xref: /openbmc/qemu/trace/simple.c (revision 8917c3bd)
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 <stdlib.h>
12 #include <stdint.h>
13 #include <stdio.h>
14 #include <time.h>
15 #ifndef _WIN32
16 #include <signal.h>
17 #include <pthread.h>
18 #endif
19 #include "qemu/timer.h"
20 #include "trace.h"
21 #include "trace/control.h"
22 
23 /** Trace file header event ID */
24 #define HEADER_EVENT_ID (~(uint64_t)0) /* avoids conflicting with TraceEventIDs */
25 
26 /** Trace file magic number */
27 #define HEADER_MAGIC 0xf2b177cb0aa429b4ULL
28 
29 /** Trace file version number, bump if format changes */
30 #define HEADER_VERSION 2
31 
32 /** Records were dropped event ID */
33 #define DROPPED_EVENT_ID (~(uint64_t)0 - 1)
34 
35 /** Trace record is valid */
36 #define TRACE_RECORD_VALID ((uint64_t)1 << 63)
37 
38 /*
39  * Trace records are written out by a dedicated thread.  The thread waits for
40  * records to become available, writes them out, and then waits again.
41  */
42 static GStaticMutex trace_lock = G_STATIC_MUTEX_INIT;
43 
44 /* g_cond_new() was deprecated in glib 2.31 but we still need to support it */
45 #if GLIB_CHECK_VERSION(2, 31, 0)
46 static GCond the_trace_available_cond;
47 static GCond the_trace_empty_cond;
48 static GCond *trace_available_cond = &the_trace_available_cond;
49 static GCond *trace_empty_cond = &the_trace_empty_cond;
50 #else
51 static GCond *trace_available_cond;
52 static GCond *trace_empty_cond;
53 #endif
54 
55 static bool trace_available;
56 static bool trace_writeout_enabled;
57 
58 enum {
59     TRACE_BUF_LEN = 4096 * 64,
60     TRACE_BUF_FLUSH_THRESHOLD = TRACE_BUF_LEN / 4,
61 };
62 
63 uint8_t trace_buf[TRACE_BUF_LEN];
64 static volatile gint trace_idx;
65 static unsigned int writeout_idx;
66 static volatile gint dropped_events;
67 static FILE *trace_fp;
68 static char *trace_file_name;
69 
70 /* * Trace buffer entry */
71 typedef struct {
72     uint64_t event; /*   TraceEventID */
73     uint64_t timestamp_ns;
74     uint32_t length;   /*    in bytes */
75     uint32_t reserved; /*    unused */
76     uint64_t arguments[];
77 } TraceRecord;
78 
79 typedef struct {
80     uint64_t header_event_id; /* HEADER_EVENT_ID */
81     uint64_t header_magic;    /* HEADER_MAGIC    */
82     uint64_t header_version;  /* HEADER_VERSION  */
83 } TraceLogHeader;
84 
85 
86 static void read_from_buffer(unsigned int idx, void *dataptr, size_t size);
87 static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size);
88 
89 static void clear_buffer_range(unsigned int idx, size_t len)
90 {
91     uint32_t num = 0;
92     while (num < len) {
93         if (idx >= TRACE_BUF_LEN) {
94             idx = idx % TRACE_BUF_LEN;
95         }
96         trace_buf[idx++] = 0;
97         num++;
98     }
99 }
100 /**
101  * Read a trace record from the trace buffer
102  *
103  * @idx         Trace buffer index
104  * @record      Trace record to fill
105  *
106  * Returns false if the record is not valid.
107  */
108 static bool get_trace_record(unsigned int idx, TraceRecord **recordptr)
109 {
110     uint64_t event_flag = 0;
111     TraceRecord record;
112     /* read the event flag to see if its a valid record */
113     read_from_buffer(idx, &record, sizeof(event_flag));
114 
115     if (!(record.event & TRACE_RECORD_VALID)) {
116         return false;
117     }
118 
119     smp_rmb(); /* read memory barrier before accessing record */
120     /* read the record header to know record length */
121     read_from_buffer(idx, &record, sizeof(TraceRecord));
122     *recordptr = malloc(record.length); /* dont use g_malloc, can deadlock when traced */
123     /* make a copy of record to avoid being overwritten */
124     read_from_buffer(idx, *recordptr, record.length);
125     smp_rmb(); /* memory barrier before clearing valid flag */
126     (*recordptr)->event &= ~TRACE_RECORD_VALID;
127     /* clear the trace buffer range for consumed record otherwise any byte
128      * with its MSB set may be considered as a valid event id when the writer
129      * thread crosses this range of buffer again.
130      */
131     clear_buffer_range(idx, record.length);
132     return true;
133 }
134 
135 /**
136  * Kick writeout thread
137  *
138  * @wait        Whether to wait for writeout thread to complete
139  */
140 static void flush_trace_file(bool wait)
141 {
142     g_static_mutex_lock(&trace_lock);
143     trace_available = true;
144     g_cond_signal(trace_available_cond);
145 
146     if (wait) {
147         g_cond_wait(trace_empty_cond, g_static_mutex_get_mutex(&trace_lock));
148     }
149 
150     g_static_mutex_unlock(&trace_lock);
151 }
152 
153 static void wait_for_trace_records_available(void)
154 {
155     g_static_mutex_lock(&trace_lock);
156     while (!(trace_available && trace_writeout_enabled)) {
157         g_cond_signal(trace_empty_cond);
158         g_cond_wait(trace_available_cond,
159                     g_static_mutex_get_mutex(&trace_lock));
160     }
161     trace_available = false;
162     g_static_mutex_unlock(&trace_lock);
163 }
164 
165 static gpointer writeout_thread(gpointer opaque)
166 {
167     TraceRecord *recordptr;
168     union {
169         TraceRecord rec;
170         uint8_t bytes[sizeof(TraceRecord) + sizeof(uint64_t)];
171     } dropped;
172     unsigned int idx = 0;
173     int dropped_count;
174     size_t unused __attribute__ ((unused));
175 
176     for (;;) {
177         wait_for_trace_records_available();
178 
179         if (g_atomic_int_get(&dropped_events)) {
180             dropped.rec.event = DROPPED_EVENT_ID,
181             dropped.rec.timestamp_ns = get_clock();
182             dropped.rec.length = sizeof(TraceRecord) + sizeof(uint64_t),
183             dropped.rec.reserved = 0;
184             do {
185                 dropped_count = g_atomic_int_get(&dropped_events);
186             } while (!g_atomic_int_compare_and_exchange(&dropped_events,
187                                                         dropped_count, 0));
188             dropped.rec.arguments[0] = dropped_count;
189             unused = fwrite(&dropped.rec, dropped.rec.length, 1, trace_fp);
190         }
191 
192         while (get_trace_record(idx, &recordptr)) {
193             unused = fwrite(recordptr, recordptr->length, 1, trace_fp);
194             writeout_idx += recordptr->length;
195             free(recordptr); /* dont use g_free, can deadlock when traced */
196             idx = writeout_idx % TRACE_BUF_LEN;
197         }
198 
199         fflush(trace_fp);
200     }
201     return NULL;
202 }
203 
204 void trace_record_write_u64(TraceBufferRecord *rec, uint64_t val)
205 {
206     rec->rec_off = write_to_buffer(rec->rec_off, &val, sizeof(uint64_t));
207 }
208 
209 void trace_record_write_str(TraceBufferRecord *rec, const char *s, uint32_t slen)
210 {
211     /* Write string length first */
212     rec->rec_off = write_to_buffer(rec->rec_off, &slen, sizeof(slen));
213     /* Write actual string now */
214     rec->rec_off = write_to_buffer(rec->rec_off, (void*)s, slen);
215 }
216 
217 int trace_record_start(TraceBufferRecord *rec, TraceEventID event, size_t datasize)
218 {
219     unsigned int idx, rec_off, old_idx, new_idx;
220     uint32_t rec_len = sizeof(TraceRecord) + datasize;
221     uint64_t timestamp_ns = get_clock();
222 
223     do {
224         old_idx = g_atomic_int_get(&trace_idx);
225         smp_rmb();
226         new_idx = old_idx + rec_len;
227 
228         if (new_idx - writeout_idx > TRACE_BUF_LEN) {
229             /* Trace Buffer Full, Event dropped ! */
230             g_atomic_int_inc(&dropped_events);
231             return -ENOSPC;
232         }
233     } while (!g_atomic_int_compare_and_exchange(&trace_idx, old_idx, new_idx));
234 
235     idx = old_idx % TRACE_BUF_LEN;
236 
237     rec_off = idx;
238     rec_off = write_to_buffer(rec_off, &event, sizeof(event));
239     rec_off = write_to_buffer(rec_off, &timestamp_ns, sizeof(timestamp_ns));
240     rec_off = write_to_buffer(rec_off, &rec_len, sizeof(rec_len));
241 
242     rec->tbuf_idx = idx;
243     rec->rec_off  = (idx + sizeof(TraceRecord)) % TRACE_BUF_LEN;
244     return 0;
245 }
246 
247 static void read_from_buffer(unsigned int idx, void *dataptr, size_t size)
248 {
249     uint8_t *data_ptr = dataptr;
250     uint32_t x = 0;
251     while (x < size) {
252         if (idx >= TRACE_BUF_LEN) {
253             idx = idx % TRACE_BUF_LEN;
254         }
255         data_ptr[x++] = trace_buf[idx++];
256     }
257 }
258 
259 static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size)
260 {
261     uint8_t *data_ptr = dataptr;
262     uint32_t x = 0;
263     while (x < size) {
264         if (idx >= TRACE_BUF_LEN) {
265             idx = idx % TRACE_BUF_LEN;
266         }
267         trace_buf[idx++] = data_ptr[x++];
268     }
269     return idx; /* most callers wants to know where to write next */
270 }
271 
272 void trace_record_finish(TraceBufferRecord *rec)
273 {
274     TraceRecord record;
275     read_from_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
276     smp_wmb(); /* write barrier before marking as valid */
277     record.event |= TRACE_RECORD_VALID;
278     write_to_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
279 
280     if (((unsigned int)g_atomic_int_get(&trace_idx) - writeout_idx)
281         > TRACE_BUF_FLUSH_THRESHOLD) {
282         flush_trace_file(false);
283     }
284 }
285 
286 void st_set_trace_file_enabled(bool enable)
287 {
288     if (enable == !!trace_fp) {
289         return; /* no change */
290     }
291 
292     /* Halt trace writeout */
293     flush_trace_file(true);
294     trace_writeout_enabled = false;
295     flush_trace_file(true);
296 
297     if (enable) {
298         static const TraceLogHeader header = {
299             .header_event_id = HEADER_EVENT_ID,
300             .header_magic = HEADER_MAGIC,
301             /* Older log readers will check for version at next location */
302             .header_version = HEADER_VERSION,
303         };
304 
305         trace_fp = fopen(trace_file_name, "wb");
306         if (!trace_fp) {
307             return;
308         }
309 
310         if (fwrite(&header, sizeof header, 1, trace_fp) != 1) {
311             fclose(trace_fp);
312             trace_fp = NULL;
313             return;
314         }
315 
316         /* Resume trace writeout */
317         trace_writeout_enabled = true;
318         flush_trace_file(false);
319     } else {
320         fclose(trace_fp);
321         trace_fp = NULL;
322     }
323 }
324 
325 /**
326  * Set the name of a trace file
327  *
328  * @file        The trace file name or NULL for the default name-<pid> set at
329  *              config time
330  */
331 bool st_set_trace_file(const char *file)
332 {
333     st_set_trace_file_enabled(false);
334 
335     g_free(trace_file_name);
336 
337     if (!file) {
338         trace_file_name = g_strdup_printf(CONFIG_TRACE_FILE, getpid());
339     } else {
340         trace_file_name = g_strdup_printf("%s", file);
341     }
342 
343     st_set_trace_file_enabled(true);
344     return true;
345 }
346 
347 void st_print_trace_file_status(FILE *stream, int (*stream_printf)(FILE *stream, const char *fmt, ...))
348 {
349     stream_printf(stream, "Trace file \"%s\" %s.\n",
350                   trace_file_name, trace_fp ? "on" : "off");
351 }
352 
353 void st_flush_trace_buffer(void)
354 {
355     flush_trace_file(true);
356 }
357 
358 void trace_print_events(FILE *stream, fprintf_function stream_printf)
359 {
360     unsigned int i;
361 
362     for (i = 0; i < NR_TRACE_EVENTS; i++) {
363         stream_printf(stream, "%s [Event ID %u] : state %u\n",
364                       trace_list[i].tp_name, i, trace_list[i].state);
365     }
366 }
367 
368 bool trace_event_set_state(const char *name, bool state)
369 {
370     unsigned int i;
371     unsigned int len;
372     bool wildcard = false;
373     bool matched = false;
374 
375     len = strlen(name);
376     if (len > 0 && name[len - 1] == '*') {
377         wildcard = true;
378         len -= 1;
379     }
380     for (i = 0; i < NR_TRACE_EVENTS; i++) {
381         if (wildcard) {
382             if (!strncmp(trace_list[i].tp_name, name, len)) {
383                 trace_list[i].state = state;
384                 matched = true;
385             }
386             continue;
387         }
388         if (!strcmp(trace_list[i].tp_name, name)) {
389             trace_list[i].state = state;
390             return true;
391         }
392     }
393     return matched;
394 }
395 
396 /* Helper function to create a thread with signals blocked.  Use glib's
397  * portable threads since QEMU abstractions cannot be used due to reentrancy in
398  * the tracer.  Also note the signal masking on POSIX hosts so that the thread
399  * does not steal signals when the rest of the program wants them blocked.
400  */
401 static GThread *trace_thread_create(GThreadFunc fn)
402 {
403     GThread *thread;
404 #ifndef _WIN32
405     sigset_t set, oldset;
406 
407     sigfillset(&set);
408     pthread_sigmask(SIG_SETMASK, &set, &oldset);
409 #endif
410 
411 #if GLIB_CHECK_VERSION(2, 31, 0)
412     thread = g_thread_new("trace-thread", fn, NULL);
413 #else
414     thread = g_thread_create(fn, NULL, FALSE, NULL);
415 #endif
416 
417 #ifndef _WIN32
418     pthread_sigmask(SIG_SETMASK, &oldset, NULL);
419 #endif
420 
421     return thread;
422 }
423 
424 bool trace_backend_init(const char *events, const char *file)
425 {
426     GThread *thread;
427 
428     if (!g_thread_supported()) {
429 #if !GLIB_CHECK_VERSION(2, 31, 0)
430         g_thread_init(NULL);
431 #else
432         fprintf(stderr, "glib threading failed to initialize.\n");
433         exit(1);
434 #endif
435     }
436 
437 #if !GLIB_CHECK_VERSION(2, 31, 0)
438     trace_available_cond = g_cond_new();
439     trace_empty_cond = g_cond_new();
440 #endif
441 
442     thread = trace_thread_create(writeout_thread);
443     if (!thread) {
444         fprintf(stderr, "warning: unable to initialize simple trace backend\n");
445         return false;
446     }
447 
448     atexit(st_flush_trace_buffer);
449     trace_backend_init_events(events);
450     st_set_trace_file(file);
451     return true;
452 }
453