drm/nouveau: consume the return of large GSP message
[drm/drm-misc.git] / tools / perf / tests / shell / test_intel_pt.sh
blobe6f0070975f664cb5cf58400b2022b0164d7b2c4
1 #!/bin/sh
2 # Miscellaneous Intel PT testing (exclusive)
3 # SPDX-License-Identifier: GPL-2.0
5 set -e
7 # Skip if no Intel PT
8 perf list pmu | grep -q 'intel_pt//' || exit 2
10 shelldir=$(dirname "$0")
11 # shellcheck source=lib/waiting.sh
12 . "${shelldir}"/lib/waiting.sh
14 skip_cnt=0
15 ok_cnt=0
16 err_cnt=0
18 temp_dir=$(mktemp -d /tmp/perf-test-intel-pt-sh.XXXXXXXXXX)
20 tmpfile="${temp_dir}/tmp-perf.data"
21 perfdatafile="${temp_dir}/test-perf.data"
22 outfile="${temp_dir}/test-out.txt"
23 errfile="${temp_dir}/test-err.txt"
24 workload="${temp_dir}/workload"
25 awkscript="${temp_dir}/awkscript"
26 jitdump_workload="${temp_dir}/jitdump_workload"
27 maxbrstack="${temp_dir}/maxbrstack.py"
29 cleanup()
31 trap - EXIT TERM INT
32 sane=$(echo "${temp_dir}" | cut -b 1-26)
33 if [ "${sane}" = "/tmp/perf-test-intel-pt-sh" ] ; then
34 echo "--- Cleaning up ---"
35 rm -f "${temp_dir}/"*
36 rmdir "${temp_dir}"
40 trap_cleanup()
42 cleanup
43 exit 1
46 trap trap_cleanup EXIT TERM INT
48 # perf record for testing without decoding
49 perf_record_no_decode()
51 # Options to speed up recording: no post-processing, no build-id cache update,
52 # and no BPF events.
53 perf record -B -N --no-bpf-event "$@"
56 # perf record for testing should not need BPF events
57 perf_record_no_bpf()
59 # Options for no BPF events
60 perf record --no-bpf-event "$@"
63 have_workload=false
64 cat << _end_of_file_ | /usr/bin/cc -o "${workload}" -xc - -pthread && have_workload=true
65 #include <time.h>
66 #include <pthread.h>
68 void work(void) {
69 struct timespec tm = {
70 .tv_nsec = 1000000,
72 int i;
74 /* Run for about 30 seconds */
75 for (i = 0; i < 30000; i++)
76 nanosleep(&tm, NULL);
79 void *threadfunc(void *arg) {
80 work();
81 return NULL;
84 int main(void) {
85 pthread_t th;
87 pthread_create(&th, NULL, threadfunc, NULL);
88 work();
89 pthread_join(th, NULL);
90 return 0;
92 _end_of_file_
94 can_cpu_wide()
96 echo "Checking for CPU-wide recording on CPU $1"
97 if ! perf_record_no_decode -o "${tmpfile}" -e dummy:u -C "$1" true >/dev/null 2>&1 ; then
98 echo "No so skipping"
99 return 2
101 echo OK
102 return 0
105 test_system_wide_side_band()
107 echo "--- Test system-wide sideband ---"
109 # Need CPU 0 and CPU 1
110 can_cpu_wide 0 || return $?
111 can_cpu_wide 1 || return $?
113 # Record on CPU 0 a task running on CPU 1
114 perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u -C 0 -- taskset --cpu-list 1 uname
116 # Should get MMAP events from CPU 1 because they can be needed to decode
117 mmap_cnt=$(perf script -i "${perfdatafile}" --no-itrace --show-mmap-events -C 1 2>/dev/null | grep -c MMAP)
119 if [ "${mmap_cnt}" -gt 0 ] ; then
120 echo OK
121 return 0
124 echo "Failed to record MMAP events on CPU 1 when tracing CPU 0"
125 return 1
128 can_kernel()
130 if [ -z "${can_kernel_trace}" ] ; then
131 can_kernel_trace=0
132 perf_record_no_decode -o "${tmpfile}" -e dummy:k true >/dev/null 2>&1 && can_kernel_trace=1
134 if [ ${can_kernel_trace} -eq 0 ] ; then
135 echo "SKIP: no kernel tracing"
136 return 2
138 return 0
141 test_per_thread()
143 k="$1"
144 desc="$2"
146 echo "--- Test per-thread ${desc}recording ---"
148 if ! $have_workload ; then
149 echo "No workload, so skipping"
150 return 2
153 if [ "${k}" = "k" ] ; then
154 can_kernel || return 2
157 cat <<- "_end_of_file_" > "${awkscript}"
158 BEGIN {
159 s = "[ ]*"
160 u = s"[0-9]+"s
161 d = s"[0-9-]+"s
162 x = s"[0-9a-fA-FxX]+"s
163 mmapping = "idx"u": mmapping fd"u
164 set_output = "idx"u": set output fd"u"->"u
165 perf_event_open = "sys_perf_event_open: pid"d"cpu"d"group_fd"d"flags"x"="u
168 /perf record opening and mmapping events/ {
169 if (!done)
170 active = 1
173 /perf record done opening and mmapping events/ {
174 active = 0
175 done = 1
178 $0 ~ perf_event_open && active {
179 match($0, perf_event_open)
180 $0 = substr($0, RSTART, RLENGTH)
181 pid = $3
182 cpu = $5
183 fd = $11
184 print "pid " pid " cpu " cpu " fd " fd " : " $0
185 fd_array[fd] = fd
186 pid_array[fd] = pid
187 cpu_array[fd] = cpu
190 $0 ~ mmapping && active {
191 match($0, mmapping)
192 $0 = substr($0, RSTART, RLENGTH)
193 fd = $5
194 print "fd " fd " : " $0
195 if (fd in fd_array) {
196 mmap_array[fd] = 1
197 } else {
198 print "Unknown fd " fd
199 exit 1
203 $0 ~ set_output && active {
204 match($0, set_output)
205 $0 = substr($0, RSTART, RLENGTH)
206 fd = $6
207 fd_to = $8
208 print "fd " fd " fd_to " fd_to " : " $0
209 if (fd in fd_array) {
210 if (fd_to in fd_array) {
211 set_output_array[fd] = fd_to
212 } else {
213 print "Unknown fd " fd_to
214 exit 1
216 } else {
217 print "Unknown fd " fd
218 exit 1
222 END {
223 print "Checking " length(fd_array) " fds"
224 for (fd in fd_array) {
225 if (fd in mmap_array) {
226 pid = pid_array[fd]
227 if (pid != -1) {
228 if (pid in pids) {
229 print "More than 1 mmap for PID " pid
230 exit 1
232 pids[pid] = 1
234 cpu = cpu_array[fd]
235 if (cpu != -1) {
236 if (cpu in cpus) {
237 print "More than 1 mmap for CPU " cpu
238 exit 1
240 cpus[cpu] = 1
242 } else if (!(fd in set_output_array)) {
243 print "No mmap for fd " fd
244 exit 1
247 n = length(pids)
248 if (n != thread_cnt) {
249 print "Expected " thread_cnt " per-thread mmaps - found " n
250 exit 1
253 _end_of_file_
255 $workload &
256 w1=$!
257 $workload &
258 w2=$!
259 echo "Workload PIDs are $w1 and $w2"
260 wait_for_threads ${w1} 2
261 wait_for_threads ${w2} 2
263 perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u"${k}" -vvv --per-thread -p "${w1},${w2}" 2>"${errfile}" >"${outfile}" &
264 ppid=$!
265 echo "perf PID is $ppid"
266 wait_for_perf_to_start ${ppid} "${errfile}" || return 1
268 kill ${w1}
269 wait_for_process_to_exit ${w1} || return 1
270 is_running ${ppid} || return 1
272 kill ${w2}
273 wait_for_process_to_exit ${w2} || return 1
274 wait_for_process_to_exit ${ppid} || return 1
276 awk -v thread_cnt=4 -f "${awkscript}" "${errfile}" || return 1
278 echo OK
279 return 0
282 test_jitdump()
284 echo "--- Test tracing self-modifying code that uses jitdump ---"
286 script_path=$(realpath "$0")
287 script_dir=$(dirname "$script_path")
288 jitdump_incl_dir="${script_dir}/../../util"
289 jitdump_h="${jitdump_incl_dir}/jitdump.h"
291 if [ ! -e "${jitdump_h}" ] ; then
292 echo "SKIP: Include file jitdump.h not found"
293 return 2
296 if [ -z "${have_jitdump_workload}" ] ; then
297 have_jitdump_workload=false
298 # Create a workload that uses self-modifying code and generates its own jitdump file
299 cat <<- "_end_of_file_" | /usr/bin/cc -o "${jitdump_workload}" -I "${jitdump_incl_dir}" -xc - -pthread && have_jitdump_workload=true
300 #define _GNU_SOURCE
301 #include <sys/mman.h>
302 #include <sys/types.h>
303 #include <stddef.h>
304 #include <stdio.h>
305 #include <stdint.h>
306 #include <unistd.h>
307 #include <string.h>
309 #include "jitdump.h"
311 #define CHK_BYTE 0x5a
313 static inline uint64_t rdtsc(void)
315 unsigned int low, high;
317 asm volatile("rdtsc" : "=a" (low), "=d" (high));
319 return low | ((uint64_t)high) << 32;
322 static FILE *open_jitdump(void)
324 struct jitheader header = {
325 .magic = JITHEADER_MAGIC,
326 .version = JITHEADER_VERSION,
327 .total_size = sizeof(header),
328 .pid = getpid(),
329 .timestamp = rdtsc(),
330 .flags = JITDUMP_FLAGS_ARCH_TIMESTAMP,
332 char filename[256];
333 FILE *f;
334 void *m;
336 snprintf(filename, sizeof(filename), "jit-%d.dump", getpid());
337 f = fopen(filename, "w+");
338 if (!f)
339 goto err;
340 /* Create an MMAP event for the jitdump file. That is how perf tool finds it. */
341 m = mmap(0, 4096, PROT_READ | PROT_EXEC, MAP_PRIVATE, fileno(f), 0);
342 if (m == MAP_FAILED)
343 goto err_close;
344 munmap(m, 4096);
345 if (fwrite(&header,sizeof(header),1,f) != 1)
346 goto err_close;
347 return f;
349 err_close:
350 fclose(f);
351 err:
352 return NULL;
355 static int write_jitdump(FILE *f, void *addr, const uint8_t *dat, size_t sz, uint64_t *idx)
357 struct jr_code_load rec = {
358 .p.id = JIT_CODE_LOAD,
359 .p.total_size = sizeof(rec) + sz,
360 .p.timestamp = rdtsc(),
361 .pid = getpid(),
362 .tid = gettid(),
363 .vma = (unsigned long)addr,
364 .code_addr = (unsigned long)addr,
365 .code_size = sz,
366 .code_index = ++*idx,
369 if (fwrite(&rec,sizeof(rec),1,f) != 1 ||
370 fwrite(dat, sz, 1, f) != 1)
371 return -1;
372 return 0;
375 static void close_jitdump(FILE *f)
377 fclose(f);
380 int main()
382 /* Get a memory page to store executable code */
383 void *addr = mmap(0, 4096, PROT_WRITE | PROT_EXEC, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
384 /* Code to execute: mov CHK_BYTE, %eax ; ret */
385 uint8_t dat[] = {0xb8, CHK_BYTE, 0x00, 0x00, 0x00, 0xc3};
386 FILE *f = open_jitdump();
387 uint64_t idx = 0;
388 int ret = 1;
390 if (!f)
391 return 1;
392 /* Copy executable code to executable memory page */
393 memcpy(addr, dat, sizeof(dat));
394 /* Record it in the jitdump file */
395 if (write_jitdump(f, addr, dat, sizeof(dat), &idx))
396 goto out_close;
397 /* Call it */
398 ret = ((int (*)(void))addr)() - CHK_BYTE;
399 out_close:
400 close_jitdump(f);
401 return ret;
403 _end_of_file_
406 if ! $have_jitdump_workload ; then
407 echo "SKIP: No jitdump workload"
408 return 2
411 # Change to temp_dir so jitdump collateral files go there
412 cd "${temp_dir}"
413 perf_record_no_bpf -o "${tmpfile}" -e intel_pt//u "${jitdump_workload}"
414 perf inject -i "${tmpfile}" -o "${perfdatafile}" --jit
415 decode_br_cnt=$(perf script -i "${perfdatafile}" --itrace=b | wc -l)
416 # Note that overflow and lost errors are suppressed for the error count
417 decode_err_cnt=$(perf script -i "${perfdatafile}" --itrace=e-o-l | grep -ci error)
418 cd -
419 # Should be thousands of branches
420 if [ "${decode_br_cnt}" -lt 1000 ] ; then
421 echo "Decode failed, only ${decode_br_cnt} branches"
422 return 1
424 # Should be no errors
425 if [ "${decode_err_cnt}" -ne 0 ] ; then
426 echo "Decode failed, ${decode_err_cnt} errors"
427 perf script -i "${perfdatafile}" --itrace=e-o-l --show-mmap-events | cat
428 return 1
431 echo OK
432 return 0
435 test_packet_filter()
437 echo "--- Test with MTC and TSC disabled ---"
438 # Disable MTC and TSC
439 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/mtc=0,tsc=0/u uname
440 # Should not get MTC packet
441 mtc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "MTC 0x")
442 if [ "${mtc_cnt}" -ne 0 ] ; then
443 echo "Failed to filter with mtc=0"
444 return 1
446 # Should not get TSC package
447 tsc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TSC 0x")
448 if [ "${tsc_cnt}" -ne 0 ] ; then
449 echo "Failed to filter with tsc=0"
450 return 1
452 echo OK
453 return 0
456 test_disable_branch()
458 echo "--- Test with branches disabled ---"
459 # Disable branch
460 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/branch=0/u uname
461 # Should not get branch related packets
462 tnt_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TNT 0x")
463 tip_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TIP 0x")
464 fup_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "FUP 0x")
465 if [ "${tnt_cnt}" -ne 0 ] || [ "${tip_cnt}" -ne 0 ] || [ "${fup_cnt}" -ne 0 ] ; then
466 echo "Failed to disable branches"
467 return 1
469 echo OK
470 return 0
473 test_time_cyc()
475 echo "--- Test with/without CYC ---"
476 # Check if CYC is supported
477 cyc=$(cat /sys/bus/event_source/devices/intel_pt/caps/psb_cyc)
478 if [ "${cyc}" != "1" ] ; then
479 echo "SKIP: CYC is not supported"
480 return 2
482 # Enable CYC
483 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/cyc/u uname
484 # should get CYC packets
485 cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")
486 if [ "${cyc_cnt}" = "0" ] ; then
487 echo "Failed to get CYC packet"
488 return 1
490 # Without CYC
491 perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u uname
492 # Should not get CYC packets
493 cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")
494 if [ "${cyc_cnt}" -gt 0 ] ; then
495 echo "Still get CYC packet without cyc"
496 return 1
498 echo OK
499 return 0
502 test_sample()
504 echo "--- Test recording with sample mode ---"
505 # Check if recording with sample mode is working
506 if ! perf_record_no_decode -o "${perfdatafile}" --aux-sample=8192 -e '{intel_pt//u,branch-misses:u}' uname ; then
507 echo "perf record failed with --aux-sample"
508 return 1
510 # Check with event with PMU name
511 if perf_record_no_decode -o "${perfdatafile}" -e br_misp_retired.all_branches:u uname ; then
512 if ! perf_record_no_decode -o "${perfdatafile}" -e '{intel_pt//,br_misp_retired.all_branches/aux-sample-size=8192/}:u' uname ; then
513 echo "perf record failed with --aux-sample-size"
514 return 1
517 echo OK
518 return 0
521 test_kernel_trace()
523 echo "--- Test with kernel trace ---"
524 # Check if recording with kernel trace is working
525 can_kernel || return 2
526 if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt//k -m1,128 uname ; then
527 echo "perf record failed with intel_pt//k"
528 return 1
530 echo OK
531 return 0
534 test_virtual_lbr()
536 echo "--- Test virtual LBR ---"
537 # Check if python script is supported
538 libpython=$(perf version --build-options | grep python | grep -cv OFF)
539 if [ "${libpython}" != "1" ] ; then
540 echo "SKIP: python scripting is not supported"
541 return 2
544 # Python script to determine the maximum size of branch stacks
545 cat << "_end_of_file_" > "${maxbrstack}"
546 from __future__ import print_function
548 bmax = 0
550 def process_event(param_dict):
551 if "brstack" in param_dict:
552 brstack = param_dict["brstack"]
553 n = len(brstack)
554 global bmax
555 if n > bmax:
556 bmax = n
558 def trace_end():
559 print("max brstack", bmax)
560 _end_of_file_
562 # Check if virtual lbr is working
563 perf_record_no_bpf -o "${perfdatafile}" --aux-sample -e '{intel_pt//,cycles}:u' uname
564 times_val=$(perf script -i "${perfdatafile}" --itrace=L -s "${maxbrstack}" 2>/dev/null | grep "max brstack " | cut -d " " -f 3)
565 case "${times_val}" in
566 [0-9]*) ;;
567 *) times_val=0;;
568 esac
569 if [ "${times_val}" -lt 2 ] ; then
570 echo "Failed with virtual lbr"
571 return 1
573 echo OK
574 return 0
577 test_power_event()
579 echo "--- Test power events ---"
580 # Check if power events are supported
581 power_event=$(cat /sys/bus/event_source/devices/intel_pt/caps/power_event_trace)
582 if [ "${power_event}" != "1" ] ; then
583 echo "SKIP: power_event_trace is not supported"
584 return 2
586 if ! perf_record_no_decode -o "${perfdatafile}" -a -e intel_pt/pwr_evt/u uname ; then
587 echo "perf record failed with pwr_evt"
588 return 1
590 echo OK
591 return 0
594 test_no_tnt()
596 echo "--- Test with TNT packets disabled ---"
597 # Check if TNT disable is supported
598 notnt=$(cat /sys/bus/event_source/devices/intel_pt/caps/tnt_disable)
599 if [ "${notnt}" != "1" ] ; then
600 echo "SKIP: tnt_disable is not supported"
601 return 2
603 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/notnt/u uname
604 # Should be no TNT packets
605 tnt_cnt=$(perf script -i "${perfdatafile}" -D | grep -c TNT)
606 if [ "${tnt_cnt}" -ne 0 ] ; then
607 echo "TNT packets still there after notnt"
608 return 1
610 echo OK
611 return 0
614 test_event_trace()
616 echo "--- Test with event_trace ---"
617 # Check if event_trace is supported
618 event_trace=$(cat /sys/bus/event_source/devices/intel_pt/caps/event_trace)
619 if [ "${event_trace}" != 1 ] ; then
620 echo "SKIP: event_trace is not supported"
621 return 2
623 if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt/event/u uname ; then
624 echo "perf record failed with event trace"
625 return 1
627 echo OK
628 return 0
631 test_pipe()
633 echo "--- Test with pipe mode ---"
634 # Check if it works with pipe
635 if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf report -q -i- --itrace=i10000 ; then
636 echo "perf record + report failed with pipe mode"
637 return 1
639 if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf inject -b > /dev/null ; then
640 echo "perf record + inject failed with pipe mode"
641 return 1
643 echo OK
644 return 0
647 count_result()
649 if [ "$1" -eq 2 ] ; then
650 skip_cnt=$((skip_cnt + 1))
651 return
653 if [ "$1" -eq 0 ] ; then
654 ok_cnt=$((ok_cnt + 1))
655 return
657 err_cnt=$((err_cnt + 1))
660 ret=0
661 test_system_wide_side_band || ret=$? ; count_result $ret ; ret=0
662 test_per_thread "" "" || ret=$? ; count_result $ret ; ret=0
663 test_per_thread "k" "(incl. kernel) " || ret=$? ; count_result $ret ; ret=0
664 test_jitdump || ret=$? ; count_result $ret ; ret=0
665 test_packet_filter || ret=$? ; count_result $ret ; ret=0
666 test_disable_branch || ret=$? ; count_result $ret ; ret=0
667 test_time_cyc || ret=$? ; count_result $ret ; ret=0
668 test_sample || ret=$? ; count_result $ret ; ret=0
669 test_kernel_trace || ret=$? ; count_result $ret ; ret=0
670 test_virtual_lbr || ret=$? ; count_result $ret ; ret=0
671 test_power_event || ret=$? ; count_result $ret ; ret=0
672 test_no_tnt || ret=$? ; count_result $ret ; ret=0
673 test_event_trace || ret=$? ; count_result $ret ; ret=0
674 test_pipe || ret=$? ; count_result $ret ; ret=0
676 cleanup
678 echo "--- Done ---"
680 if [ ${err_cnt} -gt 0 ] ; then
681 exit 1
684 if [ ${ok_cnt} -gt 0 ] ; then
685 exit 0
688 exit 2