cvs commit: src/sys/kern kern_sig.c

[Scott Long]

David Xu wrote:
> Scott Long wrote:
> 
>> David Xu wrote:
>>
>>> Julian Elischer wrote:
>>>
>>>>>
>>>>>
>>>>>>> The alternative, of course, is to just fix the code that assumes
>>>>>>> that swapping doesn't exist.
>>>>>>>
>>>>>>>     
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> First find all code written in such way, but it is not that easy.
>>>>>>   
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> True.  If we changed msleep() to disable swapping by default, then
>>>>> we wouldn't have to worry about correctness problems related to
>>>>> missing some.
>>>>>  
>>>>>
>>>>
>>>> adding the flag to ENABLE swapping would be ABI compatible.
>>>>
>>>>
>>> how about adding a PNOSWAP to msleep ?  but I won't trust the kernel
>>> under swapping, because they can not give me 100% guarantee, my
>>> machine crashes several times per-month, even when fscking at boot time,
>>> mostly it is a page fault.
>>>
>>> David Xu
>>>
>>>
>>>
>>>
>>>
>>
>> An msleep/tsleep option doesn't solve the problem because the the msleep
>> might happen several layers down from where the stack abuse is taking
>> place, and thus the caller would have no idea that it's needed.
> 
> 
> But it at least can help a piece of code managed by a single guy.
> 

It's still of limited value.  I'd rather the effort be put into
diagnostics development.  Also, regarding your objections to PHOLD,
I thought that the proc lock was a leaf mutex.  Where are you seeing
LORs with it?

>>   The fix
>> for sigwait() is easy and can be applied without hacking in new options
>> that have limited value.  I don't argue that similar problems might
>> exist elsewhere, but swappable kstacks have been part of BSD since
>> before most of us knew where the power switch was on our Ataris, so it's
>> likely not to be a wide-spread and fundamental problem in the code.  I'd
>> be in favor of adding diagnostics that help catch these problems and
>> report them, but just throwing away kstack swapping in leiu of taking
>> the 2 minutes to fix sigwait() is pretty silly.
>>
> Yes, sigwait is simple, but  my umtx code is also broken by this silly 
> swapping
> code, now I have to fill malloc/free/retry/lock_order_reversal_work_around
> all over the code.

Can you provide a reference?  It turns out the fixing sigwait() can be
done without any mallocs at all, maybe the same can be done for umtx.

Scott