Reduce memory footprint of kernel symbols (open-telemetry#194)

Reducing the kernel symbols' memory footprint by

    skipping non-text symbols when reading from /proc/kallsyms
    adjusting overcommitted symbols array

Additionally, we pre-allocate a relatively large symbols array to begin with parsing to avoid re-allocations.

libpf/symbol.go

@@ -43,6 +43,12 @@ type SymbolMap struct {
 	addressToSymbol []Symbol
 }
 
+func NewSymbolMap(capacity int) *SymbolMap {
+	return &SymbolMap{
+		addressToSymbol: make([]Symbol, 0, capacity),
+	}
+}
+
 // Add a symbol to the map
 func (symmap *SymbolMap) Add(s Symbol) {
 	symmap.addressToSymbol = append(symmap.addressToSymbol, s)
@@ -51,10 +57,16 @@ func (symmap *SymbolMap) Add(s Symbol) {
 // Finalize symbol map by sorting and constructing the nameToSymbol table after
 // all symbols are inserted via Add() calls
 func (symmap *SymbolMap) Finalize() {
+	// Adjust the overcommitted capacity
+	a := make([]Symbol, len(symmap.addressToSymbol))
+	copy(a, symmap.addressToSymbol)
+	symmap.addressToSymbol = a
+
 	sort.Slice(symmap.addressToSymbol,
 		func(i, j int) bool {
 			return symmap.addressToSymbol[i].Address > symmap.addressToSymbol[j].Address
 		})
+
 	symmap.nameToSymbol = make(map[SymbolName]*Symbol, len(symmap.addressToSymbol))
 	for i, s := range symmap.addressToSymbol {
 		symmap.nameToSymbol[s.Name] = &symmap.addressToSymbol[i]

proc/proc.go

@@ -28,7 +28,8 @@ func GetKallsyms(kallsymsPath string) (*libpf.SymbolMap, error) {
 	var address uint64
 	var symbol string
 
-	symmap := libpf.SymbolMap{}
+	// As an example, the Debian 6.10.11 kernel has ~180k text symbols.
+	symmap := libpf.NewSymbolMap(200 * 1024)
 	noSymbols := true
 
 	file, err := os.Open(kallsymsPath)
@@ -52,6 +53,12 @@ func GetKallsyms(kallsymsPath string) (*libpf.SymbolMap, error) {
 			return nil, fmt.Errorf("unexpected line in kallsyms: '%s'", line)
 		}
 
+		// Skip non-text symbols, see 'man nm'.
+		// Special case for 'etext', which can be of type `D` (data) in some kernels.
+		if strings.IndexByte("TtVvWwA", fields[1][0]) == -1 && fields[2] != "_etext" {
+			continue
+		}
+
 		if address, err = strconv.ParseUint(fields[0], 16, 64); err != nil {
 			return nil, fmt.Errorf("failed to parse address value: '%s'", fields[0])
 		}
@@ -74,7 +81,7 @@ func GetKallsyms(kallsymsPath string) (*libpf.SymbolMap, error) {
 			"all addresses from kallsyms are zero - check process permissions")
 	}
 
-	return &symmap, nil
+	return symmap, nil
 }
 
 // GetKernelModules returns SymbolMap for kernel modules from /proc/modules.