1#!/bin/sh
2# Miscellaneous Intel PT testing
3# SPDX-License-Identifier: GPL-2.0
4
5set -e
6
7# Skip if no Intel PT
8perf list | grep -q 'intel_pt//' || exit 2
9
10shelldir=$(dirname "$0")
11. "${shelldir}"/lib/waiting.sh
12
13skip_cnt=0
14ok_cnt=0
15err_cnt=0
16
17temp_dir=$(mktemp -d /tmp/perf-test-intel-pt-sh.XXXXXXXXXX)
18
19tmpfile="${temp_dir}/tmp-perf.data"
20perfdatafile="${temp_dir}/test-perf.data"
21outfile="${temp_dir}/test-out.txt"
22errfile="${temp_dir}/test-err.txt"
23workload="${temp_dir}/workload"
24awkscript="${temp_dir}/awkscript"
25jitdump_workload="${temp_dir}/jitdump_workload"
26maxbrstack="${temp_dir}/maxbrstack.py"
27
28cleanup()
29{
30	trap - EXIT TERM INT
31	sane=$(echo "${temp_dir}" | cut -b 1-26)
32	if [ "${sane}" = "/tmp/perf-test-intel-pt-sh" ] ; then
33		echo "--- Cleaning up ---"
34		rm -f "${temp_dir}/"*
35		rmdir "${temp_dir}"
36	fi
37}
38
39trap_cleanup()
40{
41	cleanup
42	exit 1
43}
44
45trap trap_cleanup EXIT TERM INT
46
47# perf record for testing without decoding
48perf_record_no_decode()
49{
50	# Options to speed up recording: no post-processing, no build-id cache update,
51	# and no BPF events.
52	perf record -B -N --no-bpf-event "$@"
53}
54
55# perf record for testing should not need BPF events
56perf_record_no_bpf()
57{
58	# Options for no BPF events
59	perf record --no-bpf-event "$@"
60}
61
62have_workload=false
63cat << _end_of_file_ | /usr/bin/cc -o "${workload}" -xc - -pthread && have_workload=true
64#include <time.h>
65#include <pthread.h>
66
67void work(void) {
68	struct timespec tm = {
69		.tv_nsec = 1000000,
70	};
71	int i;
72
73	/* Run for about 30 seconds */
74	for (i = 0; i < 30000; i++)
75		nanosleep(&tm, NULL);
76}
77
78void *threadfunc(void *arg) {
79	work();
80	return NULL;
81}
82
83int main(void) {
84	pthread_t th;
85
86	pthread_create(&th, NULL, threadfunc, NULL);
87	work();
88	pthread_join(th, NULL);
89	return 0;
90}
91_end_of_file_
92
93can_cpu_wide()
94{
95	echo "Checking for CPU-wide recording on CPU $1"
96	if ! perf_record_no_decode -o "${tmpfile}" -e dummy:u -C "$1" true >/dev/null 2>&1 ; then
97		echo "No so skipping"
98		return 2
99	fi
100	echo OK
101	return 0
102}
103
104test_system_wide_side_band()
105{
106	echo "--- Test system-wide sideband ---"
107
108	# Need CPU 0 and CPU 1
109	can_cpu_wide 0 || return $?
110	can_cpu_wide 1 || return $?
111
112	# Record on CPU 0 a task running on CPU 1
113	perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u -C 0 -- taskset --cpu-list 1 uname
114
115	# Should get MMAP events from CPU 1 because they can be needed to decode
116	mmap_cnt=$(perf script -i "${perfdatafile}" --no-itrace --show-mmap-events -C 1 2>/dev/null | grep -c MMAP)
117
118	if [ "${mmap_cnt}" -gt 0 ] ; then
119		echo OK
120		return 0
121	fi
122
123	echo "Failed to record MMAP events on CPU 1 when tracing CPU 0"
124	return 1
125}
126
127can_kernel()
128{
129	if [ -z "${can_kernel_trace}" ] ; then
130		can_kernel_trace=0
131		perf_record_no_decode -o "${tmpfile}" -e dummy:k true >/dev/null 2>&1 && can_kernel_trace=1
132	fi
133	if [ ${can_kernel_trace} -eq 0 ] ; then
134		echo "SKIP: no kernel tracing"
135		return 2
136	fi
137	return 0
138}
139
140test_per_thread()
141{
142	k="$1"
143	desc="$2"
144
145	echo "--- Test per-thread ${desc}recording ---"
146
147	if ! $have_workload ; then
148		echo "No workload, so skipping"
149		return 2
150	fi
151
152	if [ "${k}" = "k" ] ; then
153		can_kernel || return 2
154	fi
155
156	cat <<- "_end_of_file_" > "${awkscript}"
157	BEGIN {
158		s = "[ ]*"
159		u = s"[0-9]+"s
160		d = s"[0-9-]+"s
161		x = s"[0-9a-fA-FxX]+"s
162		mmapping = "idx"u": mmapping fd"u
163		set_output = "idx"u": set output fd"u"->"u
164		perf_event_open = "sys_perf_event_open: pid"d"cpu"d"group_fd"d"flags"x"="u
165	}
166
167	/perf record opening and mmapping events/ {
168		if (!done)
169			active = 1
170	}
171
172	/perf record done opening and mmapping events/ {
173		active = 0
174		done = 1
175	}
176
177	$0 ~ perf_event_open && active {
178		match($0, perf_event_open)
179		$0 = substr($0, RSTART, RLENGTH)
180		pid = $3
181		cpu = $5
182		fd = $11
183		print "pid " pid " cpu " cpu " fd " fd " : " $0
184		fd_array[fd] = fd
185		pid_array[fd] = pid
186		cpu_array[fd] = cpu
187	}
188
189	$0 ~ mmapping && active  {
190		match($0, mmapping)
191		$0 = substr($0, RSTART, RLENGTH)
192		fd = $5
193		print "fd " fd " : " $0
194		if (fd in fd_array) {
195			mmap_array[fd] = 1
196		} else {
197			print "Unknown fd " fd
198			exit 1
199		}
200	}
201
202	$0 ~ set_output && active {
203		match($0, set_output)
204		$0 = substr($0, RSTART, RLENGTH)
205		fd = $6
206		fd_to = $8
207		print "fd " fd " fd_to " fd_to " : " $0
208		if (fd in fd_array) {
209			if (fd_to in fd_array) {
210				set_output_array[fd] = fd_to
211			} else {
212				print "Unknown fd " fd_to
213				exit 1
214			}
215		} else {
216			print "Unknown fd " fd
217			exit 1
218		}
219	}
220
221	END {
222		print "Checking " length(fd_array) " fds"
223		for (fd in fd_array) {
224			if (fd in mmap_array) {
225				pid = pid_array[fd]
226				if (pid != -1) {
227					if (pid in pids) {
228						print "More than 1 mmap for PID " pid
229						exit 1
230					}
231					pids[pid] = 1
232				}
233				cpu = cpu_array[fd]
234				if (cpu != -1) {
235					if (cpu in cpus) {
236						print "More than 1 mmap for CPU " cpu
237						exit 1
238					}
239					cpus[cpu] = 1
240				}
241			} else if (!(fd in set_output_array)) {
242				print "No mmap for fd " fd
243				exit 1
244			}
245		}
246		n = length(pids)
247		if (n != thread_cnt) {
248			print "Expected " thread_cnt " per-thread mmaps - found " n
249			exit 1
250		}
251	}
252	_end_of_file_
253
254	$workload &
255	w1=$!
256	$workload &
257	w2=$!
258	echo "Workload PIDs are $w1 and $w2"
259	wait_for_threads ${w1} 2
260	wait_for_threads ${w2} 2
261
262	perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u"${k}" -vvv --per-thread -p "${w1},${w2}" 2>"${errfile}" >"${outfile}" &
263	ppid=$!
264	echo "perf PID is $ppid"
265	wait_for_perf_to_start ${ppid} "${errfile}" || return 1
266
267	kill ${w1}
268	wait_for_process_to_exit ${w1} || return 1
269	is_running ${ppid} || return 1
270
271	kill ${w2}
272	wait_for_process_to_exit ${w2} || return 1
273	wait_for_process_to_exit ${ppid} || return 1
274
275	awk -v thread_cnt=4 -f "${awkscript}" "${errfile}" || return 1
276
277	echo OK
278	return 0
279}
280
281test_jitdump()
282{
283	echo "--- Test tracing self-modifying code that uses jitdump ---"
284
285	script_path=$(realpath "$0")
286	script_dir=$(dirname "$script_path")
287	jitdump_incl_dir="${script_dir}/../../util"
288	jitdump_h="${jitdump_incl_dir}/jitdump.h"
289
290	if [ ! -e "${jitdump_h}" ] ; then
291		echo "SKIP: Include file jitdump.h not found"
292		return 2
293	fi
294
295	if [ -z "${have_jitdump_workload}" ] ; then
296		have_jitdump_workload=false
297		# Create a workload that uses self-modifying code and generates its own jitdump file
298		cat <<- "_end_of_file_" | /usr/bin/cc -o "${jitdump_workload}" -I "${jitdump_incl_dir}" -xc - -pthread && have_jitdump_workload=true
299		#define _GNU_SOURCE
300		#include <sys/mman.h>
301		#include <sys/types.h>
302		#include <stddef.h>
303		#include <stdio.h>
304		#include <stdint.h>
305		#include <unistd.h>
306		#include <string.h>
307
308		#include "jitdump.h"
309
310		#define CHK_BYTE 0x5a
311
312		static inline uint64_t rdtsc(void)
313		{
314			unsigned int low, high;
315
316			asm volatile("rdtsc" : "=a" (low), "=d" (high));
317
318			return low | ((uint64_t)high) << 32;
319		}
320
321		static FILE *open_jitdump(void)
322		{
323			struct jitheader header = {
324				.magic      = JITHEADER_MAGIC,
325				.version    = JITHEADER_VERSION,
326				.total_size = sizeof(header),
327				.pid        = getpid(),
328				.timestamp  = rdtsc(),
329				.flags      = JITDUMP_FLAGS_ARCH_TIMESTAMP,
330			};
331			char filename[256];
332			FILE *f;
333			void *m;
334
335			snprintf(filename, sizeof(filename), "jit-%d.dump", getpid());
336			f = fopen(filename, "w+");
337			if (!f)
338				goto err;
339			/* Create an MMAP event for the jitdump file. That is how perf tool finds it. */
340			m = mmap(0, 4096, PROT_READ | PROT_EXEC, MAP_PRIVATE, fileno(f), 0);
341			if (m == MAP_FAILED)
342				goto err_close;
343			munmap(m, 4096);
344			if (fwrite(&header,sizeof(header),1,f) != 1)
345				goto err_close;
346			return f;
347
348		err_close:
349			fclose(f);
350		err:
351			return NULL;
352		}
353
354		static int write_jitdump(FILE *f, void *addr, const uint8_t *dat, size_t sz, uint64_t *idx)
355		{
356			struct jr_code_load rec = {
357				.p.id          = JIT_CODE_LOAD,
358				.p.total_size  = sizeof(rec) + sz,
359				.p.timestamp   = rdtsc(),
360				.pid	       = getpid(),
361				.tid	       = gettid(),
362				.vma           = (unsigned long)addr,
363				.code_addr     = (unsigned long)addr,
364				.code_size     = sz,
365				.code_index    = ++*idx,
366			};
367
368			if (fwrite(&rec,sizeof(rec),1,f) != 1 ||
369			fwrite(dat, sz, 1, f) != 1)
370				return -1;
371			return 0;
372		}
373
374		static void close_jitdump(FILE *f)
375		{
376			fclose(f);
377		}
378
379		int main()
380		{
381			/* Get a memory page to store executable code */
382			void *addr = mmap(0, 4096, PROT_WRITE | PROT_EXEC, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
383			/* Code to execute: mov CHK_BYTE, %eax ; ret */
384			uint8_t dat[] = {0xb8, CHK_BYTE, 0x00, 0x00, 0x00, 0xc3};
385			FILE *f = open_jitdump();
386			uint64_t idx = 0;
387			int ret = 1;
388
389			if (!f)
390				return 1;
391			/* Copy executable code to executable memory page */
392			memcpy(addr, dat, sizeof(dat));
393			/* Record it in the jitdump file */
394			if (write_jitdump(f, addr, dat, sizeof(dat), &idx))
395				goto out_close;
396			/* Call it */
397			ret = ((int (*)(void))addr)() - CHK_BYTE;
398		out_close:
399			close_jitdump(f);
400			return ret;
401		}
402		_end_of_file_
403	fi
404
405	if ! $have_jitdump_workload ; then
406		echo "SKIP: No jitdump workload"
407		return 2
408	fi
409
410	# Change to temp_dir so jitdump collateral files go there
411	cd "${temp_dir}"
412	perf_record_no_bpf -o "${tmpfile}" -e intel_pt//u "${jitdump_workload}"
413	perf inject -i "${tmpfile}" -o "${perfdatafile}" --jit
414	decode_br_cnt=$(perf script -i "${perfdatafile}" --itrace=b | wc -l)
415	# Note that overflow and lost errors are suppressed for the error count
416	decode_err_cnt=$(perf script -i "${perfdatafile}" --itrace=e-o-l | grep -ci error)
417	cd -
418	# Should be thousands of branches
419	if [ "${decode_br_cnt}" -lt 1000 ] ; then
420		echo "Decode failed, only ${decode_br_cnt} branches"
421		return 1
422	fi
423	# Should be no errors
424	if [ "${decode_err_cnt}" -ne 0 ] ; then
425		echo "Decode failed, ${decode_err_cnt} errors"
426		perf script -i "${perfdatafile}" --itrace=e-o-l --show-mmap-events | cat
427		return 1
428	fi
429
430	echo OK
431	return 0
432}
433
434test_packet_filter()
435{
436	echo "--- Test with MTC and TSC disabled ---"
437	# Disable MTC and TSC
438	perf_record_no_decode -o "${perfdatafile}" -e intel_pt/mtc=0,tsc=0/u uname
439	# Should not get MTC packet
440	mtc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "MTC 0x")
441	if [ "${mtc_cnt}" -ne 0 ] ; then
442		echo "Failed to filter with mtc=0"
443		return 1
444	fi
445	# Should not get TSC package
446	tsc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TSC 0x")
447	if [ "${tsc_cnt}" -ne 0 ] ; then
448		echo "Failed to filter with tsc=0"
449		return 1
450	fi
451	echo OK
452	return 0
453}
454
455test_disable_branch()
456{
457	echo "--- Test with branches disabled ---"
458	# Disable branch
459	perf_record_no_decode -o "${perfdatafile}" -e intel_pt/branch=0/u uname
460	# Should not get branch related packets
461	tnt_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TNT 0x")
462	tip_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TIP 0x")
463	fup_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "FUP 0x")
464	if [ "${tnt_cnt}" -ne 0 ] || [ "${tip_cnt}" -ne 0 ] || [ "${fup_cnt}" -ne 0 ] ; then
465		echo "Failed to disable branches"
466		return 1
467	fi
468	echo OK
469	return 0
470}
471
472test_time_cyc()
473{
474	echo "--- Test with/without CYC ---"
475	# Check if CYC is supported
476	cyc=$(cat /sys/bus/event_source/devices/intel_pt/caps/psb_cyc)
477	if [ "${cyc}" != "1" ] ; then
478		echo "SKIP: CYC is not supported"
479		return 2
480	fi
481	# Enable CYC
482	perf_record_no_decode -o "${perfdatafile}" -e intel_pt/cyc/u uname
483	# should get CYC packets
484	cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")
485	if [ "${cyc_cnt}" = "0" ] ; then
486		echo "Failed to get CYC packet"
487		return 1
488	fi
489	# Without CYC
490	perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u uname
491	# Should not get CYC packets
492	cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")
493	if [ "${cyc_cnt}" -gt 0 ] ; then
494		echo "Still get CYC packet without cyc"
495		return 1
496	fi
497	echo OK
498	return 0
499}
500
501test_sample()
502{
503	echo "--- Test recording with sample mode ---"
504	# Check if recording with sample mode is working
505	if ! perf_record_no_decode -o "${perfdatafile}" --aux-sample=8192 -e '{intel_pt//u,branch-misses:u}' uname ; then
506		echo "perf record failed with --aux-sample"
507		return 1
508	fi
509	echo OK
510	return 0
511}
512
513test_kernel_trace()
514{
515	echo "--- Test with kernel trace ---"
516	# Check if recording with kernel trace is working
517	can_kernel || return 2
518	if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt//k -m1,128 uname ; then
519		echo "perf record failed with intel_pt//k"
520		return 1
521	fi
522	echo OK
523	return 0
524}
525
526test_virtual_lbr()
527{
528	echo "--- Test virtual LBR ---"
529	# Check if python script is supported
530	libpython=$(perf version --build-options | grep python | grep -cv OFF)
531	if [ "${libpython}" != "1" ] ; then
532		echo "SKIP: python scripting is not supported"
533		return 2
534	fi
535
536	# Python script to determine the maximum size of branch stacks
537	cat << "_end_of_file_" > "${maxbrstack}"
538from __future__ import print_function
539
540bmax = 0
541
542def process_event(param_dict):
543	if "brstack" in param_dict:
544		brstack = param_dict["brstack"]
545		n = len(brstack)
546		global bmax
547		if n > bmax:
548			bmax = n
549
550def trace_end():
551	print("max brstack", bmax)
552_end_of_file_
553
554	# Check if virtual lbr is working
555	perf_record_no_bpf -o "${perfdatafile}" --aux-sample -e '{intel_pt//,cycles}:u' uname
556	times_val=$(perf script -i "${perfdatafile}" --itrace=L -s "${maxbrstack}" 2>/dev/null | grep "max brstack " | cut -d " " -f 3)
557	case "${times_val}" in
558		[0-9]*)	;;
559		*)	times_val=0;;
560	esac
561	if [ "${times_val}" -lt 2 ] ; then
562		echo "Failed with virtual lbr"
563		return 1
564	fi
565	echo OK
566	return 0
567}
568
569test_power_event()
570{
571	echo "--- Test power events ---"
572	# Check if power events are supported
573	power_event=$(cat /sys/bus/event_source/devices/intel_pt/caps/power_event_trace)
574	if [ "${power_event}" != "1" ] ; then
575		echo "SKIP: power_event_trace is not supported"
576		return 2
577	fi
578	if ! perf_record_no_decode -o "${perfdatafile}" -a -e intel_pt/pwr_evt/u uname ; then
579		echo "perf record failed with pwr_evt"
580		return 1
581	fi
582	echo OK
583	return 0
584}
585
586test_no_tnt()
587{
588	echo "--- Test with TNT packets disabled  ---"
589	# Check if TNT disable is supported
590	notnt=$(cat /sys/bus/event_source/devices/intel_pt/caps/tnt_disable)
591	if [ "${notnt}" != "1" ] ; then
592		echo "SKIP: tnt_disable is not supported"
593		return 2
594	fi
595	perf_record_no_decode -o "${perfdatafile}" -e intel_pt/notnt/u uname
596	# Should be no TNT packets
597	tnt_cnt=$(perf script -i "${perfdatafile}" -D | grep -c TNT)
598	if [ "${tnt_cnt}" -ne 0 ] ; then
599		echo "TNT packets still there after notnt"
600		return 1
601	fi
602	echo OK
603	return 0
604}
605
606test_event_trace()
607{
608	echo "--- Test with event_trace ---"
609	# Check if event_trace is supported
610	event_trace=$(cat /sys/bus/event_source/devices/intel_pt/caps/event_trace)
611	if [ "${event_trace}" != 1 ] ; then
612		echo "SKIP: event_trace is not supported"
613		return 2
614	fi
615	if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt/event/u uname ; then
616		echo "perf record failed with event trace"
617		return 1
618	fi
619	echo OK
620	return 0
621}
622
623test_pipe()
624{
625	echo "--- Test with pipe mode ---"
626	# Check if it works with pipe
627	if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf report -q -i- --itrace=i10000 ; then
628		echo "perf record + report failed with pipe mode"
629		return 1
630	fi
631	if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf inject -b > /dev/null ; then
632		echo "perf record + inject failed with pipe mode"
633		return 1
634	fi
635	echo OK
636	return 0
637}
638
639count_result()
640{
641	if [ "$1" -eq 2 ] ; then
642		skip_cnt=$((skip_cnt + 1))
643		return
644	fi
645	if [ "$1" -eq 0 ] ; then
646		ok_cnt=$((ok_cnt + 1))
647		return
648	fi
649	err_cnt=$((err_cnt + 1))
650}
651
652ret=0
653test_system_wide_side_band		|| ret=$? ; count_result $ret ; ret=0
654test_per_thread "" ""			|| ret=$? ; count_result $ret ; ret=0
655test_per_thread "k" "(incl. kernel) "	|| ret=$? ; count_result $ret ; ret=0
656test_jitdump				|| ret=$? ; count_result $ret ; ret=0
657test_packet_filter			|| ret=$? ; count_result $ret ; ret=0
658test_disable_branch			|| ret=$? ; count_result $ret ; ret=0
659test_time_cyc				|| ret=$? ; count_result $ret ; ret=0
660test_sample				|| ret=$? ; count_result $ret ; ret=0
661test_kernel_trace			|| ret=$? ; count_result $ret ; ret=0
662test_virtual_lbr			|| ret=$? ; count_result $ret ; ret=0
663test_power_event			|| ret=$? ; count_result $ret ; ret=0
664test_no_tnt				|| ret=$? ; count_result $ret ; ret=0
665test_event_trace			|| ret=$? ; count_result $ret ; ret=0
666test_pipe				|| ret=$? ; count_result $ret ; ret=0
667
668cleanup
669
670echo "--- Done ---"
671
672if [ ${err_cnt} -gt 0 ] ; then
673	exit 1
674fi
675
676if [ ${ok_cnt} -gt 0 ] ; then
677	exit 0
678fi
679
680exit 2
681