lkml.org 
[lkml]   [2022]   [Feb]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Date
SubjectRe: [PATCH 12/12] perf test: Add docs for coresight and related tests
Hi Carsten,

I'll put a few more general comments in here as it seems to be the
place to do it!


On Mon, 31 Jan 2022 at 18:05, Carsten Haitzler
<carsten.haitzler@foss.arm.com> wrote:
>
> Hi James
>
> You've found indeed that the tests are for checking if we're even
> exporting any kind of sane data from the system into perf data files.
> The point of this was not to test all of the openocd decoding path and
> to see if, given a perf trace of sufficient completeness, that we can do
> something sensible (autofdo, coverage etc.). These involve much more
> complex decoding on the other end of "we have a perf.data file... let's
> use it". This is all about checking we get perf data files that contain
> somewhat sane content. Does our ""sandwich" have bread at either end and
> some kind of filling in the middle?"
>
> It does appear that we have bugs. These tests seem to show it. We
> regularly fail the TID checks where we provide ZERO blocks of data for
> some threads that run for an extended period of time. This repeats
> itself across various other tests.
>

I am not sure that this is what the tests are showing us.

We know that the current coresight model of multiple cores running
ETMs into a single ETR has limitations.
We cannot trigger an interrrupt when the ETR is full (other than in a
very per platform specific way, using CTIs and even then we would have
to stop every core that was tracing into the buffer - which is why we
have not attempted to solve this problem yet - if it can indeed be
solved)
The amount of trace captured is sensitive to when the user side of
perf is woken up to copy the data from the ETR. This can lead to long
trace runs wrapping the buffer - consequently trace is lost.

Things get a lot better when we have ETE + TRBE, which can interrupt
on full, and uses a 1 buffer per core model.

So these tests need to answer the question - is the trace I have
collected sensible - bearing in mind platform limitations that not all
trace may have been collected.
If the question is yes - the trace is sensible - then the tests should
pass, not fail on occasion as we are seeing now.

There is a difference between "is the trace I am capturing valid" and
"am I capturing all possible trace"


> I chose 3 specific things to probe for in the test data - the sync
> "headers" which we get every now and again in the stream to see we're
> producing some data at all. Trace info will tell us more - like what
> thread/pid and we should get at least enough of these to cover every
> thread we might run if we have long-lived threads running for a while.

There is no relationship between TRACE_INFO and PIDs.
PIDs live in the ADDR_CTXT packets - perhaps a metric on these would
be useful. While you always get ADDR_CTXT after TRACE_INFO, ADDR_CTXT
or CTXT only will appear in other places - as often as necessary to
track ContextID / PID.

> Perf data is a lossy matter here - we can't guarantee we always collect
> everything. This is why the tests also store the numbers they collect
> over multiple runs in csv files so it is possible over the long term to
> track the quality of data statistically.

Certainly there is value to this - however this should be separate
from the checks for validity of trace.
Perhaps with the TID example we could have a pass with X% of TIDs
spotted - where pass is a low number, but we record the % to track
improvements to the system - or indeed regressions.

> The documentation I added in
> the patch set covers this. The F3 atoms were simply the filling in the
> sandwich - is there content other than the "structure" of info/sync. I
> could have picked any ATOM_Fx there. As long as there is something
> there, we're collecting some amount of actual execution content.
>

As mentioned in the other patches - the amount of these can vary on a
per platform basis. Number of atoms rather than number of specific
atom packets is a better metric.



> The idea is that over time when these issues are fixed, we will see more
> of these tests pass more often and the csv data they collect should show
> a jump (on average in the long run) of data collected. Yes - the numbers
> in the tests are a bit "magical" but they can't be derived or calculated
> exactly as they will depend on the instructions a compiler may product
> for a binary as well as other system factors (races). The numbers I
> chose were a result of actually running the tests and collecting data
> and e.g. deciding "if we collect at least 20% of the lowest numbers I
> see, then we're got a pass". It's the same idea in any qualitative test
> - you decide that some value is a passing grade. Of course I could be
> very lenient and just have all the numbers be "1" (at least get one of
> each) and that is about the only derivable value we could agree on is a
> true minimum.
>
> Coverage will suffer the same problem - it will sometimes happen you
> will get no coverage. Eg.g the thread tests above. IF I formulated them
> with a different function per thread and used coverage to figure out
> which threads reported no data - we'd then need to have the coverage be
> 100% to pass and we'd be failing right now.
>
> You are right that perhaps I added too many tests to cover different
> paths (it's unusual that people are unhappy with too many tests and want
> fewer :)), but I see the point of "it takes too long to run" and I can
> certainly trim the tests down.
>
> I've been running on my dragonboard 845 and they don't all pass for me
> either - as above. 95%+ of the test time is spent dumping the perf.data
> file and parsing out the coresight traces. It's not that fast. I did
> quickly look into running perf on perf (seeing if there were some
> blindingly obvious hotspots somewhere to fix) but didn't see anything
> jump out at me. Collecting less data of course speeds it up and that's
> easy to tweak. I could definitely speed things up if I had the checker
> stop parsing the dump output once we had "enough of item X" (I_ASYNC,
> I_TRACE_INFO) which would certainly speed things up. Lowering those
> numbers to bare minimums would make it a lot faster but it would
> probably have issues with the TID checks as I'd like to give the system
> enough time to plausibly collect some data on every thread.
>
> The multiple tests with different threads were there to ensure even on
> many-core system we can saturate all cores to see we are context
> switching between threads and then still capturing data on all cores.
> The point of these tests is to stress-test and not just "bare minimum -
> do we get something" and they certainly seem to find problems in the
> process. I built them so it'd be easy to just drop more tests and tools
> into directories and expand on over time to improve quality of things
> covered. The bubble sort tests were ones that were being used for perf
> quality tests already internally in their own separate git repo, so I
> included them. I can remove them of course as they probably don't
> provide any value over other simpler tests.
>

I have no issue with the bubble sort. This will tend to generate
differing branch patterns as the sort progresses, so is a good example
for trace.

> Testing the pipeline AFTER collecting some sensible data to see if we
> can decode it and e.g. apply it to fdo or coverage is a separate set of
> tests which this was not meant to address (at this stage). Those tests
> certainly do not fit the same pattern as these which follow a simple
> design patter of build N test tools, run those test tools and collect
> data, check the data is semi-sane and seemingly collecting what it should.
>
> So I can:
>
> * Trim the number of tests down
> * Fewer versions of the same one
> * Drop the sorts
> * Optimize time spent checking data (bail early when minimum reached)
>

I would say prefer short trace runs for the majority of tests to check
quality of trace - but at least 1 long run to stress the
infrastructure

Quality of trace needs to look at other issues - are there invalid
packets in the dump. Does the decoder lose sync during the decode
process (you will see NO_SYNC packets after the first SYNC packet in
this case.)

The latest versions of the decoder offer up statistics on number of
bytes decoded compared to number of bytes processed. These can be a
measure of the quality of trace too. I think a patch to perf has been
created to dump these.


> On 1/19/22 16:43, James Clark wrote:
> >
> >
> > On 15/12/2021 16:04, carsten.haitzler@foss.arm.com wrote:
> >> From: Carsten Haitzler <carsten.haitzler@arm.com>
> >>
> >> This adds documentation about the coresight specific tests as part of
> >> perf test
> >>
> >> Signed-off-by: Carsten Haitzler <carsten.haitzler@arm.com>
> >
> > I will use this commit to comment on the set as a whole. The other comments about
> > specific issues I have left inline.
> >
> > I ran the tests twice on N1SDP (kernel version perf/core 48ffaedf017ad) and had two
> > different sets of failures, indicating some flakyness, which is one reason to not have
> > tests with magic numbers as Leo has raised. Are you able to give a complete list of
> > where you have run these so far? I know we can't check everywhere, but at least all
> > of Arm's own reference platforms should be passing.
> >
> > First run:
> >
> > 92: Coresight / Thread Loop 25 Threads : FAILED!
> > 94: Coresight / Thread Loop 10 Threads - Check TID : FAILED!
> > 95: Coresight / Thread Loop 2 Threads - Check TID : FAILED!
> > 96: Coresight / Thread Loop 250 Threads - Check TID : FAILED!
> > 97: Coresight / Unroll Loop Thread 1 : FAILED!
> > 99: Coresight / Unroll Loop Thread 2 : FAILED!
> >
> > Second run:
> >
> > 92: Coresight / Thread Loop 25 Threads : FAILED!
> > 94: Coresight / Thread Loop 10 Threads - Check TID : FAILED!
> > 95: Coresight / Thread Loop 2 Threads - Check TID : FAILED!
> > 96: Coresight / Thread Loop 250 Threads - Check TID : FAILED!
> > 97: Coresight / Unroll Loop Thread 1 : FAILED!
> >
> > I have pasted the verbose output of two of these at the very end.
> >
> > The major problem I have with the tests is the time that they take to run. Now
> > a full test run takes 31m+ on N1SDP, where previously it took 1m 9s. There are
> > quite a few reasons why this isn't a good idea. It will take extra cycles on
> > anyone's CI and that means it will now take longer for anyone to make any changes
> > if they run the tests, slowing down all Perf work from now onward. Even if you
> > want to run just the Coresight tests, it doesn't really help as that is 90% of
> > the time taken, so personally for me I'm not likely to run them while working on
> > Coresight changes. The longer the tests take the more likely it is that they are
> > skipped, reducing the benefits of the tests. Also if anyone has timeouts set on
> > CI these are likely going to have to be adjusted now.
> >
> > Is it possible to get the run time of each test down to 5-10 seconds? This will
> > make it much nicer for us to work with going forwards.
> >
> > I'd also have liked to have had a discussion on the mailing list about what the test
> > methodology should be in general before starting. For example for this patch set it
> > looks like the coverage is quite wide because of the number of tests and time it
> > takes to run, but actually they all use "cs_etm//u -- <process>" as the run command.
> > This only tests per-process userspace mode resulting in very narrow coverage.
> > And there is no mention in a cover letter about how this expands on or interacts
> > with the existing test_arm_coresight.sh test which I don't think can be ignored.
> > For example it iterates over sinks. Could that same iteration not be applied to
> > these tests? They seem to exist in isolation and it's not clear which one we should
> > be adding to going forwards.
> >
> > I'd like to see a much more exploration of the different modes and flags to justify
> > such a high run time:
> >
> > * per-CPU
> > * per-thread
> > * with and without timestamps
> > * strobing
> >
> > If this is supposed to be just an initial implementation covering one case, then
> > by the "test one thing" ethos, there should only be one test case. Each case
> > should be testing one thing with minimal overlapping. But it looks like tests like
> > "Thread loop 10" and "thread loop 250" just test the same thing, and I don't see
> > the value in maintaining that. Also "bubblesort" and "memcpy" don't have comments
> > explaining why running on those two different test programs actually makes a
> > difference. Obviously this comment is partially related to the run time, if it was
> > shorter perhaps it would be less important how many variations we ran.
> >
> > It's also not clear what failure modes the tests are looking for. There are
> > comments like "We should almost always see F3 atoms", but it's not clear how
> > F3 atoms are a signal of a high quality trace. And there are numbers like:
> >
> > perf_dump_aux_verify "$DATA" 66 6 6
> > perf_dump_aux_verify "$DATA" 4188 1630 1630
> >
> > but there are no comments to say how 66 or 4188 were judged to be "high quality" vs
> > any other number. This means that anyone encountering a failure is just going to
> > change the number until it passes, undoing any value that the test added. If the tests
> > are looking for complete loss of data, then > 1 or > 2 would be sufficient. But if
> > someone makes a mistake in the driver to produce less data, then these tests aren't
> > going to catch that anyway, since there is a chance that it falls within the allowed
> > thresholds.
> >
> > In my opinion a much better measure of quality would be to look at the coverage of
> > fully decoded trace, for example if you have 10 functions called from a loop across
> > multiple DSOs, do each of those functions show up in synthesised instruction samples.
> >
> > Which brings me to my next point, the tests only seem to look at the raw trace, rather
> > than the synthesised samples. To me this seems like it's validating what comes out of
> > the CPU more than looking at what the driver or userspace Perf are doing. I think this
> > isn't the right place for these tests to live and would be better suited to run on bare
> > metal. This is because a user isn't going to be interested in how many F3 atoms there are,
> > that's just an implementation detail. What's much more valuable is what samples are
> > synthesised. We're also skipping most of the code in Perf by only looking at the raw
> > trace, which makes it harder to justify putting the tests in the Perf code base.
> >
> > In short I think it would be much better if we had the following:
> >
> > * Much shorter run time
> > * Run each test program only once (for example it doesn't add value to run memcpy with different sizes)
> > * Only one test for each thing/failure mode
> > * Much less overlapping test code (for example F3 atoms only need to be counted for one test, not all of them)
> > * Concrete descriptions of failure modes and what each test is looking for
> > * Small magic numbers which are easy to justify (like "samples < 1, check for complete loss of data")
> > * Run through the full Perf pipeline and look at the synthesised samples rather than raw trace
> > * Some thought about how this fits in with test_arm_coresight.sh
> >
> > Thanks
> > James
> >
> >
> >
> > Test failure outputs:
> > ---
> >
> > 94: Coresight / Thread Loop 10 Threads - Check TID :
> > --- start ---
> > test child forked, pid 5821
> > Couldn't synthesize bpf events.
> > [ perf record: Woken up 1 times to write data ]
> > [ perf record: Captured and wrote 128.106 MB ./perf-thread_loop-check-tid-10th.data ]
> >
> >
> >
> > Warning:
> > AUX data lost 1 times out of 1!
> >
> > Thread IDs not found in perf AUX data

Should this not be a list of which TIDs are missing?

> > test child finished with -1
> > ---- end ----
> > Coresight / Thread Loop 10 Threads - Check TID: FAILED!
> >
> > ========================================================
> >
> >
> > 92: Coresight / Thread Loop 25 Threads :
> > --- start ---
> > test child forked, pid 5871
> > Couldn't synthesize bpf events.
> > [ perf record: Woken up 1 times to write data ]
> > [ perf record: Captured and wrote 28.849 MB ./perf-thread_loop-25th.data ]
> > Sanity check number of ATOM_F3 is too low (92062 < 388121)

This may well be an example of what I referred to in my comments on
patch 1 - there is no guarantee that the same ATOM packets will be
used in the same way across different platforms.

Regards

Mike

> > test child finished with -1
> > ---- end ----
> > Coresight / Thread Loop 25 Threads: FAILED!
> >
> >
> >> ---
> >> MAINTAINERS | 1 +
> >> tools/perf/Documentation/arm-coresight.txt | 140 +++++++++++++++++++++
> >> 2 files changed, 141 insertions(+)
> >> create mode 100644 tools/perf/Documentation/arm-coresight.txt
> >>
> >> diff --git a/MAINTAINERS b/MAINTAINERS
> >> index d46e8469c467..1a93977a0132 100644
> >> --- a/MAINTAINERS
> >> +++ b/MAINTAINERS
> >> @@ -1890,6 +1890,7 @@ F: Documentation/trace/coresight/*
> >> F: drivers/hwtracing/coresight/*
> >> F: include/dt-bindings/arm/coresight-cti-dt.h
> >> F: include/linux/coresight*
> >> +F: tools/perf/Documentation/arm-coresight.txt
> >> F: tools/perf/arch/arm/util/auxtrace.c
> >> F: tools/perf/arch/arm/util/cs-etm.c
> >> F: tools/perf/arch/arm/util/cs-etm.h
> >> diff --git a/tools/perf/Documentation/arm-coresight.txt b/tools/perf/Documentation/arm-coresight.txt
> >> new file mode 100644
> >> index 000000000000..3a9e6c573c58
> >> --- /dev/null
> >> +++ b/tools/perf/Documentation/arm-coresight.txt
> >> @@ -0,0 +1,140 @@
> >> +Arm Coresight Support
> >> +=====================
> >> +
> >> +Coresight is a feature of some Arm based processors that allows for
> >> +debugging. One of the things it can do is trace every instruction
> >> +executed and remotely expose that information in a hardware compressed
> >> +stream. Perf is able to locally access that stream and store it to the
> >> +output perf data files. This stream can then be later decoded to give the
> >> +instructions that were traced for debugging or profiling purposes. You
> >> +can log such data with a perf record command like:
> >> +
> >> + perf record -e cs_etm//u testbinary
> >> +
> >> +This would run some test binary (testbinary) until it exits and record
> >> +a perf.data trace file. That file would have AUX sections if coresight
> >> +is working correctly. You can dump the content of this file as
> >> +readable text with a command like:
> >> +
> >> + perf report --stdio --dump -i perf.data
> >> +
> >> +You should find some sections of this file have AUX data blocks like:
> >> +
> >> + 0x1e78 [0x30]: PERF_RECORD_AUXTRACE size: 0x11dd0 offset: 0 ref: 0x1b614fc1061b0ad1 idx: 0 tid: 531230 cpu: -1
> >> +
> >> + . ... CoreSight ETM Trace data: size 73168 bytes
> >> + Idx:0; ID:10; I_ASYNC : Alignment Synchronisation.
> >> + Idx:12; ID:10; I_TRACE_INFO : Trace Info.; INFO=0x0 { CC.0 }
> >> + Idx:17; ID:10; I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0x0000000000000000;
> >> + Idx:26; ID:10; I_TRACE_ON : Trace On.
> >> + Idx:27; ID:10; I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; Addr=0x0000FFFFB6069140; Ctxt: AArch64,EL0, NS;
> >> + Idx:38; ID:10; I_ATOM_F6 : Atom format 6.; EEEEEEEEEEEEEEEEEEEEEEEE
> >> + Idx:39; ID:10; I_ATOM_F6 : Atom format 6.; EEEEEEEEEEEEEEEEEEEEEEEE
> >> + Idx:40; ID:10; I_ATOM_F6 : Atom format 6.; EEEEEEEEEEEEEEEEEEEEEEEE
> >> + Idx:41; ID:10; I_ATOM_F6 : Atom format 6.; EEEEEEEEEEEN
> >> + ...
> >> +
> >> +If you see these above, then your system is tracing coresight data
> >> +correctly.
> >> +
> >> +To compile perf with coresight support in the perf directory do
> >> +
> >> + make CORESIGHT=1
> >> +
> >> +This will compile the perf tool with coresight support as well as
> >> +build some small test binaries for perf test. This requires you also
> >> +be compiling for 64bit Arm (ARM64/aarch64). The tools run as part of
> >> +perf coresight tracing are in tests/shell/tools/coresight.
> >> +
> >> +You will also want coresight support enabled in your kernel config.
> >> +Ensure it is enabled with:
> >> +
> >> + CONFIG_CORESIGHT=y
> >> +
> >> +There are various other coresight options you probably also want
> >> +enabled like:
> >> +
> >> + CONFIG_CORESIGHT_LINKS_AND_SINKS=y
> >> + CONFIG_CORESIGHT_LINK_AND_SINK_TMC=y
> >> + CONFIG_CORESIGHT_CATU=y
> >> + CONFIG_CORESIGHT_SINK_TPIU=y
> >> + CONFIG_CORESIGHT_SINK_ETBV10=y
> >> + CONFIG_CORESIGHT_SOURCE_ETM4X=y
> >> + CONFIG_CORESIGHT_STM=y
> >> + CONFIG_CORESIGHT_CPU_DEBUG=y
> >> + CONFIG_CORESIGHT_CTI=y
> >> + CONFIG_CORESIGHT_CTI_INTEGRATION_REGS=y
> >> +
> >> +Please refer to the kernel configuration help for more information.
> >> +
> >> +Perf test - Verify kernel and userspace perf coresight work
> >> +===========================================================
> >> +
> >> +When you run perf test, it will do a lot of self tests. Some of those
> >> +tests will cover Coresight (only if enabled and on ARM64). You
> >> +generally would run perf test from the tools/perf directory in the
> >> +kernel tree. Some tests will check some internal perf support like:
> >> +
> >> + Check Arm CoreSight trace data recording and synthesized samples
> >> +
> >> +Some others will actually use perf record and some test binaries that
> >> +are in tests/shell/tools/coresight and will collect traces to ensure a
> >> +minimum level of functionality is met. The scripts that launch these
> >> +tests are in tests/shell. These will all look like:
> >> +
> >> + Coresight / Memcpy 1M 25 Threads
> >> + Coresight / Unroll Loop Thread 2
> >> + ...
> >> +
> >> +These perf record tests will not run if the tool binaries do not exist
> >> +in tests/shell/tools/coresight/*/ and will be skipped. If you do not
> >> +have coresight support in hardware then either do not build perf with
> >> +coresight support or remove these binaries in order to not have these
> >> +tests fail and have them skip instead.
> >> +
> >> +These tests will log historical results in the current working
> >> +directory (e.g. tools/perf) and will be named stats-*.csv like:
> >> +
> >> + stats-asm_pure_loop-out.csv
> >> + stats-bubble_sort-random.csv
> >> + ...
> >> +
> >> +These statistic files log some aspects of the AUX data sections in
> >> +the perf data output counting some numbers of certain encodings (a
> >> +good way to know that it's working in a very simple way). One problem
> >> +with coresight is that given a large enough amount of data needing to
> >> +be logged, some of it can be lost due to the processor not waking up
> >> +in time to read out all the data from buffers etc.. You will notice
> >> +that the amount of data collected can vary a lot per run of perf test.
> >> +If you wish to see how this changes over time, simply run perf test
> >> +multiple times and all these csv files will have more and more data
> >> +appended to it that you can later examine, graph and otherwise use to
> >> +figure out if things have become worse or better.
> >> +
> >> +Be aware that amny of these tests take quite a while to run, specifically
> >> +in processing the perf data file and dumping contents to then examine what
> >> +is inside.
> >> +
> >> +You can change where these csv logs are stored by setting the
> >> +PERF_TEST_CORESIGHT_STATDIR environment variable before running perf
> >> +test like:
> >> +
> >> + export PERF_TEST_CORESIGHT_STATDIR=/var/tmp
> >> + perf test
> >> +
> >> +They will also store resulting perf output data in the current
> >> +directory for later inspection like:
> >> +
> >> + perf-memcpy-1m.data
> >> + perf-thread_loop-2th.data
> >> + ...
> >> +
> >> +You can alter where the perf data files are stored by setting the
> >> +PERF_TEST_CORESIGHT_DATADIR environment variable such as:
> >> +
> >> + PERF_TEST_CORESIGHT_DATADIR=/var/tmp
> >> + perf test
> >> +
> >> +You may wish to set these above environment variables if you which to
> >> +keep the output of tests outside of the current working directory for
> >> +longer term storage and examination.



--
Mike Leach
Principal Engineer, ARM Ltd.
Manchester Design Centre. UK

\
 
 \ /
  Last update: 2022-02-04 15:51    [W:1.474 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site