PostgreSQL Source Code  git master
pg_test_fsync.c
Go to the documentation of this file.
1 /*
2  * pg_test_fsync.c
3  * tests all supported fsync() methods
4  */
5 
6 #include "postgres_fe.h"
7 
8 #include <limits.h>
9 #include <sys/stat.h>
10 #include <sys/time.h>
11 #include <fcntl.h>
12 #include <time.h>
13 #include <unistd.h>
14 #include <signal.h>
15 
16 #include "access/xlogdefs.h"
17 #include "common/logging.h"
18 #include "getopt_long.h"
19 
20 /*
21  * put the temp files in the local directory
22  * unless the user specifies otherwise
23  */
24 #define FSYNC_FILENAME "./pg_test_fsync.out"
25 
26 #define XLOG_BLCKSZ_K (XLOG_BLCKSZ / 1024)
27 
28 #define LABEL_FORMAT " %-30s"
29 #define NA_FORMAT "%21s\n"
30 /* translator: maintain alignment with NA_FORMAT */
31 #define OPS_FORMAT gettext_noop("%13.3f ops/sec %6.0f usecs/op\n")
32 #define USECS_SEC 1000000
33 
34 /* These are macros to avoid timing the function call overhead. */
35 #ifndef WIN32
36 #define START_TIMER \
37 do { \
38  alarm_triggered = false; \
39  alarm(secs_per_test); \
40  gettimeofday(&start_t, NULL); \
41 } while (0)
42 #else
43 /* WIN32 doesn't support alarm, so we create a thread and sleep there */
44 #define START_TIMER \
45 do { \
46  alarm_triggered = false; \
47  if (CreateThread(NULL, 0, process_alarm, NULL, 0, NULL) == \
48  INVALID_HANDLE_VALUE) \
49  { \
50  pg_log_error("could not create thread for alarm"); \
51  exit(1); \
52  } \
53  gettimeofday(&start_t, NULL); \
54 } while (0)
55 #endif
56 
57 #define STOP_TIMER \
58 do { \
59  gettimeofday(&stop_t, NULL); \
60  print_elapse(start_t, stop_t, ops); \
61 } while (0)
62 
63 
64 static const char *progname;
65 
66 static unsigned int secs_per_test = 5;
67 static int needs_unlink = 0;
69  *buf,
71 static struct timeval start_t,
72  stop_t;
73 static bool alarm_triggered = false;
74 
75 
76 static void handle_args(int argc, char *argv[]);
77 static void prepare_buf(void);
78 static void test_open(void);
79 static void test_non_sync(void);
80 static void test_sync(int writes_per_op);
81 static void test_open_syncs(void);
82 static void test_open_sync(const char *msg, int writes_size);
83 static void test_file_descriptor_sync(void);
84 
85 #ifndef WIN32
86 static void process_alarm(int sig);
87 #else
88 static DWORD WINAPI process_alarm(LPVOID param);
89 #endif
90 static void signal_cleanup(int sig);
91 
92 #ifdef HAVE_FSYNC_WRITETHROUGH
93 static int pg_fsync_writethrough(int fd);
94 #endif
95 static void print_elapse(struct timeval start_t, struct timeval stop_t, int ops);
96 
97 #define die(msg) do { pg_log_error("%s: %m", _(msg)); exit(1); } while(0)
98 
99 
100 int
101 main(int argc, char *argv[])
102 {
103  pg_logging_init(argv[0]);
104  set_pglocale_pgservice(argv[0], PG_TEXTDOMAIN("pg_test_fsync"));
105  progname = get_progname(argv[0]);
106 
107  handle_args(argc, argv);
108 
109  /* Prevent leaving behind the test file */
110  pqsignal(SIGINT, signal_cleanup);
111  pqsignal(SIGTERM, signal_cleanup);
112 #ifndef WIN32
114 #endif
115 #ifdef SIGHUP
116  /* Not defined on win32 */
118 #endif
119 
120  prepare_buf();
121 
122  test_open();
123 
124  /* Test using 1 XLOG_BLCKSZ write */
125  test_sync(1);
126 
127  /* Test using 2 XLOG_BLCKSZ writes */
128  test_sync(2);
129 
130  test_open_syncs();
131 
133 
134  test_non_sync();
135 
136  unlink(filename);
137 
138  return 0;
139 }
140 
141 static void
142 handle_args(int argc, char *argv[])
143 {
144  static struct option long_options[] = {
145  {"filename", required_argument, NULL, 'f'},
146  {"secs-per-test", required_argument, NULL, 's'},
147  {NULL, 0, NULL, 0}
148  };
149 
150  int option; /* Command line option */
151  int optindex = 0; /* used by getopt_long */
152  unsigned long optval; /* used for option parsing */
153  char *endptr;
154 
155  if (argc > 1)
156  {
157  if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
158  {
159  printf(_("Usage: %s [-f FILENAME] [-s SECS-PER-TEST]\n"), progname);
160  exit(0);
161  }
162  if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
163  {
164  puts("pg_test_fsync (PostgreSQL) " PG_VERSION);
165  exit(0);
166  }
167  }
168 
169  while ((option = getopt_long(argc, argv, "f:s:",
170  long_options, &optindex)) != -1)
171  {
172  switch (option)
173  {
174  case 'f':
176  break;
177 
178  case 's':
179  errno = 0;
180  optval = strtoul(optarg, &endptr, 10);
181 
182  if (endptr == optarg || *endptr != '\0' ||
183  errno != 0 || optval != (unsigned int) optval)
184  {
185  pg_log_error("invalid argument for option %s", "--secs-per-test");
186  fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
187  exit(1);
188  }
189 
190  secs_per_test = (unsigned int) optval;
191  if (secs_per_test == 0)
192  {
193  pg_log_error("%s must be in range %u..%u",
194  "--secs-per-test", 1, UINT_MAX);
195  exit(1);
196  }
197  break;
198 
199  default:
200  fprintf(stderr, _("Try \"%s --help\" for more information.\n"),
201  progname);
202  exit(1);
203  break;
204  }
205  }
206 
207  if (argc > optind)
208  {
209  pg_log_error("too many command-line arguments (first is \"%s\")",
210  argv[optind]);
211  fprintf(stderr, _("Try \"%s --help\" for more information.\n"),
212  progname);
213  exit(1);
214  }
215 
216  printf(ngettext("%u second per test\n",
217  "%u seconds per test\n",
218  secs_per_test),
219  secs_per_test);
220 #if PG_O_DIRECT != 0
221  printf(_("O_DIRECT supported on this platform for open_datasync and open_sync.\n"));
222 #else
223  printf(_("Direct I/O is not supported on this platform.\n"));
224 #endif
225 }
226 
227 static void
229 {
230  int ops;
231 
232  /* write random data into buffer */
233  for (ops = 0; ops < DEFAULT_XLOG_SEG_SIZE; ops++)
234  full_buf[ops] = random();
235 
236  buf = (char *) TYPEALIGN(XLOG_BLCKSZ, full_buf);
237 }
238 
239 static void
241 {
242  int tmpfile;
243 
244  /*
245  * test if we can open the target file
246  */
247  if ((tmpfile = open(filename, O_RDWR | O_CREAT | PG_BINARY, S_IRUSR | S_IWUSR)) == -1)
248  die("could not open output file");
249  needs_unlink = 1;
250  if (write(tmpfile, full_buf, DEFAULT_XLOG_SEG_SIZE) !=
252  die("write failed");
253 
254  /* fsync now so that dirty buffers don't skew later tests */
255  if (fsync(tmpfile) != 0)
256  die("fsync failed");
257 
258  close(tmpfile);
259 }
260 
261 static void
262 test_sync(int writes_per_op)
263 {
264  int tmpfile,
265  ops,
266  writes;
267  bool fs_warning = false;
268 
269  if (writes_per_op == 1)
270  printf(_("\nCompare file sync methods using one %dkB write:\n"), XLOG_BLCKSZ_K);
271  else
272  printf(_("\nCompare file sync methods using two %dkB writes:\n"), XLOG_BLCKSZ_K);
273  printf(_("(in wal_sync_method preference order, except fdatasync is Linux's default)\n"));
274 
275  /*
276  * Test open_datasync if available
277  */
278  printf(LABEL_FORMAT, "open_datasync");
279  fflush(stdout);
280 
281 #ifdef OPEN_DATASYNC_FLAG
282  if ((tmpfile = open(filename, O_RDWR | O_DSYNC | PG_O_DIRECT | PG_BINARY, 0)) == -1)
283  {
284  printf(NA_FORMAT, _("n/a*"));
285  fs_warning = true;
286  }
287  else
288  {
289  START_TIMER;
290  for (ops = 0; alarm_triggered == false; ops++)
291  {
292  for (writes = 0; writes < writes_per_op; writes++)
293  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
294  die("write failed");
295  if (lseek(tmpfile, 0, SEEK_SET) == -1)
296  die("seek failed");
297  }
298  STOP_TIMER;
299  close(tmpfile);
300  }
301 #else
302  printf(NA_FORMAT, _("n/a"));
303 #endif
304 
305 /*
306  * Test fdatasync if available
307  */
308  printf(LABEL_FORMAT, "fdatasync");
309  fflush(stdout);
310 
311 #ifdef HAVE_FDATASYNC
312  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
313  die("could not open output file");
314  START_TIMER;
315  for (ops = 0; alarm_triggered == false; ops++)
316  {
317  for (writes = 0; writes < writes_per_op; writes++)
318  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
319  die("write failed");
320  fdatasync(tmpfile);
321  if (lseek(tmpfile, 0, SEEK_SET) == -1)
322  die("seek failed");
323  }
324  STOP_TIMER;
325  close(tmpfile);
326 #else
327  printf(NA_FORMAT, _("n/a"));
328 #endif
329 
330 /*
331  * Test fsync
332  */
333  printf(LABEL_FORMAT, "fsync");
334  fflush(stdout);
335 
336  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
337  die("could not open output file");
338  START_TIMER;
339  for (ops = 0; alarm_triggered == false; ops++)
340  {
341  for (writes = 0; writes < writes_per_op; writes++)
342  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
343  die("write failed");
344  if (fsync(tmpfile) != 0)
345  die("fsync failed");
346  if (lseek(tmpfile, 0, SEEK_SET) == -1)
347  die("seek failed");
348  }
349  STOP_TIMER;
350  close(tmpfile);
351 
352 /*
353  * If fsync_writethrough is available, test as well
354  */
355  printf(LABEL_FORMAT, "fsync_writethrough");
356  fflush(stdout);
357 
358 #ifdef HAVE_FSYNC_WRITETHROUGH
359  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
360  die("could not open output file");
361  START_TIMER;
362  for (ops = 0; alarm_triggered == false; ops++)
363  {
364  for (writes = 0; writes < writes_per_op; writes++)
365  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
366  die("write failed");
367  if (pg_fsync_writethrough(tmpfile) != 0)
368  die("fsync failed");
369  if (lseek(tmpfile, 0, SEEK_SET) == -1)
370  die("seek failed");
371  }
372  STOP_TIMER;
373  close(tmpfile);
374 #else
375  printf(NA_FORMAT, _("n/a"));
376 #endif
377 
378 /*
379  * Test open_sync if available
380  */
381  printf(LABEL_FORMAT, "open_sync");
382  fflush(stdout);
383 
384 #ifdef OPEN_SYNC_FLAG
385  if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG | PG_O_DIRECT | PG_BINARY, 0)) == -1)
386  {
387  printf(NA_FORMAT, _("n/a*"));
388  fs_warning = true;
389  }
390  else
391  {
392  START_TIMER;
393  for (ops = 0; alarm_triggered == false; ops++)
394  {
395  for (writes = 0; writes < writes_per_op; writes++)
396  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
397 
398  /*
399  * This can generate write failures if the filesystem has
400  * a large block size, e.g. 4k, and there is no support
401  * for O_DIRECT writes smaller than the file system block
402  * size, e.g. XFS.
403  */
404  die("write failed");
405  if (lseek(tmpfile, 0, SEEK_SET) == -1)
406  die("seek failed");
407  }
408  STOP_TIMER;
409  close(tmpfile);
410  }
411 #else
412  printf(NA_FORMAT, _("n/a"));
413 #endif
414 
415  if (fs_warning)
416  {
417  printf(_("* This file system and its mount options do not support direct\n"
418  " I/O, e.g. ext4 in journaled mode.\n"));
419  }
420 }
421 
422 static void
424 {
425  printf(_("\nCompare open_sync with different write sizes:\n"));
426  printf(_("(This is designed to compare the cost of writing 16kB in different write\n"
427  "open_sync sizes.)\n"));
428 
429  test_open_sync(_(" 1 * 16kB open_sync write"), 16);
430  test_open_sync(_(" 2 * 8kB open_sync writes"), 8);
431  test_open_sync(_(" 4 * 4kB open_sync writes"), 4);
432  test_open_sync(_(" 8 * 2kB open_sync writes"), 2);
433  test_open_sync(_("16 * 1kB open_sync writes"), 1);
434 }
435 
436 /*
437  * Test open_sync with different size files
438  */
439 static void
440 test_open_sync(const char *msg, int writes_size)
441 {
442 #ifdef OPEN_SYNC_FLAG
443  int tmpfile,
444  ops,
445  writes;
446 #endif
447 
448  printf(LABEL_FORMAT, msg);
449  fflush(stdout);
450 
451 #ifdef OPEN_SYNC_FLAG
452  if ((tmpfile = open(filename, O_RDWR | OPEN_SYNC_FLAG | PG_O_DIRECT | PG_BINARY, 0)) == -1)
453  printf(NA_FORMAT, _("n/a*"));
454  else
455  {
456  START_TIMER;
457  for (ops = 0; alarm_triggered == false; ops++)
458  {
459  for (writes = 0; writes < 16 / writes_size; writes++)
460  if (write(tmpfile, buf, writes_size * 1024) !=
461  writes_size * 1024)
462  die("write failed");
463  if (lseek(tmpfile, 0, SEEK_SET) == -1)
464  die("seek failed");
465  }
466  STOP_TIMER;
467  close(tmpfile);
468  }
469 #else
470  printf(NA_FORMAT, _("n/a"));
471 #endif
472 }
473 
474 static void
476 {
477  int tmpfile,
478  ops;
479 
480  /*
481  * Test whether fsync can sync data written on a different descriptor for
482  * the same file. This checks the efficiency of multi-process fsyncs
483  * against the same file. Possibly this should be done with writethrough
484  * on platforms which support it.
485  */
486  printf(_("\nTest if fsync on non-write file descriptor is honored:\n"));
487  printf(_("(If the times are similar, fsync() can sync data written on a different\n"
488  "descriptor.)\n"));
489 
490  /*
491  * first write, fsync and close, which is the normal behavior without
492  * multiple descriptors
493  */
494  printf(LABEL_FORMAT, "write, fsync, close");
495  fflush(stdout);
496 
497  START_TIMER;
498  for (ops = 0; alarm_triggered == false; ops++)
499  {
500  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
501  die("could not open output file");
502  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
503  die("write failed");
504  if (fsync(tmpfile) != 0)
505  die("fsync failed");
506  close(tmpfile);
507 
508  /*
509  * open and close the file again to be consistent with the following
510  * test
511  */
512  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
513  die("could not open output file");
514  close(tmpfile);
515  }
516  STOP_TIMER;
517 
518  /*
519  * Now open, write, close, open again and fsync This simulates processes
520  * fsyncing each other's writes.
521  */
522  printf(LABEL_FORMAT, "write, close, fsync");
523  fflush(stdout);
524 
525  START_TIMER;
526  for (ops = 0; alarm_triggered == false; ops++)
527  {
528  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
529  die("could not open output file");
530  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
531  die("write failed");
532  close(tmpfile);
533  /* reopen file */
534  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
535  die("could not open output file");
536  if (fsync(tmpfile) != 0)
537  die("fsync failed");
538  close(tmpfile);
539  }
540  STOP_TIMER;
541 }
542 
543 static void
545 {
546  int tmpfile,
547  ops;
548 
549  /*
550  * Test a simple write without fsync
551  */
552  printf(_("\nNon-sync'ed %dkB writes:\n"), XLOG_BLCKSZ_K);
553  printf(LABEL_FORMAT, "write");
554  fflush(stdout);
555 
556  START_TIMER;
557  for (ops = 0; alarm_triggered == false; ops++)
558  {
559  if ((tmpfile = open(filename, O_RDWR | PG_BINARY, 0)) == -1)
560  die("could not open output file");
561  if (write(tmpfile, buf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
562  die("write failed");
563  close(tmpfile);
564  }
565  STOP_TIMER;
566 }
567 
568 static void
569 signal_cleanup(int signum)
570 {
571  /* Delete the file if it exists. Ignore errors */
572  if (needs_unlink)
573  unlink(filename);
574  /* Finish incomplete line on stdout */
575  puts("");
576  exit(signum);
577 }
578 
579 #ifdef HAVE_FSYNC_WRITETHROUGH
580 
581 static int
583 {
584 #ifdef WIN32
585  return _commit(fd);
586 #elif defined(F_FULLFSYNC)
587  return (fcntl(fd, F_FULLFSYNC, 0) == -1) ? -1 : 0;
588 #else
589  errno = ENOSYS;
590  return -1;
591 #endif
592 }
593 #endif
594 
595 /*
596  * print out the writes per second for tests
597  */
598 static void
599 print_elapse(struct timeval start_t, struct timeval stop_t, int ops)
600 {
601  double total_time = (stop_t.tv_sec - start_t.tv_sec) +
602  (stop_t.tv_usec - start_t.tv_usec) * 0.000001;
603  double per_second = ops / total_time;
604  double avg_op_time_us = (total_time / ops) * USECS_SEC;
605 
606  printf(_(OPS_FORMAT), per_second, avg_op_time_us);
607 }
608 
609 #ifndef WIN32
610 static void
612 {
613  alarm_triggered = true;
614 }
615 #else
616 static DWORD WINAPI
617 process_alarm(LPVOID param)
618 {
619  /* WIN32 doesn't support alarm, so we create a thread and sleep here */
620  Sleep(secs_per_test * 1000);
621  alarm_triggered = true;
622  ExitThread(0);
623 }
624 #endif
#define PG_O_DIRECT
Definition: xlogdefs.h:72
static void test_sync(int writes_per_op)
static char * filename
Definition: pg_test_fsync.c:70
#define write(a, b, c)
Definition: win32.h:14
const char * get_progname(const char *argv0)
Definition: path.c:453
#define pg_log_error(...)
Definition: logging.h:80
int getopt_long(int argc, char *const argv[], const char *optstring, const struct option *longopts, int *longindex)
Definition: getopt_long.c:57
long random(void)
Definition: random.c:22
int pg_fsync_writethrough(int fd)
Definition: fd.c:413
void pg_logging_init(const char *argv0)
Definition: logging.c:81
#define USECS_SEC
Definition: pg_test_fsync.c:32
#define DEFAULT_XLOG_SEG_SIZE
#define printf(...)
Definition: port.h:221
#define fprintf
Definition: port.h:219
static int fd(const char *x, int i)
Definition: preproc-init.c:105
#define PG_BINARY
Definition: c.h:1213
static char full_buf[DEFAULT_XLOG_SEG_SIZE]
Definition: pg_test_fsync.c:68
static void test_open(void)
static bool alarm_triggered
Definition: pg_test_fsync.c:73
static void prepare_buf(void)
#define required_argument
Definition: getopt_long.h:25
#define fsync(fd)
Definition: win32_port.h:68
int optind
Definition: getopt.c:50
#define S_IWUSR
Definition: win32_port.h:283
static const char * progname
Definition: pg_test_fsync.c:64
static char * buf
Definition: pg_test_fsync.c:68
#define SIGHUP
Definition: win32_port.h:159
char * pg_strdup(const char *in)
Definition: fe_memutils.c:85
#define NA_FORMAT
Definition: pg_test_fsync.c:29
#define XLOG_BLCKSZ_K
Definition: pg_test_fsync.c:26
#define STOP_TIMER
Definition: pg_test_fsync.c:57
#define ngettext(s, p, n)
Definition: c.h:1124
#define PG_TEXTDOMAIN(domain)
Definition: c.h:1157
#define O_DSYNC
Definition: win32_port.h:328
#define LABEL_FORMAT
Definition: pg_test_fsync.c:28
static void print_elapse(struct timeval start_t, struct timeval stop_t, int ops)
static void test_open_syncs(void)
static void test_non_sync(void)
static unsigned int secs_per_test
Definition: pg_test_fsync.c:66
static int sig
Definition: pg_ctl.c:84
static struct timeval start_t stop_t
Definition: pg_test_fsync.c:71
#define TYPEALIGN(ALIGNVAL, LEN)
Definition: c.h:692
static void process_alarm(int sig)
pqsigfunc pqsignal(int signum, pqsigfunc handler)
Definition: signal.c:170
static void signal_cleanup(int sig)
int main(int argc, char *argv[])
#define START_TIMER
Definition: pg_test_fsync.c:36
#define OPS_FORMAT
Definition: pg_test_fsync.c:31
#define SIGALRM
Definition: win32_port.h:165
#define S_IRUSR
Definition: win32_port.h:280
static int needs_unlink
Definition: pg_test_fsync.c:67
void set_pglocale_pgservice(const char *argv0, const char *app)
Definition: exec.c:434
#define FSYNC_FILENAME
Definition: pg_test_fsync.c:24
char * optarg
Definition: getopt.c:52
static void test_file_descriptor_sync(void)
#define close(a)
Definition: win32.h:12
#define _(x)
Definition: elog.c:88
#define die(msg)
Definition: pg_test_fsync.c:97
static void test_open_sync(const char *msg, int writes_size)
static void handle_args(int argc, char *argv[])