ethdev+ethface: GTP-U RxFlow with i40e DDP

csrc/ethface/face.c

@@ -41,14 +41,14 @@ EthRxFlow_RxBurst_Checked(RxGroup* rxg, RxGroupBurstCtx* ctx) {
 
 struct rte_flow*
 EthFace_SetupFlow(EthFacePriv* priv, const uint16_t queues[], int nQueues, const EthLocator* loc,
-                  bool isolated, struct rte_flow_error* error) {
+                  bool isolated, bool prefersFlowItemGTP, struct rte_flow_error* error) {
   EthLocatorClass c = EthLocator_Classify(loc);
   NDNDPDK_ASSERT(nQueues > 0 && nQueues <= (int)RTE_DIM(priv->rxf));
 
   struct rte_flow_attr attr = {.ingress = true};
 
   EthFlowPattern pattern;
-  EthFlowPattern_Prepare(&pattern, loc);
+  EthFlowPattern_Prepare(&pattern, loc, prefersFlowItemGTP);
 
   struct rte_flow_action_queue queue = {.index = queues[0]};
   struct rte_flow_action_rss rss = {

csrc/ethface/face.h

@@ -31,7 +31,7 @@ typedef struct EthFacePriv {
 /** @brief Setup rte_flow on EthDev for hardware dispatching. */
 __attribute__((nonnull)) struct rte_flow*
 EthFace_SetupFlow(EthFacePriv* priv, const uint16_t queues[], int nQueues, const EthLocator* loc,
-                  bool isolated, struct rte_flow_error* error);
+                  bool isolated, bool prefersFlowItemGTP, struct rte_flow_error* error);
 
 /** @brief Setup RX for memif. */
 __attribute__((nonnull)) void

csrc/ethface/locator.c

@@ -349,7 +349,7 @@ EthFlowPattern_Set(EthFlowPattern* flow, size_t i, enum rte_flow_item_type typ,
 }
 
 void
-EthFlowPattern_Prepare(EthFlowPattern* flow, const EthLocator* loc) {
+EthFlowPattern_Prepare(EthFlowPattern* flow, const EthLocator* loc, bool prefersFlowItemGTP) {
   EthLocatorClass c = EthLocator_Classify(loc);
 
   *flow = (const EthFlowPattern){0};
@@ -424,10 +424,13 @@ EthFlowPattern_Prepare(EthFlowPattern* flow, const EthLocator* loc) {
       break;
     }
     case 'G': {
-      MASK(flow->gtpMask.hdr.msg_type);
       MASK(flow->gtpMask.hdr.teid);
       PutGtpHdrMinimal(&flow->gtpSpec.hdr, loc->ulTEID);
-      APPEND(GTP, gtp);
+      if (prefersFlowItemGTP) {
+        APPEND(GTP, gtp);
+      } else {
+        APPEND(GTPU, gtp);
+      }
       break;
     }
   }

csrc/ethface/locator.h

@@ -118,7 +118,7 @@ typedef struct EthFlowPattern {
 
 /** @brief Prepare rte_flow pattern from locator. */
 __attribute__((nonnull)) void
-EthFlowPattern_Prepare(EthFlowPattern* flow, const EthLocator* loc);
+EthFlowPattern_Prepare(EthFlowPattern* flow, const EthLocator* loc, bool prefersFlowItemGTP);
 
 typedef struct EthTxHdr EthTxHdr;
 

docs/hardware.md

@@ -28,15 +28,16 @@ NDN-DPDK aims to work with most Ethernet adapters supported by [DPDK Network Int
 
 The developers have tested NDN-DPDK with the following Ethernet adapters:
 
-model | speed | DPDK driver | RxFlow
--|-|-|-
-NVIDIA ConnectX-5 | 100 Gbps | mlx5 | yes
-Intel X710 | 10 Gbps | i40e | UDP only
-Intel X710 VF | 10 Gbps | iavf | untested
-Intel XXV710 | 25 Gbps | i40e | untested
-Intel X520 | 10 Gbps | ixgbe | UDP only
-Intel I350 | 1 Gbps | igb | no
-Broadcom/QLogic 57810 | 10 Gbps | bnx2x | untested
+model | speed | DPDK driver | RxFlow Ethernet | RxFlow UDP | RxFlow VXLAN | RxFlow GTP-U
+-|-|-|-|-|-|-
+NVIDIA ConnectX-5 | 100 Gbps | mlx5 | yes | yes | yes | no
+NVIDIA ConnectX-6 | 200 Gbps | mlx5 | yes | yes | yes | yes
+Intel X710 | 10 Gbps | i40e | no | yes | no | yes
+Intel X710 VF | 10 Gbps | iavf | untested | untested | untested | untested
+Intel XXV710 | 25 Gbps | i40e | untested | untested | untested | untested
+Intel X520 | 10 Gbps | ixgbe | no | yes | no | untested
+Intel I350 | 1 Gbps | igb | no | no | no | untested
+Broadcom/QLogic 57810 | 10 Gbps | bnx2x | untested | untested | untested | untested
 
 Some Ethernet adapters have more than one physical ports on the same PCI card.
 NDN-DPDK is only tested to work on the first port (lowest PCI address) of those dual-port or quad-port adapters.
@@ -217,7 +218,6 @@ To use Intel VF:
 #### RxFlow Feature
 
 Intel adapters have limited compatibility with NDN-DPDK RxFlow feature.
-As tested with i40e and ixgbe drivers, RxFlow can be used with UDP faces, but not Ethernet or VXLAN faces.
 You should test RxFlow with your specific hardware and face locators, and decide whether to use this feature.
 Example command:
 
@@ -234,6 +234,14 @@ ndndpdk-ctrl create-eth-port --pci 04:00.0 --mtu 1500 --rx-flow 16
 ndndpdk-ctrl create-eth-port --pci 04:00.0 --mtu 1500
 ```
 
+GTP-U tunnel face with RxFlow is supported with [I40E poll mode driver](https://doc.dpdk.org/guides/nics/i40e.html) on Intel Ethernet 700 series.
+It relies on [Dynamic Device Personalization (DDP)](https://www.intel.com/content/www/us/en/developer/articles/technical/dynamic-device-personalization-for-intel-ethernet-700-series.html) feature.
+You must manually download the *GTPv1 DDP profile* and place it at `/lib/firmware/intel/i40e/ddp/gtp.pkg`.
+If the profile is found, you would see "upload DDP package success" log message during Ethernet port creation.
+Without the profile, GTP-U face creation on RxFlow would fail with "GTP is not supported by default" log message.
+During NDN-DPDK service shutdown, a profile rollback will be attempted.
+In case of an abnormal shutdown, you may need to power-cycle the server to cleanup the profile.
+
 ### Broadcom/QLogic Ethernet Adapters
 
 BCM57810 has been tested with igb\_uio driver.

dpdk/ethdev/config.go

@@ -11,9 +11,8 @@ import (
 type Config struct {
 	RxQueues []RxQueueConfig
 	TxQueues []TxQueueConfig
-	MTU      int            // if non-zero, change MTU
-	Promisc  bool           // promiscuous mode
-	Conf     unsafe.Pointer // pointer to rte_eth_conf, nil means default
+	MTU      int  // if non-zero, change MTU
+	Promisc  bool // promiscuous mode
 }
 
 // AddRxQueues adds RxQueueConfig for several queues

dpdk/ethdev/ddp.go

@@ -0,0 +1,93 @@
+package ethdev
+
+/*
+#include "../../csrc/core/common.h"
+#include <rte_pmd_i40e.h>
+#cgo LDFLAGS: -lrte_net_i40e
+*/
+import "C"
+import (
+	"fmt"
+	"io"
+	"os"
+	"path"
+	"slices"
+	"unsafe"
+
+	"github.com/usnistgov/ndn-dpdk/dpdk/eal"
+	"go.uber.org/zap"
+)
+
+// DdpProfile represents a Dynamic Device Personalization profile.
+type DdpProfile struct {
+	pkg    []byte
+	info   C.struct_rte_pmd_i40e_profile_info
+	zapPkg zap.Field
+}
+
+func (dp *DdpProfile) process(dev EthDev, buf []byte, op C.enum_rte_pmd_i40e_package_op, act string) error {
+	bufC, bufSize := (*C.uint8_t)(&buf[0]), C.uint32_t(len(buf))
+	if res := C.rte_pmd_i40e_process_ddp_package(dev.(ethDev).cID(), bufC, bufSize, op); res != 0 {
+		e := eal.MakeErrno(res)
+		logger.Error(act+" DDP package error",
+			dev.ZapField("port"),
+			dp.zapPkg,
+			zap.Error(e),
+		)
+		return fmt.Errorf("rte_pmd_i40e_process_ddp_package %w", e)
+	}
+
+	logger.Info(act+" DDP package success",
+		dev.ZapField("port"),
+		dp.zapPkg,
+	)
+	return nil
+}
+
+// Upload adds the DDP profile to an i40e device.
+// The returned rollback function removes the DDP profile from the device.
+func (dp *DdpProfile) Upload(dev EthDev) (rollback func() error, e error) {
+	// OP_WR_ADD modifies the buffer; OP_WR_DEL expects the modified buffer
+	buf := slices.Clone(dp.pkg)
+	if e = dp.process(dev, buf, C.RTE_PMD_I40E_PKG_OP_WR_ADD, "upload"); e != nil {
+		return nil, e
+	}
+	return func() error {
+		return dp.process(dev, buf, C.RTE_PMD_I40E_PKG_OP_WR_DEL, "rollback")
+	}, nil
+}
+
+// OpenDdpProfile opens a DDP profile from /lib/firmware/intel/i40e/ddp/{}.pkg .
+func OpenDdpProfile(name string) (dp *DdpProfile, e error) {
+	filename := path.Join("/lib/firmware/intel/i40e/ddp", name+".pkg")
+	logEntry := logger.With(zap.String("filename", filename))
+	file, e := os.Open(filename)
+	if e != nil {
+		logEntry.Warn("open DDP profile error", zap.Error(e))
+		return nil, e
+	}
+	defer file.Close()
+
+	dp = &DdpProfile{}
+	dp.pkg, e = io.ReadAll(file)
+	if e != nil {
+		logEntry.Warn("read DDP profile error", zap.Error(e))
+		return nil, e
+	}
+
+	if res := C.rte_pmd_i40e_get_ddp_info((*C.uint8_t)(&dp.pkg[0]), C.uint32_t(len(dp.pkg)),
+		(*C.uint8_t)(unsafe.Pointer(&dp.info)), C.uint32_t(unsafe.Sizeof(dp.info)),
+		C.RTE_PMD_I40E_PKG_INFO_GLOBAL_HEADER); res != 0 {
+		e = eal.MakeErrno(res)
+		logEntry.Warn("parse DDP profile error", zap.Error(e))
+		return nil, fmt.Errorf("rte_pmd_i40e_get_ddp_info %w", e)
+	}
+
+	dp.zapPkg = zap.Dict("pkg",
+		zap.Uint32("track-id", uint32(dp.info.track_id)),
+		zap.String("name", C.GoString((*C.char)(unsafe.Pointer(&dp.info.name[0])))),
+		zap.String("version", fmt.Sprintf("%d.%d.%d.%d",
+			dp.info.version.major, dp.info.version.minor, dp.info.version.update, dp.info.version.draft)),
+	)
+	return dp, nil
+}

dpdk/ethdev/ethdev.go

@@ -143,14 +143,11 @@ func (dev ethDev) Start(cfg Config) error {
 		return fmt.Errorf("%s %w", step, e)
 	}
 
-	conf := (*C.struct_rte_eth_conf)(cfg.Conf)
-	if conf == nil {
-		conf = &C.struct_rte_eth_conf{}
-		conf.rxmode.mtu = C.uint32_t(dev.MTU())
-		conf.txmode.offloads = C.uint64_t(info.Tx_offload_capa & (txOffloadMultiSegs | txOffloadChecksum))
-	}
+	conf := C.struct_rte_eth_conf{}
+	conf.rxmode.mtu = C.uint32_t(dev.MTU())
+	conf.txmode.offloads = C.uint64_t(info.Tx_offload_capa & (txOffloadMultiSegs | txOffloadChecksum))
 
-	if res := C.rte_eth_dev_configure(dev.cID(), C.uint16_t(len(cfg.RxQueues)), C.uint16_t(len(cfg.TxQueues)), conf); res < 0 {
+	if res := C.rte_eth_dev_configure(dev.cID(), C.uint16_t(len(cfg.RxQueues)), C.uint16_t(len(cfg.TxQueues)), &conf); res < 0 {
 		return bail("rte_eth_dev_configure", res)
 	}
 

dpdk/ethdev/info.go

@@ -21,6 +21,8 @@ const (
 	DriverXDP      = "net_af_xdp"
 	DriverMemif    = "net_memif"
 	DriverRing     = "net_ring"
+	DriverMlx5     = "net_mlx5"
+	DriverI40e     = "net_i40e"
 )
 
 const (
@@ -101,6 +103,17 @@ func (info DevInfo) HasTxChecksumOffload() bool {
 	return info.Tx_offload_capa&txOffloadChecksum == txOffloadChecksum
 }
 
+// PrefersFlowItemGTP indicates the device prefers RTE_FLOW_ITEM_TYPE_GTP to RTE_FLOW_ITEM_TYPE_GTPU.
+//
+// https://doc.dpdk.org/guides/nics/overview.html rte_flow items availability in networking drivers
+func (info DevInfo) PrefersFlowItemGTP() bool {
+	switch info.Driver() {
+	case DriverMlx5:
+		return true
+	}
+	return false
+}
+
 // MarshalJSON implements json.Marshaler interface.
 func (info DevInfo) MarshalJSON() ([]byte, error) {
 	return infoJSON(info, info.DevInfoC)

iface/ethport/port.go

@@ -79,6 +79,7 @@ type Port struct {
 	logger       *zap.Logger
 	dev          ethdev.EthDev
 	devInfo      ethdev.DevInfo
+	ddpRollback  func() error
 	faces        map[iface.ID]*Face
 	rxBouncePool *pktmbuf.Pool
 	rxImpl       rxImpl
@@ -123,6 +124,11 @@ func (port *Port) closeWithPortsMutex() error {
 		port.rxImpl = nil
 	}
 
+	if port.ddpRollback != nil {
+		errs = append(errs, port.ddpRollback())
+		port.ddpRollback = nil
+	}
+
 	if port.dev != nil {
 		errs = append(errs, port.dev.Close())
 		delete(ports, port.dev)
@@ -142,13 +148,19 @@ func (port *Port) closeWithPortsMutex() error {
 	return nil
 }
 
-func (port *Port) startDev(nRxQueues int, promisc bool) error {
+func (port *Port) startDev(nRxQueues int, promisc bool) (e error) {
 	socket := port.dev.NumaSocket()
 	rxPool := port.rxBouncePool
 	if rxPool == nil {
 		rxPool = ndni.PacketMempool.Get(socket)
 	}
 
+	if port.cfg.RxFlowQueues > 0 && port.devInfo.Driver() == ethdev.DriverI40e {
+		if dp, e := ethdev.OpenDdpProfile("gtp"); e == nil {
+			port.ddpRollback, _ = dp.Upload(port.dev)
+		}
+	}
+
 	cfg := ethdev.Config{
 		MTU:     port.cfg.MTU,
 		Promisc: promisc,

iface/ethport/rxflow.go

@@ -27,9 +27,10 @@ func readFlowErr(e C.struct_rte_flow_error) error {
 }
 
 type rxFlow struct {
-	isolated        bool
-	availQueues     []uint16
-	hasDestroyError bool
+	isolated           bool
+	prefersFlowItemGTP bool
+	availQueues        []uint16
+	hasDestroyError    bool
 }
 
 func (rxFlow) String() string {
@@ -65,6 +66,7 @@ func (impl *rxFlow) Init(port *Port) error {
 	if e := impl.setIsolate(port, true); e != nil {
 		port.logger.Info("flow isolate mode unavailable", zap.Error(e))
 	}
+	impl.prefersFlowItemGTP = port.devInfo.PrefersFlowItemGTP()
 
 	maxRxQueues := int(port.devInfo.Max_rx_queues)
 	if port.cfg.RxFlowQueues > maxRxQueues {
@@ -114,7 +116,8 @@ func (impl *rxFlow) setupFlow(face *Face, queues []uint16) error {
 	queuesC := (*C.uint16_t)(unsafe.Pointer(unsafe.SliceData(queues)))
 	locC := face.loc.EthLocatorC()
 	var flowErr C.struct_rte_flow_error
-	face.flow = C.EthFace_SetupFlow(face.priv, queuesC, C.int(len(queues)), locC.ptr(), C.bool(impl.isolated), &flowErr)
+	face.flow = C.EthFace_SetupFlow(face.priv, queuesC, C.int(len(queues)),
+		locC.ptr(), C.bool(impl.isolated), C.bool(impl.prefersFlowItemGTP), &flowErr)
 	if face.flow == nil {
 		return readFlowErr(flowErr)
 	}