Fix SR-IOV test failures on HPC/GPU SKUs with InfiniBand interfaces

lisa/microsoft/testsuites/core/provisioning.py

@@ -635,8 +635,10 @@ def check_sriov(self, log: Logger, node: RemoteNode) -> None:
             else:
                 pci_nic_check = True
         if pci_nic_check:
-            log.info(f"check_sriov: PCI nic count {len(node_nic_info.get_pci_nics())}")
-            assert_that(len(node_nic_info.get_pci_nics())).described_as(
-                f"VF count inside VM is {len(node_nic_info.get_pci_nics())}, "
-                f"actual sriov nic count is {sriov_count}"
+            pci_nics = node_nic_info.get_pci_nics_except_ib()
+            pci_nic_count = len(pci_nics)
+            log.info(f"check_sriov: PCI nic count (without IB) {pci_nic_count}")
+            assert_that(pci_nic_count).described_as(
+                f"VF count inside VM (without IB) is {pci_nic_count}, "
+                f"actual sriov nic count is {sriov_count}. "
             ).is_equal_to(sriov_count)

lisa/microsoft/testsuites/network/common.py

@@ -1,6 +1,6 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
-from typing import Dict, List, cast
+from typing import Dict, List, Tuple, cast
 
 from assertpy import assert_that
 from retry import retry
@@ -53,9 +53,11 @@ def initialize_nic_info(
                 f"inside vm, it should equal to {interface_info.ip_addr}."
             ).is_true()
         if is_sriov:
-            assert_that(len(nics_info.get_device_slots())).described_as(
-                f"VF count inside VM is {len(set(nics_info.get_device_slots()))}, "
-                f"actual sriov nic count is {sriov_count}"
+            pci_nic_count = len(nics_info.get_pci_nics_except_ib())
+            assert_that(pci_nic_count).described_as(
+                f"Ethernet VF count inside VM (without IB) is {pci_nic_count}, "
+                f"actual sriov nic count is {sriov_count}. "
+                f"Total PCI NICs: {nics_info.get_pci_nics()}"
             ).is_equal_to(sriov_count)
         vm_nics[node.name] = nics_info.nics
 
@@ -70,12 +72,14 @@ def sriov_basic_test(environment: Environment) -> None:
         devices_slots = lspci.get_device_names_by_type(
             constants.DEVICE_TYPE_SRIOV, force_run=True
         )
-        if len(devices_slots) != len(set(node.nics.get_device_slots())):
+        ethernet_device_slots = set(node.nics.get_device_slots_except_ib())
+        if len(devices_slots) != len(ethernet_device_slots):
             node.nics.reload()
+            ethernet_device_slots = set(node.nics.get_device_slots_except_ib())
         assert_that(devices_slots).described_as(
             "count of sriov devices listed from lspci is not expected,"
             " please check the driver works properly"
-        ).is_length(len(set(node.nics.get_device_slots())))
+        ).is_length(len(ethernet_device_slots))
 
         # 2. Check module of sriov network device is loaded.
         for module_name in node.nics.get_used_modules(["hv_netvsc"]):
@@ -172,7 +176,7 @@ def _find_matching_dest_nic(
     source_nic_info: NicInfo,
     vm_nics: Dict[str, Dict[str, NicInfo]],
     dest_node: RemoteNode,
-) -> tuple[str, int]:
+) -> Tuple[str, int]:
     """Find destination NIC on same subnet as source.
     Returns (nic_name, skipped_count)."""
     skipped_infiniband = 0
@@ -198,7 +202,7 @@ def _setup_nic_monitoring(
     remove_module: bool,
     turn_off_lower: bool,
     source_node: RemoteNode,
-) -> tuple[str, str, str, str]:
+) -> Tuple[str, str, str, str]:
     """Setup NIC monitoring based on configuration.
     Returns (source_nic, dest_nic, source_pci_nic, dest_pci_nic)."""
     source_synthetic_nic = source_nic_info.name

lisa/microsoft/testsuites/network/sriov.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT license.
 import re
 from pathlib import Path
-from typing import Any, Dict, List, cast
+from typing import Any, Dict, List, Optional, cast
 
 from assertpy import assert_that
 from microsoft.testsuites.network.common import (
@@ -811,6 +811,9 @@ def verify_sriov_interrupts_change(self, environment: Environment) -> None:
         client_interrupt_inspector = client_node.tools[InterruptInspector]
         for _, client_nic_info in vm_nics[client_node.name].items():
             if client_nic_info.is_pci_module_enabled:
+                # Skip InfiniBand interfaces — they use RDMA, not Ethernet SR-IOV
+                if client_nic_info.name and client_nic_info.name.startswith("ib"):
+                    continue
                 # 2. Get initial interrupts sum per irq and cpu number on client node.
                 # only collect 'Completion Queue Interrupts' irqs
                 initial_pci_interrupts_by_irqs = (
@@ -834,8 +837,13 @@ def verify_sriov_interrupts_change(self, environment: Environment) -> None:
                     assert_that(len(initial_pci_interrupts_by_cpus)).described_as(
                         "initial cpu count of interrupts should be equal to cpu count"
                     ).is_equal_to(client_thread_count)
-                matched_server_nic_info: NicInfo
+                matched_server_nic_info: Optional[NicInfo] = None
                 for _, server_nic_info in vm_nics[server_node.name].items():
+                    # Skip NICs without IP addresses (IB, enslaved VFs, etc.)
+                    if not server_nic_info.ip_addr:
+                        continue
+                    if server_nic_info.name and server_nic_info.name.startswith("ib"):
+                        continue
                     if (
                         server_nic_info.ip_addr.rsplit(".", maxsplit=1)[0]
                         == client_nic_info.ip_addr.rsplit(".", maxsplit=1)[0]

lisa/microsoft/testsuites/network/stress.py

@@ -1,6 +1,6 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
-from typing import Any, cast
+from typing import Any, Optional, cast
 
 from assertpy import assert_that
 from microsoft.testsuites.network.common import (
@@ -73,15 +73,20 @@ def stress_sriov_iperf(self, environment: Environment) -> None:
         client_nic_info_list = [
             x
             for _, x in vm_nics[client_node.name].items()
-            if x.ip_addr == client_node.internal_address
+            if x.ip_addr and x.ip_addr == client_node.internal_address
         ]
         assert_that(client_nic_info_list).described_as(
             "not found the primary network interface."
-        ).is_not_none()
+        ).is_not_empty()
         client_nic_info = client_nic_info_list[0]
         isinstance(client_nic_info, NicInfo)
-        matched_server_nic_info: NicInfo
+        matched_server_nic_info: Optional[NicInfo] = None
         for _, server_nic_info in vm_nics[server_node.name].items():
+            # Skip InfiniBand and NICs without IP addresses
+            if not server_nic_info.ip_addr:
+                continue
+            if server_nic_info.name and server_nic_info.name.startswith("ib"):
+                continue
             if (
                 server_nic_info.ip_addr.rsplit(".", maxsplit=1)[0]
                 == client_nic_info.ip_addr.rsplit(".", maxsplit=1)[0]

lisa/microsoft/testsuites/performance/common.py

@@ -928,7 +928,9 @@ def check_sriov_count(node: RemoteNode, sriov_count: int) -> None:
     node_nic_info = node.nics
     node_nic_info.reload()
 
-    assert_that(len(node_nic_info.get_pci_nics())).described_as(
-        f"VF count inside VM is {len(node_nic_info.get_pci_nics())},"
-        f"actual sriov nic count is {sriov_count}"
+    pci_nics = node_nic_info.get_pci_nics_except_ib()
+    pci_nic_count = len(pci_nics)
+    assert_that(pci_nic_count).described_as(
+        f"VF count inside VM (without IB) is {pci_nic_count}, "
+        f"actual sriov nic count is {sriov_count}. "
     ).is_equal_to(sriov_count)

lisa/nic.py

@@ -185,6 +185,23 @@ def get_synthetic_devices(self) -> List[str]:
         self._node.log.debug(f"Found synthetic devices: {synthetic_devices}")
         return synthetic_devices
 
+    def get_pci_nics_except_ib(self) -> List[str]:
+        """
+        Get NIC names for PCI-associated NICs, excluding InfiniBand (ib*)
+        """
+        pci_nics = []
+        for nic in self.nics.values():
+            nic_name = nic.lower if nic.lower else nic.name
+            if nic_name and nic_name.startswith("ib"):
+                continue
+            if nic.name and nic.name.startswith("ib"):
+                continue
+            if nic.is_pci_only_nic:
+                pci_nics.append(nic.name)
+            elif nic.lower:
+                pci_nics.append(nic.lower)
+        return pci_nics
+
     def get_pci_nics(self) -> List[str]:
         pci_nics = []
         for nic in self.nics.values():
@@ -221,6 +238,20 @@ def get_used_modules(self, exclude_module_name: List[str]) -> List[str]:
     def get_device_slots(self) -> List[str]:
         return [x.pci_slot for x in self.nics.values() if x.pci_slot]
 
+    def get_device_slots_except_ib(self) -> List[str]:
+        """Get PCI slots for NICs, excluding InfiniBand (ib*) interfaces."""
+        slots = []
+        for nic in self.nics.values():
+            if not nic.pci_slot:
+                continue
+            nic_name = nic.lower if nic.lower else nic.name
+            if nic_name and nic_name.startswith("ib"):
+                continue
+            if nic.name and nic.name.startswith("ib"):
+                continue
+            slots.append(nic.pci_slot)
+        return slots
+
     def _get_nics_driver(self) -> None:
         for nic in [x.name for x in self.nics.values()]:
             self.get_nic_driver(nic)
@@ -405,7 +436,7 @@ def _get_nic_uuid(self, nic_name: str) -> str:
         readlink = self._node.tools[Readlink]
         full_dev_path = readlink.get_target(f"/sys/class/net/{nic_name}/device")
         uuid = os.path.basename(full_dev_path)
-        self._node.log.debug(f"{nic_name} UUID:{uuid}")
+        self._node.log.debug(f"{nic_name} UUID: {uuid}")
         return uuid
 
     def _get_nic_uuids(self) -> None:

lisa/tools/lspci.py

@@ -412,7 +412,7 @@ def get_pci_slot_from_device_path(self, device_path: str) -> Optional[str]:
         """
         Extract PCI slot information from a device path.
         """
-        pci_slot = get_matched_str(device_path, PATTERN_PCI_SLOT)
+        pci_slot = get_matched_str(device_path, PATTERN_PCI_SLOT, first_match=False)
         return pci_slot if pci_slot else None