qemu

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
#include "qemu/error-report.h"
19
#include "qemu/qemu-print.h"
20

21
/** Trace file header event ID, picked to avoid conflict with real event IDs */
22
#define HEADER_EVENT_ID (~(uint64_t)0)
23

24
/** Trace file magic number */
25
#define HEADER_MAGIC 0xf2b177cb0aa429b4ULL
26

27
/** Trace file version number, bump if format changes */
28
#define HEADER_VERSION 4
29

30
/** Records were dropped event ID */
31
#define DROPPED_EVENT_ID (~(uint64_t)0 - 1)
32

33
/** Trace record is valid */
34
#define TRACE_RECORD_VALID ((uint64_t)1 << 63)
35

36
/*
37
 * Trace records are written out by a dedicated thread.  The thread waits for
38
 * records to become available, writes them out, and then waits again.
39
 */
40
static GMutex trace_lock;
41
static GCond trace_available_cond;
42
static GCond trace_empty_cond;
43

44
static bool trace_available;
45
static bool trace_writeout_enabled;
46

47
enum {
48
    TRACE_BUF_LEN = 4096 * 64,
49
    TRACE_BUF_FLUSH_THRESHOLD = TRACE_BUF_LEN / 4,
50
};
51

52
uint8_t trace_buf[TRACE_BUF_LEN];
53
static volatile gint trace_idx;
54
static unsigned int writeout_idx;
55
static volatile gint dropped_events;
56
static uint32_t trace_pid;
57
static FILE *trace_fp;
58
static char *trace_file_name;
59

60
#define TRACE_RECORD_TYPE_MAPPING 0
61
#define TRACE_RECORD_TYPE_EVENT   1
62

63
/* * Trace buffer entry */
64
typedef struct {
65
    uint64_t event; /* event ID value */
66
    uint64_t timestamp_ns;
67
    uint32_t length;   /*    in bytes */
68
    uint32_t pid;
69
    uint64_t arguments[];
70
} TraceRecord;
71

72
typedef struct {
73
    uint64_t header_event_id; /* HEADER_EVENT_ID */
74
    uint64_t header_magic;    /* HEADER_MAGIC    */
75
    uint64_t header_version;  /* HEADER_VERSION  */
76
} TraceLogHeader;
77

78

79
static void read_from_buffer(unsigned int idx, void *dataptr, size_t size);
80
static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size);
81

82
static void clear_buffer_range(unsigned int idx, size_t len)
83
{
84
    uint32_t num = 0;
85
    while (num < len) {
86
        if (idx >= TRACE_BUF_LEN) {
87
            idx = idx % TRACE_BUF_LEN;
88
        }
89
        trace_buf[idx++] = 0;
90
        num++;
91
    }
92
}
93
/**
94
 * Read a trace record from the trace buffer
95
 *
96
 * @idx         Trace buffer index
97
 * @record      Trace record to fill
98
 *
99
 * Returns false if the record is not valid.
100
 */
101
static bool get_trace_record(unsigned int idx, TraceRecord **recordptr)
102
{
103
    uint64_t event_flag = 0;
104
    TraceRecord record;
105
    /* read the event flag to see if its a valid record */
106
    read_from_buffer(idx, &record, sizeof(event_flag));
107

108
    if (!(record.event & TRACE_RECORD_VALID)) {
109
        return false;
110
    }
111

112
    smp_rmb(); /* read memory barrier before accessing record */
113
    /* read the record header to know record length */
114
    read_from_buffer(idx, &record, sizeof(TraceRecord));
115
    *recordptr = malloc(record.length); /* don't use g_malloc, can deadlock when traced */
116
    /* make a copy of record to avoid being overwritten */
117
    read_from_buffer(idx, *recordptr, record.length);
118
    smp_rmb(); /* memory barrier before clearing valid flag */
119
    (*recordptr)->event &= ~TRACE_RECORD_VALID;
120
    /* clear the trace buffer range for consumed record otherwise any byte
121
     * with its MSB set may be considered as a valid event id when the writer
122
     * thread crosses this range of buffer again.
123
     */
124
    clear_buffer_range(idx, record.length);
125
    return true;
126
}
127

128
/**
129
 * Kick writeout thread
130
 *
131
 * @wait        Whether to wait for writeout thread to complete
132
 */
133
static void flush_trace_file(bool wait)
134
{
135
    g_mutex_lock(&trace_lock);
136
    trace_available = true;
137
    g_cond_signal(&trace_available_cond);
138

139
    if (wait) {
140
        g_cond_wait(&trace_empty_cond, &trace_lock);
141
    }
142

143
    g_mutex_unlock(&trace_lock);
144
}
145

146
static void wait_for_trace_records_available(void)
147
{
148
    g_mutex_lock(&trace_lock);
149
    while (!(trace_available && trace_writeout_enabled)) {
150
        g_cond_signal(&trace_empty_cond);
151
        g_cond_wait(&trace_available_cond, &trace_lock);
152
    }
153
    trace_available = false;
154
    g_mutex_unlock(&trace_lock);
155
}
156

157
static gpointer writeout_thread(gpointer opaque)
158
{
159
    TraceRecord *recordptr;
160
    union {
161
        TraceRecord rec;
162
        uint8_t bytes[sizeof(TraceRecord) + sizeof(uint64_t)];
163
    } dropped;
164
    unsigned int idx = 0;
165
    int dropped_count;
166
    size_t unused __attribute__ ((unused));
167
    uint64_t type = TRACE_RECORD_TYPE_EVENT;
168

169
    for (;;) {
170
        wait_for_trace_records_available();
171

172
        if (g_atomic_int_get(&dropped_events)) {
173
            dropped.rec.event = DROPPED_EVENT_ID;
174
            dropped.rec.timestamp_ns = get_clock();
175
            dropped.rec.length = sizeof(TraceRecord) + sizeof(uint64_t);
176
            dropped.rec.pid = trace_pid;
177
            do {
178
                dropped_count = g_atomic_int_get(&dropped_events);
179
            } while (!g_atomic_int_compare_and_exchange(&dropped_events,
180
                                                        dropped_count, 0));
181
            dropped.rec.arguments[0] = dropped_count;
182
            unused = fwrite(&type, sizeof(type), 1, trace_fp);
183
            unused = fwrite(&dropped.rec, dropped.rec.length, 1, trace_fp);
184
        }
185

186
        while (get_trace_record(idx, &recordptr)) {
187
            unused = fwrite(&type, sizeof(type), 1, trace_fp);
188
            unused = fwrite(recordptr, recordptr->length, 1, trace_fp);
189
            writeout_idx += recordptr->length;
190
            free(recordptr); /* don't use g_free, can deadlock when traced */
191
            idx = writeout_idx % TRACE_BUF_LEN;
192
        }
193

194
        fflush(trace_fp);
195
    }
196
    return NULL;
197
}
198

199
void trace_record_write_u64(TraceBufferRecord *rec, uint64_t val)
200
{
201
    rec->rec_off = write_to_buffer(rec->rec_off, &val, sizeof(uint64_t));
202
}
203

204
void trace_record_write_str(TraceBufferRecord *rec, const char *s, uint32_t slen)
205
{
206
    /* Write string length first */
207
    rec->rec_off = write_to_buffer(rec->rec_off, &slen, sizeof(slen));
208
    /* Write actual string now */
209
    rec->rec_off = write_to_buffer(rec->rec_off, (void*)s, slen);
210
}
211

212
int trace_record_start(TraceBufferRecord *rec, uint32_t event, size_t datasize)
213
{
214
    unsigned int idx, rec_off, old_idx, new_idx;
215
    uint32_t rec_len = sizeof(TraceRecord) + datasize;
216
    uint64_t event_u64 = event;
217
    uint64_t timestamp_ns = get_clock();
218

219
    do {
220
        old_idx = g_atomic_int_get(&trace_idx);
221
        smp_rmb();
222
        new_idx = old_idx + rec_len;
223

224
        if (new_idx - writeout_idx > TRACE_BUF_LEN) {
225
            /* Trace Buffer Full, Event dropped ! */
226
            g_atomic_int_inc(&dropped_events);
227
            return -ENOSPC;
228
        }
229
    } while (!g_atomic_int_compare_and_exchange(&trace_idx, old_idx, new_idx));
230

231
    idx = old_idx % TRACE_BUF_LEN;
232

233
    rec_off = idx;
234
    rec_off = write_to_buffer(rec_off, &event_u64, sizeof(event_u64));
235
    rec_off = write_to_buffer(rec_off, &timestamp_ns, sizeof(timestamp_ns));
236
    rec_off = write_to_buffer(rec_off, &rec_len, sizeof(rec_len));
237
    rec_off = write_to_buffer(rec_off, &trace_pid, sizeof(trace_pid));
238

239
    rec->tbuf_idx = idx;
240
    rec->rec_off  = (idx + sizeof(TraceRecord)) % TRACE_BUF_LEN;
241
    return 0;
242
}
243

244
static void read_from_buffer(unsigned int idx, void *dataptr, size_t size)
245
{
246
    uint8_t *data_ptr = dataptr;
247
    uint32_t x = 0;
248
    while (x < size) {
249
        if (idx >= TRACE_BUF_LEN) {
250
            idx = idx % TRACE_BUF_LEN;
251
        }
252
        data_ptr[x++] = trace_buf[idx++];
253
    }
254
}
255

256
static unsigned int write_to_buffer(unsigned int idx, void *dataptr, size_t size)
257
{
258
    uint8_t *data_ptr = dataptr;
259
    uint32_t x = 0;
260
    while (x < size) {
261
        if (idx >= TRACE_BUF_LEN) {
262
            idx = idx % TRACE_BUF_LEN;
263
        }
264
        trace_buf[idx++] = data_ptr[x++];
265
    }
266
    return idx; /* most callers wants to know where to write next */
267
}
268

269
void trace_record_finish(TraceBufferRecord *rec)
270
{
271
    TraceRecord record;
272
    read_from_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
273
    smp_wmb(); /* write barrier before marking as valid */
274
    record.event |= TRACE_RECORD_VALID;
275
    write_to_buffer(rec->tbuf_idx, &record, sizeof(TraceRecord));
276

277
    if (((unsigned int)g_atomic_int_get(&trace_idx) - writeout_idx)
278
        > TRACE_BUF_FLUSH_THRESHOLD) {
279
        flush_trace_file(false);
280
    }
281
}
282

283
static int st_write_event_mapping(TraceEventIter *iter)
284
{
285
    uint64_t type = TRACE_RECORD_TYPE_MAPPING;
286
    TraceEvent *ev;
287

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
/**
304
 * Enable / disable tracing, return whether it was enabled.
305
 *
306
 * @enable: enable if %true, else disable.
307
 */
308
bool st_set_trace_file_enabled(bool enable)
309
{
310
    TraceEventIter iter;
311
    bool was_enabled = trace_fp;
312

313
    if (enable == !!trace_fp) {
314
        return was_enabled;     /* no change */
315
    }
316

317
    /* Halt trace writeout */
318
    flush_trace_file(true);
319
    trace_writeout_enabled = false;
320
    flush_trace_file(true);
321

322
    if (enable) {
323
        static const TraceLogHeader header = {
324
            .header_event_id = HEADER_EVENT_ID,
325
            .header_magic = HEADER_MAGIC,
326
            /* Older log readers will check for version at next location */
327
            .header_version = HEADER_VERSION,
328
        };
329

330
        trace_fp = fopen(trace_file_name, "wb");
331
        if (!trace_fp) {
332
            return was_enabled;
333
        }
334

335
        trace_event_iter_init_all(&iter);
336
        if (fwrite(&header, sizeof header, 1, trace_fp) != 1 ||
337
            st_write_event_mapping(&iter) < 0) {
338
            fclose(trace_fp);
339
            trace_fp = NULL;
340
            return was_enabled;
341
        }
342

343
        /* Resume trace writeout */
344
        trace_writeout_enabled = true;
345
        flush_trace_file(false);
346
    } else {
347
        fclose(trace_fp);
348
        trace_fp = NULL;
349
    }
350
    return was_enabled;
351
}
352

353
/**
354
 * Set the name of a trace file
355
 *
356
 * @file        The trace file name or NULL for the default name-<pid> set at
357
 *              config time
358
 */
359
void st_set_trace_file(const char *file)
360
{
361
    bool saved_enable = st_set_trace_file_enabled(false);
362

363
    g_free(trace_file_name);
364

365
    if (!file) {
366
        /* Type cast needed for Windows where getpid() returns an int. */
367
        trace_file_name = g_strdup_printf(CONFIG_TRACE_FILE "-" FMT_pid, (pid_t)getpid());
368
    } else {
369
        trace_file_name = g_strdup_printf("%s", file);
370
    }
371

372
    st_set_trace_file_enabled(saved_enable);
373
}
374

375
void st_print_trace_file_status(void)
376
{
377
    qemu_printf("Trace file \"%s\" %s.\n",
378
                trace_file_name, trace_fp ? "on" : "off");
379
}
380

381
void st_flush_trace_buffer(void)
382
{
383
    flush_trace_file(true);
384
}
385

386
/* Helper function to create a thread with signals blocked.  Use glib's
387
 * portable threads since QEMU abstractions cannot be used due to reentrancy in
388
 * the tracer.  Also note the signal masking on POSIX hosts so that the thread
389
 * does not steal signals when the rest of the program wants them blocked.
390
 */
391
static GThread *trace_thread_create(GThreadFunc fn)
392
{
393
    GThread *thread;
394
#ifndef _WIN32
395
    sigset_t set, oldset;
396

397
    sigfillset(&set);
398
    pthread_sigmask(SIG_SETMASK, &set, &oldset);
399
#endif
400

401
    thread = g_thread_new("trace-thread", fn, NULL);
402

403
#ifndef _WIN32
404
    pthread_sigmask(SIG_SETMASK, &oldset, NULL);
405
#endif
406

407
    return thread;
408
}
409

410
bool st_init(void)
411
{
412
    GThread *thread;
413

414
    trace_pid = getpid();
415

416
    thread = trace_thread_create(writeout_thread);
417
    if (!thread) {
418
        warn_report("unable to initialize simple trace backend");
419
        return false;
420
    }
421

422
    atexit(st_flush_trace_buffer);
423
    return true;
424
}
425

426
void st_init_group(size_t group)
427
{
428
    TraceEventIter iter;
429

430
    if (!trace_writeout_enabled) {
431
        return;
432
    }
433

434
    trace_event_iter_init_group(&iter, group);
435
    st_write_event_mapping(&iter);
436
}
437