Hotplug latency tests

Test latency of hotplugging via udev and manual agent

Signed-off-by: James Curtis <jxcurtis@amazon.co.uk>

Author: JamesC1305

src/vmm/src/devices/pseudo/boot_timer.rs

@@ -10,7 +10,7 @@ const MAGIC_VALUE_SIGNAL_GUEST_BOOT_COMPLETE: u8 = 123;
 /// Pseudo device to record the kernel boot time.
 #[derive(Debug)]
 pub struct BootTimer {
-    start_ts: TimestampUs,
+    pub start_ts: TimestampUs,
 }
 
 impl BootTimer {

src/vmm/src/lib.rs

@@ -688,6 +688,12 @@ impl Vmm {
         self.resume_vcpu_threads(start_idx.into())?;
 
         self.acpi_device_manager.notify_cpu_container()?;
+        if let Some(devices::BusDevice::BootTimer(timer)) =
+            self.get_bus_device(DeviceType::BootTimer, "BootTimer")
+        {
+            let mut locked_timer = timer.lock().expect("Poisoned lock");
+            locked_timer.start_ts = utils::time::TimestampUs::default();
+        }
 
         Ok(new_machine_config)
     }

tests/conftest.py

@@ -361,6 +361,9 @@ def rootfs_fxt(request, record_property):
 guest_kernel_linux_5_10 = pytest.fixture(
     guest_kernel_fxt, params=kernel_params("vmlinux-5.10*")
 )
+guest_kernel_linux_acpi_only = pytest.fixture(
+    guest_kernel_fxt, params=kernel_params("vmlinux-5.10.221")
+)
 # Use the unfiltered selector, since we don't officially support 6.1 yet.
 # TODO: switch to default selector once we add full 6.1 support.
 guest_kernel_linux_6_1 = pytest.fixture(
@@ -394,6 +397,15 @@ def uvm_plain_rw(microvm_factory, guest_kernel_linux_5_10, rootfs_rw):
     return microvm_factory.build(guest_kernel_linux_5_10, rootfs_rw)
 
 
+@pytest.fixture
+def uvm_hotplug(microvm_factory, guest_kernel_linux_acpi_only, rootfs_rw):
+    """Create a VM with ACPI enabled kernels only.
+    kernel: 5.10
+    rootfs: Ubuntu 22.04
+    """
+    return microvm_factory.build(guest_kernel_linux_acpi_only, rootfs_rw)
+
+
 @pytest.fixture
 def uvm_nano(uvm_plain):
     """A preconfigured uvm with 2vCPUs and 256MiB of memory

tests/host_tools/1-cpu-hotplug.rules

@@ -0,0 +1 @@
+SUBSYSTEM=="cpu", ACTION=="add", ATTR{online}!="1", ATTR{online}="1"

tests/host_tools/hotplug.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+# Copyright 2024 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+while :; do
+  [[ -d /sys/devices/system/cpu/cpu$1 ]] && break
+done
+
+echo 1 | tee /sys/devices/system/cpu/cpu*/online
+
+/home/hotplug_time.o

tests/host_tools/hotplug_time.c

@@ -0,0 +1,33 @@
+// Copyright 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+// Init wrapper for boot timing. It points at /sbin/init.
+
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+// Base address values are defined in arch/src/lib.rs as arch::MMIO_MEM_START.
+// Values are computed in arch/src/<arch>/mod.rs from the architecture layouts.
+// Position on the bus is defined by MMIO_LEN increments, where MMIO_LEN is
+// defined as 0x1000 in vmm/src/device_manager/mmio.rs.
+#ifdef __x86_64__
+#define MAGIC_MMIO_SIGNAL_GUEST_BOOT_COMPLETE 0xd0000000
+#endif
+#ifdef __aarch64__
+#define MAGIC_MMIO_SIGNAL_GUEST_BOOT_COMPLETE 0x40000000
+#endif
+
+#define MAGIC_VALUE_SIGNAL_GUEST_BOOT_COMPLETE 123
+
+int main() {
+  int fd = open("/dev/mem", (O_RDWR | O_SYNC | O_CLOEXEC));
+  int mapped_size = getpagesize();
+
+  char *map_base = mmap(NULL, mapped_size, PROT_WRITE, MAP_SHARED, fd,
+                        MAGIC_MMIO_SIGNAL_GUEST_BOOT_COMPLETE);
+
+  *map_base = MAGIC_VALUE_SIGNAL_GUEST_BOOT_COMPLETE;
+  msync(map_base, mapped_size, MS_ASYNC);
+}

tests/host_tools/hotplug_time.o

@@ -0,0 +1,9 @@
+#!/bin/bash
+# Copyright 2024 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+while :; do
+  [[ $(nproc) == $1 ]] && break
+done
+
+/home/hotplug_time.o

tests/host_tools/hotplug_udev.sh

@@ -3,17 +3,19 @@
 
 """Integration tests for hotplugging vCPUs"""
 
+import platform
 import re
 import time
 
 import pytest
 
-from framework import microvm
 from framework.defs import MAX_SUPPORTED_VCPUS
-from framework.microvm import Serial
 from framework.utils_cpuid import check_guest_cpuid_output
 
 
+@pytest.mark.skipif(
+    platform.machine() != "x86_64", reason="Hotplug only enabled on x86_64."
+)
 @pytest.mark.parametrize("vcpu_count", [1, MAX_SUPPORTED_VCPUS - 1])
 def test_hotplug_vcpus(uvm_plain, vcpu_count):
     """Test hotplugging works as intended"""
@@ -40,6 +42,9 @@ def test_hotplug_vcpus(uvm_plain, vcpu_count):
     )
 
 
+@pytest.mark.skipif(
+    platform.machine() != "x86_64", reason="Hotplug only enabled on x86_64."
+)
 @pytest.mark.parametrize(
     "vcpu_count", [-1, 0, MAX_SUPPORTED_VCPUS, MAX_SUPPORTED_VCPUS + 1]
 )
@@ -63,7 +68,7 @@ def test_negative_hotplug_vcpus(uvm_plain, vcpu_count):
         with pytest.raises(
             RuntimeError,
             match=re.compile(
-                f"An error occurred when deserializing the json body of a request: invalid value: integer `-\\d+`, expected u8+"
+                "An error occurred when deserializing the json body of a request: invalid value: integer `-\\d+`, expected u8+"
             ),
         ):
             uvm_plain.api.hotplug.put(Vcpu={"add": vcpu_count})
@@ -75,6 +80,9 @@ def test_negative_hotplug_vcpus(uvm_plain, vcpu_count):
             uvm_plain.api.hotplug.put(Vcpu={"add": vcpu_count})
 
 
+@pytest.mark.skipif(
+    platform.machine() != "x86_64", reason="Hotplug only enabled on x86_64."
+)
 @pytest.mark.parametrize("vcpu_count", [1, MAX_SUPPORTED_VCPUS - 1])
 def test_online_hotplugged_vcpus(uvm_plain, vcpu_count):
     """Test that hotplugged CPUs can be onlined"""

tests/integration_tests/functional/test_vcpu_hotplug.py

@@ -0,0 +1,152 @@
+# Copyright 2024 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Testing hotplug performance"""
+
+import platform
+import re
+import time
+from pathlib import Path
+
+import pandas
+import pytest
+
+from framework.utils_iperf import IPerf3Test, emit_iperf3_metrics
+from host_tools.cargo_build import gcc_compile
+from host_tools.fcmetrics import FCMetricsMonitor
+
+
+@pytest.mark.skipif(
+    platform.machine() != "x86_64", reason="Hotplug only enabled on x86_64."
+)
+@pytest.mark.parametrize(
+    "vcpu_count", [2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30]
+)
+def test_custom_udev_rule_latency(
+    microvm_factory, guest_kernel_linux_acpi_only, rootfs_rw, vcpu_count, results_dir
+):
+    """Test the latency for hotplugging and booting CPUs in the guest"""
+    gcc_compile(Path("./host_tools/hotplug_time.c"), Path("host_tools/hotplug_time.o"))
+    data = []
+    for i in range(50):
+        uvm_hotplug = microvm_factory.build(guest_kernel_linux_acpi_only, rootfs_rw)
+        uvm_hotplug.jailer.extra_args.update({"boot-timer": None, "no-seccomp": None})
+        uvm_hotplug.help.enable_console()
+        uvm_hotplug.spawn()
+        uvm_hotplug.basic_config(vcpu_count=1, mem_size_mib=128)
+        uvm_hotplug.add_net_iface()
+        uvm_hotplug.start()
+        uvm_hotplug.ssh.scp_put(
+            Path("./host_tools/hotplug_udev.sh"), Path("/home/hotplug_udev.sh")
+        )
+        uvm_hotplug.ssh.scp_put(
+            Path("./host_tools/hotplug_time.o"), Path("/home/hotplug_time.o")
+        )
+        uvm_hotplug.ssh.scp_put(
+            Path("./host_tools/1-cpu-hotplug.rules"),
+            Path("/usr/lib/udev/rules.d/1-cpu-hotplug.rules"),
+        )
+        uvm_hotplug.ssh.run(
+            f"udevadm control --reload-rules && tmux new-session -d /bin/bash /home/hotplug_udev.sh {vcpu_count}"
+        )
+
+        uvm_hotplug.api.hotplug.put(Vcpu={"add": vcpu_count})
+        time.sleep(2)
+
+        # Extract API call duration
+        api_duration = (
+            float(
+                re.findall(
+                    r"Total previous API call duration: (\d+) us\.",
+                    uvm_hotplug.log_data,
+                )[-1]
+            )
+            / 1000
+        )
+        try:
+            timestamp = (
+                float(
+                    re.findall(
+                        r"Guest-boot-time\s+\=\s+(\d+)\s+us", uvm_hotplug.log_data
+                    )[0]
+                )
+                / 1000
+            )
+        except IndexError:
+            timestamp = None
+
+        data.append({"vcpus": vcpu_count, "api": api_duration, "onlining": timestamp})
+
+    output_file = results_dir / f"hotplug-{vcpu_count}.csv"
+
+    csv_data = pandas.DataFrame.from_dict(data).to_csv(
+        index=False,
+        float_format="%.3f",
+    )
+
+    output_file.write_text(csv_data)
+
+
+@pytest.mark.skipif(
+    platform.machine() != "x86_64", reason="Hotplug only enabled on x86_64."
+)
+@pytest.mark.parametrize(
+    "vcpu_count", [2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30]
+)
+def test_manual_latency(
+    microvm_factory, guest_kernel_linux_acpi_only, rootfs_rw, vcpu_count
+):
+    """Test the latency for hotplugging and booting CPUs in the guest"""
+    gcc_compile(Path("./host_tools/hotplug_time.c"), Path("host_tools/hotplug_time.o"))
+    data = []
+    for _ in range(50):
+        uvm_hotplug = microvm_factory.build(guest_kernel_linux_acpi_only, rootfs_rw)
+        uvm_hotplug.jailer.extra_args.update({"boot-timer": None, "no-seccomp": None})
+        uvm_hotplug.help.enable_console()
+        uvm_hotplug.spawn()
+        uvm_hotplug.basic_config(vcpu_count=1, mem_size_mib=128)
+        uvm_hotplug.add_net_iface()
+        uvm_hotplug.start()
+        uvm_hotplug.ssh.scp_put(
+            Path("./host_tools/hotplug.sh"), Path("/home/hotplug.sh")
+        )
+        uvm_hotplug.ssh.scp_put(
+            Path("./host_tools/hotplug_time.o"), Path("/home/hotplug_time.o")
+        )
+        uvm_hotplug.ssh.run("tmux new-session -d /bin/bash /home/hotplug.sh")
+
+        uvm_hotplug.api.hotplug.put(Vcpu={"add": vcpu_count})
+
+        time.sleep(1.5)
+        # Extract API call duration
+        api_duration = (
+            float(
+                re.findall(
+                    r"Total previous API call duration: (\d+) us\.",
+                    uvm_hotplug.log_data,
+                )[-1]
+            )
+            / 1000
+        )
+        try:
+            timestamp = (
+                float(
+                    re.findall(
+                        r"Guest-boot-time\s+\=\s+(\d+)\s+us", uvm_hotplug.log_data
+                    )[0]
+                )
+                / 1000
+            )
+        except IndexError:
+            data.append({"vcpus": vcpu_count, "api": api_duration, "onlining": None})
+            continue
+
+        data.append({"vcpus": vcpu_count, "api": api_duration, "onlining": timestamp})
+
+    df = pandas.DataFrame.from_dict(data).to_csv(
+        f"../test_results/manual-hotplug_{vcpu_count}.csv",
+        index=False,
+        float_format="%.3f",
+    )
+
+    output_file.write_text(csv_data)