Run Format

Source file src/runtime/mstats.go

Documentation: runtime

  // Copyright 2009 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  // Memory statistics
  package runtime
  import (
  // Statistics.
  // If you edit this structure, also edit type MemStats below.
  // Their layouts must match exactly.
  // For detailed descriptions see the documentation for MemStats.
  // Fields that differ from MemStats are further documented here.
  // Many of these fields are updated on the fly, while others are only
  // updated when updatememstats is called.
  type mstats struct {
  	// General statistics.
  	alloc       uint64 // bytes allocated and not yet freed
  	total_alloc uint64 // bytes allocated (even if freed)
  	sys         uint64 // bytes obtained from system (should be sum of xxx_sys below, no locking, approximate)
  	nlookup     uint64 // number of pointer lookups (unused)
  	nmalloc     uint64 // number of mallocs
  	nfree       uint64 // number of frees
  	// Statistics about malloc heap.
  	// Protected by mheap.lock
  	// Like MemStats, heap_sys and heap_inuse do not count memory
  	// in manually-managed spans.
  	heap_alloc    uint64 // bytes allocated and not yet freed (same as alloc above)
  	heap_sys      uint64 // virtual address space obtained from system for GC'd heap
  	heap_idle     uint64 // bytes in idle spans
  	heap_inuse    uint64 // bytes in _MSpanInUse spans
  	heap_released uint64 // bytes released to the os
  	heap_objects  uint64 // total number of allocated objects
  	// TODO(austin): heap_released is both useless and inaccurate
  	// in its current form. It's useless because, from the user's
  	// and OS's perspectives, there's no difference between a page
  	// that has not yet been faulted in and a page that has been
  	// released back to the OS. We could fix this by considering
  	// newly mapped spans to be "released". It's inaccurate
  	// because when we split a large span for allocation, we
  	// "unrelease" all pages in the large span and not just the
  	// ones we split off for use. This is trickier to fix because
  	// we currently don't know which pages of a span we've
  	// released. We could fix it by separating "free" and
  	// "released" spans, but then we have to allocate from runs of
  	// free and released spans.
  	// Statistics about allocation of low-level fixed-size structures.
  	// Protected by FixAlloc locks.
  	stacks_inuse uint64 // bytes in manually-managed stack spans
  	stacks_sys   uint64 // only counts newosproc0 stack in mstats; differs from MemStats.StackSys
  	mspan_inuse  uint64 // mspan structures
  	mspan_sys    uint64
  	mcache_inuse uint64 // mcache structures
  	mcache_sys   uint64
  	buckhash_sys uint64 // profiling bucket hash table
  	gc_sys       uint64
  	other_sys    uint64
  	// Statistics about garbage collector.
  	// Protected by mheap or stopping the world during GC.
  	next_gc         uint64 // goal heap_live for when next GC ends; ^0 if disabled
  	last_gc_unix    uint64 // last gc (in unix time)
  	pause_total_ns  uint64
  	pause_ns        [256]uint64 // circular buffer of recent gc pause lengths
  	pause_end       [256]uint64 // circular buffer of recent gc end times (nanoseconds since 1970)
  	numgc           uint32
  	numforcedgc     uint32  // number of user-forced GCs
  	gc_cpu_fraction float64 // fraction of CPU time used by GC
  	enablegc        bool
  	debuggc         bool
  	// Statistics about allocation size classes.
  	by_size [_NumSizeClasses]struct {
  		size    uint32
  		nmalloc uint64
  		nfree   uint64
  	// Statistics below here are not exported to MemStats directly.
  	last_gc_nanotime uint64 // last gc (monotonic time)
  	tinyallocs       uint64 // number of tiny allocations that didn't cause actual allocation; not exported to go directly
  	// triggerRatio is the heap growth ratio that triggers marking.
  	// E.g., if this is 0.6, then GC should start when the live
  	// heap has reached 1.6 times the heap size marked by the
  	// previous cycle. This should be ≤ GOGC/100 so the trigger
  	// heap size is less than the goal heap size. This is set
  	// during mark termination for the next cycle's trigger.
  	triggerRatio float64
  	// gc_trigger is the heap size that triggers marking.
  	// When heap_live ≥ gc_trigger, the mark phase will start.
  	// This is also the heap size by which proportional sweeping
  	// must be complete.
  	// This is computed from triggerRatio during mark termination
  	// for the next cycle's trigger.
  	gc_trigger uint64
  	// heap_live is the number of bytes considered live by the GC.
  	// That is: retained by the most recent GC plus allocated
  	// since then. heap_live <= heap_alloc, since heap_alloc
  	// includes unmarked objects that have not yet been swept (and
  	// hence goes up as we allocate and down as we sweep) while
  	// heap_live excludes these objects (and hence only goes up
  	// between GCs).
  	// This is updated atomically without locking. To reduce
  	// contention, this is updated only when obtaining a span from
  	// an mcentral and at this point it counts all of the
  	// unallocated slots in that span (which will be allocated
  	// before that mcache obtains another span from that
  	// mcentral). Hence, it slightly overestimates the "true" live
  	// heap size. It's better to overestimate than to
  	// underestimate because 1) this triggers the GC earlier than
  	// necessary rather than potentially too late and 2) this
  	// leads to a conservative GC rate rather than a GC rate that
  	// is potentially too low.
  	// Reads should likewise be atomic (or during STW).
  	// Whenever this is updated, call traceHeapAlloc() and
  	// gcController.revise().
  	heap_live uint64
  	// heap_scan is the number of bytes of "scannable" heap. This
  	// is the live heap (as counted by heap_live), but omitting
  	// no-scan objects and no-scan tails of objects.
  	// Whenever this is updated, call gcController.revise().
  	heap_scan uint64
  	// heap_marked is the number of bytes marked by the previous
  	// GC. After mark termination, heap_live == heap_marked, but
  	// unlike heap_live, heap_marked does not change until the
  	// next mark termination.
  	heap_marked uint64
  var memstats mstats
  // A MemStats records statistics about the memory allocator.
  type MemStats struct {
  	// General statistics.
  	// Alloc is bytes of allocated heap objects.
  	// This is the same as HeapAlloc (see below).
  	Alloc uint64
  	// TotalAlloc is cumulative bytes allocated for heap objects.
  	// TotalAlloc increases as heap objects are allocated, but
  	// unlike Alloc and HeapAlloc, it does not decrease when
  	// objects are freed.
  	TotalAlloc uint64
  	// Sys is the total bytes of memory obtained from the OS.
  	// Sys is the sum of the XSys fields below. Sys measures the
  	// virtual address space reserved by the Go runtime for the
  	// heap, stacks, and other internal data structures. It's
  	// likely that not all of the virtual address space is backed
  	// by physical memory at any given moment, though in general
  	// it all was at some point.
  	Sys uint64
  	// Lookups is the number of pointer lookups performed by the
  	// runtime.
  	// This is primarily useful for debugging runtime internals.
  	Lookups uint64
  	// Mallocs is the cumulative count of heap objects allocated.
  	// The number of live objects is Mallocs - Frees.
  	Mallocs uint64
  	// Frees is the cumulative count of heap objects freed.
  	Frees uint64
  	// Heap memory statistics.
  	// Interpreting the heap statistics requires some knowledge of
  	// how Go organizes memory. Go divides the virtual address
  	// space of the heap into "spans", which are contiguous
  	// regions of memory 8K or larger. A span may be in one of
  	// three states:
  	// An "idle" span contains no objects or other data. The
  	// physical memory backing an idle span can be released back
  	// to the OS (but the virtual address space never is), or it
  	// can be converted into an "in use" or "stack" span.
  	// An "in use" span contains at least one heap object and may
  	// have free space available to allocate more heap objects.
  	// A "stack" span is used for goroutine stacks. Stack spans
  	// are not considered part of the heap. A span can change
  	// between heap and stack memory; it is never used for both
  	// simultaneously.
  	// HeapAlloc is bytes of allocated heap objects.
  	// "Allocated" heap objects include all reachable objects, as
  	// well as unreachable objects that the garbage collector has
  	// not yet freed. Specifically, HeapAlloc increases as heap
  	// objects are allocated and decreases as the heap is swept
  	// and unreachable objects are freed. Sweeping occurs
  	// incrementally between GC cycles, so these two processes
  	// occur simultaneously, and as a result HeapAlloc tends to
  	// change smoothly (in contrast with the sawtooth that is
  	// typical of stop-the-world garbage collectors).
  	HeapAlloc uint64
  	// HeapSys is bytes of heap memory obtained from the OS.
  	// HeapSys measures the amount of virtual address space
  	// reserved for the heap. This includes virtual address space
  	// that has been reserved but not yet used, which consumes no
  	// physical memory, but tends to be small, as well as virtual
  	// address space for which the physical memory has been
  	// returned to the OS after it became unused (see HeapReleased
  	// for a measure of the latter).
  	// HeapSys estimates the largest size the heap has had.
  	HeapSys uint64
  	// HeapIdle is bytes in idle (unused) spans.
  	// Idle spans have no objects in them. These spans could be
  	// (and may already have been) returned to the OS, or they can
  	// be reused for heap allocations, or they can be reused as
  	// stack memory.
  	// HeapIdle minus HeapReleased estimates the amount of memory
  	// that could be returned to the OS, but is being retained by
  	// the runtime so it can grow the heap without requesting more
  	// memory from the OS. If this difference is significantly
  	// larger than the heap size, it indicates there was a recent
  	// transient spike in live heap size.
  	HeapIdle uint64
  	// HeapInuse is bytes in in-use spans.
  	// In-use spans have at least one object in them. These spans
  	// can only be used for other objects of roughly the same
  	// size.
  	// HeapInuse minus HeapAlloc estimates the amount of memory
  	// that has been dedicated to particular size classes, but is
  	// not currently being used. This is an upper bound on
  	// fragmentation, but in general this memory can be reused
  	// efficiently.
  	HeapInuse uint64
  	// HeapReleased is bytes of physical memory returned to the OS.
  	// This counts heap memory from idle spans that was returned
  	// to the OS and has not yet been reacquired for the heap.
  	HeapReleased uint64
  	// HeapObjects is the number of allocated heap objects.
  	// Like HeapAlloc, this increases as objects are allocated and
  	// decreases as the heap is swept and unreachable objects are
  	// freed.
  	HeapObjects uint64
  	// Stack memory statistics.
  	// Stacks are not considered part of the heap, but the runtime
  	// can reuse a span of heap memory for stack memory, and
  	// vice-versa.
  	// StackInuse is bytes in stack spans.
  	// In-use stack spans have at least one stack in them. These
  	// spans can only be used for other stacks of the same size.
  	// There is no StackIdle because unused stack spans are
  	// returned to the heap (and hence counted toward HeapIdle).
  	StackInuse uint64
  	// StackSys is bytes of stack memory obtained from the OS.
  	// StackSys is StackInuse, plus any memory obtained directly
  	// from the OS for OS thread stacks (which should be minimal).
  	StackSys uint64
  	// Off-heap memory statistics.
  	// The following statistics measure runtime-internal
  	// structures that are not allocated from heap memory (usually
  	// because they are part of implementing the heap). Unlike
  	// heap or stack memory, any memory allocated to these
  	// structures is dedicated to these structures.
  	// These are primarily useful for debugging runtime memory
  	// overheads.
  	// MSpanInuse is bytes of allocated mspan structures.
  	MSpanInuse uint64
  	// MSpanSys is bytes of memory obtained from the OS for mspan
  	// structures.
  	MSpanSys uint64
  	// MCacheInuse is bytes of allocated mcache structures.
  	MCacheInuse uint64
  	// MCacheSys is bytes of memory obtained from the OS for
  	// mcache structures.
  	MCacheSys uint64
  	// BuckHashSys is bytes of memory in profiling bucket hash tables.
  	BuckHashSys uint64
  	// GCSys is bytes of memory in garbage collection metadata.
  	GCSys uint64
  	// OtherSys is bytes of memory in miscellaneous off-heap
  	// runtime allocations.
  	OtherSys uint64
  	// Garbage collector statistics.
  	// NextGC is the target heap size of the next GC cycle.
  	// The garbage collector's goal is to keep HeapAlloc ≤ NextGC.
  	// At the end of each GC cycle, the target for the next cycle
  	// is computed based on the amount of reachable data and the
  	// value of GOGC.
  	NextGC uint64
  	// LastGC is the time the last garbage collection finished, as
  	// nanoseconds since 1970 (the UNIX epoch).
  	LastGC uint64
  	// PauseTotalNs is the cumulative nanoseconds in GC
  	// stop-the-world pauses since the program started.
  	// During a stop-the-world pause, all goroutines are paused
  	// and only the garbage collector can run.
  	PauseTotalNs uint64
  	// PauseNs is a circular buffer of recent GC stop-the-world
  	// pause times in nanoseconds.
  	// The most recent pause is at PauseNs[(NumGC+255)%256]. In
  	// general, PauseNs[N%256] records the time paused in the most
  	// recent N%256th GC cycle. There may be multiple pauses per
  	// GC cycle; this is the sum of all pauses during a cycle.
  	PauseNs [256]uint64
  	// PauseEnd is a circular buffer of recent GC pause end times,
  	// as nanoseconds since 1970 (the UNIX epoch).
  	// This buffer is filled the same way as PauseNs. There may be
  	// multiple pauses per GC cycle; this records the end of the
  	// last pause in a cycle.
  	PauseEnd [256]uint64
  	// NumGC is the number of completed GC cycles.
  	NumGC uint32
  	// NumForcedGC is the number of GC cycles that were forced by
  	// the application calling the GC function.
  	NumForcedGC uint32
  	// GCCPUFraction is the fraction of this program's available
  	// CPU time used by the GC since the program started.
  	// GCCPUFraction is expressed as a number between 0 and 1,
  	// where 0 means GC has consumed none of this program's CPU. A
  	// program's available CPU time is defined as the integral of
  	// GOMAXPROCS since the program started. That is, if
  	// GOMAXPROCS is 2 and a program has been running for 10
  	// seconds, its "available CPU" is 20 seconds. GCCPUFraction
  	// does not include CPU time used for write barrier activity.
  	// This is the same as the fraction of CPU reported by
  	// GODEBUG=gctrace=1.
  	GCCPUFraction float64
  	// EnableGC indicates that GC is enabled. It is always true,
  	// even if GOGC=off.
  	EnableGC bool
  	// DebugGC is currently unused.
  	DebugGC bool
  	// BySize reports per-size class allocation statistics.
  	// BySize[N] gives statistics for allocations of size S where
  	// BySize[N-1].Size < S ≤ BySize[N].Size.
  	// This does not report allocations larger than BySize[60].Size.
  	BySize [61]struct {
  		// Size is the maximum byte size of an object in this
  		// size class.
  		Size uint32
  		// Mallocs is the cumulative count of heap objects
  		// allocated in this size class. The cumulative bytes
  		// of allocation is Size*Mallocs. The number of live
  		// objects in this size class is Mallocs - Frees.
  		Mallocs uint64
  		// Frees is the cumulative count of heap objects freed
  		// in this size class.
  		Frees uint64
  // Size of the trailing by_size array differs between mstats and MemStats,
  // and all data after by_size is local to runtime, not exported.
  // NumSizeClasses was changed, but we cannot change MemStats because of backward compatibility.
  // sizeof_C_MStats is the size of the prefix of mstats that
  // corresponds to MemStats. It should match Sizeof(MemStats{}).
  var sizeof_C_MStats = unsafe.Offsetof(memstats.by_size) + 61*unsafe.Sizeof(memstats.by_size[0])
  func init() {
  	var memStats MemStats
  	if sizeof_C_MStats != unsafe.Sizeof(memStats) {
  		println(sizeof_C_MStats, unsafe.Sizeof(memStats))
  		throw("MStats vs MemStatsType size mismatch")
  	if unsafe.Offsetof(memstats.heap_live)%8 != 0 {
  		throw("memstats.heap_live not aligned to 8 bytes")
  // ReadMemStats populates m with memory allocator statistics.
  // The returned memory allocator statistics are up to date as of the
  // call to ReadMemStats. This is in contrast with a heap profile,
  // which is a snapshot as of the most recently completed garbage
  // collection cycle.
  func ReadMemStats(m *MemStats) {
  	stopTheWorld("read mem stats")
  	systemstack(func() {
  func readmemstats_m(stats *MemStats) {
  	// The size of the trailing by_size array differs between
  	// mstats and MemStats. NumSizeClasses was changed, but we
  	// cannot change MemStats because of backward compatibility.
  	memmove(unsafe.Pointer(stats), unsafe.Pointer(&memstats), sizeof_C_MStats)
  	// memstats.stacks_sys is only memory mapped directly for OS stacks.
  	// Add in heap-allocated stack memory for user consumption.
  	stats.StackSys += stats.StackInuse
  //go:linkname readGCStats runtime/debug.readGCStats
  func readGCStats(pauses *[]uint64) {
  	systemstack(func() {
  func readGCStats_m(pauses *[]uint64) {
  	p := *pauses
  	// Calling code in runtime/debug should make the slice large enough.
  	if cap(p) < len(memstats.pause_ns)+3 {
  		throw("short slice passed to readGCStats")
  	// Pass back: pauses, pause ends, last gc (absolute time), number of gc, total pause ns.
  	n := memstats.numgc
  	if n > uint32(len(memstats.pause_ns)) {
  		n = uint32(len(memstats.pause_ns))
  	// The pause buffer is circular. The most recent pause is at
  	// pause_ns[(numgc-1)%len(pause_ns)], and then backward
  	// from there to go back farther in time. We deliver the times
  	// most recent first (in p[0]).
  	p = p[:cap(p)]
  	for i := uint32(0); i < n; i++ {
  		j := (memstats.numgc - 1 - i) % uint32(len(memstats.pause_ns))
  		p[i] = memstats.pause_ns[j]
  		p[n+i] = memstats.pause_end[j]
  	p[n+n] = memstats.last_gc_unix
  	p[n+n+1] = uint64(memstats.numgc)
  	p[n+n+2] = memstats.pause_total_ns
  	*pauses = p[:n+n+3]
  func updatememstats() {
  	memstats.mcache_inuse = uint64(mheap_.cachealloc.inuse)
  	memstats.mspan_inuse = uint64(mheap_.spanalloc.inuse)
  	memstats.sys = memstats.heap_sys + memstats.stacks_sys + memstats.mspan_sys +
  		memstats.mcache_sys + memstats.buckhash_sys + memstats.gc_sys + memstats.other_sys
  	// We also count stacks_inuse as sys memory.
  	memstats.sys += memstats.stacks_inuse
  	// Calculate memory allocator stats.
  	// During program execution we only count number of frees and amount of freed memory.
  	// Current number of alive object in the heap and amount of alive heap memory
  	// are calculated by scanning all spans.
  	// Total number of mallocs is calculated as number of frees plus number of alive objects.
  	// Similarly, total amount of allocated memory is calculated as amount of freed memory
  	// plus amount of alive heap memory.
  	memstats.alloc = 0
  	memstats.total_alloc = 0
  	memstats.nmalloc = 0
  	memstats.nfree = 0
  	for i := 0; i < len(memstats.by_size); i++ {
  		memstats.by_size[i].nmalloc = 0
  		memstats.by_size[i].nfree = 0
  	// Flush MCache's to MCentral.
  	// Aggregate local stats.
  	// Collect allocation stats. This is safe and consistent
  	// because the world is stopped.
  	var smallFree, totalAlloc, totalFree uint64
  	// Collect per-spanclass stats.
  	for spc := range mheap_.central {
  		// The mcaches are now empty, so mcentral stats are
  		// up-to-date.
  		c := &mheap_.central[spc].mcentral
  		memstats.nmalloc += c.nmalloc
  		i := spanClass(spc).sizeclass()
  		memstats.by_size[i].nmalloc += c.nmalloc
  		totalAlloc += c.nmalloc * uint64(class_to_size[i])
  	// Collect per-sizeclass stats.
  	for i := 0; i < _NumSizeClasses; i++ {
  		if i == 0 {
  			memstats.nmalloc += mheap_.nlargealloc
  			totalAlloc += mheap_.largealloc
  			totalFree += mheap_.largefree
  			memstats.nfree += mheap_.nlargefree
  		// The mcache stats have been flushed to mheap_.
  		memstats.nfree += mheap_.nsmallfree[i]
  		memstats.by_size[i].nfree = mheap_.nsmallfree[i]
  		smallFree += mheap_.nsmallfree[i] * uint64(class_to_size[i])
  	totalFree += smallFree
  	memstats.nfree += memstats.tinyallocs
  	memstats.nmalloc += memstats.tinyallocs
  	// Calculate derived stats.
  	memstats.total_alloc = totalAlloc
  	memstats.alloc = totalAlloc - totalFree
  	memstats.heap_alloc = memstats.alloc
  	memstats.heap_objects = memstats.nmalloc - memstats.nfree
  // cachestats flushes all mcache stats.
  // The world must be stopped.
  func cachestats() {
  	for _, p := range allp {
  		c := p.mcache
  		if c == nil {
  // flushmcache flushes the mcache of allp[i].
  // The world must be stopped.
  func flushmcache(i int) {
  	p := allp[i]
  	c := p.mcache
  	if c == nil {
  // flushallmcaches flushes the mcaches of all Ps.
  // The world must be stopped.
  func flushallmcaches() {
  	for i := 0; i < int(gomaxprocs); i++ {
  func purgecachedstats(c *mcache) {
  	// Protected by either heap or GC lock.
  	h := &mheap_
  	memstats.heap_scan += uint64(c.local_scan)
  	c.local_scan = 0
  	memstats.tinyallocs += uint64(c.local_tinyallocs)
  	c.local_tinyallocs = 0
  	h.largefree += uint64(c.local_largefree)
  	c.local_largefree = 0
  	h.nlargefree += uint64(c.local_nlargefree)
  	c.local_nlargefree = 0
  	for i := 0; i < len(c.local_nsmallfree); i++ {
  		h.nsmallfree[i] += uint64(c.local_nsmallfree[i])
  		c.local_nsmallfree[i] = 0
  // Atomically increases a given *system* memory stat. We are counting on this
  // stat never overflowing a uintptr, so this function must only be used for
  // system memory stats.
  // The current implementation for little endian architectures is based on
  // xadduintptr(), which is less than ideal: xadd64() should really be used.
  // Using xadduintptr() is a stop-gap solution until arm supports xadd64() that
  // doesn't use locks.  (Locks are a problem as they require a valid G, which
  // restricts their useability.)
  // A side-effect of using xadduintptr() is that we need to check for
  // overflow errors.
  func mSysStatInc(sysStat *uint64, n uintptr) {
  	if sysStat == nil {
  	if sys.BigEndian {
  		atomic.Xadd64(sysStat, int64(n))
  	if val := atomic.Xadduintptr((*uintptr)(unsafe.Pointer(sysStat)), n); val < n {
  		print("runtime: stat overflow: val ", val, ", n ", n, "\n")
  // Atomically decreases a given *system* memory stat. Same comments as
  // mSysStatInc apply.
  func mSysStatDec(sysStat *uint64, n uintptr) {
  	if sysStat == nil {
  	if sys.BigEndian {
  		atomic.Xadd64(sysStat, -int64(n))
  	if val := atomic.Xadduintptr((*uintptr)(unsafe.Pointer(sysStat)), uintptr(-int64(n))); val+n < n {
  		print("runtime: stat underflow: val ", val, ", n ", n, "\n")

View as plain text