[Click] [PATCH 2/2] Task: Kill process_pending dead lock

[Eddie Kohler]

I suspect that the huge config problem is something else -- like the 
machine running out of memory.

The key stress test would, I think, involve multithreaded Click under 
load, with configuration swapping.

Thanks!!
Eddie


Roman Chertov wrote:
> Eddie,
>    Would this be a problem if in kernel mode a large config was dumped 
> into a write handler via proclikefs?  Maybe Jelena's crashes were caused 
> by that.
>    I can conduct a few stress tests.  What are the key requirements for 
> the test to see if the new changes work or not?
> 
> Roman
> 
> Eddie Kohler wrote:
>> Joonwoo,
>>
>>  > I took look into this lock up issue and I think I found something.
>>  >
>>  > RoutherThread::driver() calls run_tasks() with locked tasks.
>>  > But after calling run_tasks(), current processor can be changed since
>>  > schedule() might be called (eg. ScheduleLinux element)
>>  > So I think that's problem.  How do you think?
>>
>> I totally agree that this could be a problem.
>>
>> It looks like EXCLUSIVE handlers never really worked before. :(
>>
>> So my current analysis is this.  It is not appropriate for a thread to 
>> call blocking functions and/or schedule() when that thread has 
>> prevented preemption via get_cpu().  My prior patches prevented 
>> preemption.
>>
>> The solution is to separate "locking the task list" from "blocking 
>> task execution."  Clickfs, when executing an exclusive handler, 
>> "blocks task execution."  A thread that wants to examine the task list 
>> "locks" the list.
>>
>> This commit:
>> http://www.read.cs.ucla.edu/gitweb?p=click;a=commit;h=ede0c6b0a1cface05e8d8e2e3496ee7fcd5ee143 
>>
>> introduces separate APIs for locking the list and blocking task 
>> execution.  Exclusive handlers block task execution, but do not lock 
>> the task list.  I believe that task execution, in this patch, does not 
>> prevent preemption.  I believe the locking works out too.  User-level 
>> multithreading tests appear OK.
>>
>> Any willing stresstesters?  Pretty please? :)
>>
>> Eddie
>> _______________________________________________
>> click mailing list
>> click at amsterdam.lcs.mit.edu
>> https://amsterdam.lcs.mit.edu/mailman/listinfo/click
>>
>