xref: /qemu/trace/simple.c (revision 6402cbbb)
1 /*
2  * Simple trace backend
3  *
4  * Copyright IBM, Corp. 2010
5  *
6  * This work is licensed under the terms of the GNU GPL, version 2.  See
7  * the COPYING file in the top-level directory.
8  *
9  */
10 
11 #include "qemu/osdep.h"
12 #ifndef _WIN32
13 #include <pthread.h>
14 #endif
15 #include "qemu/timer.h"
16 #include "trace/control.h"
17 #include "trace/simple.h"
18 
19 /** Trace file header event ID, picked to avoid conflict with real event IDs */
20 #define HEADER_EVENT_ID (~(uint64_t)0)
21 
22 /** Trace file magic number */
23 #define HEADER_MAGIC 0xf2b177cb0aa429b4ULL
24 
25 /** Trace file version number, bump if format changes */
26 #define HEADER_VERSION 4
27 
28 /** Records were dropped event ID */
29 #define DROPPED_EVENT_ID (~(uint64_t)0 - 1)
30 
31 /** Trace record is valid */
32 #define TRACE_RECORD_VALID ((uint64_t)1 << 63)
33 
34 /*
35  * Trace records are written out by a dedicated thread.  The thread waits for
36  * records to become available, writes them out, and then waits again.
37  */
38 static CompatGMutex trace_lock;
39 static CompatGCond trace_available_cond;
40 static CompatGCond trace_empty_cond;
41 
42 static bool trace_available;
43 static bool trace_writeout_enabled;
44 
45 enum {
46     TRACE_BUF_LEN = 4096 * 64,
47     TRACE_BUF_FLUSH_THRESHOLD = TRACE_BUF_LEN / 4,
48 };
49 
50 uint8_t trace_buf[TRACE_BUF_LEN];
51 static volatile gint trace_idx;
52 static unsigned int writeout_idx;
53 static volatile gint dropped_events;
54 static uint32_t trace_pid;
55 static FILE *trace_fp;
56 static char *trace_file_name;
57 
58 #define TRACE_RECORD_TYPE_MAPPING 0
59 #define TRACE_RECORD_TYPE_EVENT   1
60 
61 /* * Trace buffer entry */
62 typedef struct {
63     uint64_t event; /* event ID value */
64     uint64_t timestamp_ns;
65     uint32_t length;   /*    in bytes */
66     uint32_t pid;
67     uint64_t arguments[];
68 } TraceRecord;
69 
70 typedef struct {
71     uint64_t header_event_id; /* HEADER_EVENT_ID */
72     uint64_t header_magic;    /* HEADER_MAGIC    */
73     uint64_t header_version;  /* HEADER_VERSION  */
74 } TraceLogHeader;
75 
76 
77 static void read_from_buffer(unsigned int idx, void *dataptr, size_t size);
78 static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size);
79 
80 static void clear_buffer_range(unsigned int idx, size_t len)
81 {
82     uint32_t num = 0;
83     while (num < len) {
84         if (idx >= TRACE_BUF_LEN) {
85             idx = idx % TRACE_BUF_LEN;
86         }
87         trace_buf[idx++] = 0;
88         num++;
89     }
90 }
91 /**
92  * Read a trace record from the trace buffer
93  *
94  * @idx         Trace buffer index
95  * @record      Trace record to fill
96  *
97  * Returns false if the record is not valid.
98  */
99 static bool get_trace_record(unsigned int idx, TraceRecord **recordptr)
100 {
101     uint64_t event_flag = 0;
102     TraceRecord record;
103     /* read the event flag to see if its a valid record */
104     read_from_buffer(idx, &record, sizeof(event_flag));
105 
106     if (!(record.event & TRACE_RECORD_VALID)) {
107         return false;
108     }
109 
110     smp_rmb(); /* read memory barrier before accessing record */
111     /* read the record header to know record length */
112     read_from_buffer(idx, &record, sizeof(TraceRecord));
113     *recordptr = malloc(record.length); /* don't use g_malloc, can deadlock when traced */
114     /* make a copy of record to avoid being overwritten */
115     read_from_buffer(idx, *recordptr, record.length);
116     smp_rmb(); /* memory barrier before clearing valid flag */
117     (*recordptr)->event &= ~TRACE_RECORD_VALID;
118     /* clear the trace buffer range for consumed record otherwise any byte
119      * with its MSB set may be considered as a valid event id when the writer
120      * thread crosses this range of buffer again.
121      */
122     clear_buffer_range(idx, record.length);
123     return true;
124 }
125 
126 /**
127  * Kick writeout thread
128  *
129  * @wait        Whether to wait for writeout thread to complete
130  */
131 static void flush_trace_file(bool wait)
132 {
133     g_mutex_lock(&trace_lock);
134     trace_available = true;
135     g_cond_signal(&trace_available_cond);
136 
137     if (wait) {
138         g_cond_wait(&trace_empty_cond, &trace_lock);
139     }
140 
141     g_mutex_unlock(&trace_lock);
142 }
143 
144 static void wait_for_trace_records_available(void)
145 {
146     g_mutex_lock(&trace_lock);
147     while (!(trace_available && trace_writeout_enabled)) {
148         g_cond_signal(&trace_empty_cond);
149         g_cond_wait(&trace_available_cond, &trace_lock);
150     }
151     trace_available = false;
152     g_mutex_unlock(&trace_lock);
153 }
154 
155 static gpointer writeout_thread(gpointer opaque)
156 {
157     TraceRecord *recordptr;
158     union {
159         TraceRecord rec;
160         uint8_t bytes[sizeof(TraceRecord) + sizeof(uint64_t)];
161     } dropped;
162     unsigned int idx = 0;
163     int dropped_count;
164     size_t unused __attribute__ ((unused));
165     uint64_t type = TRACE_RECORD_TYPE_EVENT;
166 
167     for (;;) {
168         wait_for_trace_records_available();
169 
170         if (g_atomic_int_get(&dropped_events)) {
171             dropped.rec.event = DROPPED_EVENT_ID,
172             dropped.rec.timestamp_ns = get_clock();
173             dropped.rec.length = sizeof(TraceRecord) + sizeof(uint64_t),
174             dropped.rec.pid = trace_pid;
175             do {
176                 dropped_count = g_atomic_int_get(&dropped_events);
177             } while (!g_atomic_int_compare_and_exchange(&dropped_events,
178                                                         dropped_count, 0));
179             dropped.rec.arguments[0] = dropped_count;
180             unused = fwrite(&type, sizeof(type), 1, trace_fp);
181             unused = fwrite(&dropped.rec, dropped.rec.length, 1, trace_fp);
182         }
183 
184         while (get_trace_record(idx, &recordptr)) {
185             unused = fwrite(&type, sizeof(type), 1, trace_fp);
186             unused = fwrite(recordptr, recordptr->length, 1, trace_fp);
187             writeout_idx += recordptr->length;
188             free(recordptr); /* don't use g_free, can deadlock when traced */
189             idx = writeout_idx % TRACE_BUF_LEN;
190         }
191 
192         fflush(trace_fp);
193     }
194     return NULL;
195 }
196 
197 void trace_record_write_u64(TraceBufferRecord *rec, uint64_t val)
198 {
199     rec->rec_off = write_to_buffer(rec->rec_off, &val, sizeof(uint64_t));
200 }
201 
202 void trace_record_write_str(TraceBufferRecord *rec, const char *s, uint32_t slen)
203 {
204     /* Write string length first */
205     rec->rec_off = write_to_buffer(rec->rec_off, &slen, sizeof(slen));
206     /* Write actual string now */
207     rec->rec_off = write_to_buffer(rec->rec_off, (void*)s, slen);
208 }
209 
210 int trace_record_start(TraceBufferRecord *rec, uint32_t event, size_t datasize)
211 {
212     unsigned int idx, rec_off, old_idx, new_idx;
213     uint32_t rec_len = sizeof(TraceRecord) + datasize;
214     uint64_t event_u64 = event;
215     uint64_t timestamp_ns = get_clock();
216 
217     do {
218         old_idx = g_atomic_int_get(&trace_idx);
219         smp_rmb();
220         new_idx = old_idx + rec_len;
221 
222         if (new_idx - writeout_idx > TRACE_BUF_LEN) {
223             /* Trace Buffer Full, Event dropped ! */
224             g_atomic_int_inc(&dropped_events);
225             return -ENOSPC;
226         }
227     } while (!g_atomic_int_compare_and_exchange(&trace_idx, old_idx, new_idx));
228 
229     idx = old_idx % TRACE_BUF_LEN;
230 
231     rec_off = idx;
232     rec_off = write_to_buffer(rec_off, &event_u64, sizeof(event_u64));
233     rec_off = write_to_buffer(rec_off, &timestamp_ns, sizeof(timestamp_ns));
234     rec_off = write_to_buffer(rec_off, &rec_len, sizeof(rec_len));
235     rec_off = write_to_buffer(rec_off, &trace_pid, sizeof(trace_pid));
236 
237     rec->tbuf_idx = idx;
238     rec->rec_off  = (idx + sizeof(TraceRecord)) % TRACE_BUF_LEN;
239     return 0;
240 }
241 
242 static void read_from_buffer(unsigned int idx, void *dataptr, size_t size)
243 {
244     uint8_t *data_ptr = dataptr;
245     uint32_t x = 0;
246     while (x < size) {
247         if (idx >= TRACE_BUF_LEN) {
248             idx = idx % TRACE_BUF_LEN;
249         }
250         data_ptr[x++] = trace_buf[idx++];
251     }
252 }
253 
254 static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size)
255 {
256     uint8_t *data_ptr = dataptr;
257     uint32_t x = 0;
258     while (x < size) {
259         if (idx >= TRACE_BUF_LEN) {
260             idx = idx % TRACE_BUF_LEN;
261         }
262         trace_buf[idx++] = data_ptr[x++];
263     }
264     return idx; /* most callers wants to know where to write next */
265 }
266 
267 void trace_record_finish(TraceBufferRecord *rec)
268 {
269     TraceRecord record;
270     read_from_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
271     smp_wmb(); /* write barrier before marking as valid */
272     record.event |= TRACE_RECORD_VALID;
273     write_to_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
274 
275     if (((unsigned int)g_atomic_int_get(&trace_idx) - writeout_idx)
276         > TRACE_BUF_FLUSH_THRESHOLD) {
277         flush_trace_file(false);
278     }
279 }
280 
281 static int st_write_event_mapping(void)
282 {
283     uint64_t type = TRACE_RECORD_TYPE_MAPPING;
284     TraceEventIter iter;
285     TraceEvent *ev;
286 
287     trace_event_iter_init(&iter, NULL);
288     while ((ev = trace_event_iter_next(&iter)) != NULL) {
289         uint64_t id = trace_event_get_id(ev);
290         const char *name = trace_event_get_name(ev);
291         uint32_t len = strlen(name);
292         if (fwrite(&type, sizeof(type), 1, trace_fp) != 1 ||
293             fwrite(&id, sizeof(id), 1, trace_fp) != 1 ||
294             fwrite(&len, sizeof(len), 1, trace_fp) != 1 ||
295             fwrite(name, len, 1, trace_fp) != 1) {
296             return -1;
297         }
298     }
299 
300     return 0;
301 }
302 
303 void st_set_trace_file_enabled(bool enable)
304 {
305     if (enable == !!trace_fp) {
306         return; /* no change */
307     }
308 
309     /* Halt trace writeout */
310     flush_trace_file(true);
311     trace_writeout_enabled = false;
312     flush_trace_file(true);
313 
314     if (enable) {
315         static const TraceLogHeader header = {
316             .header_event_id = HEADER_EVENT_ID,
317             .header_magic = HEADER_MAGIC,
318             /* Older log readers will check for version at next location */
319             .header_version = HEADER_VERSION,
320         };
321 
322         trace_fp = fopen(trace_file_name, "wb");
323         if (!trace_fp) {
324             return;
325         }
326 
327         if (fwrite(&header, sizeof header, 1, trace_fp) != 1 ||
328             st_write_event_mapping() < 0) {
329             fclose(trace_fp);
330             trace_fp = NULL;
331             return;
332         }
333 
334         /* Resume trace writeout */
335         trace_writeout_enabled = true;
336         flush_trace_file(false);
337     } else {
338         fclose(trace_fp);
339         trace_fp = NULL;
340     }
341 }
342 
343 /**
344  * Set the name of a trace file
345  *
346  * @file        The trace file name or NULL for the default name-<pid> set at
347  *              config time
348  */
349 void st_set_trace_file(const char *file)
350 {
351     st_set_trace_file_enabled(false);
352 
353     g_free(trace_file_name);
354 
355     if (!file) {
356         /* Type cast needed for Windows where getpid() returns an int. */
357         trace_file_name = g_strdup_printf(CONFIG_TRACE_FILE, (pid_t)getpid());
358     } else {
359         trace_file_name = g_strdup_printf("%s", file);
360     }
361 
362     st_set_trace_file_enabled(true);
363 }
364 
365 void st_print_trace_file_status(FILE *stream, int (*stream_printf)(FILE *stream, const char *fmt, ...))
366 {
367     stream_printf(stream, "Trace file \"%s\" %s.\n",
368                   trace_file_name, trace_fp ? "on" : "off");
369 }
370 
371 void st_flush_trace_buffer(void)
372 {
373     flush_trace_file(true);
374 }
375 
376 /* Helper function to create a thread with signals blocked.  Use glib's
377  * portable threads since QEMU abstractions cannot be used due to reentrancy in
378  * the tracer.  Also note the signal masking on POSIX hosts so that the thread
379  * does not steal signals when the rest of the program wants them blocked.
380  */
381 static GThread *trace_thread_create(GThreadFunc fn)
382 {
383     GThread *thread;
384 #ifndef _WIN32
385     sigset_t set, oldset;
386 
387     sigfillset(&set);
388     pthread_sigmask(SIG_SETMASK, &set, &oldset);
389 #endif
390 
391     thread = g_thread_new("trace-thread", fn, NULL);
392 
393 #ifndef _WIN32
394     pthread_sigmask(SIG_SETMASK, &oldset, NULL);
395 #endif
396 
397     return thread;
398 }
399 
400 bool st_init(void)
401 {
402     GThread *thread;
403 
404     trace_pid = getpid();
405 
406     thread = trace_thread_create(writeout_thread);
407     if (!thread) {
408         fprintf(stderr, "warning: unable to initialize simple trace backend\n");
409         return false;
410     }
411 
412     atexit(st_flush_trace_buffer);
413     return true;
414 }
415