performance-metrics: Add QCOW2 L2 cache cold miss micro benchmark

Add micro_bench_qcow_l2_cache_miss which reads one cluster from each
of num_ops distinct L2 tables in a sparsely allocated image.  Clusters
are spaced L2_ENTRIES_PER_TABLE apart so every read touches a different
L2 table, forcing eviction when num_ops exceeds the cache capacity.

Workloads: 128 and 256 L2 tables.
Signed-off-by: Anatol Belski <anbelski@linux.microsoft.com>

performance-metrics/src/main.rs

@@ -378,7 +378,7 @@ mod adjuster {
     }
 }
 
-const TEST_LIST: [PerformanceTest; 80] = [
+const TEST_LIST: [PerformanceTest; 82] = [
     PerformanceTest {
         name: "boot_time_ms",
         func_ptr: performance_boot_time,
@@ -1469,6 +1469,30 @@ const TEST_LIST: [PerformanceTest; 80] = [
         },
         unit_adjuster: adjuster::s_to_us,
     },
+    PerformanceTest {
+        name: "micro_block_qcow_l2_cache_miss_128_us",
+        func_ptr: micro_bench_block::micro_bench_qcow_l2_cache_miss,
+        control: PerformanceTestControl {
+            test_timeout: 10,
+            test_iterations: 20,
+            warmup_iterations: 5,
+            num_ops: Some(128),
+            ..PerformanceTestControl::default()
+        },
+        unit_adjuster: adjuster::s_to_us,
+    },
+    PerformanceTest {
+        name: "micro_block_qcow_l2_cache_miss_256_us",
+        func_ptr: micro_bench_block::micro_bench_qcow_l2_cache_miss,
+        control: PerformanceTestControl {
+            test_timeout: 10,
+            test_iterations: 20,
+            warmup_iterations: 5,
+            num_ops: Some(256),
+            ..PerformanceTestControl::default()
+        },
+        unit_adjuster: adjuster::s_to_us,
+    },
 ];
 
 fn run_test_with_timeout(

performance-metrics/src/micro_bench_block.rs

@@ -16,8 +16,8 @@ use block::raw_async_aio::RawFileAsyncAio;
 
 use crate::PerformanceTestControl;
 use crate::util::{
-    self, BLOCK_SIZE, QCOW_CLUSTER_SIZE, deterministic_permutation, drain_completions, read_iovec,
-    submit_reads, submit_writes, write_iovec,
+    self, BLOCK_SIZE, L2_ENTRIES_PER_TABLE, QCOW_CLUSTER_SIZE, deterministic_permutation,
+    drain_completions, read_iovec, submit_reads, submit_writes, write_iovec,
 };
 
 /// Submit num_ops AIO writes, wait for them all to land, then time
@@ -301,3 +301,30 @@ pub fn micro_bench_qcow_multi_cluster_read(control: &PerformanceTestControl) ->
 
     elapsed
 }
+
+/// Read one cluster from each of num_ops distinct L2 tables in a
+/// sparsely allocated QCOW2 image.
+///
+/// The clusters are spaced L2_ENTRIES_PER_TABLE apart so every read
+/// touches a different L2 table.  With num_ops exceeding the L2 cache
+/// capacity (100 entries), this forces eviction on nearly every read
+/// and measures the cold L2 cache miss overhead.
+///
+/// Returns the total read wall clock time in seconds.
+pub fn micro_bench_qcow_l2_cache_miss(control: &PerformanceTestControl) -> f64 {
+    let num_ops = control.num_ops.expect("num_ops required") as usize;
+    let (_tmp, disk) = util::sparse_qcow_tempfile(num_ops);
+    let mut async_io = disk.new_async_io(1).expect("new_async_io failed");
+
+    let mut buf = vec![0u8; QCOW_CLUSTER_SIZE as usize];
+    let iovec = read_iovec(&mut buf);
+
+    let stride = L2_ENTRIES_PER_TABLE as u64 * QCOW_CLUSTER_SIZE;
+    let start = Instant::now();
+    submit_reads(async_io.as_mut(), num_ops, stride, &[iovec]);
+    let elapsed = start.elapsed().as_secs_f64();
+
+    drain_completions(async_io.as_mut(), num_ops);
+
+    elapsed
+}