Perf call graphs on ARM
Hi everyone,
we are having trouble to get perf to output a call graph using frame pointers on a Raspberry Pi 2 Model B Revision 1.1 (ARM Cortex-A7) running Raspbian on its official Kernel versions 4.9 and 5.10.
Let me illustrate the problem we're having using a small test program we will call aaa.cpp:
#include <iostream>
#include <cmath>
__attribute__((noinline)) double G(double aaa) {
return sqrt(aaa);
}
__attribute__((noinline)) double doit(double aaa) {
for(int i=0; i<1000; i++)
aaa=G(aaa);
return aaa;
}
int main() {
double aaa = 12;
for(int i=0; i<10000; i++){
aaa = doit(aaa);
for(int j=0; j<1000; j++)
aaa++;
} std::cout << aaa << std::endl;
}
This program gets compiled like this using GCC Raspbian 8.3.0-6+rpi1:
g++ -O2 -fno-omit-frame-pointer aaa.cpp
Then we run perf on it like this:
perf record -e cycles --call-graph fp -- ./a.out
perf report
In the output of perf there should be a tree like the following:
main
+ doit
+ G
Instead what we are getting is all of those functions attached to a.out (but we _do_ get the runtimes correctly).
It seems the frame pointers are written to the binary, but do not work. We can see the frame pointers in a disassembly output created by compiling with -save-temps like this:
g++ -O2 -fno-omit-frame-pointer -save-temps aaa.cpp
This gives the following output for the doit() function:
_Z4doitd:
.fnstart
.LFB1758:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 1, uses_anonymous_args = 0
vmov.f64 d7, d0
push {r4, r5, fp, lr}
mov r4, #1000
add fp, sp, #12
.L7:
vmov.f64 d0, d7
bl _Z1Gd
subs r4, r4, #1
vmov.f64 d7, d0
bne .L7
pop {r4, r5, fp, pc}
.cantunwind
.fnend
One thing we did try in addition to using frame pointers is to use the DWARF format, but that has some disadvantages like e.g. using roughly 20x the space of the FP format and being much slower to record. Also the recording itself seems unstable and can simply hang the whole Raspberry Pi completely requiring a hard reset. Using Kernel 5.10 the DWARF format also did exhibit the same 'disconnectedness' of the call stack (i.e. all function directly below a.out).
We also tried running Ubuntu 20.04 using its Kernel 5.4, but there both FP and DWARF were 'disconnected' again.
We're at a loss what is going wrong here. Does someone here have an idea what we could try to further debug or even understand the problem?
Mit freundlichen Grüßen / Best regards
Sebastian Fontius
Chassis Systems Control, Image Processing 9 (XC-DA/ESV9)
Robert Bosch GmbH | Postfach 16 61 | 71226 Leonberg | GERMANY | www.bosch.com
Sitz: Stuttgart, Registergericht: Amtsgericht Stuttgart, HRB 14000;
Aufsichtsratsvorsitzender: Franz Fehrenbach; Geschäftsführung: Dr. Volkmar Denner, Prof. Dr. Stefan Asenkerschbaumer, Filiz Albrecht, Dr. Michael Bolle, Dr. Christian Fischer, Dr. Stefan Hartung, Dr. Markus Heyn, Harald Kröger, Rolf Najork, Uwe Raschke
Em Wed, Apr 14, 2021 at 01:22:32PM +0000, Fontius Sebastian (XC-DA/ESV9) escreveu:
Quoted 5 lines
Hi everyone,
we are having trouble to get perf to output a call graph using frame pointers on a Raspberry Pi 2 Model B Revision 1.1 (ARM Cortex-A7) running Raspbian on its official Kernel versions 4.9 and 5.10.
Let me illustrate the problem we're having using a small test program we will call aaa.cpp:
First question is what perf version are you using? The one that comes with the distro you're using? If so can you please try following the instructions at:
https://mirrors.edge.kernel.org/pub/linux/kernel/tools/perf/HOWTO.build.perf
And try with the latest released perf, i.e.:
https://mirrors.edge.kernel.org/pub/linux/kernel/tools/perf/v5.1.0/perf-5.1.0.tar.xz
Now lemme look at your message:
Quoted 29 lines
#include <iostream>
#include <cmath>
__attribute__((noinline)) double G(double aaa) {
return sqrt(aaa);
}
__attribute__((noinline)) double doit(double aaa) {
for(int i=0; i<1000; i++)aaa=G(aaa);return aaa;
}
intmain(){
double aaa = 12;for(int i=0; i<10000; i++){ aaa = doit(aaa); for(int j=0; j<1000; j++) aaa++; }
std::cout<< aaa <<std::endl;
}This program gets compiled like this using GCC Raspbian 8.3.0-6+rpi1:
g++ -O2 -fno-omit-frame-pointer aaa.cpp
Then we run perf on it like this:
perf record -e cycles --call-graph fp -- ./a.out
perf report
Quoted 7 lines
In the output of perf there should be a tree like the following:
main
+ doit
+ GInstead what we are getting is all of those functions attached to a.out (but we _do_ get the runtimes correctly).
Lemme try this on a x86_64, using the --no-children option:
[acme@five c++]$ perf report --no-children --stdio
# To display the perf.data header info, please use --header/--header-only options. #
#
# Total Lost Samples: 0
#
# Samples: 226 of event 'cycles:u'
# Event count (approx.): 230936728
#
# Overhead Command Shared Object Symbol
# ........ ....... ................ ........................
#
86.45% a.out a.out [.] G
|
---G
main
__libc_start_main
0x49564100002adb3d
12.96% a.out a.out [.] main
|
---main
__libc_start_main
0x49564100002adb3d
0.36% a.out [unknown] [k] 0xffffffffa7269345
0.20% a.out ld-2.32.so [.] _dl_relocate_object
0.02% a.out [unknown] [k] 0xffffffffa7555945
0.00% a.out ld-2.32.so [.] __GI___tunables_init
0.00% a.out [unknown] [k] 0xffffffffa7269206
0.00% a.out ld-2.32.so [.] _dl_start
0.00% a.out [unknown] [k] 0xffffffffa72da481
0.00% a.out [unknown] [k] 0xffffffffa7681d07
0.00% a.out [unknown] [k] 0xffffffffa76821b7#
# (Tip: Save output of perf stat using: perf stat record <target workload>) #
[acme@five c++]$
So this is the reverse, so we can:
[acme@five c++]$ perf report --call-graph=fractal,1,caller --no-children --stdio # To display the perf.data header info, please use --header/--header-only options. #
#
# Total Lost Samples: 0
#
# Samples: 226 of event 'cycles:u'
# Event count (approx.): 230936728
#
# Overhead Command Shared Object Symbol
# ........ ....... ................ ........................
#
86.45% a.out a.out [.] G
|
---0x49564100002adb3d
__libc_start_main
main
G
12.96% a.out a.out [.] main
|
---0x49564100002adb3d
__libc_start_main
maindoit doesn't appear... So I tried using -O0, i.e.:
$ g++ -O0 -fno-omit-frame-pointer aaa.cpp
And there it is:
[acme@five c++]$ perf report --call-graph=fractal,1,caller --no-children --stdio | head -25 # To display the perf.data header info, please use --header/--header-only options. #
#
# Total Lost Samples: 0
#
# Samples: 510 of event 'cycles:u'
# Event count (approx.): 536177203
#
# Overhead Command Shared Object Symbol
# ........ ....... ................ ............................
#
30.17% a.out libm-2.32.so [.] __sqrt_finite@GLIBC_2.15
|
---0x49564100002aeb3d
__libc_start_main
main
doit
__sqrt_finite@GLIBC_2.15
21.84% a.out a.out [.] main
|
---0x49564100002aeb3d
__libc_start_main
main[acme@five c++]$
That first 0x49564100002aeb3d remains a mistery, but I ran out of time, have to go drive to pick up my 5yo kid :-)
I'll try to read the remaining parts of your message later, but perhaps the above will help,
Thanks,
- Arnaldo
Quoted 43 lines
It seems the frame pointers are written to the binary, but do not work. We can see the frame pointers in a disassembly output created by compiling with -save-temps like this:
g++ -O2 -fno-omit-frame-pointer -save-temps aaa.cpp
This gives the following output for the
doit()function:
_Z4doitd:
.fnstart
.LFB1758:
@ args = 0, pretend = 0, frame = 0
@frame_needed= 1,uses_anonymous_args= 0
vmov.f64 d7, d0
push {r4, r5, fp, lr}
mov r4, #1000
add fp, sp, #12
.L7:
vmov.f64 d0, d7
bl_Z1Gd
subs r4, r4, #1
vmov.f64 d7, d0
bne .L7
pop {r4, r5, fp, pc}
.cantunwind
.fnendOne thing we did try in addition to using frame pointers is to use the DWARF format, but that has some disadvantages like e.g. using roughly 20x the space of the FP format and being much slower to record. Also the recording itself seems unstable and can simply hang the whole Raspberry Pi completely requiring a hard reset. Using Kernel 5.10 the DWARF format also did exhibit the same 'disconnectedness' of the call stack (i.e. all function directly below a.out).
We also tried running Ubuntu 20.04 using its Kernel 5.4, but there both FP and DWARF were 'disconnected' again.
We're at a loss what is going wrong here. Does someone here have an idea what we could try to further debug or even understand the problem?
Mit freundlichen Grüßen / Best regards
Sebastian Fontius
Chassis Systems Control, Image Processing 9 (XC-DA/ESV9)
Robert Bosch GmbH | Postfach 16 61 | 71226 Leonberg | GERMANY | www.bosch.comSitz: Stuttgart, Registergericht: Amtsgericht Stuttgart, HRB 14000;
Aufsichtsratsvorsitzender: Franz Fehrenbach; Geschäftsführung: Dr. Volkmar Denner, Prof. Dr. Stefan Asenkerschbaumer, Filiz Albrecht, Dr. Michael Bolle, Dr. Christian Fischer, Dr. Stefan Hartung, Dr. Markus Heyn, Harald Kröger, Rolf Najork, Uwe Raschke
Signature · 1 line
- Arnaldo
Em Wed, Apr 14, 2021 at 01:22:32PM +0000, Fontius Sebastian (XC-DA/ESV9) escreveu:
Quoted 9 lines
Quoted 6 lines
we are having trouble to get perf to output a call graph using frame pointers on a Raspberry Pi 2 Model B Revision 1.1 (ARM Cortex-A7) running Raspbian on its official Kernel versions 4.9 and 5.10.
Let me illustrate the problem we're having using a small test program we will call aaa.cpp:
First question is what perf version are you using? The one that comes with the distro you're using?
We tried the following combinations:
* Kernel 4.9 (the current version when we were installing Raspbian) with the perf binary as supplied by the 'linux-perf-4.9' Raspbian package.
* Kernel 5.10 with its accompanying perf version compiled from source on the Raspberry Pi 2 directly.
Quoted 7 lines
If so can you please try following the instructions at:
https://mirrors.edge.kernel.org/pub/linux/kernel/tools/perf/HOWTO.build.perf
And try with the latest released perf, i.e.:
https://mirrors.edge.kernel.org/pub/linux/kernel/tools/perf/v5.1.0/perf-5.1.0.tar.xz
I compiled this version as per your suggestion. The results were exactly the same as with the versions mentioned above.
Quoted 1 line
Lemme try this on a
x86_64, using the --no-children option:
On x86_64 we also do not have any problems. The problems we face are solely on ARM, and there only on the Raspberry Pi 2 Model B revision 1.1 and 1.2. We have tried the same thing on a device from a different vendor which features an ARM A53 processor like the Pi2B r1.2, but not a Broadcom SoC. The OS on the other device is a Yocto-based OS featuring kernel 4.14 built specifically for this device, in case that matters.
Quoted 3 lines
[acme@five c++]$ perf report --no-children --stdio
[...]
As far as we can tell the perf record option using frame pointers already does produce incomplete output, which cannot be used to create call graphs. All the perf report options we tried did not make any difference WRT call graphs.
Quoted 2 lines
That first
0x49564100002aeb3dremains a mistery, but I ran out of time, have to go drive to pick up my 5yo kid :-)
Thanks for your input :).
Regards,
Sebastian
Quoted 18 lines · On 14/04/2021 16:22, Fontius Sebastian (XC-DA/ESV9) wrote:
Hi everyone,
we are having trouble to get perf to output a call graph using frame pointers on a Raspberry Pi 2 Model B Revision 1.1 (ARM Cortex-A7) running Raspbian on its official Kernel versions 4.9 and 5.10.
Let me illustrate the problem we're having using a small test program we will call aaa.cpp:
#include <iostream>
#include <cmath>
__attribute__((noinline)) double G(double aaa) {
return sqrt(aaa);
}
__attribute__((noinline)) double doit(double aaa) {
for(int i=0; i<1000; i++)aaa=G(aaa);return aaa;
}
intmain(){
double aaa = 12;
Quoted 8 lines
for(int i=0; i<10000; i++){ aaa = doit(aaa); for(int j=0; j<1000; j++) aaa++; }
std::cout<< aaa <<std::endl;
}
Hi Sebastian,
Is it possible to simplify the example and hold the process in G(), with some kind of while loop for example (as long as the compiler still produces frame records for each function).
I tried this on 64bit Arm and got a sensible stack and the kernel's frame pointer unwind for 64 bit is pretty much the same as the 32 bit one. Could it be that samples are being taken in the function prologues before the frame pointer is fully setup which is causing some unexpected results?
Here is my stack which looks mostly the same with --call-graph=fp and --call-graph=dwarf:
- 99.73% 99.73% unwind-simple.o unwind-simple.out [.] G
0xfffff67decfd
_start
__libc_start_main
main
doit
G Thanks
James
Quoted 60 lines
This program gets compiled like this using GCC Raspbian 8.3.0-6+rpi1:
g++ -O2 -fno-omit-frame-pointer aaa.cpp
Then we run perf on it like this:
perf record -e cycles --call-graph fp -- ./a.out
perf reportIn the output of perf there should be a tree like the following:
main
+ doit
+ GInstead what we are getting is all of those functions attached to a.out (but we _do_ get the runtimes correctly).
It seems the frame pointers are written to the binary, but do not work. We can see the frame pointers in a disassembly output created by compiling with -save-temps like this:
g++ -O2 -fno-omit-frame-pointer -save-temps aaa.cpp
This gives the following output for the
doit()function:
_Z4doitd:
.fnstart
.LFB1758:
@ args = 0, pretend = 0, frame = 0
@frame_needed= 1,uses_anonymous_args= 0
vmov.f64 d7, d0
push {r4, r5, fp, lr}
mov r4, #1000
add fp, sp, #12
.L7:
vmov.f64 d0, d7
bl_Z1Gd
subs r4, r4, #1
vmov.f64 d7, d0
bne .L7
pop {r4, r5, fp, pc}
.cantunwind
.fnendOne thing we did try in addition to using frame pointers is to use the DWARF format, but that has some disadvantages like e.g. using roughly 20x the space of the FP format and being much slower to record. Also the recording itself seems unstable and can simply hang the whole Raspberry Pi completely requiring a hard reset. Using Kernel 5.10 the DWARF format also did exhibit the same 'disconnectedness' of the call stack (i.e. all function directly below a.out).
We also tried running Ubuntu 20.04 using its Kernel 5.4, but there both FP and DWARF were 'disconnected' again.
We're at a loss what is going wrong here. Does someone here have an idea what we could try to further debug or even understand the problem?
Mit freundlichen Grüßen / Best regards
Sebastian Fontius
Chassis Systems Control, Image Processing 9 (XC-DA/ESV9)
Robert Bosch GmbH | Postfach 16 61 | 71226 Leonberg | GERMANY | www.bosch.comSitz: Stuttgart, Registergericht: Amtsgericht Stuttgart, HRB 14000;
Aufsichtsratsvorsitzender: Franz Fehrenbach; Geschäftsführung: Dr. Volkmar Denner, Prof. Dr. Stefan Asenkerschbaumer, Filiz Albrecht, Dr. Michael Bolle, Dr. Christian Fischer, Dr. Stefan Hartung, Dr. Markus Heyn, Harald Kröger, Rolf Najork, Uwe Raschke