Merge pull request #5338 from ritika0313/T8329-azure-interface-naming-fix

T8329: Fix interface naming for Azure VF interfaces with Accelerated Networking enabled

Author: jestabro

src/etc/udev/rules.d/63-hyperv-vf-net.rules

@@ -1,5 +1,9 @@
 ATTR{[dmi/id]sys_vendor}!="Microsoft Corporation", GOTO="end_hyperv_nic"
 
-ACTION=="add", SUBSYSTEM=="net", DRIVERS=="hv_pci", NAME="vf_%k"
+# Primary rename path for Azure VF.
+# Use helper-based naming so the VFs that reuse eth0,eth1,.. names do not
+# collide with already assigned vf_ethN names. Also, prevent renaming of
+# the interfaces already renamed to vf_ethN
+ACTION=="add", SUBSYSTEM=="net", DRIVERS=="hv_pci", KERNEL=="eth*", PROGRAM="vyos_vf_name %k", NAME="%c"
 
 LABEL="end_hyperv_nic"

src/etc/udev/rules.d/65-vyos-net.rules

@@ -4,6 +4,24 @@
 ACTION!="add", 				GOTO="vyos_net_end"
 SUBSYSTEM!="net",			GOTO="vyos_net_end"
 
+# Fallback path for Azure VF interfaces: if a VF still appears as plain ethN,
+# rename it to vf_ethX before persistent synthetic interface naming runs to
+# prevent it from incorrectly being considered as a synthetic interface.
+ATTR{[dmi/id]sys_vendor}=="Microsoft Corporation", DRIVERS=="mlx*_core", KERNEL=="eth*", PROGRAM="vyos_vf_name %k", NAME="%c", GOTO="vyos_net_end"
+ATTR{[dmi/id]sys_vendor}=="Microsoft Corporation", DRIVERS=="mana", KERNEL=="eth*", PROGRAM="vyos_vf_name %k", NAME="%c", GOTO="vyos_net_end"
+
+# VF interfaces are not handled by vyos_net_name, so skip them.
+# KERNEL=="vf_*" catches already-renamed Mellanox and MANA VFs.
+# DRIVERS=="mlx*_core" and DRIVERS=="mana" on Azure catch VF interfaces
+# in an unexpected intermediate state (still named ethN/eN) that must not
+# be remapped by vyos_net_name.
+# This guard is intentionally scoped to Microsoft Azure to
+# avoid excluding bare-metal Mellanox NICs on physical hardware, which do
+# need persistent naming via vyos_net_name.
+KERNEL=="vf_*", 							GOTO="vyos_net_end"
+ATTR{[dmi/id]sys_vendor}=="Microsoft Corporation", DRIVERS=="mlx*_core",	GOTO="vyos_net_end"
+ATTR{[dmi/id]sys_vendor}=="Microsoft Corporation", DRIVERS=="mana",      	GOTO="vyos_net_end"
+
 # Do name change for ethernet and wireless devices only
 KERNEL!="eth*|wlan*|e*", 			GOTO="vyos_net_end"
 

src/system/vyos-net-name-resolve.py

@@ -178,6 +178,11 @@ def discover_physical_interfaces(sys_class_net: str = '/sys/class/net') -> dict:
     """Return {kernel_name: mac} for every interface backed by a real bus
     device - excludes lo, bridges, bonds, VLANs, veth, tunnels, etc.
 
+    Also excludes interfaces enslaved to another netdev (master symlink
+    present), which covers Azure VF datapath interfaces bound under their
+    synthetic parent. Those are acceleration children, not independent
+    primary interfaces, and must not be candidates for hw-id naming.
+
     sys_class_net is overridable for testing against a fake sysfs tree.
     """
     interfaces = {}
@@ -187,6 +192,14 @@ def discover_physical_interfaces(sys_class_net: str = '/sys/class/net') -> dict:
 
     for entry in net_dir.iterdir():
         if not (entry / 'device').exists():
+            logger.debug(
+                f"skipping '{entry.name}': no backing device in sysfs"
+            )
+            continue
+        if (entry / 'master').exists():
+            logger.debug(
+                f"skipping '{entry.name}': interface is enslaved via master link"
+            )
             continue
         mac = get_permanent_mac(entry.name, sys_class_net)
         if mac:

src/udev/vyos_vf_name

@@ -0,0 +1,88 @@
+#!/bin/sh
+# vyos_vf_name - pick a non-colliding VF interface name.
+# Called by udev rules for VF devices that still have plain ethN names.
+#
+# Written as a POSIX shell script so it works in both initramfs (no Python3)
+# and the normal rootfs udev context.
+#
+# Usage: vyos_vf_name <current_ifname>
+# Prints the target vf_ethN name to stdout.
+
+set -e
+
+if [ -z "$1" ]; then
+    exit 1
+fi
+
+IFNAME="$1"
+LOCK_DIR="/run/udev"
+LOCK_PATH="${LOCK_DIR}/vyos_vf_name.lock"
+
+# Already normalized - return as-is.
+case "$IFNAME" in
+    vf_*)
+        echo "$IFNAME"
+        exit 0
+        ;;
+esac
+
+mkdir -p "$LOCK_DIR"
+
+# Acquire a lock to avoid race conditions when multiple VF devices are being
+# renamed concurrently.
+LOCK_ACQUIRED=0
+I=0
+while [ "$I" -lt 200 ]; do
+    if mkdir "$LOCK_PATH" 2>/dev/null; then
+        LOCK_ACQUIRED=1
+        break
+    fi
+    I=$((I + 1))
+    sleep 0.01
+done
+
+[ "$LOCK_ACQUIRED" -eq 1 ] || exit 1
+
+cleanup() {
+    rmdir "$LOCK_PATH" 2>/dev/null || true
+}
+trap cleanup EXIT INT TERM
+
+# Preferred: preserve the original ethN index when vf_ethN is not yet taken.
+case "$IFNAME" in
+    eth*)
+        PREFERRED="${IFNAME#eth}"
+        # Validate it is a pure integer.
+        case "$PREFERRED" in
+            *[!0-9]*)
+                ;;
+            *)
+                if ! [ -d "/sys/class/net/vf_eth${PREFERRED}" ]; then
+                    echo "vf_eth${PREFERRED}"
+                    exit 0
+                fi
+                ;;
+        esac
+        ;;
+esac
+
+# Fallback: allocate just above the highest ethN synthetic index present to
+# avoid collisions.
+MAX=-1
+for D in /sys/class/net/eth*; do
+    [ -d "$D" ] || continue
+    N="${D##*/eth}"
+    # Skip if not a pure integer (e.g. no match expands to literal path).
+    case "$N" in *[!0-9]*) continue;; esac
+    [ "$N" -gt "$MAX" ] && MAX="$N"
+done
+
+IDX=$((MAX + 1))
+[ "$IDX" -lt 0 ] && IDX=0
+
+# Skip any indices already occupied by existing vf_ethN interfaces.
+while [ -d "/sys/class/net/vf_eth${IDX}" ]; do
+    IDX=$((IDX + 1))
+done
+
+echo "vf_eth${IDX}"