windows_exporter/pkg/perflib/perflib.go

476 lines
13 KiB
Go

package perflib
/*
Go bindings for the HKEY_PERFORMANCE_DATA perflib / Performance Counters interface.
# Overview
HKEY_PERFORMANCE_DATA is a low-level alternative to the higher-level PDH library and WMI.
It operates on blocks of counters and only returns raw values without calculating rates
or formatting them, which is exactly what you want for, say, a Prometheus exporter
(not so much for a GUI like Windows Performance Monitor).
Its overhead is much lower than the high-level libraries.
It operates on the same set of perflib providers as PDH and WMI. See this document
for more details on the relationship between the different libraries:
https://msdn.microsoft.com/en-us/library/windows/desktop/aa371643(v=vs.85).aspx
Example C++ source code:
https://msdn.microsoft.com/de-de/library/windows/desktop/aa372138(v=vs.85).aspx
For now, the API is not stable and is probably going to change in future
perflib_exporter releases. If you want to use this library, send the author an email
so we can discuss your requirements and stabilize the API.
# Names
Counter names and help texts are resolved by looking up an index in a name table.
Since Microsoft loves internalization, both names and help texts can be requested
any locally available language.
The library automatically loads the name tables and resolves all identifiers
in English ("Name" and "HelpText" struct members). You can manually resolve
identifiers in a different language by using the NameTable API.
# Performance Counters intro
Windows has a system-wide performance counter mechanism. Most performance counters
are stored as actual counters, not gauges (with some exceptions).
There's additional metadata which defines how the counter should be presented to the user
(for example, as a calculated rate). This library disregards all of the display metadata.
At the top level, there's a number of performance counter objects.
Each object has counter definitions, which contain the metadata for a particular
counter, and either zero or multiple instances. We hide the fact that there are
objects with no instances, and simply return a single null instance.
There's one counter per counter definition and instance (or the object itself, if
there are no instances).
Behind the scenes, every perflib DLL provides one or more objects.
Perflib has a registry where DLLs are dynamically registered and
unregistered. Some third party applications like VMWare provide their own counters,
but this is, sadly, a rare occurrence.
Different Windows releases have different numbers of counters.
Objects and counters are identified by well-known indices.
Here's an example object with one instance:
4320 WSMan Quota Statistics [7 counters, 1 instance(s)]
`-- "WinRMService"
`-- Total Requests/Second [4322] = 59
`-- User Quota Violations/Second [4324] = 0
`-- System Quota Violations/Second [4326] = 0
`-- Active Shells [4328] = 0
`-- Active Operations [4330] = 0
`-- Active Users [4332] = 0
`-- Process ID [4334] = 928
All "per second" metrics are counters, the rest are gauges.
Another example, with no instance:
4600 Network QoS Policy [6 counters, 1 instance(s)]
`-- (default)
`-- Packets transmitted [4602] = 1744
`-- Packets transmitted/sec [4604] = 4852
`-- Bytes transmitted [4606] = 4853
`-- Bytes transmitted/sec [4608] = 180388626632
`-- Packets dropped [4610] = 0
`-- Packets dropped/sec [4612] = 0
You can access the same values using PowerShell's Get-Counter cmdlet
or the Performance Monitor.
> Get-Counter '\WSMan Quota Statistics(WinRMService)\Process ID'
Timestamp CounterSamples
--------- --------------
1/28/2018 10:18:00 PM \\DEV\wsman quota statistics(winrmservice)\process id :
928
> (Get-Counter '\Process(Idle)\% Processor Time').CounterSamples[0] | Format-List *
[..detailed output...]
Data for some of the objects is also available through WMI:
> Get-CimInstance Win32_PerfRawData_Counters_WSManQuotaStatistics
Name : WinRMService
[...]
ActiveOperations : 0
ActiveShells : 0
ActiveUsers : 0
ProcessID : 928
SystemQuotaViolationsPerSecond : 0
TotalRequestsPerSecond : 59
UserQuotaViolationsPerSecond : 0
*/
import (
"bytes"
"encoding/binary"
"errors"
"fmt"
"io"
"strings"
"syscall"
"unsafe"
"golang.org/x/sys/windows"
)
// TODO: There's a LittleEndian field in the PERF header - we ought to check it.
var bo = binary.LittleEndian
const averageCount64Type = 1073874176
// PerfObject Top-level performance object (like "Process").
type PerfObject struct {
Name string
// NameIndex Same index you pass to QueryPerformanceData
NameIndex uint
Instances []*PerfInstance
CounterDefs []*PerfCounterDef
Frequency int64
rawData *perfObjectType
}
// PerfInstance Each object can have multiple instances. For example,
// In case the object has no instances, we return one single PerfInstance with an empty name.
type PerfInstance struct {
// *not* resolved using a name table
Name string
Counters []*PerfCounter
rawData *perfInstanceDefinition
rawCounterBlock *perfCounterBlock
}
type PerfCounterDef struct {
Name string
NameIndex uint
// For debugging - subject to removal. CounterType is a perflib
// implementation detail (see perflib.h) and should not be used outside
// of this package. We export it so we can show it on /dump.
CounterType uint32
// PERF_TYPE_COUNTER (otherwise, it's a gauge)
IsCounter bool
// PERF_COUNTER_BASE (base value of a multi-value fraction)
IsBaseValue bool
// PERF_TIMER_100NS
IsNanosecondCounter bool
HasSecondValue bool
rawData *perfCounterDefinition
}
type PerfCounter struct {
Value int64
Def *PerfCounterDef
SecondValue int64
}
var (
bufLenGlobal = uint32(400000)
bufLenCostly = uint32(2000000)
)
// queryRawData Queries the performance counter buffer using RegQueryValueEx, returning raw bytes. See:
// https://msdn.microsoft.com/de-de/library/windows/desktop/aa373219(v=vs.85).aspx
func queryRawData(query string) ([]byte, error) {
var (
valType uint32
buffer []byte
bufLen uint32
)
switch query {
case "Global":
bufLen = bufLenGlobal
case "Costly":
bufLen = bufLenCostly
default:
// TODO: depends on the number of values requested
// need make an educated guess
numCounters := len(strings.Split(query, " "))
bufLen = uint32(150000 * numCounters)
}
buffer = make([]byte, bufLen)
name, err := windows.UTF16PtrFromString(query)
if err != nil {
return nil, fmt.Errorf("failed to encode query string: %w", err)
}
for {
bufLen := uint32(len(buffer))
//nolint:forbidigo // Legacy Code
err := syscall.RegQueryValueEx(
windows.HKEY_PERFORMANCE_DATA,
name,
nil,
&valType,
(*byte)(unsafe.Pointer(&buffer[0])),
&bufLen)
if errors.Is(err, error(syscall.ERROR_MORE_DATA)) { //nolint:forbidigo // Legacy Code
newBuffer := make([]byte, len(buffer)+16384)
copy(newBuffer, buffer)
buffer = newBuffer
continue
} else if err != nil {
var errNo syscall.Errno //nolint:forbidigo // Legacy Code
if errors.As(err, &errNo) {
return nil, fmt.Errorf("ReqQueryValueEx failed: %w errno %d", err, uint(errNo))
}
return nil, err
}
buffer = buffer[:bufLen]
switch query {
case "Global":
if bufLen > bufLenGlobal {
bufLenGlobal = bufLen
}
case "Costly":
if bufLen > bufLenCostly {
bufLenCostly = bufLen
}
}
return buffer, nil
}
}
/*
QueryPerformanceData Query all performance counters that match a given query.
The query can be any of the following:
- "Global" (all performance counters except those Windows marked as costly)
- "Costly" (only the costly ones)
- One or more object indices, separated by spaces ("238 2 5")
Many objects have dependencies - if you query one of them, you often get back
more than you asked for.
*/
func QueryPerformanceData(query string) ([]*PerfObject, error) {
buffer, err := queryRawData(query)
if err != nil {
return nil, err
}
r := bytes.NewReader(buffer)
// Read global header
header := new(perfDataBlock)
err = header.BinaryReadFrom(r)
if err != nil {
return nil, fmt.Errorf("failed to read performance data block for %q with: %w", query, err)
}
// Check for "PERF" signature
if header.Signature != [4]uint16{80, 69, 82, 70} {
panic("Invalid performance block header")
}
// Parse the performance data
numObjects := int(header.NumObjectTypes)
objects := make([]*PerfObject, numObjects)
objOffset := int64(header.HeaderLength)
for i := range numObjects {
_, err := r.Seek(objOffset, io.SeekStart)
if err != nil {
return nil, err
}
obj := new(perfObjectType)
err = obj.BinaryReadFrom(r)
if err != nil {
return nil, err
}
numCounterDefs := int(obj.NumCounters)
numInstances := int(obj.NumInstances)
// Perf objects can have no instances. The perflib differentiates
// between objects with instances and without, but we just create
// an empty instance in order to simplify the interface.
if numInstances <= 0 {
numInstances = 1
}
instances := make([]*PerfInstance, numInstances)
counterDefs := make([]*PerfCounterDef, numCounterDefs)
objects[i] = &PerfObject{
Name: obj.LookupName(),
NameIndex: uint(obj.ObjectNameTitleIndex),
Instances: instances,
CounterDefs: counterDefs,
Frequency: obj.PerfFreq,
rawData: obj,
}
for i := range numCounterDefs {
def := new(perfCounterDefinition)
err := def.BinaryReadFrom(r)
if err != nil {
return nil, err
}
counterDefs[i] = &PerfCounterDef{
Name: def.LookupName(),
NameIndex: uint(def.CounterNameTitleIndex),
rawData: def,
CounterType: def.CounterType,
IsCounter: def.CounterType&0x400 == 0x400,
IsBaseValue: def.CounterType&0x00030000 == 0x00030000,
IsNanosecondCounter: def.CounterType&0x00100000 == 0x00100000,
HasSecondValue: def.CounterType == averageCount64Type,
}
}
if obj.NumInstances <= 0 { //nolint:nestif
blockOffset := objOffset + int64(obj.DefinitionLength)
if _, err := r.Seek(blockOffset, io.SeekStart); err != nil {
return nil, err
}
_, counters, err := parseCounterBlock(buffer, r, blockOffset, counterDefs)
if err != nil {
return nil, err
}
instances[0] = &PerfInstance{
Name: "",
Counters: counters,
rawData: nil,
rawCounterBlock: nil,
}
} else {
instOffset := objOffset + int64(obj.DefinitionLength)
for i := range numInstances {
if _, err := r.Seek(instOffset, io.SeekStart); err != nil {
return nil, err
}
inst := new(perfInstanceDefinition)
if err = inst.BinaryReadFrom(r); err != nil {
return nil, err
}
name, _ := readUTF16StringAtPos(r, instOffset+int64(inst.NameOffset), inst.NameLength)
pos := instOffset + int64(inst.ByteLength)
offset, counters, err := parseCounterBlock(buffer, r, pos, counterDefs)
if err != nil {
return nil, err
}
instances[i] = &PerfInstance{
Name: name,
Counters: counters,
rawData: inst,
}
instOffset = pos + offset
}
}
// Next perfObjectType
objOffset += int64(obj.TotalByteLength)
}
return objects, nil
}
func parseCounterBlock(b []byte, r io.ReadSeeker, pos int64, defs []*PerfCounterDef) (int64, []*PerfCounter, error) {
_, err := r.Seek(pos, io.SeekStart)
if err != nil {
return 0, nil, err
}
block := new(perfCounterBlock)
err = block.BinaryReadFrom(r)
if err != nil {
return 0, nil, err
}
counters := make([]*PerfCounter, len(defs))
for i, def := range defs {
valueOffset := pos + int64(def.rawData.CounterOffset)
value := convertCounterValue(def.rawData, b, valueOffset)
secondValue := int64(0)
if def.HasSecondValue {
secondValue = convertCounterValue(def.rawData, b, valueOffset+8)
}
counters[i] = &PerfCounter{
Value: value,
Def: def,
SecondValue: secondValue,
}
}
return int64(block.ByteLength), counters, nil
}
//nolint:nonamedreturns
func convertCounterValue(counterDef *perfCounterDefinition, buffer []byte, valueOffset int64) (value int64) {
/*
We can safely ignore the type since we're not interested in anything except the raw value.
We also ignore all of the other attributes (timestamp, presentation, multi counter values...)
See also: winperf.h.
Here's the most common value for CounterType:
65536 32bit counter
65792 64bit counter
272696320 32bit rate
272696576 64bit rate
*/
switch counterDef.CounterSize {
case 4:
value = int64(bo.Uint32(buffer[valueOffset:(valueOffset + 4)]))
case 8:
value = int64(bo.Uint64(buffer[valueOffset:(valueOffset + 8)]))
default:
value = int64(bo.Uint32(buffer[valueOffset:(valueOffset + 4)]))
}
return
}