Re: Fwd: no sample when profiling ARM Cortex-A9 with Linux kernel 3.3

[Maynard J. <may...@us...>]

On 05/02/2013 08:34 PM, RocChen wrote:
> Yes, I execute a 'opcontrol --deinit' after every iteration of the command sequence.
> 
> Above all, I must thank to the warmhearted people here for helping me to get the expected profiling results with oprofile.
> 
> Currently, the only problem for me seems to be that it sometimes gives the profiling results with thousands or more samples for the user application functions, libaries and the no-vmlinux (I call it 'nice results'), but some other times it only gives several tens of samples totally ('bad results'). 
> Besides, it seems to have no issue with the '--image' parameter for this problem. The profiling results are sometimes 'nice' or somtimes 'bad' or even sometimes reporting no samples no matter I set the parameter '--image=all' or '--image=<application name>' or just leave it out.

*Will*, is it possible the ARM oprofile kernel driver (or perf_events subsystem) had a bug in kernel 3.3 involving PMU hardware reservation and/or resource release?  A couple years ago, when perf_events was first being introduced, I recall seeing a similar symptom, where oprofile would simply stop getting PMU interrupts, and it had to do with some incorrect handshaking between oprofile and perf_events.

-Maynard

> Same opcontrol command sequence (the sample event may be different on different CPUs, for example, I use the 'global_power_events:6000' event on the i386/p4-ht CPU) and applications on my PCs (x86 platform) work very well and always give the nice profiling results.
> 
> On Mon, Apr 29, 2013 at 10:08 PM, Maynard Johnson <may...@us... <mailto:may...@us...>> wrote:
> 
>     On 04/27/2013 03:44 AM, RocChen wrote:
>     > Thank you for your enthusiastic guidance, sir:
>     >
>     > I could get nice profiling results with the command list. (but with another issue I can not understand, described below)
>     >
>     > Command List:
>     >> rm -rf /var/lib/oprofile
>     >> rm -rf /root/.oprofile
>     >> opcontrol --init
>     >> opcontrol --no-vmlinux
>     >> opcontrol --setup --event=CPU_CYCLES:10000 --separate=lib,kernel
>     >> opcontrol --start --image=all
>     >> ./array
>     >> ./../mpeg2dec/oprofile_results/mpeg2dec -b ../mpeg2dec/input_base/input_base_4CIF_96bps.mpg -o3  output_base_4CIF_96bps_%03d
>     >> opcontrol --dump
>     >> opreport -l array
>     >> opreport -l ./../mpeg2dec/oprofile_results/mpeg2dec
>     >
>     > Nice Results:
>     > [root]$ opreport -l array
>     > Using /var/lib/oprofile/samples/ for samples directory.
>     > warning: /no-vmlinux could not be found.
>     > CPU: ARM Cortex-A9, speed 1998 MHz (estimated)
>     > Counted CPU_CYCLES events (CPU cycle) with a unit mask of 0x00 (No unit mask) count 10000
>     > samples  %        image name               symbol name
>     > 70547    95.8858  array                    slow_multiply
>     > 1181      1.6052  array                    fast_multiply
>     > 983       1.3361  array                    main
>     > 828       1.1254  no-vmlinux               /no-vmlinux
>     > 29        0.0394  ld-2.13.so <http://ld-2.13.so> <http://ld-2.13.so>               /lib/arm-linux-gnueabi/ld-2.13.so <http://ld-2.13.so> <http://ld-2.13.so>
>     > 6         0.0082  libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>             /lib/arm-linux-gnueabi/libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>
>     > [root]$ opreport -l  mpeg2decode
>     > Using /var/lib/oprofile/samples/ for samples directory.
>     > warning: /no-vmlinux could not be found.
>     > CPU: ARM Cortex-A9, speed 1998 MHz (estimated)
>     > Counted CPU_CYCLES events (CPU cycle) with a unit mask of 0x00 (No unit mask) count 10000
>     > samples  %        image name               symbol name
>     > 23899    16.9694  mpeg2decode              conv420to422
>     > 23648    16.7912  mpeg2decode              store_ppm_tga
>     > 16695    11.8542  mpeg2decode              conv422to444
>     > 16072    11.4119  mpeg2decode              Decode_Picture
>     > 15934    11.3139  mpeg2decode              Fast_IDCT
>     > 15133    10.7451  no-vmlinux               /no-vmlinux
>     > 14614    10.3766  mpeg2decode              putbyte
>     > 9260      6.5750  mpeg2decode              form_component_prediction
>     > 1631      1.1581  mpeg2decode              Flush_Buffer
>     > 825       0.5858  mpeg2decode              Decode_MPEG2_Intra_Block
>     > 481       0.3415  mpeg2decode              form_prediction.constprop.0
>     > 415       0.2947  mpeg2decode              Decode_MPEG2_Non_Intra_Block
>     > 304       0.2159  mpeg2decode              Get_Bits
>     > 200       0.1420  mpeg2decode              Show_Bits
>     > 195       0.1385  mpeg2decode              macroblock_modes
>     > ........
>     >
>     > (*) However, there is an issue that I can hardly understand: after I repeated the commands list above several times (with the same event and sample rate, just exactly the same command sequences), it may give the different results, the total sample numbers are only several tens and no samples for user application functions.
> 
>     From the sequence of commands listed above, I don't see that you ever issue any commands to stop, shutdown or deinit oprofile.  If you simply leave oprofile running and issue another opcontrol --start' command, that should basically be a NOP, so I'm not aware of any oprofile userspace issues with what you're doing.  Have you tried any of the '--stop', '--shutdown', or '--deinit' options between profiling runs?
> 
>     -Maynard
>     >
>     > In my thought, the profiling shoud be at least similar to above nice results.
>     >
>     > Bad results:
>     > [root]$ opreport -l array
>     > Using /var/lib/oprofile/samples/ for samples directory.
>     > warning: /no-vmlinux could not be found.
>     > CPU: ARM Cortex-A9, speed 1998 MHz (estimated)
>     > Counted CPU_CYCLES events (CPU cycle) with a unit mask of 0x00 (No unit mask) count 10000
>     > samples  %        image name               symbol name
>     > 63       92.6471  no-vmlinux               /no-vmlinux
>     > 5         7.3529  libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>             /lib/arm-linux-gnueabi/libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>
>     > (0)-(linaro-chenp)-[Sat Apr 27][15:06:36]-[.=~/Workspace/Zynq/testbench-zynq/hotcode-profiling/mediabench2_video/workspace]
>     > [root]$ opreport -l mpeg2decode
>     > Using /var/lib/oprofile/samples/ for samples directory.
>     > warning: /no-vmlinux could not be found.
>     > CPU: ARM Cortex-A9, speed 1998 MHz (estimated)
>     > Counted CPU_CYCLES events (CPU cycle) with a unit mask of 0x00 (No unit mask) count 10000
>     > samples  %        image name               symbol name
>     > 69       95.8333  no-vmlinux               /no-vmlinux
>     > 3         4.1667  libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>             /lib/arm-linux-gnueabi/libc-2.13.so <http://libc-2.13.so> <http://libc-2.13.so>
>     >
> 
>