profiler: Reduce mem usage in maps with no filters (#1588)

Right now we always allocate 5k entries in `debug_pids` in kernel memory
even if the filter mode `--debug-process-names` is not enabled. Let's
only allocate the entries if we are filtering processes.

The saved memory isn't too much, around a few KB, but kernel memory is
precious and in particular this memory, as it's locked.

Test Plan
=========

**no filters**
```
[javierhonduco@fedora parca-agent]$ sudo bpftool map | grep debug_pids -A3
739: hash  name debug_pids  flags 0x0
        key 4B  value 1B  max_entries 1  memlock 4096B
        btf_id 670
        pids parca-agent-deb(154991)
```

**with filters**
```
[javierhonduco@fedora parca-agent]$ sudo bpftool map | grep debug_pids -A3
765: hash  name debug_pids  flags 0x0
        key 4B  value 1B  max_entries 5000  memlock 40960B
        btf_id 697
        pids parca-agent-deb(156286)
```

Author: kakkoyun

bpf/cpu/cpu.bpf.c

@@ -205,7 +205,7 @@ typedef struct {
 
 /*================================ MAPS =====================================*/
 
-BPF_HASH(debug_pids, int, u8, MAX_PROCESSES);
+BPF_HASH(debug_pids, int, u8, 1); // Table size will be updated in userspace.
 BPF_HASH(process_info, int, process_info_t, MAX_PROCESSES);
 
 BPF_STACK_TRACE(stack_traces, MAX_STACK_TRACES_ENTRIES);

pkg/profiler/cpu/cpu.go

@@ -258,7 +258,7 @@ func loadBpfProgram(logger log.Logger, reg prometheus.Registerer, debugEnabled,
 		}
 
 		level.Info(logger).Log("msg", "Attempting to create unwind shards", "count", unwindShards)
-		if err := bpfMaps.adjustMapSizes(unwindShards); err != nil {
+		if err := bpfMaps.adjustMapSizes(debugEnabled, unwindShards); err != nil {
 			return nil, nil, fmt.Errorf("failed to adjust map sizes: %w", err)
 		}
 

pkg/profiler/cpu/maps.go

@@ -63,6 +63,7 @@ const (
 	maxUnwindTableSize    = 250 * 1000 // Always needs to be sync with MAX_UNWIND_TABLE_SIZE in the BPF program.
 	maxMappingsPerProcess = 250        // Always need to be in sync with MAX_MAPPINGS_PER_PROCESS.
 	maxUnwindTableChunks  = 30         // Always need to be in sync with MAX_UNWIND_TABLE_CHUNKS.
+	maxProcesses          = 5000       // Always need to be in sync with MAX_PROCESSES.
 
 	/*
 		TODO: once we generate the bindings automatically, remove this.
@@ -290,20 +291,30 @@ func (m *bpfMaps) close() error {
 // adjustMapSizes updates the amount of unwind shards.
 //
 // Note: It must be called before `BPFLoadObject()`.
-func (m *bpfMaps) adjustMapSizes(unwindTableShards uint32) error {
+func (m *bpfMaps) adjustMapSizes(debugEnabled bool, unwindTableShards uint32) error {
 	unwindTables, err := m.module.GetMap(unwindTablesMapName)
 	if err != nil {
 		return fmt.Errorf("get unwind tables map: %w", err)
 	}
 
-	// Adjust unwind tables size.
+	// Adjust unwind_tables size.
 	sizeBefore := unwindTables.GetMaxEntries()
 	if err := unwindTables.Resize(unwindTableShards); err != nil {
 		return fmt.Errorf("resize unwind tables map from %d to %d elements: %w", sizeBefore, unwindTableShards, err)
 	}
 
 	m.maxUnwindShards = uint64(unwindTableShards)
 
+	// Adjust debug_pids size.
+	if debugEnabled {
+		debugPIDs, err := m.module.GetMap(debugPIDsMapName)
+		if err != nil {
+			return fmt.Errorf("get debug pids map: %w", err)
+		}
+		if err := debugPIDs.Resize(maxProcesses); err != nil {
+			return fmt.Errorf("resize debug pids map from default to %d elements: %w", maxProcesses, err)
+		}
+	}
 	return nil
 }