Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
1#!/bin/bash
2# perf record tests (exclusive)
3# SPDX-License-Identifier: GPL-2.0
4
5set -e
6
7shelldir=$(dirname "$0")
8# shellcheck source=lib/waiting.sh
9. "${shelldir}"/lib/waiting.sh
10
11# shellcheck source=lib/perf_has_symbol.sh
12. "${shelldir}"/lib/perf_has_symbol.sh
13
14testsym="test_loop"
15
16skip_test_missing_symbol ${testsym}
17
18err=0
19perfdata=$(mktemp /tmp/__perf_test.perf.data.XXXXX)
20script_output=$(mktemp /tmp/__perf_test.perf.data.XXXXX.script)
21testprog="perf test -w thloop"
22cpu_pmu_dir="/sys/bus/event_source/devices/cpu*"
23br_cntr_file="/caps/branch_counter_nr"
24br_cntr_output="branch stack counters"
25br_cntr_script_output="br_cntr: A"
26
27default_fd_limit=$(ulimit -Sn)
28# With option --threads=cpu the number of open file descriptors should be
29# equal to sum of: nmb_cpus * nmb_events (2+dummy),
30# nmb_threads for perf.data.n (equal to nmb_cpus) and
31# 2*nmb_cpus of pipes = 4*nmb_cpus (each pipe has 2 ends)
32# All together it needs 8*nmb_cpus file descriptors plus some are also used
33# outside of testing, thus raising the limit to 16*nmb_cpus
34min_fd_limit=$(($(getconf _NPROCESSORS_ONLN) * 16))
35
36cleanup() {
37 rm -f "${perfdata}"
38 rm -f "${perfdata}".old
39 rm -f "${script_output}"
40
41 trap - EXIT TERM INT
42}
43
44trap_cleanup() {
45 echo "Unexpected signal in ${FUNCNAME[1]}"
46 cleanup
47 exit 1
48}
49trap trap_cleanup EXIT TERM INT
50
51test_per_thread() {
52 echo "Basic --per-thread mode test"
53 if ! perf record -o /dev/null --quiet ${testprog} 2> /dev/null
54 then
55 echo "Per-thread record [Skipped event not supported]"
56 return
57 fi
58 if ! perf record --per-thread -o "${perfdata}" ${testprog} 2> /dev/null
59 then
60 echo "Per-thread record [Failed record]"
61 err=1
62 return
63 fi
64 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
65 then
66 echo "Per-thread record [Failed missing output]"
67 err=1
68 return
69 fi
70
71 # run the test program in background (for 30 seconds)
72 ${testprog} 30 &
73 TESTPID=$!
74
75 rm -f "${perfdata}"
76
77 wait_for_threads ${TESTPID} 2
78 perf record -p "${TESTPID}" --per-thread -o "${perfdata}" sleep 1 2> /dev/null
79 kill ${TESTPID}
80
81 if [ ! -e "${perfdata}" ]
82 then
83 echo "Per-thread record [Failed record -p]"
84 err=1
85 return
86 fi
87 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
88 then
89 echo "Per-thread record [Failed -p missing output]"
90 err=1
91 return
92 fi
93
94 echo "Basic --per-thread mode test [Success]"
95}
96
97test_register_capture() {
98 echo "Register capture test"
99 if ! perf list pmu | grep -q 'br_inst_retired.near_call'
100 then
101 echo "Register capture test [Skipped missing event]"
102 return
103 fi
104 if ! perf record --intr-regs=\? 2>&1 | grep -q 'available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10 R11 R12 R13 R14 R15'
105 then
106 echo "Register capture test [Skipped missing registers]"
107 return
108 fi
109 if ! perf record -o - --intr-regs=di,r8,dx,cx -e br_inst_retired.near_call \
110 -c 1000 --per-thread ${testprog} 2> /dev/null \
111 | perf script -F ip,sym,iregs -i - 2> /dev/null \
112 | grep -q "DI:"
113 then
114 echo "Register capture test [Failed missing output]"
115 err=1
116 return
117 fi
118 echo "Register capture test [Success]"
119}
120
121test_system_wide() {
122 echo "Basic --system-wide mode test"
123 if ! perf record -aB --synth=no -o "${perfdata}" ${testprog} 2> /dev/null
124 then
125 echo "System-wide record [Skipped not supported]"
126 return
127 fi
128 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
129 then
130 echo "System-wide record [Failed missing output]"
131 err=1
132 return
133 fi
134 if ! perf record -aB --synth=no -e cpu-clock,cs --threads=cpu \
135 -o "${perfdata}" ${testprog} 2> /dev/null
136 then
137 echo "System-wide record [Failed record --threads option]"
138 err=1
139 return
140 fi
141 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
142 then
143 echo "System-wide record [Failed --threads missing output]"
144 err=1
145 return
146 fi
147 echo "Basic --system-wide mode test [Success]"
148}
149
150test_workload() {
151 echo "Basic target workload test"
152 if ! perf record -o "${perfdata}" ${testprog} 2> /dev/null
153 then
154 echo "Workload record [Failed record]"
155 err=1
156 return
157 fi
158 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
159 then
160 echo "Workload record [Failed missing output]"
161 err=1
162 return
163 fi
164 if ! perf record -e cpu-clock,cs --threads=package \
165 -o "${perfdata}" ${testprog} 2> /dev/null
166 then
167 echo "Workload record [Failed record --threads option]"
168 err=1
169 return
170 fi
171 if ! perf report -i "${perfdata}" -q | grep -q "${testsym}"
172 then
173 echo "Workload record [Failed --threads missing output]"
174 err=1
175 return
176 fi
177 echo "Basic target workload test [Success]"
178}
179
180test_branch_counter() {
181 echo "Branch counter test"
182 # Check if the branch counter feature is supported
183 for dir in $cpu_pmu_dir
184 do
185 if [ ! -e "$dir$br_cntr_file" ]
186 then
187 echo "branch counter feature not supported on all core PMUs ($dir) [Skipped]"
188 return
189 fi
190 done
191 if ! perf record -o "${perfdata}" -e "{branches:p,instructions}" -j any,counter ${testprog} 2> /dev/null
192 then
193 echo "Branch counter record test [Failed record]"
194 err=1
195 return
196 fi
197 if ! perf report -i "${perfdata}" -D -q | grep -q "$br_cntr_output"
198 then
199 echo "Branch counter report test [Failed missing output]"
200 err=1
201 return
202 fi
203 if ! perf script -i "${perfdata}" -F +brstackinsn,+brcntr | grep -q "$br_cntr_script_output"
204 then
205 echo " Branch counter script test [Failed missing output]"
206 err=1
207 return
208 fi
209 echo "Branch counter test [Success]"
210}
211
212test_cgroup() {
213 echo "Cgroup sampling test"
214 if ! perf record -aB --synth=cgroup --all-cgroups -o "${perfdata}" ${testprog} 2> /dev/null
215 then
216 echo "Cgroup sampling [Skipped not supported]"
217 return
218 fi
219 if ! perf report -i "${perfdata}" -D | grep -q "CGROUP"
220 then
221 echo "Cgroup sampling [Failed missing output]"
222 err=1
223 return
224 fi
225 if ! perf script -i "${perfdata}" -F cgroup | grep -q -v "unknown"
226 then
227 echo "Cgroup sampling [Failed cannot resolve cgroup names]"
228 err=1
229 return
230 fi
231 echo "Cgroup sampling test [Success]"
232}
233
234test_leader_sampling() {
235 echo "Basic leader sampling test"
236 if ! perf record -o "${perfdata}" -e "{cycles,cycles}:Su" -- \
237 perf test -w brstack 2> /dev/null
238 then
239 echo "Leader sampling [Failed record]"
240 err=1
241 return
242 fi
243 perf script -i "${perfdata}" | grep brstack > $script_output
244 # Check if the two instruction counts are equal in each record.
245 # However, the throttling code doesn't consider event grouping. During throttling, only the
246 # leader is stopped, causing the slave's counts significantly higher. To temporarily solve this,
247 # let's set the tolerance rate to 80%.
248 # TODO: Revert the code for tolerance once the throttling mechanism is fixed.
249 index=0
250 valid_counts=0
251 invalid_counts=0
252 tolerance_rate=0.8
253 while IFS= read -r line
254 do
255 cycles=$(echo $line | awk '{for(i=1;i<=NF;i++) if($i=="cycles:") print $(i-1)}')
256 if [ $(($index%2)) -ne 0 ] && [ ${cycles}x != ${prev_cycles}x ]
257 then
258 invalid_counts=$(($invalid_counts+1))
259 else
260 valid_counts=$(($valid_counts+1))
261 fi
262 index=$(($index+1))
263 prev_cycles=$cycles
264 done < "${script_output}"
265 total_counts=$(bc <<< "$invalid_counts+$valid_counts")
266 if (( $(bc <<< "$total_counts <= 0") ))
267 then
268 echo "Leader sampling [No sample generated]"
269 err=1
270 return
271 fi
272 isok=$(bc <<< "scale=2; if (($invalid_counts/$total_counts) < (1-$tolerance_rate)) { 0 } else { 1 };")
273 if [ $isok -eq 1 ]
274 then
275 echo "Leader sampling [Failed inconsistent cycles count]"
276 err=1
277 else
278 echo "Basic leader sampling test [Success]"
279 fi
280}
281
282test_topdown_leader_sampling() {
283 echo "Topdown leader sampling test"
284 if ! perf stat -e "{slots,topdown-retiring}" true 2> /dev/null
285 then
286 echo "Topdown leader sampling [Skipped event parsing failed]"
287 return
288 fi
289 if ! perf record -o "${perfdata}" -e "{instructions,slots,topdown-retiring}:S" true 2> /dev/null
290 then
291 echo "Topdown leader sampling [Failed topdown events not reordered correctly]"
292 err=1
293 return
294 fi
295 echo "Topdown leader sampling test [Success]"
296}
297
298test_precise_max() {
299 local -i skipped=0
300
301 echo "precise_max attribute test"
302 # Just to make sure event cycles is supported for sampling
303 if perf record -o "${perfdata}" -e "cycles" true 2> /dev/null
304 then
305 if ! perf record -o "${perfdata}" -e "cycles:P" true 2> /dev/null
306 then
307 echo "precise_max attribute [Failed cycles:P event]"
308 err=1
309 return
310 fi
311 else
312 echo "precise_max attribute [Skipped no cycles:P event]"
313 ((skipped+=1))
314 fi
315 # On s390 event instructions is not supported for perf record
316 if perf record -o "${perfdata}" -e "instructions" true 2> /dev/null
317 then
318 # On AMD, cycles and instructions events are treated differently
319 if ! perf record -o "${perfdata}" -e "instructions:P" true 2> /dev/null
320 then
321 echo "precise_max attribute [Failed instructions:P event]"
322 err=1
323 return
324 fi
325 else
326 echo "precise_max attribute [Skipped no instructions:P event]"
327 ((skipped+=1))
328 fi
329 if [ $skipped -eq 2 ]
330 then
331 echo "precise_max attribute [Skipped no hardware events]"
332 else
333 echo "precise_max attribute test [Success]"
334 fi
335}
336
337# raise the limit of file descriptors to minimum
338if [[ $default_fd_limit -lt $min_fd_limit ]]; then
339 ulimit -Sn $min_fd_limit
340fi
341
342test_per_thread
343test_register_capture
344test_system_wide
345test_workload
346test_branch_counter
347test_cgroup
348test_leader_sampling
349test_topdown_leader_sampling
350test_precise_max
351
352# restore the default value
353ulimit -Sn $default_fd_limit
354
355cleanup
356exit $err