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.
56 typedef struct __go_open_array Slice;
57 #define array __values
59 #define cap __capacity
69 typedef struct Profile Profile;
70 typedef struct Bucket Bucket;
71 typedef struct Entry Entry;
76 uintptr stack[MaxStack];
84 bool on; // profiling is on
85 Note wait; // goroutine waits here
86 uintptr count; // tick count
87 uintptr evicts; // eviction count
88 uintptr lost; // lost ticks that need to be logged
89 uintptr totallost; // total lost ticks
91 // Active recent stack traces.
92 Bucket hash[HashSize];
94 // Log of traces evicted from hash.
95 // Signal handler has filled log[toggle][:nlog].
96 // Goroutine is writing log[1-toggle][:handoff].
97 uintptr log[2][LogSize/2];
103 // Writer maintains its own toggle to avoid races
104 // looking at signal handler's toggle.
106 bool wholding; // holding & need to release a log half
107 bool flushing; // flushing hash table - profile is over
111 static Profile *prof;
113 static void tick(uintptr*, int32);
114 static void add(Profile*, uintptr*, int32);
115 static bool evict(Profile*, Entry*);
116 static bool flushlog(Profile*);
118 // LostProfileData is a no-op function used in profiles
119 // to mark the number of profiling stack traces that were
120 // discarded due to slow data writers.
121 static void LostProfileData(void) {
124 extern void runtime_SetCPUProfileRate(int32)
125 __asm__("runtime.SetCPUProfileRate");
127 // SetCPUProfileRate sets the CPU profiling rate.
128 // The user documentation is in debug.go.
130 runtime_SetCPUProfileRate(int32 hz)
135 // Clamp hz to something reasonable.
144 prof = runtime_SysAlloc(sizeof *prof);
146 runtime_printf("runtime: cpu profiling cannot allocate memory\n");
151 if(prof->on || prof->handoff != 0) {
152 runtime_printf("runtime: cannot set cpu profile rate until previous profile has finished.\n");
159 // pprof binary header format.
160 // http://code.google.com/p/google-perftools/source/browse/trunk/src/profiledata.cc#117
161 *p++ = 0; // count for header
162 *p++ = 3; // depth for header
163 *p++ = 0; // version number
164 *p++ = 1000000 / hz; // period (microseconds)
166 prof->nlog = p - prof->log[0];
168 prof->wholding = false;
170 prof->flushing = false;
171 runtime_noteclear(&prof->wait);
173 runtime_setcpuprofilerate(tick, hz);
174 } else if(prof->on) {
175 runtime_setcpuprofilerate(nil, 0);
178 // Now add is not running anymore, and getprofile owns the entire log.
179 // Set the high bit in prof->handoff to tell getprofile.
183 runtime_printf("runtime: setcpuprofile(off) twice");
184 if(runtime_cas(&prof->handoff, n, n|0x80000000))
188 // we did the transition from 0 -> nonzero so we wake getprofile
189 runtime_notewakeup(&prof->wait);
196 tick(uintptr *pc, int32 n)
201 // add adds the stack trace to the profile.
202 // It is called from signal handlers and other limited environments
203 // and cannot allocate memory or acquire locks that might be
204 // held at the time of the signal, nor can it use substantial amounts
205 // of stack. It is allowed to call evict.
207 add(Profile *p, uintptr *pc, int32 n)
220 h = h<<8 | (h>>(8*(sizeof(h)-1)));
222 h += x*31 + x*7 + x*3;
226 // Add to entry count if already present in table.
227 b = &p->hash[h%HashSize];
228 for(i=0; i<Assoc; i++) {
230 if(e->depth != (uintptr)n)
233 if(e->stack[j] != pc[j])
240 // Evict entry with smallest count.
242 for(i=1; i<Assoc; i++)
243 if(b->entry[i].count < e->count)
247 // Could not evict entry. Record lost stack.
255 // Reuse the newly evicted entry.
262 // evict copies the given entry's data into the log, so that
263 // the entry can be reused. evict is called from add, which
264 // is called from the profiling signal handler, so it must not
265 // allocate memory or block. It is safe to call flushLog.
266 // evict returns true if the entry was copied to the log,
267 // false if there was no room available.
269 evict(Profile *p, Entry *e)
276 log = p->log[p->toggle];
277 if(p->nlog+nslot > nelem(p->log[0])) {
280 log = p->log[p->toggle];
293 // flushlog tries to flush the current log and switch to the other one.
294 // flushlog is called from evict, called from add, called from the signal handler,
295 // so it cannot allocate memory or block. It can try to swap logs with
296 // the writing goroutine, as explained in the comment at the top of this file.
302 if(!runtime_cas(&p->handoff, 0, p->nlog))
304 runtime_notewakeup(&p->wait);
306 p->toggle = 1 - p->toggle;
307 log = p->log[p->toggle];
312 *q++ = (uintptr)LostProfileData;
318 // getprofile blocks until the next block of profiling data is available
319 // and returns it as a []byte. It is called from the writing goroutine.
321 getprofile(Profile *p)
336 // Release previous log to signal handling side.
337 // Loop because we are racing against setprofile(off).
341 runtime_printf("runtime: phase error during cpu profile handoff\n");
345 p->wtoggle = 1 - p->wtoggle;
350 if(runtime_cas(&p->handoff, n, 0))
353 p->wtoggle = 1 - p->wtoggle;
360 if(!p->on && p->handoff == 0)
364 runtime_entersyscall();
365 runtime_notesleep(&p->wait);
366 runtime_exitsyscall();
367 runtime_noteclear(&p->wait);
371 runtime_printf("runtime: phase error during cpu profile wait\n");
374 if(n == 0x80000000) {
380 // Return new log to caller.
383 ret.array = (byte*)p->log[p->wtoggle];
384 ret.len = n*sizeof(uintptr);
390 // Add is no longer being called. We own the log.
391 // Also, p->handoff is non-zero, so flushlog will return false.
392 // Evict the hash table into the log and return it.
393 for(i=0; i<HashSize; i++) {
395 for(j=0; j<Assoc; j++) {
397 if(e->count > 0 && !evict(p, e)) {
398 // Filled the log. Stop the loop and return what we've got.
405 // Return pending log data.
407 // Note that we're using toggle now, not wtoggle,
408 // because we're working on the log directly.
409 ret.array = (byte*)p->log[p->toggle];
410 ret.len = p->nlog*sizeof(uintptr);
416 // Made it through the table without finding anything to log.
417 // Finally done. Clean up and return nil.
419 if(!runtime_cas(&p->handoff, p->handoff, 0))
420 runtime_printf("runtime: profile flush racing with something\n");
421 return ret; // set to nil at top of function
424 extern Slice runtime_CPUProfile(void)
425 __asm__("runtime.CPUProfile");
427 // CPUProfile returns the next cpu profile block as a []byte.
428 // The user documentation is in debug.go.
430 runtime_CPUProfile(void)
432 return getprofile(prof);