svn commit: r222060 - in user/avg/xcpu/sys: kern sys

[Andriy Gapon]

on 18/05/2011 23:50 Attilio Rao said the following:
> 2011/5/18  <mdf at freebsd.org>:
>> On Wed, May 18, 2011 at 1:06 PM, Attilio Rao <attilio at freebsd.org> wrote:
>>> 2011/5/18 Andriy Gapon <avg at freebsd.org>:
>>>> on 18/05/2011 21:24 Pawel Jakub Dawidek said the following:
>>>>> On Wed, May 18, 2011 at 03:08:30PM +0000, Andriy Gapon wrote:
>>>>> [...]
>>>>>> --- user/avg/xcpu/sys/sys/proc.h     Wed May 18 15:07:36 2011        (r222059)
>>>>>> +++ user/avg/xcpu/sys/sys/proc.h     Wed May 18 15:08:30 2011        (r222060)
>>>>>> @@ -781,7 +781,7 @@ MALLOC_DECLARE(M_SUBPROC);
>>>>>>   * Otherwise, the kernel will deadlock since the scheduler isn't
>>>>>>   * going to run the thread that holds the lock we need.
>>>>>>   */
>>>>>> -#define     THREAD_PANICED()        \
>>>>>> +#define     TD_IS_INPANIC() \
>>>>>>      (panicstr != NULL && (curthread->td_flags & TDF_INPANIC) != 0)
>>>>>
>>>>> Does TDF_INPANIC make sense without panicstr being set?
>>>>
>>>> Very good observation.  It seems that TDF_INPANIC can never be set unless
>>>> panicstr is set.  So, I guess it should be OK to simplify the macro further.
>>>> Thank you.
>>>
>>> However I think that TDF_INPANIC handling is not optimal.
>>> You should really acquire thread_lock otherwise you are going to break
>>> choosethread() concurrency.
>>>
>>> I would prefer to make TDF_INPANIC a private flag and just use it with
>>> curthread, if possible, but I still don't have a good way to resolve
>>> choosethread() (I would dig the runqueue adding path and resolve the
>>> problem later in the codeflow, I think).
>>
>> I know it's almost required now (sync on reboot?!?!), but I would
>> strongly question, from an architectural standpoint, why the scheduler
>> should be running at all in panic.  Once a thread pulls the panic
>> trigger, nothing else should run except ddb in that thread's context.

This is true.  The code that I propose makes sure that there is only one CPU and
only one thread running when TDF_INPANIC is set.  If someone decides to enable
sync-on-panic and expect it to work, I can only pray for them.  Yes, there most
likely will be a knob to turn off CPU stopping for panic, but I plan it mostly as
a quick way to switch back to legacy (present) behavior if a need arises (bugs,
tec).  It may work to the same degree as panic on SMP works now, but without any
guarantees.

> Yes, I mostly agree, so for avg's purpose he should maybe assert
> choosethread() doesn't pickup the TDF_INPANIC threads (or just skip
> the check at all, as it should be made private).

Unfortunately I can not understand your concern and proposal.
Not sure why choosethread() would be called at all, why it shouldn't pick up
TDF_INPANIC thread(s), and which check could/should be skipped and what should be
made private.  Sorry :(

-- 
Andriy Gapon