1 // Copyright 2011 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.
6 // Based on algorithms and data structures used in
7 // http://code.google.com/p/google-perftools/.
9 // The main difference between this code and the google-perftools
10 // code is that this code is written to allow copying the profile data
11 // to an arbitrary io.Writer, while the google-perftools code always
12 // writes to an operating system file.
14 // The signal handler for the profiling clock tick adds a new stack trace
15 // to a hash table tracking counts for recent traces. Most clock ticks
16 // hit in the cache. In the event of a cache miss, an entry must be
17 // evicted from the hash table, copied to a log that will eventually be
18 // written as profile data. The google-perftools code flushed the
19 // log itself during the signal handler. This code cannot do that, because
20 // the io.Writer might block or need system calls or locks that are not
21 // safe to use from within the signal handler. Instead, we split the log
22 // into two halves and let the signal handler fill one half while a goroutine
23 // is writing out the other half. When the signal handler fills its half, it
24 // offers to swap with the goroutine. If the writer is not done with its half,
25 // we lose the stack trace for this clock tick (and record that loss).
26 // The goroutine interacts with the signal handler by calling getprofile() to
27 // get the next log piece to write, implicitly handing back the last log
30 // The state of this dance between the signal handler and the goroutine
31 // is encoded in the Profile.handoff field. If handoff == 0, then the goroutine
32 // is not using either log half and is waiting (or will soon be waiting) for
33 // a new piece by calling notesleep(&p->wait). If the signal handler
34 // changes handoff from 0 to non-zero, it must call notewakeup(&p->wait)
35 // to wake the goroutine. The value indicates the number of entries in the
36 // log half being handed off. The goroutine leaves the non-zero value in
37 // place until it has finished processing the log half and then flips the number
38 // back to zero. Setting the high bit in handoff means that the profiling is over,
39 // and the goroutine is now in charge of flushing the data left in the hash table
40 // to the log and returning that data.
42 // The handoff field is manipulated using atomic operations.
43 // For the most part, the manipulation of handoff is orderly: if handoff == 0
44 // then the signal handler owns it and can change it to non-zero.
45 // If handoff != 0 then the goroutine owns it and can change it to zero.
46 // If that were the end of the story then we would not need to manipulate
47 // handoff using atomic operations. The operations are needed, however,
48 // in order to let the log closer set the high bit to indicate "EOF" safely
49 // in the situation when normally the goroutine "owns" handoff.
57 typedef struct __go_open_array Slice;
58 #define array __values
60 #define cap __capacity
70 typedef struct Profile Profile;
71 typedef struct Bucket Bucket;
72 typedef struct Entry Entry;
77 uintptr stack[MaxStack];
85 bool on; // profiling is on
86 Note wait; // goroutine waits here
87 uintptr count; // tick count
88 uintptr evicts; // eviction count
89 uintptr lost; // lost ticks that need to be logged
90 uintptr totallost; // total lost ticks
92 // Active recent stack traces.
93 Bucket hash[HashSize];
95 // Log of traces evicted from hash.
96 // Signal handler has filled log[toggle][:nlog].
97 // Goroutine is writing log[1-toggle][:handoff].
98 uintptr log[2][LogSize/2];
104 // Writer maintains its own toggle to avoid races
105 // looking at signal handler's toggle.
107 bool wholding; // holding & need to release a log half
108 bool flushing; // flushing hash table - profile is over
109 bool eod_sent; // special end-of-data record sent; => flushing
113 static Profile *prof;
115 static void tick(uintptr*, int32);
116 static void add(Profile*, uintptr*, int32);
117 static bool evict(Profile*, Entry*);
118 static bool flushlog(Profile*);
120 static uintptr eod[3] = {0, 1, 0};
122 // LostProfileData is a no-op function used in profiles
123 // to mark the number of profiling stack traces that were
124 // discarded due to slow data writers.
126 LostProfileData(void)
130 extern void runtime_SetCPUProfileRate(intgo)
131 __asm__ (GOSYM_PREFIX "runtime.SetCPUProfileRate");
133 // SetCPUProfileRate sets the CPU profiling rate.
134 // The user documentation is in debug.go.
136 runtime_SetCPUProfileRate(intgo hz)
141 // Clamp hz to something reasonable.
150 prof = runtime_SysAlloc(sizeof *prof, &mstats.other_sys);
152 runtime_printf("runtime: cpu profiling cannot allocate memory\n");
157 if(prof->on || prof->handoff != 0) {
158 runtime_printf("runtime: cannot set cpu profile rate until previous profile has finished.\n");
165 // pprof binary header format.
166 // http://code.google.com/p/google-perftools/source/browse/trunk/src/profiledata.cc#117
167 *p++ = 0; // count for header
168 *p++ = 3; // depth for header
169 *p++ = 0; // version number
170 *p++ = 1000000 / hz; // period (microseconds)
172 prof->nlog = p - prof->log[0];
174 prof->wholding = false;
176 prof->flushing = false;
177 prof->eod_sent = false;
178 runtime_noteclear(&prof->wait);
180 runtime_setcpuprofilerate(tick, hz);
181 } else if(prof != nil && prof->on) {
182 runtime_setcpuprofilerate(nil, 0);
185 // Now add is not running anymore, and getprofile owns the entire log.
186 // Set the high bit in prof->handoff to tell getprofile.
190 runtime_printf("runtime: setcpuprofile(off) twice");
191 if(runtime_cas(&prof->handoff, n, n|0x80000000))
195 // we did the transition from 0 -> nonzero so we wake getprofile
196 runtime_notewakeup(&prof->wait);
203 tick(uintptr *pc, int32 n)
208 // add adds the stack trace to the profile.
209 // It is called from signal handlers and other limited environments
210 // and cannot allocate memory or acquire locks that might be
211 // held at the time of the signal, nor can it use substantial amounts
212 // of stack. It is allowed to call evict.
214 add(Profile *p, uintptr *pc, int32 n)
227 h = h<<8 | (h>>(8*(sizeof(h)-1)));
229 h += x*31 + x*7 + x*3;
233 // Add to entry count if already present in table.
234 b = &p->hash[h%HashSize];
235 for(i=0; i<Assoc; i++) {
237 if(e->depth != (uintptr)n)
240 if(e->stack[j] != pc[j])
247 // Evict entry with smallest count.
249 for(i=1; i<Assoc; i++)
250 if(b->entry[i].count < e->count)
254 // Could not evict entry. Record lost stack.
262 // Reuse the newly evicted entry.
269 // evict copies the given entry's data into the log, so that
270 // the entry can be reused. evict is called from add, which
271 // is called from the profiling signal handler, so it must not
272 // allocate memory or block. It is safe to call flushLog.
273 // evict returns true if the entry was copied to the log,
274 // false if there was no room available.
276 evict(Profile *p, Entry *e)
283 log = p->log[p->toggle];
284 if(p->nlog+nslot > nelem(p->log[0])) {
287 log = p->log[p->toggle];
300 // flushlog tries to flush the current log and switch to the other one.
301 // flushlog is called from evict, called from add, called from the signal handler,
302 // so it cannot allocate memory or block. It can try to swap logs with
303 // the writing goroutine, as explained in the comment at the top of this file.
309 if(!runtime_cas(&p->handoff, 0, p->nlog))
311 runtime_notewakeup(&p->wait);
313 p->toggle = 1 - p->toggle;
314 log = p->log[p->toggle];
319 *q++ = (uintptr)LostProfileData;
325 // getprofile blocks until the next block of profiling data is available
326 // and returns it as a []byte. It is called from the writing goroutine.
328 getprofile(Profile *p)
343 // Release previous log to signal handling side.
344 // Loop because we are racing against SetCPUProfileRate(0).
348 runtime_printf("runtime: phase error during cpu profile handoff\n");
352 p->wtoggle = 1 - p->wtoggle;
357 if(runtime_cas(&p->handoff, n, 0))
360 p->wtoggle = 1 - p->wtoggle;
367 if(!p->on && p->handoff == 0)
371 runtime_notetsleepg(&p->wait, -1);
372 runtime_noteclear(&p->wait);
376 runtime_printf("runtime: phase error during cpu profile wait\n");
379 if(n == 0x80000000) {
385 // Return new log to caller.
388 ret.array = (byte*)p->log[p->wtoggle];
389 ret.len = n*sizeof(uintptr);
395 // Add is no longer being called. We own the log.
396 // Also, p->handoff is non-zero, so flushlog will return false.
397 // Evict the hash table into the log and return it.
398 for(i=0; i<HashSize; i++) {
400 for(j=0; j<Assoc; j++) {
402 if(e->count > 0 && !evict(p, e)) {
403 // Filled the log. Stop the loop and return what we've got.
410 // Return pending log data.
412 // Note that we're using toggle now, not wtoggle,
413 // because we're working on the log directly.
414 ret.array = (byte*)p->log[p->toggle];
415 ret.len = p->nlog*sizeof(uintptr);
421 // Made it through the table without finding anything to log.
423 // We may not have space to append this to the partial log buf,
424 // so we always return a new slice for the end-of-data marker.
426 ret.array = (byte*)eod;
427 ret.len = sizeof eod;
432 // Finally done. Clean up and return nil.
434 if(!runtime_cas(&p->handoff, p->handoff, 0))
435 runtime_printf("runtime: profile flush racing with something\n");
436 return ret; // set to nil at top of function
439 // CPUProfile returns the next cpu profile block as a []byte.
440 // The user documentation is in debug.go.
441 func CPUProfile() (ret Slice) {
442 ret = getprofile(prof);