SIS MP success

[Chris Johns]

On 15/12/18 7:17 pm, Jiri Gaisler wrote:
> 
> On 12/15/18 1:07 AM, Chris Johns wrote:
>> On 14/12/18 8:48 pm, Jiri Gaisler wrote:
>>> OK, I will look into this over Christmas. I have some long flights
>>> coming up so it will be a good way to kill time ... :-)
>> Great. I have been on a few flights recently and I have been doing the same
>> thing. Power in the seats is nice and WiFi is even better.
>>
>>> Failures:
>>>  dl07.exe
>> That should not be happening. I will check this as part of some work I am planning.
>>
>>> Invalid:
>>>  spinternalerror01.exe
>>>  sptimecounter01.exe
>> These are typically crashes.
>>
>>> Average test time: 0:00:00.498379
>>> Testing time     : 0:05:20.956184
>> That is a nice run time.
> 
> It's on my 5 years old i7-4790S (8 threads, 65W). Without the SMP tests,
> I get run times of about 2 - minutes. I use the leon3-sis target in
> rtems-test. If I use leon3-run, rtems-test hangs after a few tests
> unless I use --jobs=1.

Ah thanks, this narrows down where I need to look.

>> If you add the --mail and related options (--mail-to, --mail-from, --smtp-host)
>> the results can be posted to build at rtems.org. We love seeing results being
>> posted by any and everyone.
>>
>>> The four failing/invalid tests fail the same way on real hardware, so I
>>> am not too worried about this. 
>> Is the failure seen without rtems-test?
>>
>> I ask because I think there maybe a bug in rtems-test with a threading load when
>> running a number of simulations at once. I am not sure if this is a qemu
>> specific issue or a general one, I need to revisit it. I suspect a locking issue
>> in rtems-test and it can be seen with tests hanging and simulators showing 0
>> load in top.
>>
>> On a non-smp build, I get the same fails,
>>> so it seems to be some issue with the tests...
>> There is a ticket open on 5.1 to have tests that do fail to be tagged as
>> "Expected Fail" so we can baseline what is expected. I cannot do this alone
>> because I am not sure which ones are expected failures.
> 
> I looked into the failures again and they only occur on leon3 when
> compiled with --enable-smp. On erc32 and leon2, and on leon3 without
> --enable-smp they run OK. I am working on improving gdb debugging
> support for sis-mp, so debugging the failures of the RTEMS tests could
> be a good test case to check how well it works. Breakpoints and stepping
> in an MP system are not that obvious to handle ...

I am sure. On ARM hardware with SMP you need to use the context id register to
hold the thread id and add it as a condition of the breakpoint. This means
RTEMS's context switcher needs to update this register.

Chris