Source file src/runtime/profbuf.go

     1  // Copyright 2017 The Go Authors.  All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package runtime
     6  
     7  import (
     8  	"runtime/internal/atomic"
     9  	"unsafe"
    10  )
    11  
    12  // A profBuf is a lock-free buffer for profiling events,
    13  // safe for concurrent use by one reader and one writer.
    14  // The writer may be a signal handler running without a user g.
    15  // The reader is assumed to be a user g.
    16  //
    17  // Each logged event corresponds to a fixed size header, a list of
    18  // uintptrs (typically a stack), and exactly one unsafe.Pointer tag.
    19  // The header and uintptrs are stored in the circular buffer data and the
    20  // tag is stored in a circular buffer tags, running in parallel.
    21  // In the circular buffer data, each event takes 2+hdrsize+len(stk)
    22  // words: the value 2+hdrsize+len(stk), then the time of the event, then
    23  // hdrsize words giving the fixed-size header, and then len(stk) words
    24  // for the stack.
    25  //
    26  // The current effective offsets into the tags and data circular buffers
    27  // for reading and writing are stored in the high 30 and low 32 bits of r and w.
    28  // The bottom bits of the high 32 are additional flag bits in w, unused in r.
    29  // "Effective" offsets means the total number of reads or writes, mod 2^length.
    30  // The offset in the buffer is the effective offset mod the length of the buffer.
    31  // To make wraparound mod 2^length match wraparound mod length of the buffer,
    32  // the length of the buffer must be a power of two.
    33  //
    34  // If the reader catches up to the writer, a flag passed to read controls
    35  // whether the read blocks until more data is available. A read returns a
    36  // pointer to the buffer data itself; the caller is assumed to be done with
    37  // that data at the next read. The read offset rNext tracks the next offset to
    38  // be returned by read. By definition, r ≤ rNext ≤ w (before wraparound),
    39  // and rNext is only used by the reader, so it can be accessed without atomics.
    40  //
    41  // If the writer gets ahead of the reader, so that the buffer fills,
    42  // future writes are discarded and replaced in the output stream by an
    43  // overflow entry, which has size 2+hdrsize+1, time set to the time of
    44  // the first discarded write, a header of all zeroed words, and a "stack"
    45  // containing one word, the number of discarded writes.
    46  //
    47  // Between the time the buffer fills and the buffer becomes empty enough
    48  // to hold more data, the overflow entry is stored as a pending overflow
    49  // entry in the fields overflow and overflowTime. The pending overflow
    50  // entry can be turned into a real record by either the writer or the
    51  // reader. If the writer is called to write a new record and finds that
    52  // the output buffer has room for both the pending overflow entry and the
    53  // new record, the writer emits the pending overflow entry and the new
    54  // record into the buffer. If the reader is called to read data and finds
    55  // that the output buffer is empty but that there is a pending overflow
    56  // entry, the reader will return a synthesized record for the pending
    57  // overflow entry.
    58  //
    59  // Only the writer can create or add to a pending overflow entry, but
    60  // either the reader or the writer can clear the pending overflow entry.
    61  // A pending overflow entry is indicated by the low 32 bits of 'overflow'
    62  // holding the number of discarded writes, and overflowTime holding the
    63  // time of the first discarded write. The high 32 bits of 'overflow'
    64  // increment each time the low 32 bits transition from zero to non-zero
    65  // or vice versa. This sequence number avoids ABA problems in the use of
    66  // compare-and-swap to coordinate between reader and writer.
    67  // The overflowTime is only written when the low 32 bits of overflow are
    68  // zero, that is, only when there is no pending overflow entry, in
    69  // preparation for creating a new one. The reader can therefore fetch and
    70  // clear the entry atomically using
    71  //
    72  //	for {
    73  //		overflow = load(&b.overflow)
    74  //		if uint32(overflow) == 0 {
    75  //			// no pending entry
    76  //			break
    77  //		}
    78  //		time = load(&b.overflowTime)
    79  //		if cas(&b.overflow, overflow, ((overflow>>32)+1)<<32) {
    80  //			// pending entry cleared
    81  //			break
    82  //		}
    83  //	}
    84  //	if uint32(overflow) > 0 {
    85  //		emit entry for uint32(overflow), time
    86  //	}
    87  //
    88  type profBuf struct {
    89  	// accessed atomically
    90  	r, w         profAtomic
    91  	overflow     uint64
    92  	overflowTime uint64
    93  	eof          uint32
    94  
    95  	// immutable (excluding slice content)
    96  	hdrsize uintptr
    97  	data    []uint64
    98  	tags    []unsafe.Pointer
    99  
   100  	// owned by reader
   101  	rNext       profIndex
   102  	overflowBuf []uint64 // for use by reader to return overflow record
   103  	wait        note
   104  }
   105  
   106  // A profAtomic is the atomically-accessed word holding a profIndex.
   107  type profAtomic uint64
   108  
   109  // A profIndex is the packet tag and data counts and flags bits, described above.
   110  type profIndex uint64
   111  
   112  const (
   113  	profReaderSleeping profIndex = 1 << 32 // reader is sleeping and must be woken up
   114  	profWriteExtra     profIndex = 1 << 33 // overflow or eof waiting
   115  )
   116  
   117  func (x *profAtomic) load() profIndex {
   118  	return profIndex(atomic.Load64((*uint64)(x)))
   119  }
   120  
   121  func (x *profAtomic) store(new profIndex) {
   122  	atomic.Store64((*uint64)(x), uint64(new))
   123  }
   124  
   125  func (x *profAtomic) cas(old, new profIndex) bool {
   126  	return atomic.Cas64((*uint64)(x), uint64(old), uint64(new))
   127  }
   128  
   129  func (x profIndex) dataCount() uint32 {
   130  	return uint32(x)
   131  }
   132  
   133  func (x profIndex) tagCount() uint32 {
   134  	return uint32(x >> 34)
   135  }
   136  
   137  // countSub subtracts two counts obtained from profIndex.dataCount or profIndex.tagCount,
   138  // assuming that they are no more than 2^29 apart (guaranteed since they are never more than
   139  // len(data) or len(tags) apart, respectively).
   140  // tagCount wraps at 2^30, while dataCount wraps at 2^32.
   141  // This function works for both.
   142  func countSub(x, y uint32) int {
   143  	// x-y is 32-bit signed or 30-bit signed; sign-extend to 32 bits and convert to int.
   144  	return int(int32(x-y) << 2 >> 2)
   145  }
   146  
   147  // addCountsAndClearFlags returns the packed form of "x + (data, tag) - all flags".
   148  func (x profIndex) addCountsAndClearFlags(data, tag int) profIndex {
   149  	return profIndex((uint64(x)>>34+uint64(uint32(tag)<<2>>2))<<34 | uint64(uint32(x)+uint32(data)))
   150  }
   151  
   152  // hasOverflow reports whether b has any overflow records pending.
   153  func (b *profBuf) hasOverflow() bool {
   154  	return uint32(atomic.Load64(&b.overflow)) > 0
   155  }
   156  
   157  // takeOverflow consumes the pending overflow records, returning the overflow count
   158  // and the time of the first overflow.
   159  // When called by the reader, it is racing against incrementOverflow.
   160  func (b *profBuf) takeOverflow() (count uint32, time uint64) {
   161  	overflow := atomic.Load64(&b.overflow)
   162  	time = atomic.Load64(&b.overflowTime)
   163  	for {
   164  		count = uint32(overflow)
   165  		if count == 0 {
   166  			time = 0
   167  			break
   168  		}
   169  		// Increment generation, clear overflow count in low bits.
   170  		if atomic.Cas64(&b.overflow, overflow, ((overflow>>32)+1)<<32) {
   171  			break
   172  		}
   173  		overflow = atomic.Load64(&b.overflow)
   174  		time = atomic.Load64(&b.overflowTime)
   175  	}
   176  	return uint32(overflow), time
   177  }
   178  
   179  // incrementOverflow records a single overflow at time now.
   180  // It is racing against a possible takeOverflow in the reader.
   181  func (b *profBuf) incrementOverflow(now int64) {
   182  	for {
   183  		overflow := atomic.Load64(&b.overflow)
   184  
   185  		// Once we see b.overflow reach 0, it's stable: no one else is changing it underfoot.
   186  		// We need to set overflowTime if we're incrementing b.overflow from 0.
   187  		if uint32(overflow) == 0 {
   188  			// Store overflowTime first so it's always available when overflow != 0.
   189  			atomic.Store64(&b.overflowTime, uint64(now))
   190  			atomic.Store64(&b.overflow, (((overflow>>32)+1)<<32)+1)
   191  			break
   192  		}
   193  		// Otherwise we're racing to increment against reader
   194  		// who wants to set b.overflow to 0.
   195  		// Out of paranoia, leave 2³²-1 a sticky overflow value,
   196  		// to avoid wrapping around. Extremely unlikely.
   197  		if int32(overflow) == -1 {
   198  			break
   199  		}
   200  		if atomic.Cas64(&b.overflow, overflow, overflow+1) {
   201  			break
   202  		}
   203  	}
   204  }
   205  
   206  // newProfBuf returns a new profiling buffer with room for
   207  // a header of hdrsize words and a buffer of at least bufwords words.
   208  func newProfBuf(hdrsize, bufwords, tags int) *profBuf {
   209  	if min := 2 + hdrsize + 1; bufwords < min {
   210  		bufwords = min
   211  	}
   212  
   213  	// Buffer sizes must be power of two, so that we don't have to
   214  	// worry about uint32 wraparound changing the effective position
   215  	// within the buffers. We store 30 bits of count; limiting to 28
   216  	// gives us some room for intermediate calculations.
   217  	if bufwords >= 1<<28 || tags >= 1<<28 {
   218  		throw("newProfBuf: buffer too large")
   219  	}
   220  	var i int
   221  	for i = 1; i < bufwords; i <<= 1 {
   222  	}
   223  	bufwords = i
   224  	for i = 1; i < tags; i <<= 1 {
   225  	}
   226  	tags = i
   227  
   228  	b := new(profBuf)
   229  	b.hdrsize = uintptr(hdrsize)
   230  	b.data = make([]uint64, bufwords)
   231  	b.tags = make([]unsafe.Pointer, tags)
   232  	b.overflowBuf = make([]uint64, 2+b.hdrsize+1)
   233  	return b
   234  }
   235  
   236  // canWriteRecord reports whether the buffer has room
   237  // for a single contiguous record with a stack of length nstk.
   238  func (b *profBuf) canWriteRecord(nstk int) bool {
   239  	br := b.r.load()
   240  	bw := b.w.load()
   241  
   242  	// room for tag?
   243  	if countSub(br.tagCount(), bw.tagCount())+len(b.tags) < 1 {
   244  		return false
   245  	}
   246  
   247  	// room for data?
   248  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
   249  	want := 2 + int(b.hdrsize) + nstk
   250  	i := int(bw.dataCount() % uint32(len(b.data)))
   251  	if i+want > len(b.data) {
   252  		// Can't fit in trailing fragment of slice.
   253  		// Skip over that and start over at beginning of slice.
   254  		nd -= len(b.data) - i
   255  	}
   256  	return nd >= want
   257  }
   258  
   259  // canWriteTwoRecords reports whether the buffer has room
   260  // for two records with stack lengths nstk1, nstk2, in that order.
   261  // Each record must be contiguous on its own, but the two
   262  // records need not be contiguous (one can be at the end of the buffer
   263  // and the other can wrap around and start at the beginning of the buffer).
   264  func (b *profBuf) canWriteTwoRecords(nstk1, nstk2 int) bool {
   265  	br := b.r.load()
   266  	bw := b.w.load()
   267  
   268  	// room for tag?
   269  	if countSub(br.tagCount(), bw.tagCount())+len(b.tags) < 2 {
   270  		return false
   271  	}
   272  
   273  	// room for data?
   274  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
   275  
   276  	// first record
   277  	want := 2 + int(b.hdrsize) + nstk1
   278  	i := int(bw.dataCount() % uint32(len(b.data)))
   279  	if i+want > len(b.data) {
   280  		// Can't fit in trailing fragment of slice.
   281  		// Skip over that and start over at beginning of slice.
   282  		nd -= len(b.data) - i
   283  		i = 0
   284  	}
   285  	i += want
   286  	nd -= want
   287  
   288  	// second record
   289  	want = 2 + int(b.hdrsize) + nstk2
   290  	if i+want > len(b.data) {
   291  		// Can't fit in trailing fragment of slice.
   292  		// Skip over that and start over at beginning of slice.
   293  		nd -= len(b.data) - i
   294  		i = 0
   295  	}
   296  	return nd >= want
   297  }
   298  
   299  // write writes an entry to the profiling buffer b.
   300  // The entry begins with a fixed hdr, which must have
   301  // length b.hdrsize, followed by a variable-sized stack
   302  // and a single tag pointer *tagPtr (or nil if tagPtr is nil).
   303  // No write barriers allowed because this might be called from a signal handler.
   304  func (b *profBuf) write(tagPtr *unsafe.Pointer, now int64, hdr []uint64, stk []uintptr) {
   305  	if b == nil {
   306  		return
   307  	}
   308  	if len(hdr) > int(b.hdrsize) {
   309  		throw("misuse of profBuf.write")
   310  	}
   311  
   312  	if hasOverflow := b.hasOverflow(); hasOverflow && b.canWriteTwoRecords(1, len(stk)) {
   313  		// Room for both an overflow record and the one being written.
   314  		// Write the overflow record if the reader hasn't gotten to it yet.
   315  		// Only racing against reader, not other writers.
   316  		count, time := b.takeOverflow()
   317  		if count > 0 {
   318  			var stk [1]uintptr
   319  			stk[0] = uintptr(count)
   320  			b.write(nil, int64(time), nil, stk[:])
   321  		}
   322  	} else if hasOverflow || !b.canWriteRecord(len(stk)) {
   323  		// Pending overflow without room to write overflow and new records
   324  		// or no overflow but also no room for new record.
   325  		b.incrementOverflow(now)
   326  		b.wakeupExtra()
   327  		return
   328  	}
   329  
   330  	// There's room: write the record.
   331  	br := b.r.load()
   332  	bw := b.w.load()
   333  
   334  	// Profiling tag
   335  	//
   336  	// The tag is a pointer, but we can't run a write barrier here.
   337  	// We have interrupted the OS-level execution of gp, but the
   338  	// runtime still sees gp as executing. In effect, we are running
   339  	// in place of the real gp. Since gp is the only goroutine that
   340  	// can overwrite gp.labels, the value of gp.labels is stable during
   341  	// this signal handler: it will still be reachable from gp when
   342  	// we finish executing. If a GC is in progress right now, it must
   343  	// keep gp.labels alive, because gp.labels is reachable from gp.
   344  	// If gp were to overwrite gp.labels, the deletion barrier would
   345  	// still shade that pointer, which would preserve it for the
   346  	// in-progress GC, so all is well. Any future GC will see the
   347  	// value we copied when scanning b.tags (heap-allocated).
   348  	// We arrange that the store here is always overwriting a nil,
   349  	// so there is no need for a deletion barrier on b.tags[wt].
   350  	wt := int(bw.tagCount() % uint32(len(b.tags)))
   351  	if tagPtr != nil {
   352  		*(*uintptr)(unsafe.Pointer(&b.tags[wt])) = uintptr(unsafe.Pointer(*tagPtr))
   353  	}
   354  
   355  	// Main record.
   356  	// It has to fit in a contiguous section of the slice, so if it doesn't fit at the end,
   357  	// leave a rewind marker (0) and start over at the beginning of the slice.
   358  	wd := int(bw.dataCount() % uint32(len(b.data)))
   359  	nd := countSub(br.dataCount(), bw.dataCount()) + len(b.data)
   360  	skip := 0
   361  	if wd+2+int(b.hdrsize)+len(stk) > len(b.data) {
   362  		b.data[wd] = 0
   363  		skip = len(b.data) - wd
   364  		nd -= skip
   365  		wd = 0
   366  	}
   367  	data := b.data[wd:]
   368  	data[0] = uint64(2 + b.hdrsize + uintptr(len(stk))) // length
   369  	data[1] = uint64(now)                               // time stamp
   370  	// header, zero-padded
   371  	i := uintptr(copy(data[2:2+b.hdrsize], hdr))
   372  	for ; i < b.hdrsize; i++ {
   373  		data[2+i] = 0
   374  	}
   375  	for i, pc := range stk {
   376  		data[2+b.hdrsize+uintptr(i)] = uint64(pc)
   377  	}
   378  
   379  	for {
   380  		// Commit write.
   381  		// Racing with reader setting flag bits in b.w, to avoid lost wakeups.
   382  		old := b.w.load()
   383  		new := old.addCountsAndClearFlags(skip+2+len(stk)+int(b.hdrsize), 1)
   384  		if !b.w.cas(old, new) {
   385  			continue
   386  		}
   387  		// If there was a reader, wake it up.
   388  		if old&profReaderSleeping != 0 {
   389  			notewakeup(&b.wait)
   390  		}
   391  		break
   392  	}
   393  }
   394  
   395  // close signals that there will be no more writes on the buffer.
   396  // Once all the data has been read from the buffer, reads will return eof=true.
   397  func (b *profBuf) close() {
   398  	if atomic.Load(&b.eof) > 0 {
   399  		throw("runtime: profBuf already closed")
   400  	}
   401  	atomic.Store(&b.eof, 1)
   402  	b.wakeupExtra()
   403  }
   404  
   405  // wakeupExtra must be called after setting one of the "extra"
   406  // atomic fields b.overflow or b.eof.
   407  // It records the change in b.w and wakes up the reader if needed.
   408  func (b *profBuf) wakeupExtra() {
   409  	for {
   410  		old := b.w.load()
   411  		new := old | profWriteExtra
   412  		if !b.w.cas(old, new) {
   413  			continue
   414  		}
   415  		if old&profReaderSleeping != 0 {
   416  			notewakeup(&b.wait)
   417  		}
   418  		break
   419  	}
   420  }
   421  
   422  // profBufReadMode specifies whether to block when no data is available to read.
   423  type profBufReadMode int
   424  
   425  const (
   426  	profBufBlocking profBufReadMode = iota
   427  	profBufNonBlocking
   428  )
   429  
   430  var overflowTag [1]unsafe.Pointer // always nil
   431  
   432  func (b *profBuf) read(mode profBufReadMode) (data []uint64, tags []unsafe.Pointer, eof bool) {
   433  	if b == nil {
   434  		return nil, nil, true
   435  	}
   436  
   437  	br := b.rNext
   438  
   439  	// Commit previous read, returning that part of the ring to the writer.
   440  	// First clear tags that have now been read, both to avoid holding
   441  	// up the memory they point at for longer than necessary
   442  	// and so that b.write can assume it is always overwriting
   443  	// nil tag entries (see comment in b.write).
   444  	rPrev := b.r.load()
   445  	if rPrev != br {
   446  		ntag := countSub(br.tagCount(), rPrev.tagCount())
   447  		ti := int(rPrev.tagCount() % uint32(len(b.tags)))
   448  		for i := 0; i < ntag; i++ {
   449  			b.tags[ti] = nil
   450  			if ti++; ti == len(b.tags) {
   451  				ti = 0
   452  			}
   453  		}
   454  		b.r.store(br)
   455  	}
   456  
   457  Read:
   458  	bw := b.w.load()
   459  	numData := countSub(bw.dataCount(), br.dataCount())
   460  	if numData == 0 {
   461  		if b.hasOverflow() {
   462  			// No data to read, but there is overflow to report.
   463  			// Racing with writer flushing b.overflow into a real record.
   464  			count, time := b.takeOverflow()
   465  			if count == 0 {
   466  				// Lost the race, go around again.
   467  				goto Read
   468  			}
   469  			// Won the race, report overflow.
   470  			dst := b.overflowBuf
   471  			dst[0] = uint64(2 + b.hdrsize + 1)
   472  			dst[1] = uint64(time)
   473  			for i := uintptr(0); i < b.hdrsize; i++ {
   474  				dst[2+i] = 0
   475  			}
   476  			dst[2+b.hdrsize] = uint64(count)
   477  			return dst[:2+b.hdrsize+1], overflowTag[:1], false
   478  		}
   479  		if atomic.Load(&b.eof) > 0 {
   480  			// No data, no overflow, EOF set: done.
   481  			return nil, nil, true
   482  		}
   483  		if bw&profWriteExtra != 0 {
   484  			// Writer claims to have published extra information (overflow or eof).
   485  			// Attempt to clear notification and then check again.
   486  			// If we fail to clear the notification it means b.w changed,
   487  			// so we still need to check again.
   488  			b.w.cas(bw, bw&^profWriteExtra)
   489  			goto Read
   490  		}
   491  
   492  		// Nothing to read right now.
   493  		// Return or sleep according to mode.
   494  		if mode == profBufNonBlocking {
   495  			return nil, nil, false
   496  		}
   497  		if !b.w.cas(bw, bw|profReaderSleeping) {
   498  			goto Read
   499  		}
   500  		// Committed to sleeping.
   501  		notetsleepg(&b.wait, -1)
   502  		noteclear(&b.wait)
   503  		goto Read
   504  	}
   505  	data = b.data[br.dataCount()%uint32(len(b.data)):]
   506  	if len(data) > numData {
   507  		data = data[:numData]
   508  	} else {
   509  		numData -= len(data) // available in case of wraparound
   510  	}
   511  	skip := 0
   512  	if data[0] == 0 {
   513  		// Wraparound record. Go back to the beginning of the ring.
   514  		skip = len(data)
   515  		data = b.data
   516  		if len(data) > numData {
   517  			data = data[:numData]
   518  		}
   519  	}
   520  
   521  	ntag := countSub(bw.tagCount(), br.tagCount())
   522  	if ntag == 0 {
   523  		throw("runtime: malformed profBuf buffer - tag and data out of sync")
   524  	}
   525  	tags = b.tags[br.tagCount()%uint32(len(b.tags)):]
   526  	if len(tags) > ntag {
   527  		tags = tags[:ntag]
   528  	}
   529  
   530  	// Count out whole data records until either data or tags is done.
   531  	// They are always in sync in the buffer, but due to an end-of-slice
   532  	// wraparound we might need to stop early and return the rest
   533  	// in the next call.
   534  	di := 0
   535  	ti := 0
   536  	for di < len(data) && data[di] != 0 && ti < len(tags) {
   537  		if uintptr(di)+uintptr(data[di]) > uintptr(len(data)) {
   538  			throw("runtime: malformed profBuf buffer - invalid size")
   539  		}
   540  		di += int(data[di])
   541  		ti++
   542  	}
   543  
   544  	// Remember how much we returned, to commit read on next call.
   545  	b.rNext = br.addCountsAndClearFlags(skip+di, ti)
   546  
   547  	if raceenabled {
   548  		// Match racereleasemerge in runtime_setProfLabel,
   549  		// so that the setting of the labels in runtime_setProfLabel
   550  		// is treated as happening before any use of the labels
   551  		// by our caller. The synchronization on labelSync itself is a fiction
   552  		// for the race detector. The actual synchronization is handled
   553  		// by the fact that the signal handler only reads from the current
   554  		// goroutine and uses atomics to write the updated queue indices,
   555  		// and then the read-out from the signal handler buffer uses
   556  		// atomics to read those queue indices.
   557  		raceacquire(unsafe.Pointer(&labelSync))
   558  	}
   559  
   560  	return data[:di], tags[:ti], false
   561  }
   562  

View as plain text