is set to the same value as [meta_io](#meta_io).
   info_ru: |
-    Читать и записывать *журнал* через системный кэш Linux. Может улучшить
-    скорость чтения, если параметр [inmemory_journal](#inmemory_journal)
+    Режим ввода-вывода для *журнала*. Одно из значений "direct", "cached" или
+    "directsync".
+
+    Здесь "cached" может улучшить скорость чтения только недавно записанных
+    данных и только если параметр [inmemory_journal](#inmemory_journal)
     отключён.
 
     Если одно и то же устройство используется для метаданных и журнала,
-    включение [cached_io_meta](#cached_io_meta) также включает данный
-    параметр, при условии, что он не отключён явным образом.
+    режим ввода-вывода журнала по умолчанию устанавливается равным
+    [meta_io](#meta_io).
 - name: journal_sector_buffer_count
   type: int
   default: 32

docs/intro/features.en.md

@@ -31,6 +31,7 @@
 - [RDMA/RoCEv2 support via libibverbs](../config/network.en.md#rdma_device)
 - [Scrubbing](../config/osd.en.md#auto_scrub) (verification of copies)
 - [Checksums](../config/layout-osd.en.md#data_csum_type)
+- [Client write-back cache](../config/client.en.md#client_enable_writeback)
 
 ## Plugins and tools
 
@@ -50,13 +51,15 @@
 
 The following features are planned for the future:
 
+- File system
+- Control plane optimisation
 - Other administrative tools
 - Web GUI
 - OpenNebula plugin
-- iSCSI proxy
+- iSCSI and NVMeoF gateways
 - Multi-threaded client
 - Faster failover
+- S3
 - Tiered storage (SSD caching)
 - NVDIMM support
 - Compression (possibly)
-- Read caching using system page cache (possibly)

docs/intro/features.ru.md

@@ -33,6 +33,7 @@
 - [Поддержка RDMA/RoCEv2 через libibverbs](../config/network.ru.md#rdma_device)
 - [Фоновая проверка целостности](../config/osd.ru.md#auto_scrub) (сверка копий)
 - [Контрольные суммы](../config/layout-osd.ru.md#data_csum_type)
+- [Буферизация записи на стороне клиента](../config/client.ru.md#client_enable_writeback)
 
 ## Драйверы и инструменты
 
@@ -50,12 +51,15 @@
 
 ## Планы развития
 
+- Файловая система
+- Оптимизация слоя управления
 - Другие инструменты администрирования
 - Web-интерфейс
 - Плагин для OpenNebula
-- iSCSI-прокси
+- iSCSI и NVMeoF прокси
 - Многопоточный клиент
 - Более быстрое переключение при отказах
+- S3
 - Поддержка SSD-кэширования (tiered storage)
 - Поддержка NVDIMM
 - Возможно, сжатие

docs/usage/qemu.en.md

@@ -34,6 +34,20 @@ qemu-system-x86_64 -enable-kvm -m 1024 \
     -vnc 0.0.0.0:0
 ```
 
+With a separate I/O thread:
+
+```
+qemu-system-x86_64 -enable-kvm -m 1024 \
+    -object iothread,id=vitastor1 \
+    -blockdev '{"node-name":"drive-virtio-disk0","driver":"vitastor","image":"debian9",
+        "cache":{"direct":true,"no-flush":false},"auto-read-only":true,"discard":"unmap"}' \
+    -device 'virtio-blk-pci,iothread=vitastor1,scsi=off,bus=pci.0,addr=0x5,drive=drive-virtio-disk0,
+        id=virtio-disk0,bootindex=1,write-cache=off' \
+    -vnc 0.0.0.0:0
+```
+
+You can also specify inode ID, pool and size manually instead of `:image=<IMAGE>` option: `:pool=<POOL>:inode=<INODE>:size=<SIZE>`.
+
 ## qemu-img
 
 For qemu-img, you should use `vitastor:etcd_host=<HOST>:image=<IMAGE>` as filename.
@@ -84,6 +98,29 @@ This can be used for backups. Just note that exporting an image that is currentl
 is of course unsafe and doesn't produce a consistent result, so only export snapshots if you do this
 on a live VM.
 
+## vhost-user-blk
+
+QEMU, starting with 6.0, includes support for attaching disks via a separate
+userspace worker process, called `vhost-user-blk`. It usually has slightly (20-30 us)
+lower latency.
+
+Example commands to use it with Vitastor:
+
+```
+qemu-storage-daemon \
+    --daemonize \
+    --blockdev '{"node-name":"drive-virtio-disk1","driver":"vitastor","image":"testosd1","cache":{"direct":true,"no-flush":false},"auto-read-only":true,"discard":"unmap"}' \
+    --export type=vhost-user-blk,id=vitastor1,node-name=drive-virtio-disk1,addr.type=unix,addr.path=/run/vitastor1-user-blk.sock,writable=on,num-queues=1
+
+qemu-system-x86_64 -enable-kvm -m 2048 -M accel=kvm,memory-backend=mem \
+    -object memory-backend-memfd,id=mem,size=2G,share=on \
+    -chardev socket,id=vitastor1,reconnect=1,path=/run/vitastor1-user-blk.sock \
+    -device vhost-user-blk-pci,chardev=vitastor1,num-queues=1,config-wce=off \
+    -vnc 0.0.0.0:0
+```
+
+memfd memory-backend is crucial, vhost-user-blk does not work without it.
+
 ## VDUSE
 
 Linux kernel, starting with version 5.15, supports a new interface for attaching virtual disks

docs/usage/qemu.ru.md

@@ -36,6 +36,18 @@ qemu-system-x86_64 -enable-kvm -m 1024 \
     -vnc 0.0.0.0:0
 ```
 
+С отдельным потоком ввода-вывода:
+
+```
+qemu-system-x86_64 -enable-kvm -m 1024 \
+    -object iothread,id=vitastor1 \
+    -blockdev '{"node-name":"drive-virtio-disk0","driver":"vitastor","image":"debian9",
+        "cache":{"direct":true,"no-flush":false},"auto-read-only":true,"discard":"unmap"}' \
+    -device 'virtio-blk-pci,iothread=vitastor1,scsi=off,bus=pci.0,addr=0x5,drive=drive-virtio-disk0,
+        id=virtio-disk0,bootindex=1,write-cache=off' \
+    -vnc 0.0.0.0:0
+```
+
 Вместо `:image=<IMAGE>` также можно указывать номер инода, пул и размер: `:pool=<POOL>:inode=<INODE>:size=<SIZE>`.
 
 ## qemu-img
@@ -88,6 +100,29 @@ qemu-img rebase -u -b '' testimg.qcow2
 в то же время идёт запись, небезопасно - результат чтения не будет целостным. Так что если вы работаете
 с активными виртуальными машинами, экспортируйте только их снимки, но не сам образ.
 
+## vhost-user-blk
+
+QEMU, начиная с 6.0, позволяет подключать диски через отдельный рабочий процесс.
+Этот метод подключения называется `vhost-user-blk` и обычно имеет чуть меньшую
+задержку (ниже на 20-30 микросекунд, чем при обычном методе).
+
+Пример команд для использования vhost-user-blk с Vitastor:
+
+```
+qemu-storage-daemon \
+    --daemonize \
+    --blockdev '{"node-name":"drive-virtio-disk1","driver":"vitastor","image":"testosd1","cache":{"direct":true,"no-flush":false},"auto-read-only":true,"discard":"unmap"}' \
+    --export type=vhost-user-blk,id=vitastor1,node-name=drive-virtio-disk1,addr.type=unix,addr.path=/run/vitastor1-user-blk.sock,writable=on,num-queues=1
+
+qemu-system-x86_64 -enable-kvm -m 2048 -M accel=kvm,memory-backend=mem \
+    -object memory-backend-memfd,id=mem,size=2G,share=on \
+    -chardev socket,id=vitastor1,reconnect=1,path=/run/vitastor1-user-blk.sock \
+    -device vhost-user-blk-pci,chardev=vitastor1,num-queues=1,config-wce=off \
+    -vnc 0.0.0.0:0
+```
+
+Здесь критична опция memory-backend-memfd, vhost-user-blk без неё не работает.
+
 ## VDUSE
 
 В Linux, начиная с версии ядра 5.15, доступен новый интерфейс для подключения виртуальных дисков

mon/90-vitastor.rules

@@ -3,5 +3,5 @@ SUBSYSTEM=="block", ENV{ID_PART_ENTRY_TYPE}=="e7009fac-a5a1-4d72-af72-53de130599
     IMPORT{program}="/usr/bin/vitastor-disk udev $devnode", \
     SYMLINK+="vitastor/$env{VITASTOR_ALIAS}"
 
-ENV{VITASTOR_OSD_NUM}!="", ACTION=="add", RUN{program}+="/usr/bin/systemctl enable --now vitastor-osd@$env{VITASTOR_OSD_NUM}"
-ENV{VITASTOR_OSD_NUM}!="", ACTION=="remove", RUN{program}+="/usr/bin/systemctl disable --now vitastor-osd@$env{VITASTOR_OSD_NUM}"
+ENV{VITASTOR_OSD_NUM}!="", ACTION=="add", RUN{program}+="/usr/bin/systemctl enable --now --no-block vitastor-osd@$env{VITASTOR_OSD_NUM}"
+ENV{VITASTOR_OSD_NUM}!="", ACTION=="remove", RUN{program}+="/usr/bin/systemctl disable --now --no-block vitastor-osd@$env{VITASTOR_OSD_NUM}"

mon/mon.js

@@ -65,6 +65,7 @@ const etcd_tree = {
             // client and osd
             tcp_header_buffer_size: 65536,
             use_sync_send_recv: false,
+            use_zerocopy_send: false,
             use_rdma: true,
             rdma_device: null, // for example, "rocep5s0f0"
             rdma_port_num: 1,
@@ -78,9 +79,15 @@ const etcd_tree = {
             disk_alignment: 4096,
             bitmap_granularity: 4096,
             immediate_commit: false, // 'all' or 'small'
+            // client - configurable online
+            client_max_dirty_bytes: 33554432,
+            client_max_dirty_ops: 1024,
+            client_enable_writeback: false,
+            client_max_buffered_bytes: 33554432,
+            client_max_buffered_ops: 1024,
+            client_max_writeback_iodepth: 256,
             // client and osd - configurable online
             log_level: 0,
-            client_dirty_limit: 33554432,
             peer_connect_interval: 5, // seconds. min: 1
             peer_connect_timeout: 5, // seconds. min: 1
             osd_idle_timeout: 5, // seconds. min: 1
@@ -93,6 +100,7 @@ const etcd_tree = {
             etcd_ws_keepalive_interval: 30, // seconds
             // osd
             etcd_report_interval: 5, // seconds
+            etcd_stats_interval: 30, // seconds
             run_primary: true,
             osd_network: null, // "192.168.7.0/24" or an array of masks
             bind_address: "0.0.0.0",
@@ -390,8 +398,8 @@ class Mon
         this.etcd_prefix = this.etcd_prefix.replace(/\/\/+/g, '/').replace(/^\/?(.*[^\/])\/?$/, '/$1');
         this.etcd_start_timeout = (config.etcd_start_timeout || 5) * 1000;
         this.state = JSON.parse(JSON.stringify(this.constructor.etcd_tree));
+        this.prev_stats = { osd_stats: {}, osd_diff: {} };
         this.signals_set = false;
-        this.stat_time = Date.now();
         this.ws = null;
         this.ws_alive = false;
         this.ws_keepalive_timer = null;
@@ -539,10 +547,18 @@ class Mon
         {
             retries = 1;
         }
+        const tried = {};
         while (retries < 0 || retry < retries)
         {
             const cur_addr = this.pick_next_etcd();
             const base = 'ws'+cur_addr.substr(4);
+            let now = Date.now();
+            if (tried[base] && now-tried[base] < this.etcd_start_timeout)
+            {
+                await new Promise(ok => setTimeout(ok, this.etcd_start_timeout-(now-tried[base])));
+                now = Date.now();
+            }
+            tried[base] = now;
             const ok = await new Promise((ok, no) =>
             {
                 const timer_id = setTimeout(() =>
@@ -1148,6 +1164,33 @@ class Mon
         }
     }
 
+    filter_osds_by_block_layout(flat_tree, block_size, bitmap_granularity, immediate_commit)
+    {
+        for (const host in flat_tree)
+        {
+            let found = 0;
+            for (const osd in flat_tree[host])
+            {
+                const osd_stat = this.state.osd.stats[osd];
+                if (osd_stat && (osd_stat.bs_block_size && osd_stat.bs_block_size != block_size ||
+                    osd_stat.bitmap_granularity && osd_stat.bitmap_granularity != bitmap_granularity ||
+                    osd_stat.immediate_commit == 'small' && immediate_commit == 'all' ||
+                    osd_stat.immediate_commit == 'none' && immediate_commit != 'none'))
+                {
+                    delete flat_tree[host][osd];
+                }
+                else
+                {
+                    found++;
+                }
+            }
+            if (!found)
+            {
+                delete flat_tree[host];
+            }
+        }
+    }
+
     get_affinity_osds(pool_cfg, up_osds, osd_tree)
     {
         let aff_osds = up_osds;
@@ -1208,6 +1251,12 @@ class Mon
                 pool_tree = pool_tree ? pool_tree.children : [];
                 pool_tree = LPOptimizer.flatten_tree(pool_tree, levels, pool_cfg.failure_domain, 'osd');
                 this.filter_osds_by_tags(osd_tree, pool_tree, pool_cfg.osd_tags);
+                this.filter_osds_by_block_layout(
+                    pool_tree,
+                    pool_cfg.block_size || this.config.block_size || 131072,
+                    pool_cfg.bitmap_granularity || this.config.bitmap_granularity || 4096,
+                    pool_cfg.immediate_commit || this.config.immediate_commit || 'none'
+                );
                 // These are for the purpose of building history.osd_sets
                 const real_prev_pgs = [];
                 let pg_history = [];
@@ -1411,15 +1460,15 @@ class Mon
         }
     }
 
-    derive_osd_stats(st, prev)
+    derive_osd_stats(st, prev, prev_diff)
     {
         const zero_stats = { op: { bps: 0n, iops: 0n, lat: 0n }, subop: { iops: 0n, lat: 0n }, recovery: { bps: 0n, iops: 0n } };
-        const diff = { op_stats: {}, subop_stats: {}, recovery_stats: {} };
-        if (!st || !st.time || prev && (prev.time || this.stat_time/1000) >= st.time)
+        const diff = { op_stats: {}, subop_stats: {}, recovery_stats: {}, inode_stats: {} };
+        if (!st || !st.time || !prev || prev.time >= st.time)
         {
-            return diff;
+            return prev_diff || diff;
         }
-        const timediff = BigInt(st.time*1000 - (prev && prev.time*1000 || this.stat_time));
+        const timediff = BigInt(st.time*1000 - prev.time*1000);
         for (const op in st.op_stats||{})
         {
             const pr = prev && prev.op_stats && prev.op_stats[op];
@@ -1451,25 +1500,47 @@ class Mon
             if (n > 0)
                 diff.recovery_stats[op] = { ...c, bps: b*1000n/timediff, iops: n*1000n/timediff };
         }
+        for (const pool_id in st.inode_stats||{})
+        {
+            const pool_diff = diff.inode_stats[pool_id] = {};
+            for (const inode_num in st.inode_stats[pool_id])
+            {
+                const inode_diff = diff.inode_stats[pool_id][inode_num] = {};
+                for (const op of [ 'read', 'write', 'delete' ])
+                {
+                    const c = st.inode_stats[pool_id][inode_num][op];
+                    const pr = prev && prev.inode_stats && prev.inode_stats[pool_id] &&
+                        prev.inode_stats[pool_id][inode_num] && prev.inode_stats[pool_id][inode_num][op];
+                    const n = BigInt(c.count||0) - BigInt(pr && pr.count||0);
+                    inode_diff[op] = {
+                        bps: (BigInt(c.bytes||0) - BigInt(pr && pr.bytes||0))*1000n/timediff,
+                        iops: n*1000n/timediff,
+                        lat: (BigInt(c.usec||0) - BigInt(pr && pr.usec||0))/(n || 1n),
+                    };
+                }
+            }
+        }
         return diff;
     }
 
-    sum_op_stats(timestamp, prev_stats)
+    sum_op_stats()
     {
-        const sum_diff = { op_stats: {}, subop_stats: {}, recovery_stats: {} };
-        if (!prev_stats || prev_stats.timestamp >= timestamp)
+        for (const osd in this.state.osd.stats)
         {
-            return sum_diff;
+            const cur = { ...this.state.osd.stats[osd], inode_stats: this.state.osd.inodestats[osd]||{} };
+            this.prev_stats.osd_diff[osd] = this.derive_osd_stats(
+                cur, this.prev_stats.osd_stats[osd], this.prev_stats.osd_diff[osd]
+            );
+            this.prev_stats.osd_stats[osd] = cur;
         }
-        const tm = BigInt(timestamp - (prev_stats.timestamp || 0));
+        const sum_diff = { op_stats: {}, subop_stats: {}, recovery_stats: {} };
         // Sum derived values instead of deriving summed
         for (const osd in this.state.osd.stats)
         {
-            const derived = this.derive_osd_stats(this.state.osd.stats[osd],
-                this.prev_stats && this.prev_stats.osd_stats && this.prev_stats.osd_stats[osd]);
-            for (const type in derived)
+            const derived = this.prev_stats.osd_diff[osd];
+            for (const type in sum_diff)
             {
-                for (const op in derived[type])
+                for (const op in derived[type]||{})
                 {
                     for (const k in derived[type][op])
                     {
@@ -1526,14 +1597,14 @@ class Mon
         return { object_counts, object_bytes };
     }
 
-    sum_inode_stats(prev_stats, timestamp, prev_timestamp)
+    sum_inode_stats()
     {
         const inode_stats = {};
         const inode_stub = () => ({
             raw_used: 0n,
-            read: { count: 0n, usec: 0n, bytes: 0n },
-            write: { count: 0n, usec: 0n, bytes: 0n },
-            delete: { count: 0n, usec: 0n, bytes: 0n },
+            read: { count: 0n, usec: 0n, bytes: 0n, bps: 0n, iops: 0n, lat: 0n },
+            write: { count: 0n, usec: 0n, bytes: 0n, bps: 0n, iops: 0n, lat: 0n },
+            delete: { count: 0n, usec: 0n, bytes: 0n, bps: 0n, iops: 0n, lat: 0n },
         });
         const seen_pools = {};
         for (const pool_id in this.state.config.pools)
@@ -1585,11 +1656,25 @@ class Mon
                 }
             }
         }
-        if (prev_stats && prev_timestamp >= timestamp)
+        for (const osd in this.prev_stats.osd_diff)
         {
-            prev_stats = null;
+            for (const pool_id in this.prev_stats.osd_diff[osd].inode_stats)
+            {
+                for (const inode_num in this.prev_stats.osd_diff[osd].inode_stats[pool_id])
+                {
+                    inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();
+                    for (const op of [ 'read', 'write', 'delete' ])
+                    {
+                        const op_diff = this.prev_stats.osd_diff[osd].inode_stats[pool_id][inode_num][op] || {};
+                        const op_st = inode_stats[pool_id][inode_num][op];
+                        op_st.bps += op_diff.bps;
+                        op_st.iops += op_diff.iops;
+                        op_st.lat += op_diff.lat;
+                        op_st.n_osd = (op_st.n_osd || 0) + 1;
+                    }
+                }
+            }
         }
-        const tm = prev_stats ? BigInt(timestamp - prev_timestamp) : 0;
         for (const pool_id in inode_stats)
         {
             for (const inode_num in inode_stats[pool_id])
@@ -1598,11 +1683,12 @@ class Mon
                 for (const op of [ 'read', 'write', 'delete' ])
                 {
                     const op_st = inode_stats[pool_id][inode_num][op];
-                    const prev_st = prev_stats && prev_stats[pool_id] && prev_stats[pool_id][inode_num] && prev_stats[pool_id][inode_num][op];
-                    op_st.bps = prev_st ? (op_st.bytes - prev_st.bytes) * 1000n / tm : 0;
-                    op_st.iops = prev_st ? (op_st.count - prev_st.count) * 1000n / tm : 0;
-                    op_st.lat = prev_st ? (op_st.usec - prev_st.usec) / ((op_st.count - prev_st.count) || 1n) : 0;
-                    if (op_st.bps > 0 || op_st.iops > 0 || op_st.lat > 0)
+                    if (op_st.n_osd)
+                    {
+                        op_st.lat /= BigInt(op_st.n_osd);
+                        delete op_st.n_osd;
+                    }
+                    if (op_st.bps > 0 || op_st.iops > 0)
                         nonzero = true;
                 }
                 if (!nonzero && (!this.state.config.inode[pool_id] || !this.state.config.inode[pool_id][inode_num]))
@@ -1635,15 +1721,9 @@ class Mon
     async update_total_stats()
     {
         const txn = [];
-        const timestamp = Date.now();
         const { object_counts, object_bytes } = this.sum_object_counts();
-        let stats = this.sum_op_stats(timestamp, this.prev_stats);
-        let { inode_stats, seen_pools } = this.sum_inode_stats(
-            this.prev_stats ? this.prev_stats.inode_stats : null,
-            timestamp, this.prev_stats ? this.prev_stats.timestamp : null
-        );
-        this.prev_stats = { timestamp, inode_stats, osd_stats: { ...this.state.osd.stats } };
-        this.stat_time = Date.now();
+        let stats = this.sum_op_stats();
+        let { inode_stats, seen_pools } = this.sum_inode_stats();
         stats.object_counts = object_counts;
         stats.object_bytes = object_bytes;
         stats = this.serialize_bigints(stats);
@@ -1788,10 +1868,18 @@ class Mon
         {
             retries = 1;
         }
+        const tried = {};
         while (retries < 0 || retry < retries)
         {
             retry++;
             const base = this.pick_next_etcd();
+            let now = Date.now();
+            if (tried[base] && now-tried[base] < timeout)
+            {
+                await new Promise(ok => setTimeout(ok, timeout-(now-tried[base])));
+                now = Date.now();
+            }
+            tried[base] = now;
             const res = await POST(base+path, body, timeout);
             if (res.error)
             {

mon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vitastor-mon",
-  "version": "1.0.0",
+  "version": "1.1.0",
   "description": "Vitastor SDS monitor service",
   "main": "mon-main.js",
   "scripts": {

patches/cinder-vitastor.py

@@ -50,7 +50,7 @@ from cinder.volume import configuration
 from cinder.volume import driver
 from cinder.volume import volume_utils
 
-VERSION = '1.0.0'
+VERSION = '1.1.0'
 
 LOG = logging.getLogger(__name__)
 

rpm/build-tarball.sh

@@ -24,4 +24,4 @@ rm fio
 mv fio-copy fio
 FIO=`rpm -qi fio | perl -e 'while(<>) { /^Epoch[\s:]+(\S+)/ && print "$1:"; /^Version[\s:]+(\S+)/ && print $1; /^Release[\s:]+(\S+)/ && print "-$1"; }'`
 perl -i -pe 's/(Requires:\s*fio)([^\n]+)?/$1 = '$FIO'/' $VITASTOR/rpm/vitastor-el$EL.spec
-tar --transform 's#^#vitastor-1.0.0/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-1.0.0$(rpm --eval '%dist').tar.gz *
+tar --transform 's#^#vitastor-1.1.0/#' --exclude 'rpm/*.rpm' -czf $VITASTOR/../vitastor-1.1.0$(rpm --eval '%dist').tar.gz *

rpm/vitastor-el7.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-1.0.0.el7.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.1.0.el7.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el7.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el7.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.0.0
+Version:        1.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.0.0.el7.tar.gz
+Source0:        vitastor-1.1.0.el7.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el8.Dockerfile

@@ -35,7 +35,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-1.0.0.el8.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.1.0.el8.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el8.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el8.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.0.0
+Version:        1.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.0.0.el8.tar.gz
+Source0:        vitastor-1.1.0.el8.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

rpm/vitastor-el9.Dockerfile

@@ -18,7 +18,7 @@ ADD . /root/vitastor
 RUN set -e; \
     cd /root/vitastor/rpm; \
     sh build-tarball.sh; \
-    cp /root/vitastor-1.0.0.el9.tar.gz ~/rpmbuild/SOURCES; \
+    cp /root/vitastor-1.1.0.el9.tar.gz ~/rpmbuild/SOURCES; \
     cp vitastor-el9.spec ~/rpmbuild/SPECS/vitastor.spec; \
     cd ~/rpmbuild/SPECS/; \
     rpmbuild -ba vitastor.spec; \

rpm/vitastor-el9.spec

@@ -1,11 +1,11 @@
 Name:           vitastor
-Version:        1.0.0
+Version:        1.1.0
 Release:        1%{?dist}
 Summary:        Vitastor, a fast software-defined clustered block storage
 
 License:        Vitastor Network Public License 1.1
 URL:            https://vitastor.io/
-Source0:        vitastor-1.0.0.el9.tar.gz
+Source0:        vitastor-1.1.0.el9.tar.gz
 
 BuildRequires:  liburing-devel >= 0.6
 BuildRequires:  gperftools-devel

src/CMakeLists.txt

@@ -16,7 +16,7 @@ if("${CMAKE_INSTALL_PREFIX}" MATCHES "^/usr/local/?$")
 	set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
 endif()
 
-add_definitions(-DVERSION="1.0.0")
+add_definitions(-DVERSION="1.1.0")
 add_definitions(-Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fdiagnostics-color=always -I ${CMAKE_SOURCE_DIR}/src)
 if (${WITH_ASAN})
 	add_definitions(-fsanitize=address -fno-omit-frame-pointer)
@@ -137,6 +137,7 @@ endif (${WITH_FIO})
 add_library(vitastor_client SHARED
 	cluster_client.cpp
 	cluster_client_list.cpp
+	cluster_client_wb.cpp
 	vitastor_c.cpp
 	cli_common.cpp
 	cli_alloc_osd.cpp
@@ -300,7 +301,7 @@ target_link_libraries(test_crc32
 add_executable(test_cluster_client
 	EXCLUDE_FROM_ALL
 	test_cluster_client.cpp
-	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
+	pg_states.cpp osd_ops.cpp cluster_client.cpp cluster_client_list.cpp cluster_client_wb.cpp msgr_op.cpp mock/messenger.cpp msgr_stop.cpp
 	etcd_state_client.cpp timerfd_manager.cpp str_util.cpp ../json11/json11.cpp
 )
 target_compile_definitions(test_cluster_client PUBLIC -D__MOCK__)

src/addr_util.cpp

@@ -19,8 +19,8 @@ bool string_to_addr(std::string str, bool parse_port, int default_port, struct s
         if (p != std::string::npos && !(str.length() > 0 && str[p-1] == ']')) // "[ipv6]" which contains ':'
         {
             char null_byte = 0;
-            int n = sscanf(str.c_str()+p+1, "%d%c", &default_port, &null_byte);
-            if (n != 1 || default_port >= 0x10000)
+            int scanned = sscanf(str.c_str()+p+1, "%d%c", &default_port, &null_byte);
+            if (scanned != 1 || default_port >= 0x10000)
                 return false;
             str = str.substr(0, p);
         }

src/blockstore_disk.cpp

@@ -45,13 +45,31 @@ void blockstore_disk_t::parse_config(std::map<std::string, std::string> & config
     meta_block_size = parse_size(config["meta_block_size"]);
     bitmap_granularity = parse_size(config["bitmap_granularity"]);
     meta_format = stoull_full(config["meta_format"]);
-    cached_io_data = config["cached_io_data"] == "true" || config["cached_io_data"] == "yes" || config["cached_io_data"] == "1";
-    cached_io_meta = cached_io_data && (meta_device == data_device || meta_device == "") &&
-        config.find("cached_io_meta") == config.end() ||
-        config["cached_io_meta"] == "true" || config["cached_io_meta"] == "yes" || config["cached_io_meta"] == "1";
-    cached_io_journal = cached_io_meta && (journal_device == meta_device || journal_device == "") &&
-        config.find("cached_io_journal") == config.end() ||
-        config["cached_io_journal"] == "true" || config["cached_io_journal"] == "yes" || config["cached_io_journal"] == "1";
+    if (config.find("data_io") == config.end() &&
+        config.find("meta_io") == config.end() &&
+        config.find("journal_io") == config.end())
+    {
+        bool cached_io_data = config["cached_io_data"] == "true" || config["cached_io_data"] == "yes" || config["cached_io_data"] == "1";
+        bool cached_io_meta = cached_io_data && (meta_device == data_device || meta_device == "") &&
+            config.find("cached_io_meta") == config.end() ||
+            config["cached_io_meta"] == "true" || config["cached_io_meta"] == "yes" || config["cached_io_meta"] == "1";
+        bool cached_io_journal = cached_io_meta && (journal_device == meta_device || journal_device == "") &&
+            config.find("cached_io_journal") == config.end() ||
+            config["cached_io_journal"] == "true" || config["cached_io_journal"] == "yes" || config["cached_io_journal"] == "1";
+        data_io = cached_io_data ? "cached" : "direct";
+        meta_io = cached_io_meta ? "cached" : "direct";
+        journal_io = cached_io_journal ? "cached" : "direct";
+    }
+    else
+    {
+        data_io = config.find("data_io") != config.end() ? config["data_io"] : "direct";
+        meta_io = config.find("meta_io") != config.end()
+            ? config["meta_io"]
+            : (meta_device == data_device || meta_device == "" ? data_io : "direct");
+        journal_io = config.find("journal_io") != config.end()
+            ? config["journal_io"]
+            : (journal_device == meta_device || journal_device == "" ? meta_io : "direct");
+    }
     if (config["data_csum_type"] == "crc32c")
     {
         data_csum_type = BLOCKSTORE_CSUM_CRC32C;
@@ -272,9 +290,19 @@ static void check_size(int fd, uint64_t *size, uint64_t *sectsize, std::string n
     }
 }
 
+static int bs_openmode(const std::string & mode)
+{
+    if (mode == "directsync")
+        return O_DIRECT|O_SYNC;
+    else if (mode == "cached")
+        return O_SYNC;
+    else
+        return O_DIRECT;
+}
+
 void blockstore_disk_t::open_data()
 {
-    data_fd = open(data_device.c_str(), (cached_io_data ? O_SYNC : O_DIRECT) | O_RDWR);
+    data_fd = open(data_device.c_str(), bs_openmode(data_io) | O_RDWR);
     if (data_fd == -1)
     {
         throw std::runtime_error("Failed to open data device "+data_device+": "+std::string(strerror(errno)));
@@ -299,9 +327,9 @@ void blockstore_disk_t::open_data()
 
 void blockstore_disk_t::open_meta()
 {
-    if (meta_device != data_device || cached_io_meta != cached_io_data)
+    if (meta_device != data_device || meta_io != data_io)
     {
-        meta_fd = open(meta_device.c_str(), (cached_io_meta ? O_SYNC : O_DIRECT) | O_RDWR);
+        meta_fd = open(meta_device.c_str(), bs_openmode(meta_io) | O_RDWR);
         if (meta_fd == -1)
         {
             throw std::runtime_error("Failed to open metadata device "+meta_device+": "+std::string(strerror(errno)));
@@ -337,9 +365,9 @@ void blockstore_disk_t::open_meta()
 
 void blockstore_disk_t::open_journal()
 {
-    if (journal_device != meta_device || cached_io_journal != cached_io_meta)
+    if (journal_device != meta_device || journal_io != meta_io)
     {
-        journal_fd = open(journal_device.c_str(), (cached_io_journal ? O_SYNC : O_DIRECT) | O_RDWR);
+        journal_fd = open(journal_device.c_str(), bs_openmode(journal_io) | O_RDWR);
         if (journal_fd == -1)
         {
             throw std::runtime_error("Failed to open journal device "+journal_device+": "+std::string(strerror(errno)));

src/blockstore_disk.h

@@ -31,8 +31,9 @@ struct blockstore_disk_t
     uint32_t csum_block_size = 4096;
     // By default, Blockstore locks all opened devices exclusively. This option can be used to disable locking
     bool disable_flock = false;
-    // Use Linux page cache for reads and writes, i.e. open FDs with O_SYNC instead of O_DIRECT
-    bool cached_io_data = false, cached_io_meta = false, cached_io_journal = false;
+    // I/O modes for data, metadata and journal: direct or "" = O_DIRECT, cached = O_SYNC, directsync = O_DIRECT|O_SYNC
+    // O_SYNC without O_DIRECT = use Linux page cache for reads and writes
+    std::string data_io, meta_io, journal_io;
 
     int meta_fd = -1, data_fd = -1, journal_fd = -1;
     uint64_t meta_offset, meta_device_sect, meta_device_size, meta_len, meta_format = 0;

src/blockstore_impl.cpp

@@ -384,6 +384,10 @@ void blockstore_impl_t::enqueue_op(blockstore_op_t *op)
         ringloop->set_immediate([op]() { std::function<void (blockstore_op_t*)>(op->callback)(op); });
         return;
     }
+    if (op->opcode == BS_OP_SYNC)
+    {
+        unsynced_queued_ops = 0;
+    }
     init_op(op);
     submit_queue.push_back(op);
     ringloop->wakeup();
@@ -393,6 +397,7 @@ void blockstore_impl_t::init_op(blockstore_op_t *op)
 {
     // Call constructor without allocating memory. We'll call destructor before returning op back
     new ((void*)op->private_data) blockstore_op_private_t;
+    PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
     PRIV(op)->wait_for = 0;
     PRIV(op)->op_state = 0;
     PRIV(op)->pending_ops = 0;

src/blockstore_impl.h

@@ -210,7 +210,7 @@ struct blockstore_op_private_t
     std::vector<copy_buffer_t> read_vec;
 
     // Sync, write
-    int min_flushed_journal_sector, max_flushed_journal_sector;
+    uint64_t min_flushed_journal_sector, max_flushed_journal_sector;
 
     // Write
     struct iovec iov_zerofill[3];
@@ -220,7 +220,6 @@ struct blockstore_op_private_t
 
     // Sync
     std::vector<obj_ver_id> sync_big_writes, sync_small_writes;
-    int sync_small_checked, sync_big_checked;
 };
 
 typedef uint32_t pool_id_t;
@@ -263,6 +262,8 @@ class blockstore_impl_t
     int throttle_target_parallelism = 1;
     // Minimum difference in microseconds between target and real execution times to throttle the response
     int throttle_threshold_us = 50;
+    // Maximum writes between automatically added fsync operations
+    uint64_t autosync_writes = 128;
     /******* END OF OPTIONS *******/
 
     struct ring_consumer_t ring_consumer;
@@ -274,6 +275,7 @@ class blockstore_impl_t
     std::vector<blockstore_op_t*> submit_queue;
     std::vector<obj_ver_id> unsynced_big_writes, unsynced_small_writes;
     int unsynced_big_write_count = 0;
+    int unsynced_queued_ops = 0;
     allocator *data_alloc = NULL;
     uint8_t *zero_object;
 

src/blockstore_journal.cpp

@@ -198,6 +198,7 @@ void blockstore_impl_t::prepare_journal_sector_write(int cur_sector, blockstore_
     priv->pending_ops++;
     if (!priv->min_flushed_journal_sector)
         priv->min_flushed_journal_sector = 1+cur_sector;
+    assert(priv->min_flushed_journal_sector <= journal.sector_count);
     priv->max_flushed_journal_sector = 1+cur_sector;
 }
 

src/blockstore_open.cpp

@@ -19,6 +19,10 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config, bool init)
     throttle_target_mbs = strtoull(config["throttle_target_mbs"].c_str(), NULL, 10);
     throttle_target_parallelism = strtoull(config["throttle_target_parallelism"].c_str(), NULL, 10);
     throttle_threshold_us = strtoull(config["throttle_threshold_us"].c_str(), NULL, 10);
+    if (config.find("autosync_writes") != config.end())
+    {
+        autosync_writes = strtoull(config["autosync_writes"].c_str(), NULL, 10);
+    }
     if (!max_flusher_count)
     {
         max_flusher_count = 256;

src/blockstore_sync.cpp

@@ -27,8 +27,6 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
         unsynced_big_write_count -= unsynced_big_writes.size();
         PRIV(op)->sync_big_writes.swap(unsynced_big_writes);
         PRIV(op)->sync_small_writes.swap(unsynced_small_writes);
-        PRIV(op)->sync_small_checked = 0;
-        PRIV(op)->sync_big_checked = 0;
         unsynced_big_writes.clear();
         unsynced_small_writes.clear();
         if (PRIV(op)->sync_big_writes.size() > 0)

src/blockstore_write.cpp

@@ -127,8 +127,9 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
             return false;
         }
     }
-    if (wait_big && !is_del && !deleted && op->len < dsk.data_block_size &&
-        immediate_commit != IMMEDIATE_ALL)
+    bool imm = (op->len < dsk.data_block_size ? (immediate_commit != IMMEDIATE_NONE) : (immediate_commit == IMMEDIATE_ALL));
+    if (wait_big && !is_del && !deleted && op->len < dsk.data_block_size && !imm ||
+        !imm && unsynced_queued_ops >= autosync_writes)
     {
         // Issue an additional sync so that the previous big write can reach the journal
         blockstore_op_t *sync_op = new blockstore_op_t;
@@ -139,6 +140,8 @@ bool blockstore_impl_t::enqueue_write(blockstore_op_t *op)
         };
         enqueue_op(sync_op);
     }
+    else if (!imm)
+        unsynced_queued_ops++;
 #ifdef BLOCKSTORE_DEBUG
     if (is_del)
         printf("Delete %lx:%lx v%lu\n", op->oid.inode, op->oid.stripe, op->version);
@@ -286,13 +289,18 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         printf("Restoring %lx:%lx version: v%lu -> v%lu\n", op->oid.inode, op->oid.stripe, op->version, PRIV(op)->real_version);
 #endif
         auto prev_it = dirty_it;
-        prev_it--;
-        if (prev_it->first.oid == op->oid && prev_it->first.version >= PRIV(op)->real_version)
+        if (prev_it != dirty_db.begin())
         {
-            // Original version is still invalid
-            // All subsequent writes to the same object must be canceled too
-            cancel_all_writes(op, dirty_it, -EEXIST);
-            return 2;
+            prev_it--;
+            if (prev_it->first.oid == op->oid && prev_it->first.version >= PRIV(op)->real_version)
+            {
+                // Original version is still invalid
+                // All subsequent writes to the same object must be canceled too
+                printf("Tried to write %lx:%lx v%lu after delete (old version v%lu), but already have v%lu\n",
+                    op->oid.inode, op->oid.stripe, PRIV(op)->real_version, op->version, prev_it->first.version);
+                cancel_all_writes(op, dirty_it, -EEXIST);
+                return 2;
+            }
         }
         op->version = PRIV(op)->real_version;
         PRIV(op)->real_version = 0;
@@ -378,7 +386,6 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             sqe, dsk.data_fd, PRIV(op)->iov_zerofill, vcnt, dsk.data_offset + (loc << dsk.block_order) + op->offset - stripe_offset
         );
         PRIV(op)->pending_ops = 1;
-        PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
         if (immediate_commit != IMMEDIATE_ALL)
         {
             // Increase the counter, but don't save into unsynced_writes yet (can't sync until the write is finished)
@@ -415,16 +422,10 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         write_iodepth++;
         // Got SQEs. Prepare previous journal sector write if required
         auto cb = [this, op](ring_data_t *data) { handle_write_event(data, op); };
-        if (immediate_commit == IMMEDIATE_NONE)
+        if (immediate_commit == IMMEDIATE_NONE &&
+            !journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size))
         {
-            if (!journal.entry_fits(sizeof(journal_entry_small_write) + dyn_size))
-            {
-                prepare_journal_sector_write(journal.cur_sector, op);
-            }
-            else
-            {
-                PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-            }
+            prepare_journal_sector_write(journal.cur_sector, op);
         }
         // Then pre-fill journal entry
         journal_entry_small_write *je = (journal_entry_small_write*)prefill_single_journal_entry(
@@ -750,17 +751,11 @@ int blockstore_impl_t::dequeue_del(blockstore_op_t *op)
     }
     write_iodepth++;
     // Prepare journal sector write
-    if (immediate_commit == IMMEDIATE_NONE)
+    if (immediate_commit == IMMEDIATE_NONE &&
+        (dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
+        journal.sector_info[journal.cur_sector].dirty)
     {
-        if ((dsk.journal_block_size - journal.in_sector_pos) < sizeof(journal_entry_del) &&
-            journal.sector_info[journal.cur_sector].dirty)
-        {
-            prepare_journal_sector_write(journal.cur_sector, op);
-        }
-        else
-        {
-            PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
-        }
+        prepare_journal_sector_write(journal.cur_sector, op);
     }
     // Pre-fill journal entry
     journal_entry_del *je = (journal_entry_del*)prefill_single_journal_entry(

src/cli.cpp

@@ -349,6 +349,7 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
                 p->ringloop->wait();
         }
         // Destroy the client
+        p->cli->flush();
         delete p->cli;
         delete p->epmgr;
         delete p->ringloop;
@@ -357,6 +358,8 @@ static int run(cli_tool_t *p, json11::Json::object cfg)
         p->ringloop = NULL;
     }
     // Print result
+    fflush(stderr);
+    fflush(stdout);
     if (p->json_output && !result.data.is_null())
     {
         printf("%s\n", result.data.dump().c_str());

src/cli_alloc_osd.cpp

@@ -77,8 +77,8 @@ struct alloc_osd_t
                     std::string key = base64_decode(kv["key"].string_value());
                     osd_num_t cur_osd;
                     char null_byte = 0;
-                    sscanf(key.c_str() + parent->cli->st_cli.etcd_prefix.length(), "/osd/stats/%lu%c", &cur_osd, &null_byte);
-                    if (!cur_osd || null_byte != 0)
+                    int scanned = sscanf(key.c_str() + parent->cli->st_cli.etcd_prefix.length(), "/osd/stats/%lu%c", &cur_osd, &null_byte);
+                    if (scanned != 1 || !cur_osd)
                     {
                         fprintf(stderr, "Invalid key in etcd: %s\n", key.c_str());
                         continue;

src/cli_df.cpp

@@ -67,8 +67,8 @@ resume_1:
             // pool ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -82,8 +82,8 @@ resume_1:
             // osd ID
             osd_num_t osd_num;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/osd/stats/%lu%c", &osd_num, &null_byte);
-            if (!osd_num || osd_num >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/osd/stats/%lu%c", &osd_num, &null_byte);
+            if (scanned != 1 || !osd_num || osd_num >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -124,6 +124,7 @@ resume_1:
                 pool_avail *= (pool_cfg.pg_size - pool_cfg.parity_chunks);
             }
             pool_stats[pool_cfg.id] = json11::Json::object {
+                { "id", (uint64_t)pool_cfg.id },
                 { "name", pool_cfg.name },
                 { "pg_count", pool_cfg.pg_count },
                 { "scheme", pool_cfg.scheme == POOL_SCHEME_REPLICATED ? "replicated" : "ec" },

src/cli_ls.cpp

@@ -133,8 +133,8 @@ resume_1:
             // pool ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(), "/pool/stats/%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -149,9 +149,9 @@ resume_1:
             pool_id_t pool_id;
             inode_t only_inode_num;
             char null_byte = 0;
-            sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(),
+            int scanned = sscanf(kv.key.substr(parent->cli->st_cli.etcd_prefix.length()).c_str(),
                 "/inode/stats/%u/%lu%c", &pool_id, &only_inode_num, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || INODE_POOL(only_inode_num) != 0 || null_byte != 0)
+            if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || INODE_POOL(only_inode_num) != 0)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;
@@ -174,7 +174,7 @@ resume_1:
                     { "size", 0 },
                     { "readonly", false },
                     { "pool_id", (uint64_t)INODE_POOL(inode_num) },
-                    { "pool_name", pool_it == parent->cli->st_cli.pool_config.end()
+                    { "pool_name", pool_it != parent->cli->st_cli.pool_config.end()
                         ? (pool_it->second.name == "" ? "<Unnamed>" : pool_it->second.name) : "?" },
                     { "inode_num", INODE_NO_POOL(inode_num) },
                     { "inode_id", inode_num },

src/cli_merge.cpp

@@ -53,6 +53,7 @@ struct snap_merger_t
     std::map<inode_t, std::vector<uint64_t>> layer_lists;
     std::map<inode_t, uint64_t> layer_block_size;
     std::map<inode_t, uint64_t> layer_list_pos;
+    std::vector<snap_rw_op_t*> continue_rwo, continue_rwo2;
     int in_flight = 0;
     uint64_t last_fsync_offset = 0;
     uint64_t last_written_offset = 0;
@@ -304,6 +305,12 @@ struct snap_merger_t
         oit = merge_offsets.begin();
     resume_5:
         // Now read, overwrite and optionally delete offsets one by one
+        continue_rwo2.swap(continue_rwo);
+        for (auto rwo: continue_rwo2)
+        {
+            next_write(rwo);
+        }
+        continue_rwo2.clear();
         while (in_flight < parent->iodepth*parent->parallel_osds &&
             oit != merge_offsets.end() && !rwo_error.size())
         {
@@ -464,7 +471,8 @@ struct snap_merger_t
                 rwo->error_offset = op->offset;
                 rwo->error_read = true;
             }
-            next_write(rwo);
+            continue_rwo.push_back(rwo);
+            parent->ringloop->wakeup();
         };
         parent->cli->execute(op);
     }
@@ -544,11 +552,9 @@ struct snap_merger_t
             }
             // Increment CAS version
             rwo->op.version = subop->version;
-            if (use_cas)
-                next_write(rwo);
-            else
-                autofree_op(rwo);
             delete subop;
+            continue_rwo.push_back(rwo);
+            parent->ringloop->wakeup();
         };
         parent->cli->execute(subop);
     }

src/cli_rm.cpp

@@ -384,8 +384,8 @@ resume_100:
                 pool_id_t pool_id = 0;
                 inode_t inode = 0;
                 char null_byte = 0;
-                sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode, &null_byte);
-                if (!inode || null_byte != 0)
+                int scanned = sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode, &null_byte);
+                if (scanned != 2 || !inode)
                 {
                     result = (cli_result_t){ .err = EIO, .text = "Bad key returned from etcd: "+kv.key };
                     state = 100;

src/cli_status.cpp

@@ -132,8 +132,8 @@ resume_2:
             auto kv = parent->cli->st_cli.parse_etcd_kv(osd_stats[i]);
             osd_num_t stat_osd_num = 0;
             char null_byte = 0;
-            sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.size(), "/osd/stats/%lu%c", &stat_osd_num, &null_byte);
-            if (!stat_osd_num || null_byte != 0)
+            int scanned = sscanf(kv.key.c_str() + parent->cli->st_cli.etcd_prefix.size(), "/osd/stats/%lu%c", &stat_osd_num, &null_byte);
+            if (scanned != 1 || !stat_osd_num)
             {
                 fprintf(stderr, "Invalid key in etcd: %s\n", kv.key.c_str());
                 continue;

src/cluster_client.cpp

@@ -3,21 +3,13 @@
 
 #include <stdexcept>
 #include <assert.h>
-#include "cluster_client.h"
-
-#define SCRAP_BUFFER_SIZE 4*1024*1024
-#define PART_SENT 1
-#define PART_DONE 2
-#define PART_ERROR 4
-#define PART_RETRY 8
-#define CACHE_DIRTY 1
-#define CACHE_FLUSHING 2
-#define CACHE_REPEATING 3
-#define OP_FLUSH_BUFFER 0x02
-#define OP_IMMEDIATE_COMMIT 0x04
+#include "cluster_client_impl.h"
+#include "http_client.h" // json_is_true
 
 cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config)
 {
+    wb = new writeback_cache_t();
+
     cli_config = config.object_items();
     file_config = osd_messenger_t::read_config(config);
     config = osd_messenger_t::merge_configs(cli_config, file_config, etcd_global_config, {});
@@ -37,20 +29,14 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
             continue_lists();
             continue_raw_ops(peer_osd);
         }
-        else if (dirty_buffers.size())
+        else
         {
             // peer_osd just dropped connection
             // determine WHICH dirty_buffers are now obsolete and repeat them
-            for (auto & wr: dirty_buffers)
+            if (wb->repeat_ops_for(this, peer_osd) > 0)
             {
-                if (affects_osd(wr.first.inode, wr.first.stripe, wr.second.len, peer_osd) &&
-                    wr.second.state != CACHE_REPEATING)
-                {
-                    // FIXME: Flush in larger parts
-                    flush_buffer(wr.first, &wr.second);
-                }
+                continue_ops();
             }
-            continue_ops();
         }
     };
     msgr.exec_op = [this](osd_op_t *op)
@@ -78,16 +64,14 @@ cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd
 
 cluster_client_t::~cluster_client_t()
 {
-    for (auto bp: dirty_buffers)
-    {
-        free(bp.second.buf);
-    }
-    dirty_buffers.clear();
+    msgr.repeer_pgs = [this](osd_num_t){};
     if (ringloop)
     {
         ringloop->unregister_consumer(&consumer);
     }
     free(scrap_buffer);
+    delete wb;
+    wb = NULL;
 }
 
 cluster_op_t::~cluster_op_t()
@@ -136,6 +120,19 @@ void cluster_client_t::init_msgr()
     }
 }
 
+void cluster_client_t::unshift_op(cluster_op_t *op)
+{
+    op->next = op_queue_head;
+    if (op_queue_head)
+    {
+        op_queue_head->prev = op;
+        op_queue_head = op;
+    }
+    else
+        op_queue_tail = op_queue_head = op;
+    inc_wait(op->opcode, op->flags, op->next, 1);
+}
+
 void cluster_client_t::calc_wait(cluster_op_t *op)
 {
     op->prev_wait = 0;
@@ -156,7 +153,7 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
     {
         for (auto prev = op->prev; prev; prev = prev->prev)
         {
-            if (prev->opcode == OSD_OP_SYNC || prev->opcode == OSD_OP_WRITE && !(prev->flags & OP_IMMEDIATE_COMMIT))
+            if (prev->opcode == OSD_OP_SYNC || prev->opcode == OSD_OP_WRITE && (!(prev->flags & OP_IMMEDIATE_COMMIT) || enable_writeback))
             {
                 op->prev_wait++;
             }
@@ -166,68 +163,58 @@ void cluster_client_t::calc_wait(cluster_op_t *op)
     }
     else /* if (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP) */
     {
-        for (auto prev = op_queue_head; prev && prev != op; prev = prev->next)
-        {
-            if (prev->opcode == OSD_OP_WRITE && (prev->flags & OP_FLUSH_BUFFER))
-            {
-                op->prev_wait++;
-            }
-            else if (prev->opcode == OSD_OP_WRITE || prev->opcode == OSD_OP_READ ||
-                prev->opcode == OSD_OP_READ_BITMAP || prev->opcode == OSD_OP_READ_CHAIN_BITMAP)
-            {
-                // Flushes are always in the beginning (we're scanning from the beginning of the queue)
-                break;
-            }
-        }
-        if (!op->prev_wait)
-            continue_rw(op);
+        continue_rw(op);
     }
 }
 
 void cluster_client_t::inc_wait(uint64_t opcode, uint64_t flags, cluster_op_t *next, int inc)
 {
-    if (opcode == OSD_OP_WRITE)
+    if (opcode != OSD_OP_WRITE && opcode != OSD_OP_SYNC)
     {
-        while (next)
-        {
-            auto n2 = next->next;
-            if (next->opcode == OSD_OP_SYNC && !(flags & OP_IMMEDIATE_COMMIT) ||
-                next->opcode == OSD_OP_WRITE && (flags & OP_FLUSH_BUFFER) && !(next->flags & OP_FLUSH_BUFFER) ||
-                (next->opcode == OSD_OP_READ || next->opcode == OSD_OP_READ_BITMAP ||
-                    next->opcode == OSD_OP_READ_CHAIN_BITMAP) && (flags & OP_FLUSH_BUFFER))
-            {
-                next->prev_wait += inc;
-                assert(next->prev_wait >= 0);
-                if (!next->prev_wait)
-                {
-                    if (next->opcode == OSD_OP_SYNC)
-                        continue_sync(next);
-                    else
-                        continue_rw(next);
-                }
-            }
-            next = n2;
-        }
+        return;
     }
-    else if (opcode == OSD_OP_SYNC)
+    cluster_op_t *bh_ops_local[32], **bh_ops = bh_ops_local;
+    int bh_op_count = 0, bh_op_max = 32;
+    while (next)
     {
-        while (next)
+        auto n2 = next->next;
+        if (opcode == OSD_OP_WRITE
+            ? (next->opcode == OSD_OP_SYNC && (!(flags & OP_IMMEDIATE_COMMIT) || enable_writeback) ||
+                next->opcode == OSD_OP_WRITE && (flags & OP_FLUSH_BUFFER) && !(next->flags & OP_FLUSH_BUFFER))
+            : (next->opcode == OSD_OP_SYNC || next->opcode == OSD_OP_WRITE))
         {
-            auto n2 = next->next;
-            if (next->opcode == OSD_OP_SYNC || next->opcode == OSD_OP_WRITE)
+            next->prev_wait += inc;
+            assert(next->prev_wait >= 0);
+            if (!next->prev_wait)
             {
-                next->prev_wait += inc;
-                assert(next->prev_wait >= 0);
-                if (!next->prev_wait)
+                // Kind of std::vector with local "small vector optimisation"
+                if (bh_op_count >= bh_op_max)
                 {
-                    if (next->opcode == OSD_OP_SYNC)
-                        continue_sync(next);
-                    else
-                        continue_rw(next);
+                    bh_op_max *= 2;
+                    cluster_op_t **n = (cluster_op_t**)malloc_or_die(sizeof(cluster_op_t*) * bh_op_max);
+                    memcpy(n, bh_ops, sizeof(cluster_op_t*) * bh_op_count);
+                    if (bh_ops != bh_ops_local)
+                    {
+                        free(bh_ops);
+                    }
+                    bh_ops = n;
                 }
+                bh_ops[bh_op_count++] = next;
             }
-            next = n2;
         }
+        next = n2;
+    }
+    for (int i = 0; i < bh_op_count; i++)
+    {
+        cluster_op_t *next = bh_ops[i];
+        if (next->opcode == OSD_OP_SYNC)
+            continue_sync(next);
+        else
+            continue_rw(next);
+    }
+    if (bh_ops != bh_ops_local)
+    {
+        free(bh_ops);
     }
 }
 
@@ -245,13 +232,37 @@ void cluster_client_t::erase_op(cluster_op_t *op)
         op_queue_tail = op->prev;
     op->next = op->prev = NULL;
     if (flags & OP_FLUSH_BUFFER)
+    {
+        // Completed flushes change writeback buffer states,
+        // so the callback should be run before inc_wait()
+        // which may continue following SYNCs, but these SYNCs
+        // should know about the changed buffer state
+        // This is ugly but this is the way we do it
         std::function<void(cluster_op_t*)>(op->callback)(op);
-    if (!(flags & OP_IMMEDIATE_COMMIT))
+    }
+    if (!(flags & OP_IMMEDIATE_COMMIT) || enable_writeback)
+    {
         inc_wait(opcode, flags, next, -1);
-    // Call callback at the end to avoid inconsistencies in prev_wait
-    // if the callback adds more operations itself
+    }
     if (!(flags & OP_FLUSH_BUFFER))
+    {
+        // Call callback at the end to avoid inconsistencies in prev_wait
+        // if the callback adds more operations itself
         std::function<void(cluster_op_t*)>(op->callback)(op);
+    }
+    if (flags & OP_FLUSH_BUFFER)
+    {
+        int i = 0;
+        while (i < wb->writeback_overflow.size() && wb->writebacks_active < client_max_writeback_iodepth)
+        {
+            execute_internal(wb->writeback_overflow[i]);
+            i++;
+        }
+        if (i > 0)
+        {
+            wb->writeback_overflow.erase(wb->writeback_overflow.begin(), wb->writeback_overflow.begin()+i);
+        }
+    }
 }
 
 void cluster_client_t::continue_ops(bool up_retry)
@@ -295,6 +306,7 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & etcd_global_co
 {
     this->etcd_global_config = etcd_global_config;
     config = osd_messenger_t::merge_configs(cli_config, file_config, etcd_global_config, {});
+    // client_max_dirty_bytes/client_dirty_limit
     if (config.find("client_max_dirty_bytes") != config.end())
     {
         client_max_dirty_bytes = config["client_max_dirty_bytes"].uint64_value();
@@ -310,11 +322,34 @@ void cluster_client_t::on_load_config_hook(json11::Json::object & etcd_global_co
     {
         client_max_dirty_bytes = DEFAULT_CLIENT_MAX_DIRTY_BYTES;
     }
+    // client_max_dirty_ops
     client_max_dirty_ops = config["client_max_dirty_ops"].uint64_value();
     if (!client_max_dirty_ops)
     {
         client_max_dirty_ops = DEFAULT_CLIENT_MAX_DIRTY_OPS;
     }
+    // client_enable_writeback
+    enable_writeback = json_is_true(config["client_enable_writeback"]) &&
+        json_is_true(config["client_writeback_allowed"]);
+    // client_max_buffered_bytes
+    client_max_buffered_bytes = config["client_max_buffered_bytes"].uint64_value();
+    if (!client_max_buffered_bytes)
+    {
+        client_max_buffered_bytes = DEFAULT_CLIENT_MAX_BUFFERED_BYTES;
+    }
+    // client_max_buffered_ops
+    client_max_buffered_ops = config["client_max_buffered_ops"].uint64_value();
+    if (!client_max_buffered_ops)
+    {
+        client_max_buffered_ops = DEFAULT_CLIENT_MAX_BUFFERED_OPS;
+    }
+    // client_max_writeback_iodepth
+    client_max_writeback_iodepth = config["client_max_writeback_iodepth"].uint64_value();
+    if (!client_max_writeback_iodepth)
+    {
+        client_max_writeback_iodepth = DEFAULT_CLIENT_MAX_WRITEBACK_IODEPTH;
+    }
+    // up_wait_retry_interval
     up_wait_retry_interval = config["up_wait_retry_interval"].uint64_value();
     if (!up_wait_retry_interval)
     {
@@ -374,6 +409,8 @@ void cluster_client_t::on_change_hook(std::map<std::string, etcd_kv_t> & changes
 
 bool cluster_client_t::get_immediate_commit(uint64_t inode)
 {
+    if (enable_writeback)
+        return false;
     pool_id_t pool_id = INODE_POOL(inode);
     if (!pool_id)
         return true;
@@ -408,6 +445,41 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
     }
 }
 
+bool cluster_client_t::flush()
+{
+    if (!ringloop)
+    {
+        if (wb->writeback_queue.size())
+        {
+            wb->start_writebacks(this, 0);
+            cluster_op_t *sync = new cluster_op_t;
+            sync->opcode = OSD_OP_SYNC;
+            sync->callback = [this](cluster_op_t *sync)
+            {
+                delete sync;
+            };
+            execute(sync);
+        }
+        return op_queue_head == NULL;
+    }
+    bool sync_done = false;
+    cluster_op_t *sync = new cluster_op_t;
+    sync->opcode = OSD_OP_SYNC;
+    sync->callback = [this, &sync_done](cluster_op_t *sync)
+    {
+        delete sync;
+        sync_done = true;
+    };
+    execute(sync);
+    while (!sync_done)
+    {
+        ringloop->loop();
+        if (!sync_done)
+            ringloop->wait();
+    }
+    return true;
+}
+
 /**
  * How writes are synced when immediate_commit is false
  *
@@ -428,6 +500,9 @@ void cluster_client_t::on_ready(std::function<void(void)> fn)
  * 3) if yes, send all SYNCs. otherwise, leave current SYNC as is.
  * 4) if any of them fail due to disconnected peers, repeat SYNC after repeating all writes
  * 5) if any of them fail due to other errors, fail the SYNC operation
+ *
+ * If writeback caching is turned on and writeback limit is not exhausted:
+ * data is just copied and the write is confirmed to the client.
  */
 void cluster_client_t::execute(cluster_op_t *op)
 {
@@ -443,67 +518,73 @@ void cluster_client_t::execute(cluster_op_t *op)
         offline_ops.push_back(op);
         return;
     }
+    op->flags = op->flags & OSD_OP_IGNORE_READONLY; // the only allowed flag
+    execute_internal(op);
+}
+
+void cluster_client_t::execute_internal(cluster_op_t *op)
+{
     op->cur_inode = op->inode;
     op->retval = 0;
-    op->flags = op->flags & OSD_OP_IGNORE_READONLY; // single allowed flag
-    if (op->opcode != OSD_OP_SYNC)
+    // check alignment, readonly flag and so on
+    if (!check_rw(op))
     {
-        pool_id_t pool_id = INODE_POOL(op->cur_inode);
-        if (!pool_id)
+        return;
+    }
+    if (op->opcode == OSD_OP_WRITE && enable_writeback && !(op->flags & OP_FLUSH_BUFFER) &&
+        !op->version /* FIXME no CAS writeback */)
+    {
+        if (wb->writebacks_active >= client_max_writeback_iodepth)
         {
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
+            // Writeback queue is full, postpone the operation
+            wb->writeback_overflow.push_back(op);
             return;
         }
-        auto pool_it = st_cli.pool_config.find(pool_id);
-        if (pool_it == st_cli.pool_config.end() || pool_it->second.real_pg_count == 0)
+        // Just copy and acknowledge the operation
+        wb->copy_write(op, CACHE_DIRTY);
+        while (wb->writeback_bytes + op->len > client_max_buffered_bytes || wb->writeback_queue_size > client_max_buffered_ops)
         {
-            // Pools are loaded, but this one is unknown
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
-            return;
-        }
-        // Check alignment
-        if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
-            op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
-        {
-            op->retval = -EINVAL;
-            std::function<void(cluster_op_t*)>(op->callback)(op);
-            return;
-        }
-        if (pool_it->second.immediate_commit == IMMEDIATE_ALL)
-        {
-            op->flags |= OP_IMMEDIATE_COMMIT;
+            // Initiate some writeback (asynchronously)
+            wb->start_writebacks(this, 1);
         }
+        op->retval = op->len;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return;
     }
     if (op->opcode == OSD_OP_WRITE && !(op->flags & OP_IMMEDIATE_COMMIT))
     {
+        if (!(op->flags & OP_FLUSH_BUFFER))
+        {
+            wb->copy_write(op, CACHE_WRITTEN);
+        }
         if (dirty_bytes >= client_max_dirty_bytes || dirty_ops >= client_max_dirty_ops)
         {
             // Push an extra SYNC operation to flush previous writes
             cluster_op_t *sync_op = new cluster_op_t;
             sync_op->opcode = OSD_OP_SYNC;
+            sync_op->flags = OP_FLUSH_BUFFER;
             sync_op->callback = [](cluster_op_t* sync_op)
             {
                 delete sync_op;
             };
-            sync_op->prev = op_queue_tail;
-            if (op_queue_tail)
-            {
-                op_queue_tail->next = sync_op;
-                op_queue_tail = sync_op;
-            }
-            else
-                op_queue_tail = op_queue_head = sync_op;
-            dirty_bytes = 0;
-            dirty_ops = 0;
-            calc_wait(sync_op);
+            execute_internal(sync_op);
         }
         dirty_bytes += op->len;
         dirty_ops++;
     }
     else if (op->opcode == OSD_OP_SYNC)
     {
+        // Flush the whole write-back queue first
+        if (!(op->flags & OP_FLUSH_BUFFER) && wb->writeback_overflow.size() > 0)
+        {
+            // Writeback queue is full, postpone the operation
+            wb->writeback_overflow.push_back(op);
+            return;
+        }
+        if (wb->writeback_queue.size())
+        {
+            wb->start_writebacks(this, 0);
+        }
         dirty_bytes = 0;
         dirty_ops = 0;
     }
@@ -515,7 +596,7 @@ void cluster_client_t::execute(cluster_op_t *op)
     }
     else
         op_queue_tail = op_queue_head = op;
-    if (!(op->flags & OP_IMMEDIATE_COMMIT))
+    if (!(op->flags & OP_IMMEDIATE_COMMIT) || enable_writeback)
         calc_wait(op);
     else
     {
@@ -526,6 +607,52 @@ void cluster_client_t::execute(cluster_op_t *op)
     }
 }
 
+bool cluster_client_t::check_rw(cluster_op_t *op)
+{
+    if (op->opcode == OSD_OP_SYNC)
+    {
+        return true;
+    }
+    pool_id_t pool_id = INODE_POOL(op->cur_inode);
+    if (!pool_id)
+    {
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    auto pool_it = st_cli.pool_config.find(pool_id);
+    if (pool_it == st_cli.pool_config.end() || pool_it->second.real_pg_count == 0)
+    {
+        // Pools are loaded, but this one is unknown
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    // Check alignment
+    if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
+        op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
+    {
+        op->retval = -EINVAL;
+        std::function<void(cluster_op_t*)>(op->callback)(op);
+        return false;
+    }
+    if (pool_it->second.immediate_commit == IMMEDIATE_ALL)
+    {
+        op->flags |= OP_IMMEDIATE_COMMIT;
+    }
+    if ((op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE) && !(op->flags & OSD_OP_IGNORE_READONLY))
+    {
+        auto ino_it = st_cli.inode_config.find(op->inode);
+        if (ino_it != st_cli.inode_config.end() && ino_it->second.readonly)
+        {
+            op->retval = -EROFS;
+            std::function<void(cluster_op_t*)>(op->callback)(op);
+            return false;
+        }
+    }
+    return true;
+}
+
 void cluster_client_t::execute_raw(osd_num_t osd_num, osd_op_t *op)
 {
     auto fd_it = msgr.osd_peer_fds.find(osd_num);
@@ -543,114 +670,6 @@ void cluster_client_t::execute_raw(osd_num_t osd_num, osd_op_t *op)
     }
 }
 
-void cluster_client_t::copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers)
-{
-    // Save operation for replay when one of PGs goes out of sync
-    // (primary OSD drops our connection in this case)
-    auto dirty_it = dirty_buffers.lower_bound((object_id){
-        .inode = op->inode,
-        .stripe = op->offset,
-    });
-    while (dirty_it != dirty_buffers.begin())
-    {
-        dirty_it--;
-        if (dirty_it->first.inode != op->inode ||
-            (dirty_it->first.stripe + dirty_it->second.len) <= op->offset)
-        {
-            dirty_it++;
-            break;
-        }
-    }
-    uint64_t pos = op->offset, len = op->len, iov_idx = 0, iov_pos = 0;
-    while (len > 0)
-    {
-        uint64_t new_len = 0;
-        if (dirty_it == dirty_buffers.end())
-        {
-            new_len = len;
-        }
-        else if (dirty_it->first.inode != op->inode || dirty_it->first.stripe > pos)
-        {
-            new_len = dirty_it->first.stripe - pos;
-            if (new_len > len)
-            {
-                new_len = len;
-            }
-        }
-        if (new_len > 0)
-        {
-            dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
-                .inode = op->inode,
-                .stripe = pos,
-            }, (cluster_buffer_t){
-                .buf = malloc_or_die(new_len),
-                .len = new_len,
-            });
-        }
-        // FIXME: Split big buffers into smaller ones on overwrites. But this will require refcounting
-        dirty_it->second.state = CACHE_DIRTY;
-        uint64_t cur_len = (dirty_it->first.stripe + dirty_it->second.len - pos);
-        if (cur_len > len)
-        {
-            cur_len = len;
-        }
-        while (cur_len > 0 && iov_idx < op->iov.count)
-        {
-            unsigned iov_len = (op->iov.buf[iov_idx].iov_len - iov_pos);
-            if (iov_len <= cur_len)
-            {
-                memcpy((uint8_t*)dirty_it->second.buf + pos - dirty_it->first.stripe,
-                    (uint8_t*)op->iov.buf[iov_idx].iov_base + iov_pos, iov_len);
-                pos += iov_len;
-                len -= iov_len;
-                cur_len -= iov_len;
-                iov_pos = 0;
-                iov_idx++;
-            }
-            else
-            {
-                memcpy((uint8_t*)dirty_it->second.buf + pos - dirty_it->first.stripe,
-                    (uint8_t*)op->iov.buf[iov_idx].iov_base + iov_pos, cur_len);
-                pos += cur_len;
-                len -= cur_len;
-                iov_pos += cur_len;
-                cur_len = 0;
-            }
-        }
-        dirty_it++;
-    }
-}
-
-void cluster_client_t::flush_buffer(const object_id & oid, cluster_buffer_t *wr)
-{
-    wr->state = CACHE_REPEATING;
-    cluster_op_t *op = new cluster_op_t;
-    op->flags = OSD_OP_IGNORE_READONLY|OP_FLUSH_BUFFER;
-    op->opcode = OSD_OP_WRITE;
-    op->cur_inode = op->inode = oid.inode;
-    op->offset = oid.stripe;
-    op->len = wr->len;
-    op->iov.push_back(wr->buf, wr->len);
-    op->callback = [wr](cluster_op_t* op)
-    {
-        if (wr->state == CACHE_REPEATING)
-        {
-            wr->state = CACHE_DIRTY;
-        }
-        delete op;
-    };
-    op->next = op_queue_head;
-    if (op_queue_head)
-    {
-        op_queue_head->prev = op;
-        op_queue_head = op;
-    }
-    else
-        op_queue_tail = op_queue_head = op;
-    inc_wait(op->opcode, op->flags, op->next, 1);
-    continue_rw(op);
-}
-
 int cluster_client_t::continue_rw(cluster_op_t *op)
 {
     if (op->state == 0)
@@ -659,27 +678,7 @@ int cluster_client_t::continue_rw(cluster_op_t *op)
         goto resume_1;
     else if (op->state == 2)
         goto resume_2;
-    else if (op->state == 3)
-        goto resume_3;
 resume_0:
-    if (op->opcode == OSD_OP_WRITE || op->opcode == OSD_OP_DELETE)
-    {
-        if (!(op->flags & OSD_OP_IGNORE_READONLY))
-        {
-            auto ino_it = st_cli.inode_config.find(op->inode);
-            if (ino_it != st_cli.inode_config.end() && ino_it->second.readonly)
-            {
-                op->retval = -EINVAL;
-                erase_op(op);
-                return 1;
-            }
-        }
-        if (op->opcode == OSD_OP_WRITE && !(op->flags & OP_IMMEDIATE_COMMIT) && !(op->flags & OP_FLUSH_BUFFER))
-        {
-            copy_write(op, dirty_buffers);
-        }
-    }
-resume_1:
     // Slice the operation into parts
     slice_rw(op);
     op->needs_reslice = false;
@@ -690,9 +689,9 @@ resume_1:
         erase_op(op);
         return 1;
     }
-resume_2:
+resume_1:
     // Send unsent parts, if they're not subject to change
-    op->state = 3;
+    op->state = 2;
     if (op->needs_reslice)
     {
         for (int i = 0; i < op->parts.size(); i++)
@@ -702,7 +701,7 @@ resume_2:
                 op->retval = -EPIPE;
             }
         }
-        goto resume_3;
+        goto resume_2;
     }
     for (int i = 0; i < op->parts.size(); i++)
     {
@@ -723,18 +722,18 @@ resume_2:
                         });
                     }
                 }
-                op->state = 2;
+                op->state = 1;
             }
         }
     }
-    if (op->state == 2)
+    if (op->state == 1)
     {
         return 0;
     }
-resume_3:
+resume_2:
     if (op->inflight_count > 0)
     {
-        op->state = 3;
+        op->state = 2;
         return 0;
     }
     if (op->done_count >= op->parts.size())
@@ -762,7 +761,7 @@ resume_3:
                 op->cur_inode = ino_it->second.parent_id;
                 op->parts.clear();
                 op->done_count = 0;
-                goto resume_1;
+                goto resume_0;
             }
         }
         op->retval = op->len;
@@ -774,7 +773,8 @@ resume_3:
         erase_op(op);
         return 1;
     }
-    else if (op->retval != 0 && op->retval != -EPIPE && op->retval != -EIO && op->retval != -ENOSPC)
+    else if (op->retval != 0 && !(op->flags & OP_FLUSH_BUFFER) &&
+        op->retval != -EPIPE && op->retval != -EIO && op->retval != -ENOSPC)
     {
         // Fatal error (neither -EPIPE, -EIO nor -ENOSPC)
         // FIXME: Add a parameter to allow to not wait for EIOs (incomplete or corrupted objects) to heal
@@ -789,7 +789,7 @@ resume_3:
         {
             op->parts.clear();
             op->done_count = 0;
-            goto resume_1;
+            goto resume_0;
         }
         else
         {
@@ -800,7 +800,7 @@ resume_3:
                     op->parts[i].flags = PART_RETRY;
                 }
             }
-            goto resume_2;
+            goto resume_1;
         }
     }
     return 0;
@@ -874,6 +874,11 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
     int iov_idx = 0;
     size_t iov_pos = 0;
     int i = 0;
+    // We also have to return reads from CACHE_REPEATING buffers - they are not
+    // guaranteed to be present on target OSDs at the moment of repeating
+    // And we're also free to return data from other cached buffers just
+    // because it's faster
+    bool dirty_copied = wb->read_from_cache(op, pool_cfg.bitmap_granularity);
     for (uint64_t stripe = first_stripe; stripe <= last_stripe; stripe += pg_block_size)
     {
         pg_num_t pg_num = (stripe/pool_cfg.pg_stripe_size) % pool_cfg.real_pg_count + 1; // like map_to_pg()
@@ -881,7 +886,8 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
         uint64_t end = (op->offset + op->len) > (stripe + pg_block_size)
             ? (stripe + pg_block_size) : (op->offset + op->len);
         op->parts[i].iov.reset();
-        if (op->cur_inode != op->inode)
+        op->parts[i].flags = 0;
+        if (op->cur_inode != op->inode || op->opcode == OSD_OP_READ && dirty_copied)
         {
             // Read remaining parts from upper layers
             uint64_t prev = begin, cur = begin;
@@ -918,7 +924,10 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             else
                 add_iov(cur-prev, skip_prev, op, iov_idx, iov_pos, op->parts[i].iov, scrap_buffer, scrap_buffer_size);
             if (end == begin)
+            {
                 op->done_count++;
+                op->parts[i].flags = PART_DONE;
+            }
         }
         else if (op->opcode != OSD_OP_READ_BITMAP && op->opcode != OSD_OP_READ_CHAIN_BITMAP && op->opcode != OSD_OP_DELETE)
         {
@@ -930,7 +939,6 @@ void cluster_client_t::slice_rw(cluster_op_t *op)
             op->opcode == OSD_OP_DELETE ? 0 : (uint32_t)(end - begin);
         op->parts[i].pg_num = pg_num;
         op->parts[i].osd_num = 0;
-        op->parts[i].flags = 0;
         i++;
     }
 }
@@ -1042,13 +1050,7 @@ int cluster_client_t::continue_sync(cluster_op_t *op)
             do_it++;
     }
     // Post sync to affected OSDs
-    for (auto & prev_op: dirty_buffers)
-    {
-        if (prev_op.second.state == CACHE_DIRTY)
-        {
-            prev_op.second.state = CACHE_FLUSHING;
-        }
-    }
+    wb->fsync_start();
     op->parts.resize(dirty_osds.size());
     op->retval = 0;
     {
@@ -1073,13 +1075,7 @@ resume_1:
     }
     if (op->retval != 0)
     {
-        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); uw_it++)
-        {
-            if (uw_it->second.state == CACHE_FLUSHING)
-            {
-                uw_it->second.state = CACHE_DIRTY;
-            }
-        }
+        wb->fsync_error();
         if (op->retval == -EPIPE || op->retval == -EIO || op->retval == -ENOSPC)
         {
             // Retry later
@@ -1093,16 +1089,7 @@ resume_1:
     }
     else
     {
-        for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); )
-        {
-            if (uw_it->second.state == CACHE_FLUSHING)
-            {
-                free(uw_it->second.buf);
-                dirty_buffers.erase(uw_it++);
-            }
-            else
-                uw_it++;
-        }
+        wb->fsync_ok();
     }
     erase_op(op);
     return 1;

src/cluster_client.h

@@ -8,6 +8,9 @@
 
 #define DEFAULT_CLIENT_MAX_DIRTY_BYTES 32*1024*1024
 #define DEFAULT_CLIENT_MAX_DIRTY_OPS 1024
+#define DEFAULT_CLIENT_MAX_BUFFERED_BYTES 32*1024*1024
+#define DEFAULT_CLIENT_MAX_BUFFERED_OPS 1024
+#define DEFAULT_CLIENT_MAX_WRITEBACK_IODEPTH 256
 #define INODE_LIST_DONE 1
 #define INODE_LIST_HAS_UNSTABLE 2
 #define OSD_OP_READ_BITMAP OSD_OP_SEC_READ_BMP
@@ -64,17 +67,12 @@ protected:
     cluster_op_t *prev = NULL, *next = NULL;
     int prev_wait = 0;
     friend class cluster_client_t;
-};
-
-struct cluster_buffer_t
-{
-    void *buf;
-    uint64_t len;
-    int state;
+    friend class writeback_cache_t;
 };
 
 struct inode_list_t;
 struct inode_list_osd_t;
+class writeback_cache_t;
 
 // FIXME: Split into public and private interfaces
 class cluster_client_t
@@ -83,16 +81,23 @@ class cluster_client_t
     ring_loop_t *ringloop;
 
     std::map<pool_id_t, uint64_t> pg_counts;
-    // FIXME: Implement inmemory_commit mode. Note that it requires to return overlapping reads from memory.
+    // client_max_dirty_* is actually "max unsynced", for the case when immediate_commit is off
     uint64_t client_max_dirty_bytes = 0;
     uint64_t client_max_dirty_ops = 0;
+    // writeback improves (1) small consecutive writes and (2) Q1 writes without fsync
+    bool enable_writeback = false;
+    // client_max_buffered_* is the real "dirty limit" - maximum amount of writes buffered in memory
+    uint64_t client_max_buffered_bytes = 0;
+    uint64_t client_max_buffered_ops = 0;
+    uint64_t client_max_writeback_iodepth = 0;
+
     int log_level;
     int up_wait_retry_interval = 500; // ms
 
     int retry_timeout_id = 0;
     std::vector<cluster_op_t*> offline_ops;
     cluster_op_t *op_queue_head = NULL, *op_queue_tail = NULL;
-    std::map<object_id, cluster_buffer_t> dirty_buffers;
+    writeback_cache_t *wb = NULL;
     std::set<osd_num_t> dirty_osds;
     uint64_t dirty_bytes = 0, dirty_ops = 0;
 
@@ -122,10 +127,10 @@ public:
     void execute_raw(osd_num_t osd_num, osd_op_t *op);
     bool is_ready();
     void on_ready(std::function<void(void)> fn);
+    bool flush();
 
     bool get_immediate_commit(uint64_t inode);
 
-    static void copy_write(cluster_op_t *op, std::map<object_id, cluster_buffer_t> & dirty_buffers);
     void continue_ops(bool up_retry = false);
     inode_list_t *list_inode_start(inode_t inode,
         std::function<void(inode_list_t* lst, std::set<object_id>&& objects, pg_num_t pg_num, osd_num_t primary_osd, int status)> callback);
@@ -138,12 +143,14 @@ public:
 
 protected:
     bool affects_osd(uint64_t inode, uint64_t offset, uint64_t len, osd_num_t osd);
-    void flush_buffer(const object_id & oid, cluster_buffer_t *wr);
     void on_load_config_hook(json11::Json::object & config);
     void on_load_pgs_hook(bool success);
     void on_change_hook(std::map<std::string, etcd_kv_t> & changes);
     void on_change_osd_state_hook(uint64_t peer_osd);
+    void execute_internal(cluster_op_t *op);
+    void unshift_op(cluster_op_t *op);
     int continue_rw(cluster_op_t *op);
+    bool check_rw(cluster_op_t *op);
     void slice_rw(cluster_op_t *op);
     bool try_send(cluster_op_t *op, int i);
     int continue_sync(cluster_op_t *op);
@@ -157,4 +164,6 @@ protected:
     void continue_listing(inode_list_t *lst);
     void send_list(inode_list_osd_t *cur_list);
     void continue_raw_ops(osd_num_t peer_osd);
+
+    friend class writeback_cache_t;
 };

src/cluster_client_impl.h

@@ -0,0 +1,57 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#pragma once
+
+#include "cluster_client.h"
+
+#define SCRAP_BUFFER_SIZE 4*1024*1024
+#define PART_SENT 1
+#define PART_DONE 2
+#define PART_ERROR 4
+#define PART_RETRY 8
+#define CACHE_DIRTY 1
+#define CACHE_WRITTEN 2
+#define CACHE_FLUSHING 3
+#define CACHE_REPEATING 4
+#define OP_FLUSH_BUFFER 0x02
+#define OP_IMMEDIATE_COMMIT 0x04
+
+struct cluster_buffer_t
+{
+    uint8_t *buf;
+    uint64_t len;
+    int state;
+    uint64_t flush_id;
+    uint64_t *refcnt;
+};
+
+typedef std::map<object_id, cluster_buffer_t>::iterator dirty_buf_it_t;
+
+class writeback_cache_t
+{
+public:
+    uint64_t writeback_bytes = 0;
+    int writeback_queue_size = 0;
+    int writebacks_active = 0;
+    uint64_t last_flush_id = 0;
+
+    std::map<object_id, cluster_buffer_t> dirty_buffers;
+    std::vector<cluster_op_t*> writeback_overflow;
+    std::vector<object_id> writeback_queue;
+    std::multimap<uint64_t, uint64_t*> flushed_buffers; // flush_id => refcnt
+
+    ~writeback_cache_t();
+    dirty_buf_it_t find_dirty(uint64_t inode, uint64_t offset);
+    bool is_left_merged(dirty_buf_it_t dirty_it);
+    bool is_right_merged(dirty_buf_it_t dirty_it);
+    bool is_merged(const dirty_buf_it_t & dirty_it);
+    void copy_write(cluster_op_t *op, int state);
+    int repeat_ops_for(cluster_client_t *cli, osd_num_t peer_osd);
+    void start_writebacks(cluster_client_t *cli, int count);
+    bool read_from_cache(cluster_op_t *op, uint32_t bitmap_granularity);
+    void flush_buffers(cluster_client_t *cli, dirty_buf_it_t from_it, dirty_buf_it_t to_it);
+    void fsync_start();
+    void fsync_error();
+    void fsync_ok();
+};

src/cluster_client_wb.cpp

@@ -0,0 +1,498 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <cassert>
+
+#include "cluster_client_impl.h"
+
+writeback_cache_t::~writeback_cache_t()
+{
+    for (auto & bp: dirty_buffers)
+    {
+        if (!--(*bp.second.refcnt))
+        {
+            free(bp.second.refcnt); // refcnt is allocated with the buffer
+        }
+    }
+    dirty_buffers.clear();
+}
+
+dirty_buf_it_t writeback_cache_t::find_dirty(uint64_t inode, uint64_t offset)
+{
+    auto dirty_it = dirty_buffers.lower_bound((object_id){
+        .inode = inode,
+        .stripe = offset,
+    });
+    while (dirty_it != dirty_buffers.begin())
+    {
+        dirty_it--;
+        if (dirty_it->first.inode != inode ||
+            (dirty_it->first.stripe + dirty_it->second.len) <= offset)
+        {
+            dirty_it++;
+            break;
+        }
+    }
+    return dirty_it;
+}
+
+bool writeback_cache_t::is_left_merged(dirty_buf_it_t dirty_it)
+{
+    if (dirty_it != dirty_buffers.begin())
+    {
+        auto prev_it = dirty_it;
+        prev_it--;
+        if (prev_it->first.inode == dirty_it->first.inode &&
+            prev_it->first.stripe+prev_it->second.len == dirty_it->first.stripe &&
+            prev_it->second.state == CACHE_DIRTY)
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+bool writeback_cache_t::is_right_merged(dirty_buf_it_t dirty_it)
+{
+    auto next_it = dirty_it;
+    next_it++;
+    if (next_it != dirty_buffers.end() &&
+        next_it->first.inode == dirty_it->first.inode &&
+        next_it->first.stripe == dirty_it->first.stripe+dirty_it->second.len &&
+        next_it->second.state == CACHE_DIRTY)
+    {
+        return true;
+    }
+    return false;
+}
+
+bool writeback_cache_t::is_merged(const dirty_buf_it_t & dirty_it)
+{
+    return is_left_merged(dirty_it) || is_right_merged(dirty_it);
+}
+
+void writeback_cache_t::copy_write(cluster_op_t *op, int state)
+{
+    // Save operation for replay when one of PGs goes out of sync
+    // (primary OSD drops our connection in this case)
+    // ...or just save it for writeback if write buffering is enabled
+    if (op->len == 0)
+    {
+        return;
+    }
+    auto dirty_it = find_dirty(op->inode, op->offset);
+    auto new_end = op->offset + op->len;
+    while (dirty_it != dirty_buffers.end() &&
+        dirty_it->first.inode == op->inode &&
+        dirty_it->first.stripe < op->offset+op->len)
+    {
+        assert(dirty_it->first.stripe + dirty_it->second.len > op->offset);
+        // Remove overlapping part(s) of buffers
+        auto old_end = dirty_it->first.stripe + dirty_it->second.len;
+        if (dirty_it->first.stripe < op->offset)
+        {
+            if (old_end > new_end)
+            {
+                // Split into end and start
+                dirty_it->second.len = op->offset - dirty_it->first.stripe;
+                dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+                    .inode = op->inode,
+                    .stripe = new_end,
+                }, (cluster_buffer_t){
+                    .buf = dirty_it->second.buf + new_end - dirty_it->first.stripe,
+                    .len = old_end - new_end,
+                    .state = dirty_it->second.state,
+                    .flush_id = dirty_it->second.flush_id,
+                    .refcnt = dirty_it->second.refcnt,
+                });
+                (*dirty_it->second.refcnt)++;
+                if (dirty_it->second.state == CACHE_DIRTY)
+                {
+                    writeback_bytes -= op->len;
+                    writeback_queue_size++;
+                }
+                break;
+            }
+            else
+            {
+                // Only leave the beginning
+                if (dirty_it->second.state == CACHE_DIRTY)
+                {
+                    writeback_bytes -= old_end - op->offset;
+                    if (is_left_merged(dirty_it) && !is_right_merged(dirty_it))
+                    {
+                        writeback_queue_size++;
+                    }
+                }
+                dirty_it->second.len = op->offset - dirty_it->first.stripe;
+                dirty_it++;
+            }
+        }
+        else if (old_end > new_end)
+        {
+            // Only leave the end
+            if (dirty_it->second.state == CACHE_DIRTY)
+            {
+                writeback_bytes -= new_end - dirty_it->first.stripe;
+                if (!is_left_merged(dirty_it) && is_right_merged(dirty_it))
+                {
+                    writeback_queue_size++;
+                }
+            }
+            auto new_dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+                .inode = op->inode,
+                .stripe = new_end,
+            }, (cluster_buffer_t){
+                .buf = dirty_it->second.buf + new_end - dirty_it->first.stripe,
+                .len = old_end - new_end,
+                .state = dirty_it->second.state,
+                .flush_id = dirty_it->second.flush_id,
+                .refcnt = dirty_it->second.refcnt,
+            });
+            dirty_buffers.erase(dirty_it);
+            dirty_it = new_dirty_it;
+            break;
+        }
+        else
+        {
+            // Remove the whole buffer
+            if (dirty_it->second.state == CACHE_DIRTY && !is_merged(dirty_it))
+            {
+                writeback_bytes -= dirty_it->second.len;
+                assert(writeback_queue_size > 0);
+                writeback_queue_size--;
+            }
+            if (!--(*dirty_it->second.refcnt))
+            {
+                free(dirty_it->second.refcnt);
+            }
+            dirty_buffers.erase(dirty_it++);
+        }
+    }
+    // Overlapping buffers are removed, just insert the new one
+    uint64_t *refcnt = (uint64_t*)malloc_or_die(sizeof(uint64_t) + op->len);
+    uint8_t *buf = (uint8_t*)refcnt + sizeof(uint64_t);
+    *refcnt = 1;
+    dirty_it = dirty_buffers.emplace_hint(dirty_it, (object_id){
+        .inode = op->inode,
+        .stripe = op->offset,
+    }, (cluster_buffer_t){
+        .buf = buf,
+        .len = op->len,
+        .state = state,
+        .refcnt = refcnt,
+    });
+    if (state == CACHE_DIRTY)
+    {
+        writeback_bytes += op->len;
+        // Track consecutive write-back operations
+        if (!is_merged(dirty_it))
+        {
+            // <writeback_queue> is OK to contain more than actual number of consecutive
+            // requests as long as it doesn't miss anything. But <writeback_queue_size>
+            // is always calculated correctly.
+            writeback_queue_size++;
+            writeback_queue.push_back((object_id){
+                .inode = op->inode,
+                .stripe = op->offset,
+            });
+        }
+    }
+    uint64_t pos = 0, len = op->len, iov_idx = 0;
+    while (len > 0 && iov_idx < op->iov.count)
+    {
+        auto & iov = op->iov.buf[iov_idx];
+        memcpy(buf + pos, iov.iov_base, iov.iov_len);
+        pos += iov.iov_len;
+        iov_idx++;
+    }
+}
+
+int writeback_cache_t::repeat_ops_for(cluster_client_t *cli, osd_num_t peer_osd)
+{
+    int repeated = 0;
+    if (dirty_buffers.size())
+    {
+        // peer_osd just dropped connection
+        // determine WHICH dirty_buffers are now obsolete and repeat them
+        for (auto wr_it = dirty_buffers.begin(), flush_it = wr_it, last_it = wr_it; ; )
+        {
+            bool end = wr_it == dirty_buffers.end();
+            bool flush_this = !end && wr_it->second.state != CACHE_REPEATING &&
+                cli->affects_osd(wr_it->first.inode, wr_it->first.stripe, wr_it->second.len, peer_osd);
+            if (flush_it != wr_it && (end || !flush_this ||
+                wr_it->first.inode != flush_it->first.inode ||
+                wr_it->first.stripe != last_it->first.stripe+last_it->second.len))
+            {
+                repeated++;
+                flush_buffers(cli, flush_it, wr_it);
+                flush_it = wr_it;
+            }
+            if (end)
+                break;
+            last_it = wr_it;
+            wr_it++;
+            if (!flush_this)
+                flush_it = wr_it;
+        }
+    }
+    return repeated;
+}
+
+void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from_it, dirty_buf_it_t to_it)
+{
+    auto prev_it = to_it;
+    prev_it--;
+    bool is_writeback = from_it->second.state == CACHE_DIRTY;
+    cluster_op_t *op = new cluster_op_t;
+    op->flags = OSD_OP_IGNORE_READONLY|OP_FLUSH_BUFFER;
+    op->opcode = OSD_OP_WRITE;
+    op->cur_inode = op->inode = from_it->first.inode;
+    op->offset = from_it->first.stripe;
+    op->len = prev_it->first.stripe + prev_it->second.len - from_it->first.stripe;
+    uint32_t calc_len = 0;
+    uint64_t flush_id = ++last_flush_id;
+    for (auto it = from_it; it != to_it; it++)
+    {
+        it->second.state = CACHE_REPEATING;
+        it->second.flush_id = flush_id;
+        (*it->second.refcnt)++;
+        flushed_buffers.emplace(flush_id, it->second.refcnt);
+        op->iov.push_back(it->second.buf, it->second.len);
+        calc_len += it->second.len;
+    }
+    assert(calc_len == op->len);
+    writebacks_active++;
+    op->callback = [this, cli, flush_id](cluster_op_t* op)
+    {
+        // Buffer flushes should be always retried, regardless of the error,
+        // so they should never result in an error here
+        assert(op->retval == op->len);
+        for (auto fl_it = flushed_buffers.find(flush_id);
+            fl_it != flushed_buffers.end() && fl_it->first == flush_id; )
+        {
+            if (!--(*fl_it->second)) // refcnt
+            {
+                free(fl_it->second);
+            }
+            flushed_buffers.erase(fl_it++);
+        }
+        for (auto dirty_it = find_dirty(op->inode, op->offset);
+            dirty_it != dirty_buffers.end() && dirty_it->first.inode == op->inode &&
+            dirty_it->first.stripe < op->offset+op->len; dirty_it++)
+        {
+            if (dirty_it->second.flush_id == flush_id && dirty_it->second.state == CACHE_REPEATING)
+            {
+                dirty_it->second.flush_id = 0;
+                dirty_it->second.state = CACHE_WRITTEN;
+            }
+        }
+        delete op;
+        writebacks_active--;
+        // We can't call execute_internal because it affects an invalid copy of the list here
+        // (erase_op remembers `next` after writeback callback)
+    };
+    if (is_writeback)
+    {
+        cli->execute_internal(op);
+    }
+    else
+    {
+        // Insert repeated flushes into the beginning
+        cli->unshift_op(op);
+        cli->continue_rw(op);
+    }
+}
+
+void writeback_cache_t::start_writebacks(cluster_client_t *cli, int count)
+{
+    if (!writeback_queue.size())
+    {
+        return;
+    }
+    std::vector<object_id> queue_copy;
+    queue_copy.swap(writeback_queue);
+    int started = 0, i = 0;
+    for (i = 0; i < queue_copy.size() && (!count || started < count); i++)
+    {
+        object_id & req = queue_copy[i];
+        auto dirty_it = find_dirty(req.inode, req.stripe);
+        if (dirty_it == dirty_buffers.end() ||
+            dirty_it->first.inode != req.inode ||
+            dirty_it->second.state != CACHE_DIRTY)
+        {
+            continue;
+        }
+        auto from_it = dirty_it;
+        uint64_t off = dirty_it->first.stripe;
+        while (from_it != dirty_buffers.begin())
+        {
+            from_it--;
+            if (from_it->second.state != CACHE_DIRTY ||
+                from_it->first.inode != req.inode ||
+                from_it->first.stripe+from_it->second.len != off)
+            {
+                from_it++;
+                break;
+            }
+            off = from_it->first.stripe;
+        }
+        off = dirty_it->first.stripe + dirty_it->second.len;
+        auto to_it = dirty_it;
+        to_it++;
+        while (to_it != dirty_buffers.end())
+        {
+            if (to_it->second.state != CACHE_DIRTY ||
+                to_it->first.inode != req.inode ||
+                to_it->first.stripe != off)
+            {
+                break;
+            }
+            off = to_it->first.stripe + to_it->second.len;
+            to_it++;
+        }
+        started++;
+        assert(writeback_queue_size > 0);
+        writeback_queue_size--;
+        writeback_bytes -= off - from_it->first.stripe;
+        flush_buffers(cli, from_it, to_it);
+    }
+    queue_copy.erase(queue_copy.begin(), queue_copy.begin()+i);
+    if (writeback_queue.size())
+    {
+        queue_copy.insert(queue_copy.end(), writeback_queue.begin(), writeback_queue.end());
+    }
+    queue_copy.swap(writeback_queue);
+}
+
+static void copy_to_op(cluster_op_t *op, uint64_t offset, uint8_t *buf, uint64_t len, uint32_t bitmap_granularity)
+{
+    if (op->opcode == OSD_OP_READ)
+    {
+        // Not OSD_OP_READ_BITMAP or OSD_OP_READ_CHAIN_BITMAP
+        int iov_idx = 0;
+        uint64_t cur_offset = op->offset;
+        while (iov_idx < op->iov.count && cur_offset+op->iov.buf[iov_idx].iov_len <= offset)
+        {
+            cur_offset += op->iov.buf[iov_idx].iov_len;
+            iov_idx++;
+        }
+        while (iov_idx < op->iov.count && cur_offset < offset+len)
+        {
+            auto & v = op->iov.buf[iov_idx];
+            auto begin = (cur_offset < offset ? offset : cur_offset);
+            auto end = (cur_offset+v.iov_len > offset+len ? offset+len : cur_offset+v.iov_len);
+            memcpy(
+                v.iov_base + begin - cur_offset,
+                buf + (cur_offset <= offset ? 0 : cur_offset-offset),
+                end - begin
+            );
+            cur_offset += v.iov_len;
+            iov_idx++;
+        }
+    }
+    // Set bitmap bits
+    int start_bit = (offset-op->offset)/bitmap_granularity;
+    int end_bit = (offset-op->offset+len)/bitmap_granularity;
+    for (int bit = start_bit; bit < end_bit;)
+    {
+        if (!(bit%8) && bit <= end_bit-8)
+        {
+            ((uint8_t*)op->bitmap_buf)[bit/8] = 0xFF;
+            bit += 8;
+        }
+        else
+        {
+            ((uint8_t*)op->bitmap_buf)[bit/8] |= (1 << (bit%8));
+            bit++;
+        }
+    }
+}
+
+bool writeback_cache_t::read_from_cache(cluster_op_t *op, uint32_t bitmap_granularity)
+{
+    bool dirty_copied = false;
+    if (dirty_buffers.size() && (op->opcode == OSD_OP_READ ||
+        op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP))
+    {
+        // We also have to return reads from CACHE_REPEATING buffers - they are not
+        // guaranteed to be present on target OSDs at the moment of repeating
+        // And we're also free to return data from other cached buffers just
+        // because it's faster
+        auto dirty_it = find_dirty(op->cur_inode, op->offset);
+        while (dirty_it != dirty_buffers.end() && dirty_it->first.inode == op->cur_inode &&
+            dirty_it->first.stripe < op->offset+op->len)
+        {
+            uint64_t begin = dirty_it->first.stripe, end = dirty_it->first.stripe + dirty_it->second.len;
+            if (begin < op->offset)
+                begin = op->offset;
+            if (end > op->offset+op->len)
+                end = op->offset+op->len;
+            bool skip_prev = true;
+            uint64_t cur = begin, prev = begin;
+            while (cur < end)
+            {
+                unsigned bmp_loc = (cur - op->offset)/bitmap_granularity;
+                bool skip = (((*((uint8_t*)op->bitmap_buf + bmp_loc/8)) >> (bmp_loc%8)) & 0x1);
+                if (skip_prev != skip)
+                {
+                    if (cur > prev && !skip)
+                    {
+                        // Copy data
+                        dirty_copied = true;
+                        copy_to_op(op, prev, dirty_it->second.buf + prev - dirty_it->first.stripe, cur-prev, bitmap_granularity);
+                    }
+                    skip_prev = skip;
+                    prev = cur;
+                }
+                cur += bitmap_granularity;
+            }
+            assert(cur > prev);
+            if (!skip_prev)
+            {
+                // Copy data
+                dirty_copied = true;
+                copy_to_op(op, prev, dirty_it->second.buf + prev - dirty_it->first.stripe, cur-prev, bitmap_granularity);
+            }
+            dirty_it++;
+        }
+    }
+    return dirty_copied;
+}
+
+void writeback_cache_t::fsync_start()
+{
+    for (auto & prev_op: dirty_buffers)
+    {
+        if (prev_op.second.state == CACHE_WRITTEN)
+        {
+            prev_op.second.state = CACHE_FLUSHING;
+        }
+    }
+}
+
+void writeback_cache_t::fsync_error()
+{
+    for (auto & prev_op: dirty_buffers)
+    {
+        if (prev_op.second.state == CACHE_FLUSHING)
+        {
+            prev_op.second.state = CACHE_WRITTEN;
+        }
+    }
+}
+
+void writeback_cache_t::fsync_ok()
+{
+    for (auto uw_it = dirty_buffers.begin(); uw_it != dirty_buffers.end(); )
+    {
+        if (uw_it->second.state == CACHE_FLUSHING)
+        {
+            if (!--(*uw_it->second.refcnt))
+                free(uw_it->second.refcnt);
+            dirty_buffers.erase(uw_it++);
+        }
+        else
+            uw_it++;
+    }
+}

src/disk_tool.cpp

@@ -74,7 +74,7 @@ static const char *help_text =
     "  If it doesn't succeed it issues a warning in the system log.\n"
     "  \n"
     "  You can also pass other OSD options here as arguments and they'll be persisted\n"
-    "  in the superblock: cached_io_data, cached_io_meta, cached_io_journal,\n"
+    "  in the superblock: data_io, meta_io, journal_io,\n"
     "  inmemory_metadata, inmemory_journal, max_write_iodepth,\n"
     "  min_flusher_count, max_flusher_count, journal_sector_buffer_count,\n"
     "  journal_no_same_sector_overwrites, throttle_small_writes, throttle_target_iops,\n"

src/disk_tool_prepare.cpp

@@ -8,9 +8,9 @@
 int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_hdd)
 {
     static const char *allow_additional_params[] = {
-        "cached_io_data",
-        "cached_io_meta",
-        "cached_io_journal",
+        "data_io",
+        "meta_io",
+        "journal_io",
         "max_write_iodepth",
         "max_write_iodepth",
         "min_flusher_count",
@@ -119,7 +119,7 @@ int disk_tool_t::prepare_one(std::map<std::string, std::string> options, int is_
     try
     {
         dsk.parse_config(options);
-        dsk.cached_io_data = dsk.cached_io_meta = dsk.cached_io_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();
@@ -483,7 +483,7 @@ int disk_tool_t::get_meta_partition(std::vector<vitastor_dev_info_t> & ssds, std
     {
         blockstore_disk_t dsk;
         dsk.parse_config(options);
-        dsk.cached_io_data = dsk.cached_io_meta = dsk.cached_io_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();

src/disk_tool_resize.cpp

@@ -91,7 +91,7 @@ int disk_tool_t::resize_parse_params()
     try
     {
         dsk.parse_config(options);
-        dsk.cached_io_data = dsk.cached_io_meta = dsk.cached_io_journal = false;
+        dsk.data_io = dsk.meta_io = dsk.journal_io = "direct";
         dsk.open_data();
         dsk.open_meta();
         dsk.open_journal();

src/epoll_manager.cpp

@@ -23,19 +23,24 @@ epoll_manager_t::epoll_manager_t(ring_loop_t *ringloop)
 
     tfd = new timerfd_manager_t([this](int fd, bool wr, std::function<void(int, int)> handler) { set_fd_handler(fd, wr, handler); });
 
-    consumer.loop = [this]()
+    if (ringloop)
     {
-        if (pending)
-            handle_epoll_events();
-    };
-    ringloop->register_consumer(&consumer);
-
-    handle_epoll_events();
+        consumer.loop = [this]()
+        {
+            if (pending)
+                handle_uring_event();
+        };
+        ringloop->register_consumer(&consumer);
+        handle_uring_event();
+    }
 }
 
 epoll_manager_t::~epoll_manager_t()
 {
-    ringloop->unregister_consumer(&consumer);
+    if (ringloop)
+    {
+        ringloop->unregister_consumer(&consumer);
+    }
     if (tfd)
     {
         delete tfd;
@@ -44,6 +49,11 @@ epoll_manager_t::~epoll_manager_t()
     close(epoll_fd);
 }
 
+int epoll_manager_t::get_fd()
+{
+    return epoll_fd;
+}
+
 void epoll_manager_t::set_fd_handler(int fd, bool wr, std::function<void(int, int)> handler)
 {
     if (handler != NULL)
@@ -75,7 +85,7 @@ void epoll_manager_t::set_fd_handler(int fd, bool wr, std::function<void(int, in
     }
 }
 
-void epoll_manager_t::handle_epoll_events()
+void epoll_manager_t::handle_uring_event()
 {
     io_uring_sqe *sqe = ringloop->get_sqe();
     if (!sqe)
@@ -95,14 +105,20 @@ void epoll_manager_t::handle_epoll_events()
         {
             throw std::runtime_error(std::string("epoll failed: ") + strerror(-data->res));
         }
-        handle_epoll_events();
+        handle_uring_event();
     };
     ringloop->submit();
+    handle_events(0);
+}
+
+void epoll_manager_t::handle_events(int timeout)
+{
     int nfds;
     epoll_event events[MAX_EPOLL_EVENTS];
     do
     {
-        nfds = epoll_wait(epoll_fd, events, MAX_EPOLL_EVENTS, 0);
+        nfds = epoll_wait(epoll_fd, events, MAX_EPOLL_EVENTS, timeout);
+        timeout = 0;
         for (int i = 0; i < nfds; i++)
         {
             auto cb_it = epoll_handlers.find(events[i].data.fd);

src/epoll_manager.h

@@ -15,11 +15,14 @@ class epoll_manager_t
     ring_consumer_t consumer;
     ring_loop_t *ringloop;
     std::map<int, std::function<void(int, int)>> epoll_handlers;
+
+    void handle_uring_event();
 public:
     epoll_manager_t(ring_loop_t *ringloop);
     ~epoll_manager_t();
+    int get_fd();
     void set_fd_handler(int fd, bool wr, std::function<void(int, int)> handler);
-    void handle_epoll_events();
+    void handle_events(int timeout);
 
     timerfd_manager_t *tfd;
 };

src/etcd_state_client.cpp

@@ -684,8 +684,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             // ID
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Pool ID %s is invalid (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
@@ -829,8 +829,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         {
             pool_id_t pool_id;
             char null_byte = 0;
-            sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
-            if (!pool_id || pool_id >= POOL_ID_MAX || null_byte != 0)
+            int scanned = sscanf(pool_item.first.c_str(), "%u%c", &pool_id, &null_byte);
+            if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
             {
                 fprintf(stderr, "Pool ID %s is invalid in PG configuration (must be a number less than 0x%x), skipping pool\n", pool_item.first.c_str(), POOL_ID_MAX);
                 continue;
@@ -838,8 +838,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             for (auto & pg_item: pool_item.second.object_items())
             {
                 pg_num_t pg_num = 0;
-                sscanf(pg_item.first.c_str(), "%u%c", &pg_num, &null_byte);
-                if (!pg_num || null_byte != 0)
+                int scanned = sscanf(pg_item.first.c_str(), "%u%c", &pg_num, &null_byte);
+                if (scanned != 1 || !pg_num)
                 {
                     fprintf(stderr, "Bad key in pool %u PG configuration: %s (must be a number), skipped\n", pool_id, pg_item.first.c_str());
                     continue;
@@ -889,8 +889,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         pg_num_t pg_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+12, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !pg_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+12, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !pg_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -944,8 +944,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         pg_num_t pg_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !pg_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !pg_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -1015,8 +1015,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
         uint64_t pool_id = 0;
         uint64_t inode_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + etcd_prefix.length()+14, "%lu/%lu%c", &pool_id, &inode_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !inode_num || (inode_num >> (64-POOL_ID_BITS)) || null_byte != 0)
+        int scanned = sscanf(key.c_str() + etcd_prefix.length()+14, "%lu/%lu%c", &pool_id, &inode_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !inode_num || (inode_num >> (64-POOL_ID_BITS)))
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }

src/fio_cluster.cpp

@@ -24,6 +24,7 @@
 #include <netinet/tcp.h>
 
 #include <vector>
+#include <string>
 
 #include "vitastor_c.h"
 #include "fio_headers.h"
@@ -31,6 +32,7 @@
 struct sec_data
 {
     vitastor_c *cli = NULL;
+    bool epoll_based = false;
     void *watch = NULL;
     bool last_sync = false;
     /* The list of completed io_u structs. */
@@ -57,6 +59,7 @@ struct sec_options
     int rdma_port_num = 0;
     int rdma_gid_index = 0;
     int rdma_mtu = 0;
+    int no_io_uring = 0;
 };
 
 static struct fio_option options[] = {
@@ -192,6 +195,16 @@ static struct fio_option options[] = {
         .category = FIO_OPT_C_ENGINE,
         .group  = FIO_OPT_G_FILENAME,
     },
+    {
+        .name   = "no_io_uring",
+        .lname  = "Disable io_uring",
+        .type   = FIO_OPT_BOOL,
+        .off1   = offsetof(struct sec_options, no_io_uring),
+        .help   = "Use epoll and plain sendmsg/recvmsg instead of io_uring (slower)",
+        .def    = "0",
+        .category = FIO_OPT_C_ENGINE,
+        .group  = FIO_OPT_G_FILENAME,
+    },
     {
         .name = NULL,
     },
@@ -203,6 +216,15 @@ static void watch_callback(void *opaque, long watch)
     bsd->watch = (void*)watch;
 }
 
+static void opt_push(std::vector<char *> & options, const char *opt, const char *value)
+{
+    if (value)
+    {
+        options.push_back(strdup(opt));
+        options.push_back(strdup(value));
+    }
+}
+
 static int sec_setup(struct thread_data *td)
 {
     sec_options *o = (sec_options*)td->eo;
@@ -254,18 +276,59 @@ static int sec_setup(struct thread_data *td)
     {
         o->inode = 0;
     }
-    bsd->cli = vitastor_c_create_uring(o->config_path, o->etcd_host, o->etcd_prefix,
-        o->use_rdma, o->rdma_device, o->rdma_port_num, o->rdma_gid_index, o->rdma_mtu, o->cluster_log);
+    std::vector<char *> options;
+    opt_push(options, "config_path", o->config_path);
+    opt_push(options, "etcd_address", o->etcd_host);
+    opt_push(options, "etcd_prefix", o->etcd_prefix);
+    if (o->use_rdma != -1)
+        opt_push(options, "use_rdma", std::to_string(o->use_rdma).c_str());
+    opt_push(options, "rdma_device", o->rdma_device);
+    if (o->rdma_port_num)
+        opt_push(options, "rdma_port_num", std::to_string(o->rdma_port_num).c_str());
+    if (o->rdma_gid_index)
+        opt_push(options, "rdma_gid_index", std::to_string(o->rdma_gid_index).c_str());
+    if (o->rdma_mtu)
+        opt_push(options, "rdma_mtu", std::to_string(o->rdma_mtu).c_str());
+    if (o->cluster_log)
+        opt_push(options, "log_level", std::to_string(o->cluster_log).c_str());
+    // allow writeback caching if -direct is not set
+    opt_push(options, "client_writeback_allowed", td->o.odirect ? "0" : "1");
+    bsd->cli = o->no_io_uring ? NULL : vitastor_c_create_uring_json((const char**)options.data(), options.size());
+    bsd->epoll_based = false;
+    if (!bsd->cli)
+    {
+        if (o->no_io_uring)
+            fprintf(stderr, "vitastor: io_uring disabled - I/O will be slower\n");
+        else
+            fprintf(stderr, "vitastor: failed to create io_uring: %s - I/O will be slower\n", strerror(errno));
+        bsd->cli = vitastor_c_create_epoll_json((const char**)options.data(), options.size());
+        bsd->epoll_based = true;
+    }
+    for (auto opt: options)
+        free(opt);
+    options.clear();
     if (o->image)
     {
         bsd->watch = NULL;
         vitastor_c_watch_inode(bsd->cli, o->image, watch_callback, bsd);
-        while (true)
+        if (!bsd->epoll_based)
         {
-            vitastor_c_uring_handle_events(bsd->cli);
-            if (bsd->watch)
-                break;
-            vitastor_c_uring_wait_events(bsd->cli);
+            while (true)
+            {
+                vitastor_c_uring_handle_events(bsd->cli);
+                if (bsd->watch)
+                    break;
+                vitastor_c_uring_wait_events(bsd->cli);
+            }
+        }
+        else
+        {
+            while (true)
+            {
+                if (bsd->watch)
+                    break;
+                vitastor_c_epoll_handle_events(bsd->cli, 1000);
+            }
         }
         td->files[0]->real_file_size = vitastor_c_inode_get_size(bsd->watch);
         if (!vitastor_c_inode_get_num(bsd->watch) ||
@@ -408,12 +471,24 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
 static int sec_getevents(struct thread_data *td, unsigned int min, unsigned int max, const struct timespec *t)
 {
     sec_data *bsd = (sec_data*)td->io_ops_data;
-    while (true)
+    if (!bsd->epoll_based)
     {
-        vitastor_c_uring_handle_events(bsd->cli);
-        if (bsd->completed.size() >= min)
-            break;
-        vitastor_c_uring_wait_events(bsd->cli);
+        while (true)
+        {
+            vitastor_c_uring_handle_events(bsd->cli);
+            if (bsd->completed.size() >= min)
+                break;
+            vitastor_c_uring_wait_events(bsd->cli);
+        }
+    }
+    else
+    {
+        while (true)
+        {
+            if (bsd->completed.size() >= min)
+                break;
+            vitastor_c_epoll_handle_events(bsd->cli, 1000);
+        }
     }
     return bsd->completed.size();
 }

src/fio_sec_osd.cpp

@@ -242,6 +242,7 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
             op.sec_rw.version = UINT64_MAX; // last unstable
             op.sec_rw.offset = io->offset % bsd->block_size;
             op.sec_rw.len = io->xfer_buflen;
+            op.sec_rw.attr_len = 0;
         }
         else
         {
@@ -263,6 +264,7 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
             op.sec_rw.version = 0; // assign automatically
             op.sec_rw.offset = io->offset % bsd->block_size;
             op.sec_rw.len = io->xfer_buflen;
+            op.sec_rw.attr_len = 0;
         }
         else
         {

src/messenger.cpp

@@ -11,6 +11,9 @@
 
 #include "addr_util.h"
 #include "messenger.h"
+#ifdef WITH_RDMA
+#include "msgr_rdma.h"
+#endif
 
 void osd_messenger_t::init()
 {
@@ -39,6 +42,12 @@ void osd_messenger_t::init()
             handle_rdma_events();
         }
     }
+#endif
+#ifndef SO_ZEROCOPY
+    if (log_level > 0)
+    {
+        fprintf(stderr, "Zero-copy TCP send is not supported in this build, ignoring\n");
+    }
 #endif
     keepalive_timer_id = tfd->set_timer(1000, true, [this](int)
     {
@@ -170,6 +179,8 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         this->receive_buffer_size = 65536;
     this->use_sync_send_recv = config["use_sync_send_recv"].bool_value() ||
         config["use_sync_send_recv"].uint64_value();
+    this->use_zerocopy_send = config["use_zerocopy_send"].bool_value() ||
+        config["use_zerocopy_send"].uint64_value();
     this->peer_connect_interval = config["peer_connect_interval"].uint64_value();
     if (!this->peer_connect_interval)
         this->peer_connect_interval = 5;
@@ -300,8 +311,7 @@ void osd_messenger_t::handle_connect_epoll(int peer_fd)
         on_connect_peer(peer_osd, -result);
         return;
     }
-    int one = 1;
-    setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
+    set_socket_options(cl);
     cl->peer_state = PEER_CONNECTED;
     tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
     {
@@ -311,6 +321,23 @@ void osd_messenger_t::handle_connect_epoll(int peer_fd)
     check_peer_config(cl);
 }
 
+void osd_messenger_t::set_socket_options(osd_client_t *cl)
+{
+    int one = 1;
+    setsockopt(cl->peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
+#ifdef SO_ZEROCOPY
+    if (!use_zerocopy_send)
+        cl->zerocopy_send = false;
+    else if (setsockopt(cl->peer_fd, SOL_SOCKET, SO_ZEROCOPY, &one, sizeof(one)) != 0)
+    {
+        if (log_level > 0)
+            fprintf(stderr, "[OSD %lu] Failed to enable zero-copy send for client %d: %s\n", this->osd_num, cl->peer_fd, strerror(errno));
+    }
+    else
+        cl->zerocopy_send = true;
+#endif
+}
+
 void osd_messenger_t::handle_peer_epoll(int peer_fd, int epoll_events)
 {
     // Mark client as ready (i.e. some data is available)
@@ -512,14 +539,13 @@ void osd_messenger_t::accept_connections(int listen_fd)
         fprintf(stderr, "[OSD %lu] new client %d: connection from %s\n", this->osd_num, peer_fd,
             addr_to_string(addr).c_str());
         fcntl(peer_fd, F_SETFL, fcntl(peer_fd, F_GETFL, 0) | O_NONBLOCK);
-        int one = 1;
-        setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-        clients[peer_fd] = new osd_client_t();
-        clients[peer_fd]->peer_addr = addr;
-        clients[peer_fd]->peer_port = ntohs(((sockaddr_in*)&addr)->sin_port);
-        clients[peer_fd]->peer_fd = peer_fd;
-        clients[peer_fd]->peer_state = PEER_CONNECTED;
-        clients[peer_fd]->in_buf = malloc_or_die(receive_buffer_size);
+        auto cl = clients[peer_fd] = new osd_client_t();
+        cl->peer_addr = addr;
+        cl->peer_port = ntohs(((sockaddr_in*)&addr)->sin_port);
+        cl->peer_fd = peer_fd;
+        cl->peer_state = PEER_CONNECTED;
+        cl->in_buf = malloc_or_die(receive_buffer_size);
+        set_socket_options(cl);
         // Add FD to epoll
         tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
         {

src/messenger.h

@@ -18,10 +18,6 @@
 #include "timerfd_manager.h"
 #include <ringloop.h>
 
-#ifdef WITH_RDMA
-#include "msgr_rdma.h"
-#endif
-
 #define CL_READ_HDR 1
 #define CL_READ_DATA 2
 #define CL_READ_REPLY_DATA 3
@@ -44,6 +40,17 @@ struct msgr_sendp_t
     int flags;
 };
 
+#ifdef WITH_RDMA
+struct msgr_rdma_connection_t;
+struct msgr_rdma_context_t;
+#endif
+
+struct msgr_zc_not_t
+{
+    osd_op_t *op;
+    uint32_t nsend;
+};
+
 struct osd_client_t
 {
     int refs = 0;
@@ -56,6 +63,7 @@ struct osd_client_t
     int ping_time_remaining = 0;
     int idle_time_remaining = 0;
     osd_num_t osd_num = 0;
+    bool zerocopy_send = false;
 
     void *in_buf = NULL;
 
@@ -86,6 +94,12 @@ struct osd_client_t
     int write_state = 0;
     std::vector<iovec> send_list, next_send_list;
     std::vector<msgr_sendp_t> outbox, next_outbox;
+    std::vector<msgr_zc_not_t> zerocopy_sent;
+    uint64_t outbox_size = 0, next_outbox_size = 0;
+    uint32_t zerocopy_notification_idx = 0;
+    uint32_t zerocopy_notification_prev = 0;
+    uint8_t zerocopy_notification_buf[256];
+    struct msghdr zerocopy_notification_msg;
 
     ~osd_client_t();
 };
@@ -122,6 +136,7 @@ protected:
     int osd_ping_timeout = 0;
     int log_level = 0;
     bool use_sync_send_recv = false;
+    bool use_zerocopy_send = false;
 
 #ifdef WITH_RDMA
     bool use_rdma = true;
@@ -184,9 +199,11 @@ protected:
     void check_peer_config(osd_client_t *cl);
     void cancel_osd_ops(osd_client_t *cl);
     void cancel_op(osd_op_t *op);
+    void set_socket_options(osd_client_t *cl);
 
     bool try_send(osd_client_t *cl);
     void handle_send(int result, osd_client_t *cl);
+    void handle_zerocopy_notification(osd_client_t *cl, int res);
 
     bool handle_read(int result, osd_client_t *cl);
     bool handle_read_buffer(osd_client_t *cl, void *curbuf, int remain);

src/mock/messenger.cpp

@@ -55,3 +55,10 @@ json11::Json::object osd_messenger_t::merge_configs(const json11::Json::object &
 {
     return cli_config;
 }
+
+bool json_is_true(const json11::Json & val)
+{
+    if (val.is_string())
+        return val == "true" || val == "yes" || val == "1";
+    return val.bool_value();
+}

src/mock/ringloop.h

@@ -22,4 +22,10 @@ public:
     void submit()
     {
     }
+    void wait()
+    {
+    }
+    void loop()
+    {
+    }
 };

src/msgr_rdma.cpp

@@ -19,12 +19,12 @@ std::string msgr_rdma_address_t::to_string()
 bool msgr_rdma_address_t::from_string(const char *str, msgr_rdma_address_t *dest)
 {
     uint64_t* gid = (uint64_t*)&dest->gid;
-    int n = sscanf(
+    int scanned = sscanf(
         str, "%hx:%x:%x:%16lx%16lx", &dest->lid, &dest->qpn, &dest->psn, gid, gid+1
     );
     gid[0] = be64toh(gid[0]);
     gid[1] = be64toh(gid[1]);
-    return n == 5;
+    return scanned == 5;
 }
 
 msgr_rdma_context_t::~msgr_rdma_context_t()

src/msgr_send.cpp

@@ -6,6 +6,12 @@
 
 #include "messenger.h"
 
+#include <linux/errqueue.h>
+
+#ifndef MSG_ZEROCOPY
+#define MSG_ZEROCOPY 0
+#endif
+
 void osd_messenger_t::outbox_push(osd_op_t *cur_op)
 {
     assert(cur_op->peer_fd);
@@ -36,6 +42,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
     }
     auto & to_send_list = cl->write_msg.msg_iovlen ? cl->next_send_list : cl->send_list;
     auto & to_outbox = cl->write_msg.msg_iovlen ? cl->next_outbox : cl->outbox;
+    auto & to_size = cl->write_msg.msg_iovlen ? cl->next_outbox_size : cl->outbox_size;
     if (cur_op->op_type == OSD_OP_IN)
     {
         measure_exec(cur_op);
@@ -46,6 +53,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
         to_send_list.push_back((iovec){ .iov_base = cur_op->req.buf, .iov_len = OSD_PACKET_SIZE });
         cl->sent_ops[cur_op->req.hdr.id] = cur_op;
     }
+    to_size += OSD_PACKET_SIZE;
     to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = MSGR_SENDP_HDR });
     // Bitmap
     if (cur_op->op_type == OSD_OP_IN &&
@@ -57,6 +65,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             .iov_len = cur_op->reply.sec_rw.attr_len,
         });
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        to_size += cur_op->reply.sec_rw.attr_len;
     }
     else if (cur_op->op_type == OSD_OP_OUT &&
         (cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE || cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE) &&
@@ -67,6 +76,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             .iov_len = cur_op->req.sec_rw.attr_len,
         });
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        to_size += cur_op->req.sec_rw.attr_len;
     }
     // Operation data
     if ((cur_op->op_type == OSD_OP_IN
@@ -89,15 +99,22 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
                 assert(cur_op->iov.buf[i].iov_base);
                 to_send_list.push_back(cur_op->iov.buf[i]);
                 to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+                to_size += cur_op->iov.buf[i].iov_len;
             }
         }
     }
     if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
     {
         if (cur_op->op_type == OSD_OP_IN && cur_op->reply.hdr.retval > 0)
+        {
             to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->reply.hdr.retval });
+            to_size += cur_op->reply.hdr.retval;
+        }
         else if (cur_op->op_type == OSD_OP_OUT && cur_op->req.sec_read_bmp.len > 0)
+        {
             to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len });
+            to_size += cur_op->req.sec_read_bmp.len;
+        }
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
     }
     if (cur_op->op_type == OSD_OP_IN)
@@ -119,9 +136,9 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             try_send(cl);
         }
     }
-    else if (cl->write_msg.msg_iovlen > 0 || !try_send(cl))
+    else
     {
-        if (cl->write_state == 0)
+        if ((cl->write_msg.msg_iovlen > 0 || !try_send(cl)) && (cl->write_state == 0))
         {
             cl->write_state = CL_WRITE_READY;
             write_ready_clients.push_back(cur_op->peer_fd);
@@ -183,17 +200,19 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
         }
         cl->write_msg.msg_iov = cl->send_list.data();
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+        cl->write_msg.msg_flags = (cl->zerocopy_send && (cl->outbox_size/cl->send_list.size()) >= 4096 ? MSG_ZEROCOPY : 0);
         cl->refs++;
         ring_data_t* data = ((ring_data_t*)sqe->user_data);
         data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, cl); };
-        my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, 0);
+        my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, cl->write_msg.msg_flags);
     }
     else
     {
         cl->write_msg.msg_iov = cl->send_list.data();
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+        cl->write_msg.msg_flags = (cl->zerocopy_send && (cl->outbox_size/cl->send_list.size()) >= 4096 ? MSG_ZEROCOPY : 0);
         cl->refs++;
-        int result = sendmsg(peer_fd, &cl->write_msg, MSG_NOSIGNAL);
+        int result = sendmsg(peer_fd, &cl->write_msg, MSG_NOSIGNAL | cl->write_msg.msg_flags);
         if (result < 0)
         {
             result = -errno;
@@ -203,6 +222,62 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
     return true;
 }
 
+void osd_messenger_t::handle_zerocopy_notification(osd_client_t *cl, int res)
+{
+    cl->refs--;
+    if (cl->peer_state == PEER_STOPPED)
+    {
+        if (cl->refs <= 0)
+        {
+            delete cl;
+        }
+        return;
+    }
+    if (res != 0)
+    {
+        return;
+    }
+    if (cl->zerocopy_notification_msg.msg_flags & MSG_CTRUNC)
+    {
+        fprintf(stderr, "zero-copy send notification truncated on client socket %d\n", cl->peer_fd);
+        return;
+    }
+    for (struct cmsghdr *cm = CMSG_FIRSTHDR(&cl->zerocopy_notification_msg); cm; cm = CMSG_NXTHDR(&cl->zerocopy_notification_msg, cm))
+    {
+        if (cm->cmsg_level == SOL_IP && cm->cmsg_type == IP_RECVERR)
+        {
+            struct sock_extended_err *serr = (struct sock_extended_err*)CMSG_DATA(cm);
+            if (serr->ee_errno == 0 && serr->ee_origin == SO_EE_ORIGIN_ZEROCOPY)
+            {
+                // completed sends numbered serr->ee_info .. serr->ee_data
+                int start = 0;
+                while (start < cl->zerocopy_sent.size() && cl->zerocopy_sent[start].nsend < serr->ee_info)
+                    start++;
+                int end = start;
+                if (serr->ee_data < serr->ee_info)
+                {
+                    // counter has wrapped around
+                    while (end < cl->zerocopy_sent.size() && cl->zerocopy_sent[end].nsend >= cl->zerocopy_sent[start].nsend)
+                        end++;
+                }
+                while (end < cl->zerocopy_sent.size() && cl->zerocopy_sent[end].nsend <= serr->ee_data)
+                    end++;
+                if (end > start)
+                {
+                    for (int i = start; i < end; i++)
+                    {
+                        delete cl->zerocopy_sent[i].op;
+                    }
+                    cl->zerocopy_sent.erase(
+                        cl->zerocopy_sent.begin() + start,
+                        cl->zerocopy_sent.begin() + end
+                    );
+                }
+            }
+        }
+    }
+}
+
 void osd_messenger_t::send_replies()
 {
     for (int i = 0; i < write_ready_clients.size(); i++)
@@ -230,16 +305,19 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
         }
         return;
     }
-    if (result < 0 && result != -EAGAIN && result != -EINTR)
+    if (result < 0 && result != -EAGAIN && result != -EINTR && result != -ENOBUFS)
     {
         // this is a client socket, so don't panic. just disconnect it
         fprintf(stderr, "Client %d socket write error: %d (%s). Disconnecting client\n", cl->peer_fd, -result, strerror(-result));
         stop_client(cl->peer_fd);
         return;
     }
+    bool used_zerocopy = false;
     if (result >= 0)
     {
+        used_zerocopy = (cl->write_msg.msg_flags & MSG_ZEROCOPY) ? true : false;
         int done = 0;
+        int bytes_written = result;
         while (result > 0 && done < cl->send_list.size())
         {
             iovec & iov = cl->send_list[done];
@@ -248,7 +326,19 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 if (cl->outbox[done].flags & MSGR_SENDP_FREE)
                 {
                     // Reply fully sent
-                    delete cl->outbox[done].op;
+                    if (!used_zerocopy)
+                    {
+                        delete cl->outbox[done].op;
+                    }
+                    else
+                    {
+                        // With zero-copy send the difference is that we must keep the buffer (i.e. the operation)
+                        // allocated until we get send notification from MSG_ERRQUEUE
+                        cl->zerocopy_sent.push_back((msgr_zc_not_t){
+                            .op = cl->outbox[done].op,
+                            .nsend = cl->zerocopy_notification_idx,
+                        });
+                    }
                 }
                 result -= iov.iov_len;
                 done++;
@@ -260,6 +350,11 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 break;
             }
         }
+        if (used_zerocopy)
+        {
+            cl->zerocopy_notification_idx++;
+        }
+        cl->outbox_size -= bytes_written;
         if (done > 0)
         {
             cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+done);
@@ -269,8 +364,10 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
         {
             cl->send_list.insert(cl->send_list.end(), cl->next_send_list.begin(), cl->next_send_list.end());
             cl->outbox.insert(cl->outbox.end(), cl->next_outbox.begin(), cl->next_outbox.end());
+            cl->outbox_size += cl->next_outbox_size;
             cl->next_send_list.clear();
             cl->next_outbox.clear();
+            cl->next_outbox_size = 0;
         }
         cl->write_state = cl->outbox.size() > 0 ? CL_WRITE_READY : 0;
 #ifdef WITH_RDMA
@@ -293,4 +390,34 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
     {
         write_ready_clients.push_back(cl->peer_fd);
     }
+    if (used_zerocopy && (cl->zerocopy_notification_idx-cl->zerocopy_notification_prev) >= 16 &&
+        cl->zerocopy_sent.size() > 0)
+    {
+        cl->zerocopy_notification_prev = cl->zerocopy_notification_idx;
+        cl->zerocopy_notification_msg = {
+            .msg_control = cl->zerocopy_notification_buf,
+            .msg_controllen = sizeof(cl->zerocopy_notification_buf),
+        };
+        cl->refs++;
+        io_uring_sqe* sqe = NULL;
+        if (ringloop && !use_sync_send_recv)
+        {
+            sqe = ringloop->get_sqe();
+        }
+        if (!sqe)
+        {
+            int res = recvmsg(cl->peer_fd, &cl->zerocopy_notification_msg, MSG_ERRQUEUE|MSG_DONTWAIT);
+            if (res < 0)
+            {
+                res = -errno;
+            }
+            handle_zerocopy_notification(cl, res);
+        }
+        else
+        {
+            ring_data_t* data = ((ring_data_t*)sqe->user_data);
+            data->callback = [this, cl](ring_data_t *data) { handle_zerocopy_notification(cl, data->res); };
+            my_uring_prep_recvmsg(sqe, cl->peer_fd, &cl->zerocopy_notification_msg, MSG_ERRQUEUE);
+        }
+    }
 }

src/msgr_stop.cpp

@@ -5,6 +5,9 @@
 #include <assert.h>
 
 #include "messenger.h"
+#ifdef WITH_RDMA
+#include "msgr_rdma.h"
+#endif
 
 void osd_messenger_t::cancel_osd_ops(osd_client_t *cl)
 {

src/nbd_proxy.cpp

@@ -216,6 +216,14 @@ public:
         {
             nbd_timeout = cfg["nbd_timeout"].uint64_value();
         }
+        if (cfg["client_writeback_allowed"].is_null())
+        {
+            // NBD is always aware of fsync, so we allow write-back cache
+            // by default if it's enabled
+            auto obj = cfg.object_items();
+            obj["client_writeback_allowed"] = true;
+            cfg = obj;
+        }
         // Create client
         ringloop = new ring_loop_t(512);
         epmgr = new epoll_manager_t(ringloop);
@@ -341,6 +349,7 @@ public:
             ringloop->loop();
             ringloop->wait();
         }
+        cli->flush();
         delete cli;
         delete epmgr;
         delete ringloop;

src/nfs_proxy.cpp

@@ -34,7 +34,10 @@ nfs_proxy_t::~nfs_proxy_t()
     if (cmd)
         delete cmd;
     if (cli)
+    {
+        cli->flush();
         delete cli;
+    }
     if (epmgr)
         delete epmgr;
     if (ringloop)
@@ -62,8 +65,9 @@ json11::Json::object nfs_proxy_t::parse_args(int narg, const char *args[])
                 "  --pool <POOL>     use <POOL> as default pool for new files (images)\n"
                 "  --foreground 1    stay in foreground, do not daemonize\n"
                 "\n"
-                "NFS proxy is stateless if you use immediate_commit=all in your cluster, so\n"
-                "you can freely use multiple NFS proxies with L3 load balancing in this case.\n"
+                "NFS proxy is stateless if you use immediate_commit=all in your cluster and if\n"
+                "you do not use client_enable_writeback=true, so you can freely use multiple\n"
+                "NFS proxies with L3 load balancing in this case.\n"
                 "\n"
                 "Example start and mount commands for a custom NFS port:\n"
                 "  %s --etcd_address 192.168.5.10:2379 --portmap 0 --port 2050 --pool testpool\n"
@@ -111,6 +115,14 @@ void nfs_proxy_t::run(json11::Json cfg)
         if (name_prefix.size())
             name_prefix += "/";
     }
+    if (cfg["client_writeback_allowed"].is_null())
+    {
+        // NFS is always aware of fsync, so we allow write-back cache
+        // by default if it's enabled
+        auto obj = cfg.object_items();
+        obj["client_writeback_allowed"] = true;
+        cfg = obj;
+    }
     // Create client
     ringloop = new ring_loop_t(512);
     epmgr = new epoll_manager_t(ringloop);
@@ -261,16 +273,8 @@ void nfs_proxy_t::run(json11::Json cfg)
         ringloop->loop();
         ringloop->wait();
     }
-    /*// Sync at the end
-    cluster_op_t *close_sync = new cluster_op_t;
-    close_sync->opcode = OSD_OP_SYNC;
-    close_sync->callback = [&stop](cluster_op_t *op)
-    {
-        stop = true;
-        delete op;
-    };
-    cli->execute(close_sync);*/
     // Destroy the client
+    cli->flush();
     delete cli;
     delete epmgr;
     delete ringloop;
@@ -346,8 +350,8 @@ void nfs_proxy_t::parse_stats(etcd_kv_t & kv)
         pool_id_t pool_id = 0;
         inode_t inode_num = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode_num, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX || !inode_num || null_byte != 0)
+        int scanned = sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+13, "%u/%lu%c", &pool_id, &inode_num, &null_byte);
+        if (scanned != 2 || !pool_id || pool_id >= POOL_ID_MAX || !inode_num)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }
@@ -360,8 +364,8 @@ void nfs_proxy_t::parse_stats(etcd_kv_t & kv)
     {
         pool_id_t pool_id = 0;
         char null_byte = 0;
-        sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+12, "%u%c", &pool_id, &null_byte);
-        if (!pool_id || pool_id >= POOL_ID_MAX)
+        int scanned = sscanf(key.c_str() + cli->st_cli.etcd_prefix.length()+12, "%u%c", &pool_id, &null_byte);
+        if (scanned != 1 || !pool_id || pool_id >= POOL_ID_MAX)
         {
             fprintf(stderr, "Bad etcd key %s, ignoring\n", key.c_str());
         }

src/osd.cpp

@@ -160,6 +160,9 @@ void osd_t::parse_config(bool init)
         etcd_report_interval = config["etcd_report_interval"].uint64_value();
         if (etcd_report_interval <= 0)
             etcd_report_interval = 5;
+        etcd_stats_interval = config["etcd_stats_interval"].uint64_value();
+        if (etcd_stats_interval <= 0)
+            etcd_stats_interval = 30;
         readonly = json_is_true(config["readonly"]);
         run_primary = !json_is_false(config["run_primary"]);
         allow_test_ops = json_is_true(config["allow_test_ops"]);

src/osd.h

@@ -93,6 +93,7 @@ class osd_t
 
     json11::Json::object cli_config, file_config, etcd_global_config, etcd_osd_config, config;
     int etcd_report_interval = 5;
+    int etcd_stats_interval = 30;
 
     bool readonly = false;
     osd_num_t osd_num = 1; // OSD numbers start with 1

src/osd_cluster.cpp

@@ -186,10 +186,12 @@ json11::Json osd_t::get_statistics()
     if (bs)
     {
         st["blockstore_ready"] = bs->is_started();
-        st["data_block_size"] = (uint64_t)bs->get_block_size();
         st["size"] = bs->get_block_count() * bs->get_block_size();
         st["free"] = bs->get_free_block_count() * bs->get_block_size();
     }
+    st["data_block_size"] = (uint64_t)bs_block_size;
+    st["bitmap_granularity"] = (uint64_t)bs_bitmap_granularity;
+    st["immediate_commit"] = immediate_commit == IMMEDIATE_ALL ? "all" : (immediate_commit == IMMEDIATE_SMALL ? "small" : "none");
     st["host"] = self_state["host"];
     json11::Json::object op_stats, subop_stats;
     for (int i = OSD_OP_MIN; i <= OSD_OP_MAX; i++)
@@ -427,14 +429,18 @@ void osd_t::acquire_lease()
         create_osd_state();
     });
     printf(
-        "[OSD %lu] reporting to etcd at %s every %d seconds\n", this->osd_num,
+        "[OSD %lu] reporting to etcd at %s every %d seconds (statistics every %d seconds)\n", this->osd_num,
         (config["etcd_address"].is_string() ? config["etcd_address"].string_value() : config["etcd_address"].dump()).c_str(),
-        etcd_report_interval
+        etcd_report_interval, etcd_stats_interval
     );
     tfd->set_timer(etcd_report_interval*1000, true, [this](int timer_id)
     {
         renew_lease(false);
     });
+    tfd->set_timer(etcd_stats_interval*1000, true, [this](int timer_id)
+    {
+        report_statistics();
+    });
 }
 
 // Report "up" state once, then keep it alive using the lease
@@ -539,7 +545,6 @@ void osd_t::renew_lease(bool reload)
         else
         {
             etcd_failed_attempts = 0;
-            report_statistics();
             // Reload PGs
             if (reload && run_primary)
             {
@@ -649,7 +654,7 @@ void osd_t::apply_pg_config()
             auto pg_it = this->pgs.find({ .pool_id = pool_id, .pg_num = pg_num });
             bool currently_taken = pg_it != this->pgs.end() && pg_it->second.state != PG_OFFLINE;
             // Check pool block size and bitmap granularity
-            if (this->bs_block_size != pool_item.second.data_block_size ||
+            if (take && this->bs_block_size != pool_item.second.data_block_size ||
                 this->bs_bitmap_granularity != pool_item.second.bitmap_granularity)
             {
                 if (!warned_block_size)
@@ -967,8 +972,8 @@ void osd_t::report_pg_states()
                         pool_id_t pool_id = 0;
                         pg_num_t pg_num = 0;
                         char null_byte = 0;
-                        sscanf(kv.key.c_str() + st_cli.etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
-                        if (null_byte == 0)
+                        int scanned = sscanf(kv.key.c_str() + st_cli.etcd_prefix.length()+10, "%u/%u%c", &pool_id, &pg_num, &null_byte);
+                        if (scanned == 2)
                         {
                             auto pg_it = pgs.find({ .pool_id = pool_id, .pg_num = pg_num });
                             if (pg_it != pgs.end() && pg_it->second.state != PG_OFFLINE && pg_it->second.state != PG_STARTING &&