add per-cgroup cpu perf metrics (iopsystems#395)

Adds per-cgroup cpu perf metrics.

src/common/bpf/builder.rs

@@ -62,6 +62,7 @@ pub struct Builder<T: 'static + SkelBuilder<'static>> {
         Vec<&'static CounterGroup>,
     )>,
     perf_events: Vec<(&'static str, PerfEvent)>,
+    packed_counters: Vec<(&'static str, &'static CounterGroup)>,
 }
 
 impl<T: 'static> Builder<T>
@@ -78,6 +79,7 @@ where
             maps: Vec::new(),
             cpu_counters: Vec::new(),
             perf_events: Vec::new(),
+            packed_counters: Vec::new(),
         }
     }
 
@@ -177,6 +179,12 @@ where
                 perf_events.len()
             );
 
+            let mut packed_counters: Vec<PackedCounters> = self
+                .packed_counters
+                .into_iter()
+                .map(|(name, counters)| PackedCounters::new(skel.map(name), counters))
+                .collect();
+
             // load any data from userspace into BPF maps
             for (name, values) in self.maps.into_iter() {
                 let fd = skel.map(name).as_fd().as_raw_fd();
@@ -223,6 +231,10 @@ where
                     v.refresh();
                 }
 
+                for v in &mut packed_counters {
+                    v.refresh();
+                }
+
                 // notify that we have finished running
                 sync.notify();
             }
@@ -295,4 +307,13 @@ where
         self.perf_events.push((name, event));
         self
     }
+
+    /// Register a set of packed counters. The `name` is the BPF map name and
+    /// the `counters` are a set of userspace dynamic counters. The BPF map is
+    /// expected to be densely packed, meaning there is no padding. The order of
+    /// the `counters` must exactly match the order in the BPF map.
+    pub fn packed_counters(mut self, name: &'static str, counters: &'static CounterGroup) -> Self {
+        self.packed_counters.push((name, counters));
+        self
+    }
 }

src/common/bpf/counters.rs

@@ -184,3 +184,57 @@ impl<'a> CpuCounters<'a> {
         }
     }
 }
+
+/// Represents a set of counters where the BPF map is a dense set of counters,
+/// meaning there is no padding. No aggregation is performed, and the values are
+/// updated into a single `RwLockCounterGroup`.
+pub struct PackedCounters<'a> {
+    _map: &'a Map<'a>,
+    mmap: MmapMut,
+    counters: &'static CounterGroup,
+}
+
+impl<'a> PackedCounters<'a> {
+    /// Create a new set of counters from the provided BPF map and collection of
+    /// counter metrics.
+    ///
+    /// The map layout is not cacheline padded. The ordering of the dynamic
+    /// counters must exactly match the layout in the BPF map.
+    pub fn new(map: &'a Map, counters: &'static CounterGroup) -> Self {
+        let total_bytes = counters.len() * std::mem::size_of::<u64>();
+
+        let fd = map.as_fd().as_raw_fd();
+        let file = unsafe { std::fs::File::from_raw_fd(fd as _) };
+        let mmap: MmapMut = unsafe {
+            MmapOptions::new()
+                .len(total_bytes)
+                .map_mut(&file)
+                .expect("failed to mmap() bpf counterset")
+        };
+
+        let (_prefix, values, _suffix) = unsafe { mmap.align_to::<u64>() };
+
+        if values.len() != counters.len() {
+            panic!("mmap region not aligned or width doesn't match");
+        }
+
+        Self {
+            _map: map,
+            mmap,
+            counters,
+        }
+    }
+
+    /// Refreshes the counters by reading from the BPF map and setting each
+    /// counter metric to the current value.
+    pub fn refresh(&mut self) {
+        let (_prefix, values, _suffix) = unsafe { self.mmap.align_to::<u64>() };
+
+        // update all individual counters
+        for (idx, value) in values.iter().enumerate() {
+            if *value != 0 {
+                let _ = self.counters.set(idx, *value);
+            }
+        }
+    }
+}

src/common/bpf/mod.rs

@@ -23,7 +23,10 @@ const CACHELINE_SIZE: usize = 64;
 const PAGE_SIZE: usize = 4096;
 
 // This is the maximum number of CPUs we track with BPF counters.
-const MAX_CPUS: usize = 1024;
+pub const MAX_CPUS: usize = 1024;
+
+// This is the maximum number of cgroups we track with BPF counters.
+pub const MAX_CGROUPS: usize = 4096;
 
 const COUNTER_SIZE: usize = std::mem::size_of::<u64>();
 
@@ -37,7 +40,7 @@ fn whole_pages<T>(count: usize) -> usize {
     ((count * std::mem::size_of::<T>()) + PAGE_SIZE - 1) / PAGE_SIZE
 }
 
-use counters::{Counters, CpuCounters};
+use counters::{Counters, CpuCounters, PackedCounters};
 use histogram::Histogram;
 use sync_primitive::SyncPrimitive;
 

src/samplers/cpu/linux/perf/mod.bpf.c

@@ -5,18 +5,18 @@
 #include "../../../common/bpf/helpers.h"
 #include <bpf/bpf_core_read.h>
 #include <bpf/bpf_helpers.h>
+#include <bpf/bpf_tracing.h>
 
+#define COUNTERS 2
 #define COUNTER_GROUP_WIDTH 8
 #define MAX_CPUS 1024
+#define MAX_CGROUPS 4096
 
 #define TASK_RUNNING 0
 
 // counter positions
 #define CYCLES 0
 #define INSTRUCTIONS 1
-#define TSC 2
-#define APERF 3
-#define MPERF 4
 
 // counters (see constants defined at top)
 struct {
@@ -27,6 +27,38 @@ struct {
 	__uint(max_entries, MAX_CPUS * COUNTER_GROUP_WIDTH);
 } counters SEC(".maps");
 
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(map_flags, BPF_F_MMAPABLE);
+	__type(key, u32);
+	__type(value, u64);
+	__uint(max_entries, MAX_CGROUPS);
+} cgroup_cycles SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(map_flags, BPF_F_MMAPABLE);
+	__type(key, u32);
+	__type(value, u64);
+	__uint(max_entries, MAX_CGROUPS);
+} cgroup_instructions SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(map_flags, BPF_F_MMAPABLE);
+	__type(key, u32);
+	__type(value, u64);
+	__uint(max_entries, MAX_CGROUPS);
+} cycles_prev SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(map_flags, BPF_F_MMAPABLE);
+	__type(key, u32);
+	__type(value, u64);
+	__uint(max_entries, MAX_CGROUPS);
+} instructions_prev SEC(".maps");
+
 /**
  * perf event arrays
  */
@@ -35,14 +67,12 @@ struct {
 	__uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
 	__type(key, u32);
 	__type(value, u32);
-	// __uint(max_entries, MAX_CPUS);
 } cycles SEC(".maps");
 
 struct {
 	__uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
 	__type(key, u32);
 	__type(value, u32);
-	// __uint(max_entries, MAX_CPUS);
 } instructions SEC(".maps");
 
 /**
@@ -68,6 +98,26 @@ static __always_inline __s64 get_task_state(void *task)
 	return BPF_CORE_READ((struct task_struct___o *)task, state);
 }
 
+// attach a kprobe cpuacct to update per-cpu counters
+
+SEC("kprobe/cpuacct_account_field")
+int BPF_KPROBE(cpuacct_account_field_kprobe, void *task, u32 index, u64 delta)
+{
+	u32 idx;
+	u32 processor_id = bpf_get_smp_processor_id();
+
+	u64 c = bpf_perf_event_read(&cycles, BPF_F_CURRENT_CPU);
+	u64 i = bpf_perf_event_read(&instructions, BPF_F_CURRENT_CPU);
+
+	idx = processor_id * COUNTER_GROUP_WIDTH + CYCLES;
+	bpf_map_update_elem(&counters, &idx, &c, BPF_ANY);
+
+	idx = processor_id * COUNTER_GROUP_WIDTH + INSTRUCTIONS;
+	bpf_map_update_elem(&counters, &idx, &i, BPF_ANY);
+}
+
+// attach a tracepoint on sched_switch for per-cgroup accounting
+
 SEC("tp_btf/sched_switch")
 int handle__sched_switch(u64 *ctx)
 {
@@ -78,22 +128,48 @@ int handle__sched_switch(u64 *ctx)
 	struct task_struct *next = (struct task_struct *)ctx[2];
 
 	u32 idx;
-	u64 *elem;
+	u64 *elem, delta_c, delta_i;
 
 	u32 processor_id = bpf_get_smp_processor_id();
-	u64 ts = bpf_ktime_get_ns();
-
-	u64 flags = processor_id & BPF_F_INDEX_MASK;
 
-	u64 c = bpf_perf_event_read(&cycles, flags);
-	u64 i = bpf_perf_event_read(&instructions, flags);
+	u64 c = bpf_perf_event_read(&cycles, BPF_F_CURRENT_CPU);
+	u64 i = bpf_perf_event_read(&instructions, BPF_F_CURRENT_CPU);
 
 	idx = processor_id * COUNTER_GROUP_WIDTH + CYCLES;
 	bpf_map_update_elem(&counters, &idx, &c, BPF_ANY);
 
 	idx = processor_id * COUNTER_GROUP_WIDTH + INSTRUCTIONS;
 	bpf_map_update_elem(&counters, &idx, &i, BPF_ANY);
 
+	if (bpf_core_field_exists(prev->sched_task_group)) {
+		int cgroup_id = prev->sched_task_group->css.id;
+
+		if (cgroup_id && cgroup_id < MAX_CGROUPS) {
+			// update cgroup cycles
+
+			elem = bpf_map_lookup_elem(&cycles_prev, &processor_id);
+
+			if (elem) {
+				delta_c = c - *elem;
+
+				array_add(&cgroup_cycles, cgroup_id, delta_c);
+			}
+
+			// update cgroup instructions
+
+			elem = bpf_map_lookup_elem(&instructions_prev, &processor_id);
+
+			if (elem) {
+				delta_i = i - *elem;
+
+				array_add(&cgroup_instructions, cgroup_id, delta_i);
+			}
+		}
+	}
+
+	bpf_map_update_elem(&cycles_prev, &processor_id, &c, BPF_ANY);
+	bpf_map_update_elem(&instructions_prev, &processor_id, &i, BPF_ANY);
+
 	return 0;
 }
 

src/samplers/cpu/linux/perf/mod.rs

@@ -37,6 +37,8 @@ fn init(config: Arc<Config>) -> SamplerResult {
         .perf_event("cycles", PerfEvent::cpu_cycles())
         .perf_event("instructions", PerfEvent::instructions())
         .cpu_counters("counters", totals, individual)
+        .packed_counters("cgroup_cycles", &CGROUP_CPU_CYCLES)
+        .packed_counters("cgroup_instructions", &CGROUP_CPU_INSTRUCTIONS)
         .build()?;
 
     Ok(Some(Box::new(bpf)))
@@ -45,6 +47,8 @@ fn init(config: Arc<Config>) -> SamplerResult {
 impl SkelExt for ModSkel<'_> {
     fn map(&self, name: &str) -> &libbpf_rs::Map {
         match name {
+            "cgroup_cycles" => &self.maps.cgroup_cycles,
+            "cgroup_instructions" => &self.maps.cgroup_instructions,
             "counters" => &self.maps.counters,
             "cycles" => &self.maps.cycles,
             "instructions" => &self.maps.instructions,
@@ -55,6 +59,10 @@ impl SkelExt for ModSkel<'_> {
 
 impl OpenSkelExt for ModSkel<'_> {
     fn log_prog_instructions(&self) {
+        debug!(
+            "{NAME} cpuacct_account_field() BPF instruction count: {}",
+            self.progs.cpuacct_account_field_kprobe.insn_cnt()
+        );
         debug!(
             "{NAME} handle__sched_switch() BPF instruction count: {}",
             self.progs.handle__sched_switch.insn_cnt()

src/samplers/cpu/linux/stats.rs

@@ -1,10 +1,8 @@
-use crate::common::CounterGroup;
+use crate::common::*;
 use crate::samplers::cpu::stats::*;
 
 use metriken::*;
 
-pub static MAX_CPUS: usize = 1024;
-
 #[metric(
     name = "cpu/usage/total",
     description = "The amount of CPU time spent servicing interrupts",
@@ -174,6 +172,22 @@ pub static CPU_MPERF_PERCORE: CounterGroup = CounterGroup::new(MAX_CPUS);
 )]
 pub static CPU_TSC_PERCORE: CounterGroup = CounterGroup::new(MAX_CPUS);
 
+#[metric(
+    name = "cgroup/cpu/cycles",
+    description = "The number of elapsed CPU cycles on a per-cgroup basis",
+    formatter = cpu_metric_cgroup_formatter,
+    metadata = { unit = "cycles" }
+)]
+pub static CGROUP_CPU_CYCLES: CounterGroup = CounterGroup::new(MAX_CGROUPS);
+
+#[metric(
+    name = "cgroup/cpu/instructions",
+    description = "The number of elapsed CPU cycles on a per-cgroup basis",
+    formatter = cpu_metric_cgroup_formatter,
+    metadata = { unit = "cycles" }
+)]
+pub static CGROUP_CPU_INSTRUCTIONS: CounterGroup = CounterGroup::new(MAX_CGROUPS);
+
 pub fn cpu_metric_percore_formatter(metric: &MetricEntry, format: Format) -> String {
     match format {
         Format::Simple => {
@@ -195,3 +209,12 @@ pub fn cpu_usage_percore_formatter(metric: &MetricEntry, format: Format) -> Stri
         _ => metric.name().to_string(),
     }
 }
+
+pub fn cpu_metric_cgroup_formatter(metric: &MetricEntry, format: Format) -> String {
+    match format {
+        Format::Simple => {
+            format!("{}/cgroup", metric.name())
+        }
+        _ => metric.name().to_string(),
+    }
+}