1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * 2003, Copyright, Hewlett-Packard Development Compnay, LP.
6 * Developed under the sponsorship of the U.S. Government
7 * under Subcontract No. B514193
11 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
12 * Use is subject to license terms.
18 #include <sys/types.h>
29 #include <sys/ioctl.h>
35 #include <lustre/liblustreapi.h> /* for O_LOV_DELAY_CREATE */
37 #define CHECK_COUNT 10000
38 #define DISPLAY_COUNT (CHECK_COUNT * 10)
39 #define DISPLAY_TIME 100
67 struct option longOpts[] = {
68 {"create", 0, NULL, CREATE },
69 {"lookup", 0, NULL, LOOKUP },
70 {"mknod", 0, NULL, MKNOD },
71 {"open", 0, NULL, OPEN },
72 {"stat", 0, NULL, STAT },
73 {"unlink", 0, NULL, UNLINK },
74 {"begin", 1, NULL, BEGIN },
75 {"iters", 1, NULL, ITERS },
76 {"time", 1, NULL, TIME }, /* seconds */
77 {"dirfmt", 1, NULL, DIRFMT },
78 {"ndirs", 1, NULL, NDIRS },
79 {"filefmt", 1, NULL, FILEFMT },
80 {"nfiles", 1, NULL, NFILES },
81 {"noexcl", 0, NULL, NOEXCL },
82 {"stripes", 1, NULL, STRIPES },
83 {"seed", 1, NULL, SEED },
84 {"seedfile", 1, NULL, SEEDFILE },
85 {"random_order", 0, NULL, RANDOM },
86 {"readdir_order", 0, NULL, READDIR },
87 {"recreate", 0, NULL, RECREATE },
88 {"verbose", 0, NULL, VERBOSE },
89 {"debug", 0, NULL, DEBUG },
90 {"help", 0, NULL, HELP },
100 char hostname[512] = "unknown";
103 int openflags = O_RDWR|O_CREAT|O_EXCL;
107 char mkdir_cmd[PATH_MAX+14];
111 struct dirent *dir_entry;
113 char filefmt[PATH_MAX];
114 char filename[PATH_MAX];
122 struct sigaction act;
130 #define dmesg if (debug) printf
132 #define DISPLAY_PROGRESS() { \
133 if ((++nops % CHECK_COUNT) == 0 && verbose) { \
135 interval = curTime - lastTime; \
136 if (interval > DISPLAY_TIME || nops % DISPLAY_COUNT == 0) { \
137 rate = (float)(nops - lastOps); \
139 rate /= (float)interval; \
140 printf("Rank %d: %.2f %ss/sec %lu secs " \
141 "(total: %d %ss %lu secs)\n", \
142 myrank, rate, cmd, interval, \
143 nops, cmd, curTime - startTime); \
145 lastTime = curTime; \
150 char *usage_msg = "usage: %s\n"
151 " { --create [ --noexcl ] | --lookup | --mknod |\n"
152 " --open | --stat | --unlink [ --recreate ] }\n"
153 " [ --help ] [ --verbose ] [ --debug ]\n"
154 " { [ --begin <num> ] --nfiles <num> }\n"
155 " [ --iters <num> ] [ --time <secs> ]\n"
156 " [ --dirfmt <str> ] [ --ndirs <num> ]\n"
157 " [ --filefmt <str> ] [ --stripes <num> ]\n"
158 " [ --random_order [--seed <num> | --seedfile <file>] ]\n"
159 " [ --readdir_order ]\n";
162 usage(FILE *stream, char *fmt, ...)
168 fprintf(stream, "%s: ", prog);
170 vfprintf(stderr, fmt, ap);
173 fprintf(stream, usage_msg, prog);
177 exit(stream == stderr);
180 /* Print process myrank and message, and exit (i.e. a fatal error) */
182 fatal(int rank, const char *fmt, ...)
184 if (rank == myrank) {
187 fprintf(stderr, "rank %d: ", rank);
189 vfprintf(stderr, fmt, ap);
193 MPI_Abort(MPI_COMM_WORLD, 1);
198 sigalrm_handler(int signum)
203 /* HAVE_LLAPI_FILE_LOOKUP is defined by liblustreapi.h if this function is
204 * defined therein. Otherwise we can do the equivalent operation via ioctl
205 * if we have access to a complete lustre build tree to get the various
206 * definitions - then compile with USE_MDC_LOOKUP defined. */
207 #if defined(HAVE_LLAPI_FILE_LOOKUP)
208 #define HAVE_MDC_LOOKUP
209 #elif defined(USE_MDC_LOOKUP)
211 #include <liblustre.h>
212 #include <linux/lustre_lib.h>
214 int llapi_file_lookup(int dirfd, const char *name)
216 struct obd_ioctl_data data = { 0 };
221 if (dirfd < 0 || name == NULL)
224 data.ioc_version = OBD_IOCTL_VERSION;
225 data.ioc_len = sizeof(data);
226 data.ioc_inlbuf1 = name;
227 data.ioc_inllen1 = strlen(name) + 1;
229 rc = obd_ioctl_pack(&data, &buf, sizeof(rawbuf));
231 fatal(myrank, "ioctl_pack failed: rc = %d\n", rc);
235 return ioctl(fd, IOC_MDC_LOOKUP, buf);
237 #define HAVE_MDC_LOOKUP
241 process_args(int argc, char *argv[])
243 char c, *cp, *endptr;
244 int i, index, offset, tmpend, rc;
251 prog = basename(argv[0]);
252 strcpy(filefmt, "f%d");
253 gethostname(hostname, sizeof(hostname));
255 /* auto create shortOpts rather than maintaining a static string. */
256 for (opt = longOpts, cp = shortOpts; opt->name != NULL; opt++, cp++) {
262 while ((c = getopt_long(argc,argv, shortOpts, longOpts,&index)) != -1) {
265 openflags &= ~(O_CREAT|O_EXCL);
267 #ifdef HAVE_MDC_LOOKUP
274 fatal(0, "Invalid - more than one operation "
276 longOpts[index].name);
279 cmd = (char *)longOpts[index].name;
282 if (mode != CREATE && mode != MKNOD) {
283 usage(stderr, "--noexcl only applies to "
284 "--create or --mknod.\n");
286 openflags &= ~O_EXCL;
289 if (mode != UNLINK) {
290 usage(stderr, "--recreate only makes sense"
296 begin = strtol(optarg, &endptr, 0);
297 if ((*endptr != 0) || (begin < 0)) {
298 fatal(0, "Invalid --start value.\n");
302 iters = strtol(optarg, &endptr, 0);
303 if ((*endptr != 0) || (iters <= 0)) {
304 fatal(0, "Invalid --iters value.\n");
306 if (mode != LOOKUP && mode != OPEN && mode != STAT) {
307 usage(stderr, "--iters only makes sense with "
308 "--lookup, --open, or --stat.\n");
312 seconds = strtol(optarg, &endptr, 0);
313 if ((*endptr != 0) || (seconds <= 0)) {
314 fatal(0, "Invalid --time value.\n");
318 if (strlen(optarg) > (PATH_MAX - 16)) {
319 fatal(0, "--dirfmt too long\n");
324 ndirs = strtol(optarg, &endptr, 0);
325 if ((*endptr != 0) || (ndirs <= 0)) {
326 fatal(0, "Invalid --ndirs value.\n");
328 if ((ndirs > nthreads) &&
329 ((mode == CREATE) || (mode == MKNOD))) {
330 fatal(0, "--ndirs=%d must be less than or "
331 "equal to the number of threads (%d).\n",
336 if (strlen(optarg) > 4080) {
337 fatal(0, "--filefmt too long\n");
340 /* Use %%d where you want the file # in the name. */
341 sprintf(filefmt, optarg, myrank);
344 nfiles = strtol(optarg, &endptr, 0);
345 if ((*endptr != 0) || (nfiles <= 0)) {
346 fatal(0, "Invalid --nfiles value.\n");
350 stripes = strtol(optarg, &endptr, 0);
351 if ((*endptr != 0) || (stripes < 0)) {
352 fatal(0, "Invalid --stripes value.\n");
356 openflags |= O_LOV_DELAY_CREATE;
358 fatal(0, "non-zero --stripes value "
359 "not yet supported.\n");
364 seed = strtoul(optarg, &endptr, 0);
366 fatal(0, "bad --seed option %s\n", optarg);
370 seed_file = fopen(optarg, "r");
372 fatal(myrank, "fopen(%s) error: %s\n",
373 optarg, strerror(errno));
376 for (i = -1; fgets(tmp, 16, seed_file) != NULL;) {
382 rc = sscanf(tmp, "%d", &seed);
383 if ((rc != 1) || (seed < 0)) {
384 fatal(myrank, "Invalid seed value '%s' "
385 "at line %d in %s.\n",
389 fatal(myrank, "File '%s' too short. Does not "
390 "contain a seed for thread %d.\n",
398 if (mode != LOOKUP && mode != OPEN && mode != STAT) {
399 fatal(0, "--%s can only be specified with "
400 "--lookup, --open, or --stat.\n",
401 (char *)longOpts[index].name);
413 usage(stderr, "unrecognized option: '%c'.\n", optopt);
418 usage(stderr, "too many arguments %d >= %d.\n", optind, argc);
421 if (mode == CREATE || mode == MKNOD || mode == UNLINK) {
425 } else if (nfiles == 0) {
426 usage(stderr, "--nfiles or --time must be specified "
429 } else if (mode == LOOKUP || mode == OPEN || mode == STAT) {
433 } else if (iters == 0) {
434 usage(stderr, "--iters or --time must be specifed "
439 usage(stderr, "--nfiles must be specifed with --%s.\n",
444 int fd = open("/dev/urandom", O_RDONLY);
447 if (read(fd, &seed, sizeof(seed)) <
458 dmesg("%s: rank %d seed %d (%s).\n", prog, myrank, seed,
459 (order == RANDOM) ? "random_order" : "readdir_order");
461 usage(stderr, "one --create, --mknod, --open, --stat,"
462 #ifdef HAVE_MDC_LOOKUP
465 " or --unlink must be specifed.");
468 /* support for multiple threads in a dir, set begin/end appropriately.*/
469 dirnum = myrank % ndirs;
470 dirthreads = nthreads / ndirs;
471 if (nthreads > (ndirs * dirthreads + dirnum))
474 offset = myrank / ndirs;
476 tmpend = begin + nfiles - 1;
480 end = begin + (nfiles / dirthreads) * dirthreads + offset;
481 if ((end > tmpend) || (end <= 0))
490 dmesg("%d: iters %d nfiles %d time %d begin %d end %d dirthreads %d."
491 "\n", myrank, iters, nfiles, seconds, begin, end, dirthreads);
493 if (dirfmt == NULL) {
496 sprintf(dir, dirfmt, dirnum);
498 sprintf(mkdir_cmd, "/bin/mkdir -p %s", dir);
499 #ifdef _LIGHTWEIGHT_KERNEL
500 printf("NOTICE: not running system(%s)\n", mkdir_cmd);
502 rc = system(mkdir_cmd);
504 fatal(myrank, "'%s' failed.\n", mkdir_cmd);
510 fatal(myrank, "unable to chdir to '%s'.\n", dir);
515 static inline char *next_file()
517 if (order == RANDOM) {
518 sprintf(filename, filefmt, random() % nfiles);
524 dir_entry = readdir(directory);
525 if (dir_entry == NULL) {
526 rewinddir(directory);
527 while ((dir_entry = readdir(directory)) != NULL) {
528 if (dir_entry->d_name[0] != '.')
529 return(dir_entry->d_name);
532 fatal(myrank, "unable to read directory %s (%s).\n",
533 dir, strerror(errno));
536 return(dir_entry->d_name);
540 main(int argc, char *argv[])
542 int i, j, fd, rc, nops, lastOps, ag_ops;
544 time_t startTime, lastTime, curTime, interval;
547 rc = MPI_Init(&argc, &argv);
548 if (rc != MPI_SUCCESS)
549 fatal(myrank, "MPI_Init failed: %d\n", rc);
551 rc = MPI_Comm_size(MPI_COMM_WORLD, &nthreads);
552 if (rc != MPI_SUCCESS)
553 fatal(myrank, "MPI_Comm_size failed: %d\n", rc);
555 rc = MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
556 if (rc != MPI_SUCCESS)
557 fatal(myrank, "MPI_Comm_rank failed: %d\n", rc);
559 process_args(argc, argv);
562 if ((myrank == 0) || debug) {
563 printf("%d: %s starting at %s",
564 myrank, hostname, ctime(&startTime));
567 /* if we're not measuring creation rates then precreate
568 * the files we're operating on. */
569 if ((mode != CREATE) && (mode != MKNOD)) {
570 /* create the files in reverse order. When we encounter
571 * a file that already exists, assume the remainder of
572 * the files exist to save time. The timed performance
573 * test scripts make use of this behavior. */
574 for (i = end, j = 0; i >= begin; i -= dirthreads) {
575 sprintf(filename, filefmt, i);
576 fd = open(filename, openflags, 0644);
581 fatal(myrank, "precreate open(%s) error: %s\n",
582 filename, strerror(rc));
587 dmesg("%d: %s pre-created %d files.\n",myrank,hostname,j);
589 rc = MPI_Barrier(MPI_COMM_WORLD);
590 if (rc != MPI_SUCCESS)
591 fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
594 if (order == READDIR) {
595 directory = opendir(dir);
596 if (directory == NULL) {
598 fatal(myrank, "opendir(%s) error: %s\n",
603 j = random() % nfiles;
604 dmesg("%d: %s initializing dir offset %u: %s",
605 myrank, hostname, j, ctime(&startTime));
607 for (i = 0; i <= j; i++) {
608 if ((dir_entry = readdir(directory)) == NULL) {
609 fatal(myrank, "could not read entry number %d "
610 "in directory %s.\n", i, dir);
615 dmesg("%d: index %d, filename %s, offset %ld: "
616 "%s initialization complete: %s",
617 myrank, i, dir_entry->d_name, telldir(directory),
618 hostname, ctime(&lastTime));
621 rc = MPI_Barrier(MPI_COMM_WORLD);
622 if (rc != MPI_SUCCESS)
623 fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
626 act.sa_handler = sigalrm_handler;
627 (void)sigemptyset(&act.sa_mask);
629 sigaction(SIGALRM, &act, NULL);
633 startTime = lastTime = time(0);
638 for (; begin <= end && !alarm_caught; begin += dirthreads) {
639 sprintf(filename, filefmt, begin);
640 if ((fd = open(filename, openflags, 0644)) < 0) {
641 if (((rc = errno) == EINTR) && alarm_caught)
643 fatal(myrank, "open(%s) error: %s\n",
644 filename, strerror(rc));
651 dmesg("%d: created %d files, last file '%s'.\n",
652 myrank, nops, filename);
654 #ifdef HAVE_MDC_LOOKUP
656 fd = open(dir, O_RDONLY);
658 fatal(myrank, "open(dir == '%s') error: %s\n",
659 dir, strerror(errno));
662 for (; nops < iters && !alarm_caught;) {
663 char *filename = next_file();
664 rc = llapi_file_lookup(fd, filename);
666 if (((rc = errno) == EINTR) && alarm_caught)
668 fatal(myrank, "llapi_file_lookup(%s) "
669 "error: %s\n", filename, strerror(rc));
677 for (; begin <= end && !alarm_caught; begin += dirthreads) {
678 sprintf(filename, filefmt, begin);
679 rc = mknod(filename, S_IFREG| 0644, 0);
681 if (((rc = errno) == EINTR) && alarm_caught)
683 fatal(myrank, "mknod(%s) error: %s\n",
684 filename, strerror(rc));
691 for (; nops < iters && !alarm_caught;) {
693 if ((fd = open(file, openflags, 0644)) < 0) {
694 if (((rc = errno) == EINTR) && alarm_caught)
696 fatal(myrank, "open(%s) error: %s\n",
706 for (; nops < iters && !alarm_caught;) {
707 rc = stat(file = next_file(), &statbuf);
709 if (((rc = errno) == EINTR) && alarm_caught)
711 fatal(myrank, "stat(%s) error: %s\n",
719 for (; begin <= end && !alarm_caught; begin += dirthreads) {
720 sprintf(filename, filefmt, begin);
721 rc = unlink(filename);
723 if (((rc = errno) == EINTR) && alarm_caught)
725 fatal(myrank, "unlink(%s) error: %s\n",
726 filename, strerror(rc));
735 interval = curTime - startTime;
736 rate = (float)(nops);
738 rate /= (float)interval;
740 rc = MPI_Reduce(&nops, &ag_ops, 1, MPI_INT, MPI_SUM, 0,
742 if (rc != MPI_SUCCESS) {
743 fatal(myrank, "Failure in MPI_Reduce of total ops.\n");
746 rc = MPI_Reduce(&rate, &ag_rate, 1, MPI_FLOAT, MPI_SUM, 0,
748 if (rc != MPI_SUCCESS) {
749 fatal(myrank, "Failure in MPI_Reduce of aggregated rate.\n");
753 printf("Rate: %.2f %ss/sec (total: %d threads %d %ss %lu secs)"
754 "\n", ag_rate, cmd, nthreads, ag_ops, cmd, interval);
758 for (begin = beginsave; begin <= end; begin += dirthreads) {
759 sprintf(filename, filefmt, begin);
760 if ((fd = open(filename, openflags, 0644)) < 0) {
764 fatal(myrank, "recreate open(%s) error: %s\n",
765 filename, strerror(rc));
773 if ((myrank == 0) || debug) {
774 printf("%d: %s finished at %s",
775 myrank, hostname, ctime(&curTime));