...
Run Format

Source file src/runtime/profbuf.go

Documentation: runtime

  // Copyright 2017 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.
  
  package runtime
  
  import (
  	"runtime/internal/atomic"
  	"unsafe"
  )
  
  // A profBuf is a lock-free buffer for profiling events,
  // safe for concurrent use by one reader and one writer.
  // The writer may be a signal handler running without a user g.
  // The reader is assumed to be a user g.
  //
  // Each logged event corresponds to a fixed size header, a list of
  // uintptrs (typically a stack), and exactly one unsafe.Pointer tag.
  // The header and uintptrs are stored in the circular buffer data and the
  // tag is stored in a circular buffer tags, running in parallel.
  // In the circular buffer data, each event takes 2+hdrsize+len(stk)
  // words: the value 2+hdrsize+len(stk), then the time of the event, then
  // hdrsize words giving the fixed-size header, and then len(stk) words
  // for the stack.
  //
  // The current effective offsets into the tags and data circular buffers
  // for reading and writing are stored in the high 30 and low 32 bits of r and w.
  // The bottom bits of the high 32 are additional flag bits in w, unused in r.
  // "Effective" offsets means the total number of reads or writes, mod 2^length.
  // The offset in the buffer is the effective offset mod the length of the buffer.
  // To make wraparound mod 2^length match wraparound mod length of the buffer,
  // the length of the buffer must be a power of two.
  //
  // If the reader catches up to the writer, a flag passed to read controls
  // whether the read blocks until more data is available. A read returns a
  // pointer to the buffer data itself; the caller is assumed to be done with
  // that data at the next read. The read offset rNext tracks the next offset to
  // be returned by read. By definition, r ≤ rNext ≤ w (before wraparound),
  // and rNext is only used by the reader, so it can be accessed without atomics.
  //
  // If the writer gets ahead of the reader, so that the buffer fills,
  // future writes are discarded and replaced in the output stream by an
  // overflow entry, which has size 2+hdrsize+1, time set to the time of
  // the first discarded write, a header of all zeroed words, and a "stack"
  // containing one word, the number of discarded writes.
  //
  // Between the time the buffer fills and the buffer becomes empty enough
  // to hold more data, the overflow entry is stored as a pending overflow
  // entry in the fields overflow and overflowTime. The pending overflow
  // entry can be turned into a real record by either the writer or the
  // reader. If the writer is called to write a new record and finds that
  // the output buffer has room for both the pending overflow entry and the
  // new record, the writer emits the pending overflow entry and the new
  // record into the buffer. If the reader is called to read data and finds
  // that the output buffer is empty but that there is a pending overflow
  // entry, the reader will return a synthesized record for the pending
  // overflow entry.
  //
  // Only the writer can create or add to a pending overflow entry, but
  // either the reader or the writer can clear the pending overflow entry.
  // A pending overflow entry is indicated by the low 32 bits of 'overflow'
  // holding the number of discarded writes, and overflowTime holding the
  // time of the first discarded write. The high 32 bits of 'overflow'
  // increment each time the low 32 bits transition from zero to non-zero
  // or vice versa. This sequence number avoids ABA problems in the use of
  // compare-and-swap to coordinate between reader and writer.
  // The overflowTime is only written when the low 32 bits of overflow are
  // zero, that is, only when there is no pending overflow entry, in
  // preparation for creating a new one. The reader can therefore fetch and
  // clear the entry atomically using
  //
  //	for {
  //		overflow = load(&b.overflow)
  //		if uint32(overflow) == 0 {
  //			// no pending entry
  //			break
  //		}
  //		time = load(&b.overflowTime)
  //		if cas(&b.overflow, overflow, ((overflow>>32)+1)<<32) {
  //			// pending entry cleared
  //			break
  //		}
  //	}
  //	if uint32(overflow) > 0 {
  //		emit entry for uint32(overflow), time
  //	}
  //
  type profBuf struct {
  	// accessed atomically
  	r, w         profAtomic
  	overflow     uint64
  	overflowTime uint64
  	eof          uint32
  
  	// immutable (excluding slice content)
  	hdrsize uintptr
  	data    []uint64
  	tags    []unsafe.Pointer
  
  	// owned by reader
  	rNext       profIndex
  	overflowBuf []uint64 // for use by reader to return overflow record
  	wait        note
  }
  
  // A profAtomic is the atomically-accessed word holding a profIndex.
  type profAtomic uint64
  
  // A profIndex is the packet tag and data counts and flags bits, described above.
  type profIndex uint64
  
  const (
  	profReaderSleeping profIndex = 1 << 32 // reader is sleeping and must be woken up
  	profWriteExtra     profIndex = 1 << 33 // overflow or eof waiting
  )
  
  func (x *profAtomic) load() profIndex {
  	return profIndex(atomic.Load64((*uint64)(x)))
  }
  
  func (x *profAtomic) store(new profIndex) {
  	atomic.Store64((*uint64)(x), uint64(new))
  }
  
  func (x *profAtomic) cas(old, new profIndex) bool {
  	return atomic.Cas64((*uint64)(x), uint64(old), uint64(new))
  }
  
  func (x profIndex) dataCount() uint32 {
  	return uint32(x)
  }
  
  func (x profIndex) tagCount() uint32 {
  	return uint32(x >> 34)
  }
  
  // countSub subtracts two counts obtained from profIndex.dataCount or profIndex.tagCount,
  // assuming that they are no more than 2^29 apart (guaranteed since they are never more than
  // len(data) or len(tags) apart, respectively).
  // tagCount wraps at 2^30, while dataCount wraps at 2^32.
  // This function works for both.
  func countSub(x, y uint32) int {
  	// x-y is 32-bit signed or 30-bit signed; sign-extend to 32 bits and convert to int.
  	return int(int32(x-y) << 2 >> 2)
  }
  
  // addCountsAndClearFlags returns the packed form of "x + (data, tag) - all flags".
  func (x profIndex) addCountsAndClearFlags(data, tag int) profIndex {
  	return profIndex((uint64(x)>>34+uint64(uint32(tag)<<2>>2))<<34 | uint64(uint32(x)+uint32(data)))
  }
  
  // hasOverflow reports whether b has any overflow records pending.
  func (b *profBuf) hasOverflow() bool {
  	return uint32(atomic.Load64(&b.overflow)) > 0
  }
  
  // takeOverflow consumes the pending overflow records, returning the overflow count
  // and the time of the first overflow.
  // When called by the reader, it is racing against incrementOverflow.
  func (b *profBuf) takeOverflow() (count uint32, time uint64) {
  	overflow := atomic.Load64(&b.overflow)
  	time = atomic.Load64(&b.overflowTime)
  	for {
  		count = uint32(overflow)
  		if count == 0 {
  			time = 0
  			break
  		}
  		// Increment generation, clear overflow count in low bits.
  		if atomic.Cas64(&b.overflow, overflow, ((overflow>>32)+1)<<32) {
  			break
  		}
  		overflow = atomic.Load64(&b.overflow)
  		time = atomic.Load64(&b.overflowTime)
  	}
  	return uint32(overflow), time
  }
  
  // incrementOverflow records a single overflow at time now.
  // It is racing against a possible takeOverflow in the reader.
  func (b *profBuf) incrementOverflow(now int64) {
  	for {
  		overflow := atomic.Load64(&b.overflow)
  
  		// Once we see b.overflow reach 0, it's stable: no one else is changing it underfoot.
  		// We need to set overflowTime if we're incrementing b.overflow from 0.
  		if uint32(overflow) == 0 {
  			// Store overflowTime first so it's always available when overflow != 0.
  			atomic.Store64(&b.overflowTime, uint64(now))
  			atomic.Store64(&b.overflow, (((overflow>>32)+1)<<32)+1)
  			break
  		}
  		// Otherwise we're racing to increment against reader
  		// who wants to set b.overflow to 0.
  		// Out of paranoia, leave 2³²-1 a sticky overflow value,
  		// to avoid wrapping around. Extremely unlikely.
  		if int32(overflow) == -1 {
  			break
  		}
  		if atomic.Cas64(&b.overflow, overflow, overflow+1) {
  			break
  		}
  	}
  }
  
  // newProfBuf returns a new profiling buffer with room for
  // a header of hdrsize words and a buffer of at least bufwords words.
  func newProfBuf(hdrsize, bufwords, tags int) *profBuf {
  	if min := 2 + hdrsize + 1; bufwords < min {
  		bufwords = min
  	}
  
  	// Buffer sizes must be power of two, so that we don't have to
  	// worry about uint32 wraparound changing the effective position
  	// within the buffers. We store 30 bits of count; limiting to 28
  	// gives us some room for intermediate calculations.
  	if bufwords >= 1<<28 || tags >= 1<<28 {
  		throw("newProfBuf: buffer too large")
  	}
  	var i int
  	for i = 1; i < bufwords; i <<= 1 {
  	}
  	bufwords = i
  	for i = 1; i < tags; i <<= 1 {
  	}
  	tags = i
  
  	b := new(profBuf)
  	b.hdrsize = uintptr(hdrsize)
  	b.data = make([]uint64, bufwords)
  	b.tags = make([]unsafe.Pointer, tags)
  	b.overflowBuf = make([]uint64, 2+b.hdrsize+1)
  	return b
  }
  
  // canWriteRecord reports whether the buffer has room
  // for a single contiguous record with a stack of length nstk.
  func (b *profBuf) canWriteRecord(nstk int) bool {
  	br := b.r.load()
  	bw := b.w.load()
  
  	// room for tag?
  	if countSub(br.tagCount(), bw.tagCount())+len(b.tags) < 1 {
  		return false
  	}
  
  	// room for data?
  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
  	want := 2 + int(b.hdrsize) + nstk
  	i := int(bw.dataCount() % uint32(len(b.data)))
  	if i+want > len(b.data) {
  		// Can't fit in trailing fragment of slice.
  		// Skip over that and start over at beginning of slice.
  		nd -= len(b.data) - i
  	}
  	return nd >= want
  }
  
  // canWriteTwoRecords reports whether the buffer has room
  // for two records with stack lengths nstk1, nstk2, in that order.
  // Each record must be contiguous on its own, but the two
  // records need not be contiguous (one can be at the end of the buffer
  // and the other can wrap around and start at the beginning of the buffer).
  func (b *profBuf) canWriteTwoRecords(nstk1, nstk2 int) bool {
  	br := b.r.load()
  	bw := b.w.load()
  
  	// room for tag?
  	if countSub(br.tagCount(), bw.tagCount())+len(b.tags) < 2 {
  		return false
  	}
  
  	// room for data?
  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
  
  	// first record
  	want := 2 + int(b.hdrsize) + nstk1
  	i := int(bw.dataCount() % uint32(len(b.data)))
  	if i+want > len(b.data) {
  		// Can't fit in trailing fragment of slice.
  		// Skip over that and start over at beginning of slice.
  		nd -= len(b.data) - i
  		i = 0
  	}
  	i += want
  	nd -= want
  
  	// second record
  	want = 2 + int(b.hdrsize) + nstk2
  	if i+want > len(b.data) {
  		// Can't fit in trailing fragment of slice.
  		// Skip over that and start over at beginning of slice.
  		nd -= len(b.data) - i
  		i = 0
  	}
  	return nd >= want
  }
  
  // write writes an entry to the profiling buffer b.
  // The entry begins with a fixed hdr, which must have
  // length b.hdrsize, followed by a variable-sized stack
  // and a single tag pointer *tagPtr (or nil if tagPtr is nil).
  // No write barriers allowed because this might be called from a signal handler.
  func (b *profBuf) write(tagPtr *unsafe.Pointer, now int64, hdr []uint64, stk []uintptr) {
  	if b == nil {
  		return
  	}
  	if len(hdr) > int(b.hdrsize) {
  		throw("misuse of profBuf.write")
  	}
  
  	if hasOverflow := b.hasOverflow(); hasOverflow && b.canWriteTwoRecords(1, len(stk)) {
  		// Room for both an overflow record and the one being written.
  		// Write the overflow record if the reader hasn't gotten to it yet.
  		// Only racing against reader, not other writers.
  		count, time := b.takeOverflow()
  		if count > 0 {
  			var stk [1]uintptr
  			stk[0] = uintptr(count)
  			b.write(nil, int64(time), nil, stk[:])
  		}
  	} else if hasOverflow || !b.canWriteRecord(len(stk)) {
  		// Pending overflow without room to write overflow and new records
  		// or no overflow but also no room for new record.
  		b.incrementOverflow(now)
  		b.wakeupExtra()
  		return
  	}
  
  	// There's room: write the record.
  	br := b.r.load()
  	bw := b.w.load()
  
  	// Profiling tag
  	//
  	// The tag is a pointer, but we can't run a write barrier here.
  	// We have interrupted the OS-level execution of gp, but the
  	// runtime still sees gp as executing. In effect, we are running
  	// in place of the real gp. Since gp is the only goroutine that
  	// can overwrite gp.labels, the value of gp.labels is stable during
  	// this signal handler: it will still be reachable from gp when
  	// we finish executing. If a GC is in progress right now, it must
  	// keep gp.labels alive, because gp.labels is reachable from gp.
  	// If gp were to overwrite gp.labels, the deletion barrier would
  	// still shade that pointer, which would preserve it for the
  	// in-progress GC, so all is well. Any future GC will see the
  	// value we copied when scanning b.tags (heap-allocated).
  	// We arrange that the store here is always overwriting a nil,
  	// so there is no need for a deletion barrier on b.tags[wt].
  	wt := int(bw.tagCount() % uint32(len(b.tags)))
  	if tagPtr != nil {
  		*(*uintptr)(unsafe.Pointer(&b.tags[wt])) = uintptr(unsafe.Pointer(*tagPtr))
  	}
  
  	// Main record.
  	// It has to fit in a contiguous section of the slice, so if it doesn't fit at the end,
  	// leave a rewind marker (0) and start over at the beginning of the slice.
  	wd := int(bw.dataCount() % uint32(len(b.data)))
  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
  	skip := 0
  	if wd+2+int(b.hdrsize)+len(stk) > len(b.data) {
  		b.data[wd] = 0
  		skip = len(b.data) - wd
  		nd -= skip
  		wd = 0
  	}
  	data := b.data[wd:]
  	data[0] = uint64(2 + b.hdrsize + uintptr(len(stk))) // length
  	data[1] = uint64(now)                               // time stamp
  	// header, zero-padded
  	i := uintptr(copy(data[2:2+b.hdrsize], hdr))
  	for ; i < b.hdrsize; i++ {
  		data[2+i] = 0
  	}
  	for i, pc := range stk {
  		data[2+b.hdrsize+uintptr(i)] = uint64(pc)
  	}
  
  	for {
  		// Commit write.
  		// Racing with reader setting flag bits in b.w, to avoid lost wakeups.
  		old := b.w.load()
  		new := old.addCountsAndClearFlags(skip+2+len(stk)+int(b.hdrsize), 1)
  		if !b.w.cas(old, new) {
  			continue
  		}
  		// If there was a reader, wake it up.
  		if old&profReaderSleeping != 0 {
  			notewakeup(&b.wait)
  		}
  		break
  	}
  }
  
  // close signals that there will be no more writes on the buffer.
  // Once all the data has been read from the buffer, reads will return eof=true.
  func (b *profBuf) close() {
  	if atomic.Load(&b.eof) > 0 {
  		throw("runtime: profBuf already closed")
  	}
  	atomic.Store(&b.eof, 1)
  	b.wakeupExtra()
  }
  
  // wakeupExtra must be called after setting one of the "extra"
  // atomic fields b.overflow or b.eof.
  // It records the change in b.w and wakes up the reader if needed.
  func (b *profBuf) wakeupExtra() {
  	for {
  		old := b.w.load()
  		new := old | profWriteExtra
  		if !b.w.cas(old, new) {
  			continue
  		}
  		if old&profReaderSleeping != 0 {
  			notewakeup(&b.wait)
  		}
  		break
  	}
  }
  
  // profBufReadMode specifies whether to block when no data is available to read.
  type profBufReadMode int
  
  const (
  	profBufBlocking profBufReadMode = iota
  	profBufNonBlocking
  )
  
  var overflowTag [1]unsafe.Pointer // always nil
  
  func (b *profBuf) read(mode profBufReadMode) (data []uint64, tags []unsafe.Pointer, eof bool) {
  	if b == nil {
  		return nil, nil, true
  	}
  
  	br := b.rNext
  
  	// Commit previous read, returning that part of the ring to the writer.
  	// First clear tags that have now been read, both to avoid holding
  	// up the memory they point at for longer than necessary
  	// and so that b.write can assume it is always overwriting
  	// nil tag entries (see comment in b.write).
  	rPrev := b.r.load()
  	if rPrev != br {
  		ntag := countSub(br.tagCount(), rPrev.tagCount())
  		ti := int(rPrev.tagCount() % uint32(len(b.tags)))
  		for i := 0; i < ntag; i++ {
  			b.tags[ti] = nil
  			if ti++; ti == len(b.tags) {
  				ti = 0
  			}
  		}
  		b.r.store(br)
  	}
  
  Read:
  	bw := b.w.load()
  	numData := countSub(bw.dataCount(), br.dataCount())
  	if numData == 0 {
  		if b.hasOverflow() {
  			// No data to read, but there is overflow to report.
  			// Racing with writer flushing b.overflow into a real record.
  			count, time := b.takeOverflow()
  			if count == 0 {
  				// Lost the race, go around again.
  				goto Read
  			}
  			// Won the race, report overflow.
  			dst := b.overflowBuf
  			dst[0] = uint64(2 + b.hdrsize + 1)
  			dst[1] = uint64(time)
  			for i := uintptr(0); i < b.hdrsize; i++ {
  				dst[2+i] = 0
  			}
  			dst[2+b.hdrsize] = uint64(count)
  			return dst[:2+b.hdrsize+1], overflowTag[:1], false
  		}
  		if atomic.Load(&b.eof) > 0 {
  			// No data, no overflow, EOF set: done.
  			return nil, nil, true
  		}
  		if bw&profWriteExtra != 0 {
  			// Writer claims to have published extra information (overflow or eof).
  			// Attempt to clear notification and then check again.
  			// If we fail to clear the notification it means b.w changed,
  			// so we still need to check again.
  			b.w.cas(bw, bw&^profWriteExtra)
  			goto Read
  		}
  
  		// Nothing to read right now.
  		// Return or sleep according to mode.
  		if mode == profBufNonBlocking {
  			return nil, nil, false
  		}
  		if !b.w.cas(bw, bw|profReaderSleeping) {
  			goto Read
  		}
  		// Committed to sleeping.
  		notetsleepg(&b.wait, -1)
  		noteclear(&b.wait)
  		goto Read
  	}
  	data = b.data[br.dataCount()%uint32(len(b.data)):]
  	if len(data) > numData {
  		data = data[:numData]
  	} else {
  		numData -= len(data) // available in case of wraparound
  	}
  	skip := 0
  	if data[0] == 0 {
  		// Wraparound record. Go back to the beginning of the ring.
  		skip = len(data)
  		data = b.data
  		if len(data) > numData {
  			data = data[:numData]
  		}
  	}
  
  	ntag := countSub(bw.tagCount(), br.tagCount())
  	if ntag == 0 {
  		throw("runtime: malformed profBuf buffer - tag and data out of sync")
  	}
  	tags = b.tags[br.tagCount()%uint32(len(b.tags)):]
  	if len(tags) > ntag {
  		tags = tags[:ntag]
  	}
  
  	// Count out whole data records until either data or tags is done.
  	// They are always in sync in the buffer, but due to an end-of-slice
  	// wraparound we might need to stop early and return the rest
  	// in the next call.
  	di := 0
  	ti := 0
  	for di < len(data) && data[di] != 0 && ti < len(tags) {
  		if uintptr(di)+uintptr(data[di]) > uintptr(len(data)) {
  			throw("runtime: malformed profBuf buffer - invalid size")
  		}
  		di += int(data[di])
  		ti++
  	}
  
  	// Remember how much we returned, to commit read on next call.
  	b.rNext = br.addCountsAndClearFlags(skip+di, ti)
  
  	if raceenabled {
  		// Match racereleasemerge in runtime_setProfLabel,
  		// so that the setting of the labels in runtime_setProfLabel
  		// is treated as happening before any use of the labels
  		// by our caller. The synchronization on labelSync itself is a fiction
  		// for the race detector. The actual synchronization is handled
  		// by the fact that the signal handler only reads from the current
  		// goroutine and uses atomics to write the updated queue indices,
  		// and then the read-out from the signal handler buffer uses
  		// atomics to read those queue indices.
  		raceacquire(unsafe.Pointer(&labelSync))
  	}
  
  	return data[:di], tags[:ti], false
  }
  

View as plain text