lockstat.c

Sun Solaris 10 中的 DTrace 组件的源代码。请参看: http://www.sun.com/software/solaris/observability.jsp

			break;

		case 'R':
			g_rates = 1;
			break;

		case 'p':
			g_pflag = 1;
			break;

		case 'P':
			g_Pflag = 1;
			break;

		case 'o':
			if ((out = fopen(optarg, "w")) == NULL)
				fail(1, "error opening file");
			break;

		case 'V':
			g_Vflag = 1;
			break;

		default:
			usage();
		}
	}

	if (filt != NULL) {
		predicate_add(&g_predicate, filt, NULL, 0);
		filter_destroy(&filt);
	}

	if (ifilt != NULL) {
		predicate_add(&g_ipredicate, ifilt, NULL, 0);
		filter_destroy(&ifilt);
	}

	argc -= optind;
	argv += optind;

	if (g_recsize == 0) {
		if (g_gflag) {
			g_stkdepth = LS_MAX_STACK_DEPTH;
			g_recsize = LS_STACK(g_stkdepth);
		} else {
			g_recsize = LS_TIME;
		}
	}

	if (g_gflag && g_recsize <= LS_STACK(0))
		fail(0, "'-g' requires at least '-s 1' data gathering");

	/*
	 * Make sure the alignment is reasonable
	 */
	g_recsize = -(-g_recsize & -sizeof (uint64_t));

	for (i = 0; i < LS_MAX_EVENTS; i++) {
		/*
		 * If no events were specified, enable -C.
		 */
		if (!events_specified && g_event_info[i].ev_type == 'C')
			g_enabled[i] = 1;
	}

	for (i = 0; i < LS_MAX_EVENTS; i++) {
		if (g_enabled[i])
			dprog_addevent(i);
	}

	/*
	 * Make sure there is a child command to execute
	 */
	if (argc <= 0)
		usage();

	data_buf_size = (g_nrecs + 1) * g_recsize;

	/*
	 * By default, we set our aggregation size to be the number of records
	 * times a slop factor.
	 */
	(void) sprintf(data_buf_str, "%ld", (long)(g_nrecs * 512));

	if (!g_tracing) {
		if (dtrace_setopt(g_dtp, "bufsize", "4k") == -1)
			dfail("failed to set 'bufsize'");

		if (dtrace_setopt(g_dtp, "aggsize", data_buf_str) == -1)
			dfail("failed to set 'aggsize'");
	} else {
		if (dtrace_setopt(g_dtp, "bufsize", data_buf_str) == -1)
			dfail("failed to set 'bufsize'");
	}

	if (dtrace_setopt(g_dtp, "statusrate", "10sec") == -1)
		dfail("failed to set 'statusrate'");

	if ((data_buf = memalign(sizeof (uint64_t), data_buf_size)) == NULL)
		fail(1, "Memory allocation failed");

	dprog_compile();
	status_init();

	g_elapsed = -gethrtime();

	/*
	 * Spawn the specified command and wait for it to complete.
	 */
	child = fork();
	if (child == -1)
		fail(1, "cannot fork");
	if (child == 0) {
		(void) dtrace_close(g_dtp);
		(void) execvp(argv[0], &argv[0]);
		exec_errno = errno;
		exit(127);
	}

	while (waitpid(child, &status, WEXITED) != child)
		status_check();

	g_elapsed += gethrtime();

	if (WIFEXITED(status)) {
		if (WEXITSTATUS(status) != 0) {
			if (exec_errno != 0) {
				errno = exec_errno;
				fail(1, "could not execute %s", argv[0]);
			}
			(void) fprintf(stderr,
			    "lockstat: warning: %s exited with code %d\n",
				argv[0], WEXITSTATUS(status));
		}
	} else {
		(void) fprintf(stderr,
		    "lockstat: warning: %s died on signal %d\n",
			argv[0], WTERMSIG(status));
	}

	if (dtrace_stop(g_dtp) == -1)
		dfail("failed to stop dtrace");

	/*
	 * Read out the DTrace data.
	 */
	g_nrecs_used = process_data(out, data_buf);

	if (g_nrecs_used >= g_nrecs || g_dropped)
		(void) fprintf(stderr, "lockstat: warning: "
		    "ran out of data records (use -n for more)\n");

	/* LINTED - alignment */
	for (i = 0, lsp = (lsrec_t *)data_buf; i < g_nrecs_used; i++,
	    /* LINTED - alignment */
	    lsp = (lsrec_t *)((char *)lsp + g_recsize)) {
		ev_count[lsp->ls_event] += lsp->ls_count;
		ev_time[lsp->ls_event] += lsp->ls_time;
	}

	/*
	 * If -g was specified, convert stacks into individual records.
	 */
	if (g_gflag) {
		lsrec_t *newlsp, *oldlsp;

		newlsp = memalign(sizeof (uint64_t),
		    g_nrecs_used * LS_TIME * (g_stkdepth + 1));
		if (newlsp == NULL)
			fail(1, "Cannot allocate space for -g processing");
		lsp = newlsp;
		/* LINTED - alignment */
		for (i = 0, oldlsp = (lsrec_t *)data_buf; i < g_nrecs_used; i++,
		    /* LINTED - alignment */
		    oldlsp = (lsrec_t *)((char *)oldlsp + g_recsize)) {
			int fr;
			int caller_in_stack = 0;

			if (oldlsp->ls_count == 0)
				continue;

			for (fr = 0; fr < g_stkdepth; fr++) {
				if (oldlsp->ls_stack[fr] == 0)
					break;
				if (oldlsp->ls_stack[fr] == oldlsp->ls_caller)
					caller_in_stack = 1;
				bcopy(oldlsp, lsp, LS_TIME);
				lsp->ls_caller = oldlsp->ls_stack[fr];
				/* LINTED - alignment */
				lsp = (lsrec_t *)((char *)lsp + LS_TIME);
			}
			if (!caller_in_stack) {
				bcopy(oldlsp, lsp, LS_TIME);
				/* LINTED - alignment */
				lsp = (lsrec_t *)((char *)lsp + LS_TIME);
			}
		}
		g_nrecs = g_nrecs_used =
		    ((uintptr_t)lsp - (uintptr_t)newlsp) / LS_TIME;
		g_recsize = LS_TIME;
		g_stkdepth = 0;
		free(data_buf);
		data_buf = (char *)newlsp;
	}

	if ((sort_buf = calloc(2 * (g_nrecs + 1),
	    sizeof (void *))) == NULL)
		fail(1, "Sort buffer allocation failed");
	merge_buf = sort_buf + (g_nrecs + 1);

	/*
	 * Build the sort buffer, discarding zero-count records along the way.
	 */
	/* LINTED - alignment */
	for (i = 0, lsp = (lsrec_t *)data_buf; i < g_nrecs_used; i++,
	    /* LINTED - alignment */
	    lsp = (lsrec_t *)((char *)lsp + g_recsize)) {
		if (lsp->ls_count == 0)
			lsp->ls_event = LS_MAX_EVENTS;
		sort_buf[i] = lsp;
	}

	if (g_nrecs_used == 0)
		exit(0);

	/*
	 * Add a sentinel after the last record
	 */
	sort_buf[i] = lsp;
	lsp->ls_event = LS_MAX_EVENTS;

	if (g_tracing) {
		report_trace(out, sort_buf);
		return (0);
	}

	/*
	 * Application of -g may have resulted in multiple records
	 * with the same signature; coalesce them.
	 */
	if (g_gflag) {
		mergesort(lockcmp, sort_buf, merge_buf, g_nrecs_used);
		coalesce(lockcmp, sort_buf, g_nrecs_used);
	}

	/*
	 * Coalesce locks within the same symbol if -c option specified.
	 * Coalesce PCs within the same function if -k option specified.
	 */
	if (g_cflag || g_kflag) {
		for (i = 0; i < g_nrecs_used; i++) {
			int fr;
			lsp = sort_buf[i];
			if (g_cflag)
				coalesce_symbol(&lsp->ls_lock);
			if (g_kflag) {
				for (fr = 0; fr < g_stkdepth; fr++)
					coalesce_symbol(&lsp->ls_stack[fr]);
				coalesce_symbol(&lsp->ls_caller);
			}
		}
		mergesort(lockcmp, sort_buf, merge_buf, g_nrecs_used);
		coalesce(lockcmp, sort_buf, g_nrecs_used);
	}

	/*
	 * Coalesce callers if -w option specified
	 */
	if (g_wflag) {
		mergesort(lock_and_count_cmp_anywhere,
		    sort_buf, merge_buf, g_nrecs_used);
		coalesce(lockcmp_anywhere, sort_buf, g_nrecs_used);
	}

	/*
	 * Coalesce locks if -W option specified
	 */
	if (g_Wflag) {
		mergesort(site_and_count_cmp_anylock,
		    sort_buf, merge_buf, g_nrecs_used);
		coalesce(sitecmp_anylock, sort_buf, g_nrecs_used);
	}

	/*
	 * Sort data by contention count (ls_count) or total time (ls_time),
	 * depending on g_Pflag.  Override g_Pflag if time wasn't measured.
	 */
	if (g_recsize < LS_TIME)
		g_Pflag = 0;

	if (g_Pflag)
		mergesort(timecmp, sort_buf, merge_buf, g_nrecs_used);
	else
		mergesort(countcmp, sort_buf, merge_buf, g_nrecs_used);

	/*
	 * Display data by event type
	 */
	first = &sort_buf[0];
	while ((event = (*first)->ls_event) < LS_MAX_EVENTS) {
		current = first;
		while ((lsp = *current)->ls_event == event)
			current++;
		report_stats(out, first, current - first, ev_count[event],
		    ev_time[event]);
		first = current;
	}

	return (0);
}

static char *
format_symbol(char *buf, uintptr_t addr, int show_size)
{
	uintptr_t symoff;
	char *symname;
	size_t symsize;

	symname = addr_to_sym(addr, &symoff, &symsize);

	if (show_size && symoff == 0)
		(void) sprintf(buf, "%s[%ld]", symname, (long)symsize);
	else if (symoff == 0)
		(void) sprintf(buf, "%s", symname);
	else if (symoff < 16 && bcmp(symname, "cpu[", 4) == 0)	/* CPU+PIL */
		(void) sprintf(buf, "%s+%ld", symname, (long)symoff);
	else if (symoff <= symsize || (symoff < 256 && addr != symoff))
		(void) sprintf(buf, "%s+0x%llx", symname,
		    (unsigned long long)symoff);
	else
		(void) sprintf(buf, "0x%llx", (unsigned long long)addr);
	return (buf);
}

static void
report_stats(FILE *out, lsrec_t **sort_buf, size_t nrecs, uint64_t total_count,
	uint64_t total_time)
{
	uint32_t event = sort_buf[0]->ls_event;
	lsrec_t *lsp;
	double ptotal = 0.0;
	double percent;
	int i, j, fr;
	int displayed;
	int first_bin, last_bin, max_bin_count, total_bin_count;
	int rectype;
	char buf[256];
	char lhdr[80], chdr[80];

	rectype = g_recsize;

	if (g_topn == 0) {
		(void) fprintf(out, "%20llu %s\n",
		    g_rates == 0 ? total_count :
		    ((unsigned long long)total_count * NANOSEC) / g_elapsed,
		    g_event_info[event].ev_desc);
		return;
	}

	(void) sprintf(lhdr, "%s%s",
	    g_Wflag ? "Hottest " : "", g_event_info[event].ev_lhdr);
	(void) sprintf(chdr, "%s%s",
	    g_wflag ? "Hottest " : "", "Caller");

	if (!g_pflag)
		(void) fprintf(out,
		    "\n%s: %.0f events in %.3f seconds (%.0f events/sec)\n\n",
		    g_event_info[event].ev_desc, (double)total_count,
		    (double)g_elapsed / NANOSEC,
		    (double)total_count * NANOSEC / g_elapsed);

	if (!g_pflag && rectype < LS_HIST) {
		(void) sprintf(buf, "%s", g_event_info[event].ev_units);
		(void) fprintf(out, "%5s %4s %4s %4s %8s %-22s %-24s\n",
		    g_rates ? "ops/s" : "Count",
		    g_gflag ? "genr" : "indv",
		    "cuml", "rcnt", rectype >= LS_TIME ? buf : "", lhdr, chdr);
		(void) fprintf(out, "---------------------------------"
		    "----------------------------------------------\n");
	}

	displayed = 0;
	for (i = 0; i < nrecs; i++) {
		lsp = sort_buf[i];

		if (displayed++ >= g_topn)
			break;

		if (g_pflag) {
			int j;

			(void) fprintf(out, "%u %u",
			    lsp->ls_event, lsp->ls_count);
			(void) fprintf(out, " %s",
			    format_symbol(buf, lsp->ls_lock, g_cflag));
			(void) fprintf(out, " %s",
			    format_symbol(buf, lsp->ls_caller, 0));
			(void) fprintf(out, " %f",
			    (double)lsp->ls_refcnt / lsp->ls_count);
			if (rectype >= LS_TIME)
				(void) fprintf(out, " %llu",
				    (unsigned long long)lsp->ls_time);
			if (rectype >= LS_HIST) {
				for (j = 0; j < 64; j++)
					(void) fprintf(out, " %u",
					    lsp->ls_hist[j]);
			}
			for (j = 0; j < LS_MAX_STACK_DEPTH; j++) {
				if (rectype <= LS_STACK(j) ||
				    lsp->ls_stack[j] == 0)
					break;
				(void) fprintf(out, " %s",
				    format_symbol(buf, lsp->ls_stack[j], 0));
			}
			(void) fprintf(out, "\n");
			continue;
		}

		if (rectype >= LS_HIST) {
			(void) fprintf(out, "---------------------------------"
			    "----------------------------------------------\n");
			(void) sprintf(buf, "%s",
			    g_event_info[event].ev_units);
			(void) fprintf(out, "%5s %4s %4s %4s %8s %-22s %-24s\n",
			    g_rates ? "ops/s" : "Count",
			    g_gflag ? "genr" : "indv",
			    "cuml", "rcnt", buf, lhdr, chdr);
		}

		if (g_Pflag && total_time != 0)
			percent = (lsp->ls_time * 100.00) / total_time;
		else
			percent = (lsp->ls_count * 100.00) / total_count;

		ptotal += percent;

		if (rectype >= LS_TIME)
			(void) sprintf(buf, "%llu",
			    (unsigned long long)(lsp->ls_time / lsp->ls_count));
		else
			buf[0] = '\0';

		(void) fprintf(out, "%5llu ",
		    g_rates == 0 ? lsp->ls_count :
		    ((uint64_t)lsp->ls_count * NANOSEC) / g_elapsed);

		(void) fprintf(out, "%3.0f%% ", percent);

		if (g_gflag)
			(void) fprintf(out, "---- ");
		else
			(void) fprintf(out, "%3.0f%% ", ptotal);

		(void) fprintf(out, "%4.2f %8s ",
		    (double)lsp->ls_refcnt / lsp->ls_count, buf);

		(void) fprintf(out, "%-22s ",
		    format_symbol(buf, lsp->ls_lock, g_cflag));

		(void) fprintf(out, "%-24s\n",
		    format_symbol(buf, lsp->ls_caller, 0));

		if (rectype < LS_HIST)
			continue;

		(void) fprintf(out, "\n");
		(void) fprintf(out, "%10s %31s %-9s %-24s\n",
			g_event_info[event].ev_units,
			"------ Time Distribution ------",
			g_rates ? "ops/s" : "count",
			rectype > LS_STACK(0) ? "Stack" : "");

		first_bin = 0;
		while (lsp->ls_hist[first_bin] == 0)
			first_bin++;

		last_bin = 63;
		while (lsp->ls_hist[last_bin] == 0)
			last_bin--;

		max_bin_count = 0;
		total_bin_count = 0;
		for (j = first_bin; j <= last_bin; j++) {
			total_bin_count += lsp->ls_hist[j];
			if (lsp->ls_hist[j] > max_bin_count)
				max_bin_count = lsp->ls_hist[j];
		}

		/*
		 * If we went a few frames below the caller, ignore them
		 */
		for (fr = 3; fr > 0; fr--)
			if (lsp->ls_stack[fr] == lsp->ls_caller)
				break;

		for (j = first_bin; j <= last_bin; j++) {
			uint_t depth = (lsp->ls_hist[j] * 30) / total_bin_count;
			(void) fprintf(out, "%10llu |%s%s %-9u ",
			    1ULL << j,
			    "@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@" + 30 - depth,
			    "                              " + depth,
			    g_rates == 0 ? lsp->ls_hist[j] :
			    (uint_t)(((uint64_t)lsp->ls_hist[j] * NANOSEC) /
			    g_elapsed));
			if (rectype <= LS_STACK(fr) || lsp->ls_stack[fr] == 0) {
				(void) fprintf(out, "\n");
				continue;
			}
			(void) fprintf(out, "%-24s\n",
			    format_symbol(buf, lsp->ls_stack[fr], 0));
			fr++;
		}
		while (rectype > LS_STACK(fr) && lsp->ls_stack[fr] != 0) {
			(void) fprintf(out, "%15s %-36s %-24s\n", "", "",
			    format_symbol(buf, lsp->ls_stack[fr], 0));
			fr++;
		}
	}

	if (!g_pflag)
		(void) fprintf(out, "---------------------------------"
		    "----------------------------------------------\n");

	(void) fflush(out);
}

static void
report_trace(FILE *out, lsrec_t **sort_buf)
{
	lsrec_t *lsp;
	int i, fr;
	int rectype;
	char buf[256], buf2[256];

	rectype = g_recsize;

	if (!g_pflag) {
		(void) fprintf(out, "%5s  %7s  %11s  %-24s  %-24s\n",
		    "Event", "Time", "Owner", "Lock", "Caller");
		(void) fprintf(out, "---------------------------------"
		    "----------------------------------------------\n");
	}

	for (i = 0; i < g_nrecs_used; i++) {

		lsp = sort_buf[i];

		if (lsp->ls_event >= LS_MAX_EVENTS || lsp->ls_count == 0)
			continue;

		(void) fprintf(out, "%2d  %10llu  %11p  %-24s  %-24s\n",
		    lsp->ls_event, (unsigned long long)lsp->ls_time,
		    (void *)lsp->ls_next,
		    format_symbol(buf, lsp->ls_lock, 0),
		    format_symbol(buf2, lsp->ls_caller, 0));

		if (rectype <= LS_STACK(0))
			continue;

		/*
		 * If we went a few frames below the caller, ignore them
		 */
		for (fr = 3; fr > 0; fr--)
			if (lsp->ls_stack[fr] == lsp->ls_caller)
				break;

		while (rectype > LS_STACK(fr) && lsp->ls_stack[fr] != 0) {
			(void) fprintf(out, "%53s  %-24s\n", "",
			    format_symbol(buf, lsp->ls_stack[fr], 0));
			fr++;
		}
		(void) fprintf(out, "\n");
	}

	(void) fflush(out);
}