perf tests: Fix data symbol test with LTO builds
Ian Rogers
reviewerPATCH 1/1 perf tests: Fix data symbol test with LTO builds
With LTO builds, although regular builds could also see this as all the code is in one file, the datasym workload can realize the buf1.reserved data is never accessed. The compiler moves the variable to bss and only keeps the data1 and data2 parts as separate variables. This causes the symbol check to fail in the test. Make the variable volatile to disable the more aggressive optimization. Rename the variable to make which buf1 in perf is being referred to.
Before:
``` $ perf test -vv "data symbol" 126: Test data symbol: --- start --- test child forked, pid 299808 perf does not have symbol 'buf1' perf is missing symbols - skipping test ---- end(-2) ---- 126: Test data symbol : Skip $ nm perf|grep buf1 0000000000a5fa40 b buf1.0 0000000000a5fa48 b buf1.1 ```
After:
``` $ nm perf|grep buf1 0000000000a53a00 d buf1 $ perf test -vv "data symbol"126: Test data symbol: --- start --- test child forked, pid 302166 a53a00-a53a39 l buf1 perf does have symbol 'buf1' Recording workload... Waiting for "perf record has started" message OK Cleaning up files... ---- end(0) ---- 126: Test data symbol : Ok ```
--- tools/perf/tests/shell/test_data_symbol.sh | 17 +++++++++-------- tools/perf/tests/workloads/datasym.c | 11 ++++++----- 2 files changed, 15 insertions(+), 13 deletions(-)
tools/perf/tests/shell/test_data_symbol.sh +9 -8
@@ -18,7 +18,7 @@ skip_if_no_mem_event() { skip_if_no_mem_event || exit 2 -skip_test_missing_symbol buf1 +skip_test_missing_symbol workload_datasym_buf1 TEST_PROGRAM="perf test -w datasym" PERF_DATA=$(mktemp /tmp/__perf_test.perf.data.XXXXX)
@@ -26,18 +26,19 @@ ERR_FILE=$(mktemp /tmp/__perf_test.stderr.XXXXX) check_result() { # The memory report format is as below: - # 99.92% ... [.] buf1+0x38 + # 99.92% ... [.] workload_datasym_buf1+0x38 result=$(perf mem report -i ${PERF_DATA} -s symbol_daddr -q 2>&1 | - awk '/buf1/ { print $4 }') + awk '/workload_datasym_buf1/ { print $4 }') - # Testing is failed if has no any sample for "buf1" + # Testing is failed if has no any sample for "workload_datasym_buf1" [ -z "$result" ] && return 1 while IFS= read -r line; do - # The "data1" and "data2" fields in structure "buf1" have - # offset "0x0" and "0x38", returns failure if detect any - # other offset value. - if [ "$line" != "buf1+0x0" ] && [ "$line" != "buf1+0x38" ]; then + # The "data1" and "data2" fields in structure + # "workload_datasym_buf1" have offset "0x0" and "0x38", returns + # failure if detect any other offset value. + if [ "$line" != "workload_datasym_buf1+0x0" ] && \ + [ "$line" != "workload_datasym_buf1+0x38" ]; then return 1 fi done <<< "$result"
tools/perf/tests/workloads/datasym.c +6 -5
@@ -7,7 +7,8 @@ typedef struct _buf {char data2; } buf __attribute__((aligned(64))); -static buf buf1 = { +/* volatile to try to avoid the compiler seeing reserved as unused. */ +static volatile buf workload_datasym_buf1 = {/* to have this in the data section */ .reserved[0 ] =1 , };
@@ -15,8 +16,8 @@ static buf buf1 = {static int datasym (int argc__maybe_unused ,const char **argv__maybe_unused ) {for (;;) { - buf1.data1++; -if (buf1.data1 ==123 ) { + workload_datasym_buf1.data1++; +if (workload_datasym_buf1.data1 ==123 ) {/* * Add some 'noise' in the loop to work around errata * 1694299 on Arm N1.
@@ -30,9 +31,9 @@ static int datasym(int argc __maybe_unused, const char **argv __maybe_unused)* longer a continuous repeating pattern that interacts * badly with the bias. */ - buf1.data1++; + workload_datasym_buf1.data1++; } - buf1.data2 += buf1.data1; + workload_datasym_buf1.data2 += workload_datasym_buf1.data1; }return 0 ; }
Signature · 1 line
2.48.1.711.g2feabab25a-goog
Ian Rogers
reviewerQuoted 42 lines · Ian Rogers
With LTO builds, although regular builds could also see this as all the code is in one file, the datasym workload can realize the buf1.reserved data is never accessed. The compiler moves the variable to bss and only keeps the data1 and data2 parts as separate variables. This causes the symbol check to fail in the test. Make the variable volatile to disable the more aggressive optimization. Rename the variable to make which buf1 in perf is being referred to.
Before:
``` $ perf test -vv "data symbol" 126: Test data symbol: --- start --- test child forked, pid 299808 perf does not have symbol 'buf1' perf is missing symbols - skipping test ---- end(-2) ---- 126: Test data symbol : Skip $ nm perf|grep buf1 0000000000a5fa40 b buf1.0 0000000000a5fa48 b buf1.1 ```After:
``` $ nm perf|grep buf1 0000000000a53a00 d buf1 $ perf test -vv "data symbol"126: Test data symbol: --- start --- test child forked, pid 302166 a53a00-a53a39 l buf1 perf does have symbol 'buf1' Recording workload... Waiting for "perf record has started" message OK Cleaning up files... ---- end(0) ---- 126: Test data symbol : Ok ```
Quoted 2 lines
Fixes 3dfc01fe9d12 ("perf test: Add 'datasym' test workload")Signed-off-by Ian Rogers <[email protected]>
Ah, I see we're trying to force -O0 and -fno-inline in the Makefile (git.kernel.org is giving 403s):
https://github.com/torvalds/linux/blob/master/tools/perf/tests/workloads/Build#L11
Which LTO later undoes. I'm not seeing LTO breakages for brstack.c and the shell test "Check branch stack sampling". I think LTO is able to optimize this case as it is a variable/struct being optimized, so the "-O0" and "-fno-inline" mustn't be being made to apply. Not a wholly satisfactory reason to add the volatile, but I'm short on alternatives.
Thanks,
Ian
Quoted 85 lines
--- tools/perf/tests/shell/test_data_symbol.sh | 17 +++++++++-------- tools/perf/tests/workloads/datasym.c | 11 ++++++----- 2 files changed, 15 insertions(+), 13 deletions(-)tools/perf/tests/shell/test_data_symbol.sh +9
-8@@ -18,7 +18,7 @@ skip_if_no_mem_event() { skip_if_no_mem_event || exit 2 -skip_test_missing_symbol buf1 +skip_test_missing_symbol workload_datasym_buf1 TEST_PROGRAM="perf test -w datasym" PERF_DATA=$(mktemp /tmp/__perf_test.perf.data.XXXXX)@@ -26,18 +26,19 @@ ERR_FILE=$(mktemp /tmp/__perf_test.stderr.XXXXX) check_result() { # The memory report format is as below: - # 99.92% ... [.] buf1+0x38 + # 99.92% ... [.] workload_datasym_buf1+0x38 result=$(perf mem report -i ${PERF_DATA} -s symbol_daddr -q 2>&1 | - awk '/buf1/ { print $4 }') + awk '/workload_datasym_buf1/ { print $4 }') - # Testing is failed if has no any sample for "buf1" + # Testing is failed if has no any sample for "workload_datasym_buf1" [ -z "$result" ] && return 1 while IFS= read -r line; do - # The "data1" and "data2" fields in structure "buf1" have - # offset "0x0" and "0x38", returns failure if detect any - # other offset value. - if [ "$line" != "buf1+0x0" ] && [ "$line" != "buf1+0x38" ]; then + # The "data1" and "data2" fields in structure + # "workload_datasym_buf1" have offset "0x0" and "0x38", returns + # failure if detect any other offset value. + if [ "$line" != "workload_datasym_buf1+0x0" ] && \ + [ "$line" != "workload_datasym_buf1+0x38" ]; then return 1 fi done <<< "$result"tools/perf/tests/workloads/datasym.c +6
-6@@ -7,7 +7,8 @@ typedef struct _buf {char data2; } buf __attribute__((aligned(64))); -static buf buf1 = { +/* volatile to try to avoid the compiler seeing reserved as unused. */ +static volatile buf workload_datasym_buf1 = {/* to have this in the data section */ .reserved[0 ] =1 , };@@ -15,8 +16,8 @@ static buf buf1 = {static int datasym (int argc__maybe_unused ,const char **argv__maybe_unused ) {for (;;) { - buf1.data1++; -if (buf1.data1 ==123 ) { + workload_datasym_buf1.data1++; +if (workload_datasym_buf1.data1 ==123 ) {/* * Add some 'noise' in the loop to work around errata * 1694299 on Arm N1. @@ -30,9 +31,9 @@ static int datasym(int argc __maybe_unused, const char **argv __maybe_unused)* longer a continuous repeating pattern that interacts * badly with the bias. */ - buf1.data1++; + workload_datasym_buf1.data1++; } - buf1.data2 += buf1.data1; + workload_datasym_buf1.data2 += workload_datasym_buf1.data1; }return 0 ; } --2.48.1.711.g2feabab25a-goog
James Clark
reviewerQuoted 56 lines · Ian Rogers
Quoted 42 lines · Ian Rogers
With LTO builds, although regular builds could also see this as all the code is in one file, the datasym workload can realize the buf1.reserved data is never accessed. The compiler moves the variable to bss and only keeps the data1 and data2 parts as separate variables. This causes the symbol check to fail in the test. Make the variable volatile to disable the more aggressive optimization. Rename the variable to make which buf1 in perf is being referred to.
Before:
``` $ perf test -vv "data symbol" 126: Test data symbol: --- start --- test child forked, pid 299808 perf does not have symbol 'buf1' perf is missing symbols - skipping test ---- end(-2) ---- 126: Test data symbol : Skip $ nm perf|grep buf1 0000000000a5fa40 b buf1.0 0000000000a5fa48 b buf1.1 ```After:
``` $ nm perf|grep buf1 0000000000a53a00 d buf1 $ perf test -vv "data symbol"126: Test data symbol: --- start --- test child forked, pid 302166 a53a00-a53a39 l buf1 perf does have symbol 'buf1' Recording workload... Waiting for "perf record has started" message OK Cleaning up files... ---- end(0) ---- 126: Test data symbol : Ok ```Quoted 2 lines
Fixes 3dfc01fe9d12 ("perf test: Add 'datasym' test workload")Signed-off-by Ian Rogers <[email protected]>Ah, I see we're trying to force -O0 and -fno-inline in the
Makefile(git.kernel.org is giving 403s):
https://github.com/torvalds/linux/blob/master/tools/perf/tests/workloads/Build#L11
Which LTO later undoes. I'm not seeing LTO breakages forbrstack.cand the shell test "Check branch stack sampling". I think LTO is able to optimize this case as it is a variable/struct being optimized, so the "-O0" and "-fno-inline" mustn't be being made to apply. Not a wholly satisfactory reason to add the volatile, but I'm short on alternatives.
Quoted 5 lines
Thanks,
Ian
Adding -fno-lto to that file works for me (llvm-15):
CFLAGS_datasym.o = -fno-lto -g -O0 -fno-inline ...
In fact it looks like we should add this to everywhere we're already doing -g and -O0. If lto is just another form of optimisation it should be disabled as well.
Namhyung Kim
maintainerHello,
Quoted 65 lines · James Clark
Quoted 56 lines · Ian Rogers
Quoted 42 lines · Ian Rogers
With LTO builds, although regular builds could also see this as all the code is in one file, the datasym workload can realize the buf1.reserved data is never accessed. The compiler moves the variable to bss and only keeps the data1 and data2 parts as separate variables. This causes the symbol check to fail in the test. Make the variable volatile to disable the more aggressive optimization. Rename the variable to make which buf1 in perf is being referred to.
Before:
``` $ perf test -vv "data symbol" 126: Test data symbol: --- start --- test child forked, pid 299808 perf does not have symbol 'buf1' perf is missing symbols - skipping test ---- end(-2) ---- 126: Test data symbol : Skip $ nm perf|grep buf1 0000000000a5fa40 b buf1.0 0000000000a5fa48 b buf1.1 ```After:
``` $ nm perf|grep buf1 0000000000a53a00 d buf1 $ perf test -vv "data symbol"126: Test data symbol: --- start --- test child forked, pid 302166 a53a00-a53a39 l buf1 perf does have symbol 'buf1' Recording workload... Waiting for "perf record has started" message OK Cleaning up files... ---- end(0) ---- 126: Test data symbol : Ok ```Quoted 2 lines
Fixes 3dfc01fe9d12 ("perf test: Add 'datasym' test workload")Signed-off-by Ian Rogers <[email protected]>Ah, I see we're trying to force -O0 and -fno-inline in the
Makefile(git.kernel.org is giving 403s):
https://github.com/torvalds/linux/blob/master/tools/perf/tests/workloads/Build#L11
Which LTO later undoes. I'm not seeing LTO breakages forbrstack.cand the shell test "Check branch stack sampling". I think LTO is able to optimize this case as it is a variable/struct being optimized, so the "-O0" and "-fno-inline" mustn't be being made to apply. Not a wholly satisfactory reason to add the volatile, but I'm short on alternatives.Quoted 5 lines
Thanks,
Ian
Quoted 7 lines
Adding -fno-lto to that file works for me (llvm-15):
CFLAGS_datasym.o = -fno-lto -g -O0 -fno-inline ...
In fact it looks like we should add this to everywhere we're already doing -g and -O0. If lto is just another form of optimisation it should be disabled as well.
Make sense. But it turned out that disabling LTO caused some subtle troubles in my work environment unfortunately. :(
Adding 'volatile' seems to be a simple and reasonable alternative and I agree with the renaming.
Thanks,
Namhyung
Quoted 10 lines · Ian Rogers
With LTO builds, although regular builds could also see this as all the code is in one file, the datasym workload can realize the buf1.reserved data is never accessed. The compiler moves the variable to bss and only keeps the data1 and data2 parts as separate variables. This causes the symbol check to fail in the test. Make the variable volatile to disable the more aggressive optimization. Rename the variable to make which buf1 in perf is being referred to.
[...]
Applied to perf-tools-next, thanks!
Best regards,
Namhyung