Before going any further please provide a pastebin of the command:
$ "perf report --dump"
On 27 September 2017 at 19:27, 임영재 youngjae24.lim@samsung.com wrote:
Hello.
I backported 4.13 coresight driver to my kernel to use the latest CoreSight drivers.
I checked whether issues that I mentioned is resolved or not.
- ETR issue
You will have to give us more information than simply it "doesn't work". There are 2 problems. First, etr driver don't support 64bit physical address. My model has
64bit physical address.
=> This issue was resolved.
- ETR issue
I saw auxtrace data area is all 0. Maybe I think the problem occured during copying from etr dma buffer
area to mmaped memory
0x29c8b8 [0x30]: PERF_RECORD_AUXTRACE size: 0x122d30 offset: 0x29abe0
ref: 0x68d3ed7647ee5271 idx: 0 tid: 12108 cpu: -1
. ... CoreSight ETM Trace data: size 1191216 bytes
=> I execute like below.
./perf record -e cs_etm/timestamp,cycacc,@1e00a000.etr/u --per-threadtaskset 1 uname
But, I got the same result. I saw auxtrace data area is all 0.
- injection issue
0 0 0x1158 [0x40]: PERF_RECORD_FORK(2:2):(0:0)
0x1198 [0x40]: event: 3 . ... raw event: size 64 bytes ...skipping... ... branch stack: nr:0 ... thread: :-1:-1 ...... dso: <not found>
=> I execute like below.
./perf record -e cs_etm/timestamp,cycacc,@1e005000.etf1/u -C 0--per-thread taskset 1 uname
I got the same result. I got 0 in branch stack nr. I debugged it to give information. Branch stack nr is 0 because decoder->packet_count is 0. The perf data must include OCSD_GEN_TRC_ELEM_INSTR_RANGE to getpackt_count.
But my perf data includes OCSD_GEN_TRC_ELEM_ADDR_NACC. Can you explain about this situation?Thank you.
Best Regard
Youngjae
--------- *Original Message* ---------
*Sender* : Mathieu Poirier mathieu.poirier@linaro.org
*Date* : 2017-09-23 00:13 (GMT+9)
*Title* : Re: Re: FW: RE: Re: FW: Re: Question for openCSD - open soruce Coresight Trace Decode
On 22 September 2017 at 03:10, 임영재 youngjae24.lim@samsung.com wrote:
Hello .Mathieu Poirier
You will have to give us more information than simply it "doesn't work".
=> There are 2 problems. First, etr driver don't support 64bit physical address. My model has 64bit physical address. Below code, there is problem. writel_relaxed(drvdata->paddr, drvdata->base + TMC_DBALO); writel_relaxed(0x0, drvdata->base + TMC_DBAHI);
Suzuki wrote a patchset to fix that. On github, branch perf-opencsd-4.13:
98b62f69c4af coresight: Add support for Coresight SoC 600 components c9b590717a3a coresight tmc: Add support for Coresight SoC 600 TMC 29f38854110e coresight tmc: Support for save-restore in ETR 0863f7b1f278 coresight tmc etr: Setup AXI cache encoding for read transfers f7bab28b7f16 coresight tmc etr: Cleanup AXICTL register handling d81598f029b5 coresight tmc etr: Detect address width at runtime 3c97c4f2566a coresight tmc: Detect support for scatter gather ca684de0ba58 coresight tmc etr: Add capabilitiy information b8b9228d7160 coresight tmc: Handle configuration types properly cc442642fcf5 coresight replicator: Expose replicator management registers 16c65a29684e coresight tmc: Expose DBA and AXICTL 560c49a5aa57 coresight tmc: Add helpers for accessing 64bit registers
I suggest you backport the relevant bit to your kernel 4.9.
Anyway, I fixed it like below.
drivers/base/dma-contiguous.c +static phys_addr_t size_cmdline = 0x1400000; +static phys_addr_t base_cmdline = 0xE8000000; +static phys_addr_t limit_cmdline = 0xE8000000+0x1400000; I checked rrp of etr. it's ok. And I tried to read etr dma buffer area. Etm information was saved normally.
Second, I recorded etm information using perf like below. echo 1 > /sys/devices/platform/1e00a000.etr/1e00a000.etr/enable_sink
The above is not required.
./perf record -e cs_etm/timestamp,cycacc,@1e00a000.etr/u --per-thread taskset 1 uname
I checked the dump like below. ./perf report --stdio --dump
I saw auxtrace data area is all 0.
Things look good - everything seems to be in order. There isn't much else I can do here... Things do work on both reference platform (DB401c and Juno). I always tell people in your situation to get a DB410c and see where the discrepancies are. Clocks could be an issue but I'm just guessing here.
Maybe I think the problem occured during copying from etr dma buffer area to mmaped memory
0x29c8b8 [0x30]: PERF_RECORD_AUXTRACE size: 0x122d30 offset: 0x29abe0 ref: 0x68d3ed7647ee5271 idx: 0 tid: 12108 cpu: -1 . ... CoreSight ETM Trace data: size 1191216 bytes
Also, on gitHub I see close to 30 patches on top of kernel 4.9, where
below you have only 10. Please add all the patches so that your environment is as close to ours.
=> Of course, there are 30 patches on top of kernel 4.9. 20 patches is relative to tools/perf. Because I use perf in opencsd-4.9, I don't need to apply it.
Anyway, I compared my kernel source with opencsd-perf-4.9. It is thesame..
b50067a cs-etm: Update to perf cs-etm decoder for OpenCSD v0.5 tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 2 ++ 2308063 cs-etm: fixing map__load() parameter
tools/perf/util/cs-etm.c | 2 +- ae6ba28 cs-etm: Update to perf cs-etm decode for OpenCSD v0.4 tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 30 +++++++++++++++++------ 29ff0bf perf scripts: Fixing file name extension
tools/perf/scripts/python/cs-trace-disasm.py | 2 +- 9074ff4 perf scripts: Fix binary filename and path tools/perf/scripts/python/cs-trace-disasm.py | 12 +++++++++++- 5f5c594 cs-etm: Update to cs-etm-decoder to handle new packet type from OpenCSD. tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 6 +----- a02ed95 cs-etm: associating output packet with CPU they executed on tools/perf/... 9e34dee cs-etm: removing unecessary structure field tools/perf/util/cs-etm.c | 10 +--------- 677a287 cs-etm: account for each trace buffer in the queue tools/perf/util/cs-etm.c | 17 +++++++++-------- 593a948 cs-etm: avoid casting variable
tools/perf/util/cs-etm.c | 6 ++++-- 26cee10 cs-etm: fixing uninitialised cpumode
tools/perf/util/cs-etm.c | 1 + 328948b perf tools: fixing Makefile problems
tools/perf/Makefile.config | 4 ++-- 73d4294 perf tools: new naming convention for openCSD tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 138 ++++++++++++----------- 451ca19 perf scripts: Add python scripts for CoreSight traces tools/perf/scripts/python/... 8f510fb perf tools: decoding capabilitity for CoreSight traces tools/perf/ .. b5d782e perf symbols: Check before overwriting build_id tools/perf/... dad8531 coresight: updating documentation to reflect integration with perf Documentation/trace/coresight.txt | 138 +++++++++++++++++++++++++++++++++---- ea817df coresight: tmc: implementing TMC-ETR AUX space API drivers/hwtracing/coresight/coresight-tmc-etr.c | 243 +++++++++++++++++++++++ 0a189e4 perf tools: Removing miscellaneous debug output tools/perf/arch/arm/util/cs-etm.c | 2 -- 285fb22 coresight: perf: Add a missing call to etm_free_aux drivers/hwtracing/coresight/coresight-etm-perf.c | 2 +- 3bfa9ae coresight: Add support for ARM Coresight STM-500 drivers/hwtracing/coresight/coresight-stm.c | 5 +++++ 8044597 coresight: tmc: Remove duplicate memset drivers/hwtracing/coresight/coresight-tmc-etr.c | 2 -- 63c0930 coresight: tmc: Get rid of mode parameter for helper routines drivers/hwtracing/coresight/... 034cd3a coresight: tmc: Cleanup operation mode handling drivers/hwtracing/coresight/... 2866655 coresight: reset "enable_sink" flag when need be drivers/hwtracing/coresight/ 4cba396 coresight: etm3x: Adding missing features of Coresight PTM components .../hwtracing/coresight/... 6f6514b coresight: etm3x: indentation fix (extra space removed) .../hwtracing/coresight/coresight-etm3x-sysfs.c | 2 +- d3167d3 coresight: stm: return error code instead of zero in .packet() drivers/hwtracing/coresight/coresight-stm.c | 4 ++--
--------- *Original Message* ---------
*Sender* : Mathieu Poirier mathieu.poirier@linaro.org
*Date* : 2017-09-16 01:09 (GMT+9)
*Title* : Re: FW: RE: Re: FW: Re: Question for openCSD - open soruce Coresight Trace Decode
On 15 September 2017 at 00:22, 임영재 youngjae24.lim@samsung.com wrote:
Hello Mike.
I wonder whether opencsd-perf-4.9 supports injection of branch stack or not.
Thank you.
--------- *Original Message* ---------
*Sender* : 임영재 youngjae24.lim@samsung.com Senior Engineer/System S/W개발2그룹(무선)/삼성전자
*Date* : 2017-09-15 14:32 (GMT+9)
*Title* : RE: Re: FW: Re: Question for openCSD - open soruce Coresight Trace Decode
Hello Mike,
I cannot build a kernel using opencsd-perf-4.13.
My project has kernel version 4.9.
I don't want to change kernel version.
So, I just want to apply patches of coresight drvier in opencsd-perf-4.9.
I applyed pathes of coresight like below.
But, ETR didn't have normal operation.
You will have to give us more information than simply it "doesn't work".
Also, on gitHub I see close to 30 patches on top of kernel 4.9, where below you have only 10. Please add all the patches so that your environment is as close to ours.
0001-coresight-stm-return-error-code-instead-of-zero-in-..patch 0002-coresight-etm3x-indentation-fix-extra-space-removed.patch 0003-coresight-etm3x-Adding-missing-features-of-Coresight.patch 0004-coresight-reset-enable_sink-flag-when-need-be.patch 0005-coresight-tmc-Cleanup-operation-mode-handling.patch 0006-coresight-tmc-Get-rid-of-mode-parameter-for-helper-r.patch 0007-coresight-tmc-Remove-duplicate-memset.patch 0008-coresight-Add-support-for-ARM-Coresight-STM-500.patch 0009-coresight-perf-Add-a-missing-call-to-etm_free_aux.patch 0010-coresight-tmc-implementing-TMC-ETR-AUX-space-API.patch
--------- *Original Message* ---------
*Sender* : Mike Leach mike.leach@linaro.org
*Date* : 2017-09-13 19:37 (GMT+9)
*Title* : Re: FW: Re: Question for openCSD - open soruce Coresight Trace Decode
Hello Youngkae Lim,
V8A cores should not be producing the conditional packets mentioned earlier in this e-mail. Please build a kernel using the opencsd-perf-4.13 or opencsd-perf-master branches which contain the latest CoreSight drivers.
I cannot comment on the inject issue - Sebastian may be of more help here. There are two additional patches submitted to this list which are not integrated into the opencsd-perf branches. Please add these patches before building perf because they correct some issues.
Best Regards
Mike
On 13 September 2017 at 08:38, 임영재 youngjae24.lim@samsung.com wrote:
Dear. Mike Leach.
My cores are v8-a cores.
Perf datas that I gathered don't include unsupported packet.
But, injection is failed below.
0 0 0x1158 [0x40]: PERF_RECORD_FORK(2:2):(0:0)
0x1198 [0x40]: event: 3 . . ... raw event: size 64 bytes ...skipping... ... branch stack: nr:0 ... thread: :-1:-1 ...... dso: <not found>
Best regards.
Youngjae Lim
--------- *Original Message* ---------
*Sender* : Mike Leach mike.leach@linaro.org
*Date* : 2017-09-05 01:14 (GMT+9)
*Title* : Re: Question for openCSD - open soruce Coresight Trace Decode
Hi Sebastian and Youngjae.
The decoder does not support at present the packet types for tracing of Conditional non-branches (i.e. the /* conditional instruction tracing */ block above). Looking at the TRCIDR0 for each core from the perf dump above, the cores in this system do not support conditional instruction tracing, so these packets should never be seen. Architecturally it is not permitted in ETMv4 for A class cores to trace conditional none branch instructions, so these values should never be seen on any A class trace output.
Can you confirm that these cores are in fact v7-A or v8-A cores?
These unsupported packets can be due to perf concatenating wrapped trace data into a single buffer which throws out the decoder. The latest drivers in the opencsd-perf-master branch insert barrier packets to overcome this issue. Can you confirm that the kernel you are using uses the latest driver set?
Conditional instruction trace of none-branches is associated with data trace - this is only permitted on R and M class cores, if implemented at all.
Regarding the change above - I agree this is a good change. I can make it myself or you can send a patch - either works for me.
Regards
Mike
On 1 September 2017 at 15:23, Sebastian Pop sebpop@gmail.com wrote:
The next problem that I see is an unhandled packet in decoder/source/etmv4/trc_pkt_decode_etmv4i.cpp: TrcPktDecodeEtmV4I::decodePacket()
(gdb) p m_curr_packet_in->getType() $267 = ETM4_PKT_I_COND_RES_F1
/*** presently unsupported packets ***/ /* conditional instruction tracing */ case ETM4_PKT_I_COND_FLUSH: case ETM4_PKT_I_COND_I_F1: case ETM4_PKT_I_COND_I_F2: case ETM4_PKT_I_COND_I_F3: case ETM4_PKT_I_COND_RES_F1: case ETM4_PKT_I_COND_RES_F2: case ETM4_PKT_I_COND_RES_F3: case ETM4_PKT_I_COND_RES_F4: // speculation case ETM4_PKT_I_CANCEL_F1: case ETM4_PKT_I_CANCEL_F2: case ETM4_PKT_I_CANCEL_F3: case ETM4_PKT_I_COMMIT: case ETM4_PKT_I_MISPREDICT: case ETM4_PKT_I_DISCARD: // data synchronisation markers case ETM4_PKT_I_NUM_DS_MKR: case ETM4_PKT_I_UNNUM_DS_MKR: /* Q packets */ case ETM4_PKT_I_Q: resp = OCSD_RESP_FATAL_INVALID_DATA;LogError(ocsdError(OCSD_ERR_SEV_ERROR,OCSD_ERR_BAD_DECODE_PKT,"Unsupported packet type.")); break;
-- Mike Leach Principal Engineer, ARM Ltd. Blackburn Design Centre. UK
-- Mike Leach Principal Engineer, ARM Ltd. Blackburn Design Centre. UK
