689 lines · bash
1#!/bin/sh2# Miscellaneous Intel PT testing3# SPDX-License-Identifier: GPL-2.04 5set -e6 7# Skip if no Intel PT8perf list | grep -q 'intel_pt//' || exit 29 10shelldir=$(dirname "$0")11# shellcheck source=lib/waiting.sh12. "${shelldir}"/lib/waiting.sh13 14skip_cnt=015ok_cnt=016err_cnt=017 18temp_dir=$(mktemp -d /tmp/perf-test-intel-pt-sh.XXXXXXXXXX)19 20tmpfile="${temp_dir}/tmp-perf.data"21perfdatafile="${temp_dir}/test-perf.data"22outfile="${temp_dir}/test-out.txt"23errfile="${temp_dir}/test-err.txt"24workload="${temp_dir}/workload"25awkscript="${temp_dir}/awkscript"26jitdump_workload="${temp_dir}/jitdump_workload"27maxbrstack="${temp_dir}/maxbrstack.py"28 29cleanup()30{31 trap - EXIT TERM INT32 sane=$(echo "${temp_dir}" | cut -b 1-26)33 if [ "${sane}" = "/tmp/perf-test-intel-pt-sh" ] ; then34 echo "--- Cleaning up ---"35 rm -f "${temp_dir}/"*36 rmdir "${temp_dir}"37 fi38}39 40trap_cleanup()41{42 cleanup43 exit 144}45 46trap trap_cleanup EXIT TERM INT47 48# perf record for testing without decoding49perf_record_no_decode()50{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 "$@"54}55 56# perf record for testing should not need BPF events57perf_record_no_bpf()58{59 # Options for no BPF events60 perf record --no-bpf-event "$@"61}62 63have_workload=false64cat << _end_of_file_ | /usr/bin/cc -o "${workload}" -xc - -pthread && have_workload=true65#include <time.h>66#include <pthread.h>67 68void work(void) {69 struct timespec tm = {70 .tv_nsec = 1000000,71 };72 int i;73 74 /* Run for about 30 seconds */75 for (i = 0; i < 30000; i++)76 nanosleep(&tm, NULL);77}78 79void *threadfunc(void *arg) {80 work();81 return NULL;82}83 84int main(void) {85 pthread_t th;86 87 pthread_create(&th, NULL, threadfunc, NULL);88 work();89 pthread_join(th, NULL);90 return 0;91}92_end_of_file_93 94can_cpu_wide()95{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 ; then98 echo "No so skipping"99 return 2100 fi101 echo OK102 return 0103}104 105test_system_wide_side_band()106{107 echo "--- Test system-wide sideband ---"108 109 # Need CPU 0 and CPU 1110 can_cpu_wide 0 || return $?111 can_cpu_wide 1 || return $?112 113 # Record on CPU 0 a task running on CPU 1114 perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u -C 0 -- taskset --cpu-list 1 uname115 116 # Should get MMAP events from CPU 1 because they can be needed to decode117 mmap_cnt=$(perf script -i "${perfdatafile}" --no-itrace --show-mmap-events -C 1 2>/dev/null | grep -c MMAP)118 119 if [ "${mmap_cnt}" -gt 0 ] ; then120 echo OK121 return 0122 fi123 124 echo "Failed to record MMAP events on CPU 1 when tracing CPU 0"125 return 1126}127 128can_kernel()129{130 if [ -z "${can_kernel_trace}" ] ; then131 can_kernel_trace=0132 perf_record_no_decode -o "${tmpfile}" -e dummy:k true >/dev/null 2>&1 && can_kernel_trace=1133 fi134 if [ ${can_kernel_trace} -eq 0 ] ; then135 echo "SKIP: no kernel tracing"136 return 2137 fi138 return 0139}140 141test_per_thread()142{143 k="$1"144 desc="$2"145 146 echo "--- Test per-thread ${desc}recording ---"147 148 if ! $have_workload ; then149 echo "No workload, so skipping"150 return 2151 fi152 153 if [ "${k}" = "k" ] ; then154 can_kernel || return 2155 fi156 157 cat <<- "_end_of_file_" > "${awkscript}"158 BEGIN {159 s = "[ ]*"160 u = s"[0-9]+"s161 d = s"[0-9-]+"s162 x = s"[0-9a-fA-FxX]+"s163 mmapping = "idx"u": mmapping fd"u164 set_output = "idx"u": set output fd"u"->"u165 perf_event_open = "sys_perf_event_open: pid"d"cpu"d"group_fd"d"flags"x"="u166 }167 168 /perf record opening and mmapping events/ {169 if (!done)170 active = 1171 }172 173 /perf record done opening and mmapping events/ {174 active = 0175 done = 1176 }177 178 $0 ~ perf_event_open && active {179 match($0, perf_event_open)180 $0 = substr($0, RSTART, RLENGTH)181 pid = $3182 cpu = $5183 fd = $11184 print "pid " pid " cpu " cpu " fd " fd " : " $0185 fd_array[fd] = fd186 pid_array[fd] = pid187 cpu_array[fd] = cpu188 }189 190 $0 ~ mmapping && active {191 match($0, mmapping)192 $0 = substr($0, RSTART, RLENGTH)193 fd = $5194 print "fd " fd " : " $0195 if (fd in fd_array) {196 mmap_array[fd] = 1197 } else {198 print "Unknown fd " fd199 exit 1200 }201 }202 203 $0 ~ set_output && active {204 match($0, set_output)205 $0 = substr($0, RSTART, RLENGTH)206 fd = $6207 fd_to = $8208 print "fd " fd " fd_to " fd_to " : " $0209 if (fd in fd_array) {210 if (fd_to in fd_array) {211 set_output_array[fd] = fd_to212 } else {213 print "Unknown fd " fd_to214 exit 1215 }216 } else {217 print "Unknown fd " fd218 exit 1219 }220 }221 222 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 " pid230 exit 1231 }232 pids[pid] = 1233 }234 cpu = cpu_array[fd]235 if (cpu != -1) {236 if (cpu in cpus) {237 print "More than 1 mmap for CPU " cpu238 exit 1239 }240 cpus[cpu] = 1241 }242 } else if (!(fd in set_output_array)) {243 print "No mmap for fd " fd244 exit 1245 }246 }247 n = length(pids)248 if (n != thread_cnt) {249 print "Expected " thread_cnt " per-thread mmaps - found " n250 exit 1251 }252 }253 _end_of_file_254 255 $workload &256 w1=$!257 $workload &258 w2=$!259 echo "Workload PIDs are $w1 and $w2"260 wait_for_threads ${w1} 2261 wait_for_threads ${w2} 2262 263 perf_record_no_decode -o "${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 1267 268 kill ${w1}269 wait_for_process_to_exit ${w1} || return 1270 is_running ${ppid} || return 1271 272 kill ${w2}273 wait_for_process_to_exit ${w2} || return 1274 wait_for_process_to_exit ${ppid} || return 1275 276 awk -v thread_cnt=4 -f "${awkscript}" "${errfile}" || return 1277 278 echo OK279 return 0280}281 282test_jitdump()283{284 echo "--- Test tracing self-modifying code that uses jitdump ---"285 286 script_path=$(realpath "$0")287 script_dir=$(dirname "$script_path")288 jitdump_incl_dir="${script_dir}/../../util"289 jitdump_h="${jitdump_incl_dir}/jitdump.h"290 291 if [ ! -e "${jitdump_h}" ] ; then292 echo "SKIP: Include file jitdump.h not found"293 return 2294 fi295 296 if [ -z "${have_jitdump_workload}" ] ; then297 have_jitdump_workload=false298 # Create a workload that uses self-modifying code and generates its own jitdump file299 cat <<- "_end_of_file_" | /usr/bin/cc -o "${jitdump_workload}" -I "${jitdump_incl_dir}" -xc - -pthread && have_jitdump_workload=true300 #define _GNU_SOURCE301 #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>308 309 #include "jitdump.h"310 311 #define CHK_BYTE 0x5a312 313 static inline uint64_t rdtsc(void)314 {315 unsigned int low, high;316 317 asm volatile("rdtsc" : "=a" (low), "=d" (high));318 319 return low | ((uint64_t)high) << 32;320 }321 322 static FILE *open_jitdump(void)323 {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,331 };332 char filename[256];333 FILE *f;334 void *m;335 336 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;348 349 err_close:350 fclose(f);351 err:352 return NULL;353 }354 355 static int write_jitdump(FILE *f, void *addr, const uint8_t *dat, size_t sz, uint64_t *idx)356 {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,367 };368 369 if (fwrite(&rec,sizeof(rec),1,f) != 1 ||370 fwrite(dat, sz, 1, f) != 1)371 return -1;372 return 0;373 }374 375 static void close_jitdump(FILE *f)376 {377 fclose(f);378 }379 380 int main()381 {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;389 390 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;402 }403 _end_of_file_404 fi405 406 if ! $have_jitdump_workload ; then407 echo "SKIP: No jitdump workload"408 return 2409 fi410 411 # Change to temp_dir so jitdump collateral files go there412 cd "${temp_dir}"413 perf_record_no_bpf -o "${tmpfile}" -e intel_pt//u "${jitdump_workload}"414 perf inject -i "${tmpfile}" -o "${perfdatafile}" --jit415 decode_br_cnt=$(perf script -i "${perfdatafile}" --itrace=b | wc -l)416 # Note that overflow and lost errors are suppressed for the error count417 decode_err_cnt=$(perf script -i "${perfdatafile}" --itrace=e-o-l | grep -ci error)418 cd -419 # Should be thousands of branches420 if [ "${decode_br_cnt}" -lt 1000 ] ; then421 echo "Decode failed, only ${decode_br_cnt} branches"422 return 1423 fi424 # Should be no errors425 if [ "${decode_err_cnt}" -ne 0 ] ; then426 echo "Decode failed, ${decode_err_cnt} errors"427 perf script -i "${perfdatafile}" --itrace=e-o-l --show-mmap-events | cat428 return 1429 fi430 431 echo OK432 return 0433}434 435test_packet_filter()436{437 echo "--- Test with MTC and TSC disabled ---"438 # Disable MTC and TSC439 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/mtc=0,tsc=0/u uname440 # Should not get MTC packet441 mtc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "MTC 0x")442 if [ "${mtc_cnt}" -ne 0 ] ; then443 echo "Failed to filter with mtc=0"444 return 1445 fi446 # Should not get TSC package447 tsc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "TSC 0x")448 if [ "${tsc_cnt}" -ne 0 ] ; then449 echo "Failed to filter with tsc=0"450 return 1451 fi452 echo OK453 return 0454}455 456test_disable_branch()457{458 echo "--- Test with branches disabled ---"459 # Disable branch460 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/branch=0/u uname461 # Should not get branch related packets462 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 ] || [ "${fup_cnt}" -ne 0 ] ; then466 echo "Failed to disable branches"467 return 1468 fi469 echo OK470 return 0471}472 473test_time_cyc()474{475 echo "--- Test with/without CYC ---"476 # Check if CYC is supported477 cyc=$(cat /sys/bus/event_source/devices/intel_pt/caps/psb_cyc)478 if [ "${cyc}" != "1" ] ; then479 echo "SKIP: CYC is not supported"480 return 2481 fi482 # Enable CYC483 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/cyc/u uname484 # should get CYC packets485 cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")486 if [ "${cyc_cnt}" = "0" ] ; then487 echo "Failed to get CYC packet"488 return 1489 fi490 # Without CYC491 perf_record_no_decode -o "${perfdatafile}" -e intel_pt//u uname492 # Should not get CYC packets493 cyc_cnt=$(perf script -i "${perfdatafile}" -D 2>/dev/null | grep -c "CYC 0x")494 if [ "${cyc_cnt}" -gt 0 ] ; then495 echo "Still get CYC packet without cyc"496 return 1497 fi498 echo OK499 return 0500}501 502test_sample()503{504 echo "--- Test recording with sample mode ---"505 # Check if recording with sample mode is working506 if ! perf_record_no_decode -o "${perfdatafile}" --aux-sample=8192 -e '{intel_pt//u,branch-misses:u}' uname ; then507 echo "perf record failed with --aux-sample"508 return 1509 fi510 # Check with event with PMU name511 if perf_record_no_decode -o "${perfdatafile}" -e br_misp_retired.all_branches:u uname ; then512 if ! perf_record_no_decode -o "${perfdatafile}" -e '{intel_pt//,br_misp_retired.all_branches/aux-sample-size=8192/}:u' uname ; then513 echo "perf record failed with --aux-sample-size"514 return 1515 fi516 fi517 echo OK518 return 0519}520 521test_kernel_trace()522{523 echo "--- Test with kernel trace ---"524 # Check if recording with kernel trace is working525 can_kernel || return 2526 if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt//k -m1,128 uname ; then527 echo "perf record failed with intel_pt//k"528 return 1529 fi530 echo OK531 return 0532}533 534test_virtual_lbr()535{536 echo "--- Test virtual LBR ---"537 # Check if python script is supported538 libpython=$(perf version --build-options | grep python | grep -cv OFF)539 if [ "${libpython}" != "1" ] ; then540 echo "SKIP: python scripting is not supported"541 return 2542 fi543 544 # Python script to determine the maximum size of branch stacks545 cat << "_end_of_file_" > "${maxbrstack}"546from __future__ import print_function547 548bmax = 0549 550def process_event(param_dict):551 if "brstack" in param_dict:552 brstack = param_dict["brstack"]553 n = len(brstack)554 global bmax555 if n > bmax:556 bmax = n557 558def trace_end():559 print("max brstack", bmax)560_end_of_file_561 562 # Check if virtual lbr is working563 perf_record_no_bpf -o "${perfdatafile}" --aux-sample -e '{intel_pt//,cycles}:u' uname564 times_val=$(perf script -i "${perfdatafile}" --itrace=L -s "${maxbrstack}" 2>/dev/null | grep "max brstack " | cut -d " " -f 3)565 case "${times_val}" in566 [0-9]*) ;;567 *) times_val=0;;568 esac569 if [ "${times_val}" -lt 2 ] ; then570 echo "Failed with virtual lbr"571 return 1572 fi573 echo OK574 return 0575}576 577test_power_event()578{579 echo "--- Test power events ---"580 # Check if power events are supported581 power_event=$(cat /sys/bus/event_source/devices/intel_pt/caps/power_event_trace)582 if [ "${power_event}" != "1" ] ; then583 echo "SKIP: power_event_trace is not supported"584 return 2585 fi586 if ! perf_record_no_decode -o "${perfdatafile}" -a -e intel_pt/pwr_evt/u uname ; then587 echo "perf record failed with pwr_evt"588 return 1589 fi590 echo OK591 return 0592}593 594test_no_tnt()595{596 echo "--- Test with TNT packets disabled ---"597 # Check if TNT disable is supported598 notnt=$(cat /sys/bus/event_source/devices/intel_pt/caps/tnt_disable)599 if [ "${notnt}" != "1" ] ; then600 echo "SKIP: tnt_disable is not supported"601 return 2602 fi603 perf_record_no_decode -o "${perfdatafile}" -e intel_pt/notnt/u uname604 # Should be no TNT packets605 tnt_cnt=$(perf script -i "${perfdatafile}" -D | grep -c TNT)606 if [ "${tnt_cnt}" -ne 0 ] ; then607 echo "TNT packets still there after notnt"608 return 1609 fi610 echo OK611 return 0612}613 614test_event_trace()615{616 echo "--- Test with event_trace ---"617 # Check if event_trace is supported618 event_trace=$(cat /sys/bus/event_source/devices/intel_pt/caps/event_trace)619 if [ "${event_trace}" != 1 ] ; then620 echo "SKIP: event_trace is not supported"621 return 2622 fi623 if ! perf_record_no_decode -o "${perfdatafile}" -e intel_pt/event/u uname ; then624 echo "perf record failed with event trace"625 return 1626 fi627 echo OK628 return 0629}630 631test_pipe()632{633 echo "--- Test with pipe mode ---"634 # Check if it works with pipe635 if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf report -q -i- --itrace=i10000 ; then636 echo "perf record + report failed with pipe mode"637 return 1638 fi639 if ! perf_record_no_bpf -o- -e intel_pt//u uname | perf inject -b > /dev/null ; then640 echo "perf record + inject failed with pipe mode"641 return 1642 fi643 echo OK644 return 0645}646 647count_result()648{649 if [ "$1" -eq 2 ] ; then650 skip_cnt=$((skip_cnt + 1))651 return652 fi653 if [ "$1" -eq 0 ] ; then654 ok_cnt=$((ok_cnt + 1))655 return656 fi657 err_cnt=$((err_cnt + 1))658}659 660ret=0661test_system_wide_side_band || ret=$? ; count_result $ret ; ret=0662test_per_thread "" "" || ret=$? ; count_result $ret ; ret=0663test_per_thread "k" "(incl. kernel) " || ret=$? ; count_result $ret ; ret=0664test_jitdump || ret=$? ; count_result $ret ; ret=0665test_packet_filter || ret=$? ; count_result $ret ; ret=0666test_disable_branch || ret=$? ; count_result $ret ; ret=0667test_time_cyc || ret=$? ; count_result $ret ; ret=0668test_sample || ret=$? ; count_result $ret ; ret=0669test_kernel_trace || ret=$? ; count_result $ret ; ret=0670test_virtual_lbr || ret=$? ; count_result $ret ; ret=0671test_power_event || ret=$? ; count_result $ret ; ret=0672test_no_tnt || ret=$? ; count_result $ret ; ret=0673test_event_trace || ret=$? ; count_result $ret ; ret=0674test_pipe || ret=$? ; count_result $ret ; ret=0675 676cleanup677 678echo "--- Done ---"679 680if [ ${err_cnt} -gt 0 ] ; then681 exit 1682fi683 684if [ ${ok_cnt} -gt 0 ] ; then685 exit 0686fi687 688exit 2689