cvs commit: src/sys/kern kern_thread.c

[David Xu]

Scott Long wrote:

> Poul-Henning Kamp wrote:
>
>> In message <423C0A0F.6090409 at freebsd.org>, David Xu writes:
>>
>>> Poul-Henning Kamp wrote:
>>>
>>>
>>>> phk         2005-03-19 08:22:13 UTC
>>>>
>>>> FreeBSD src repository
>>>>
>>>> Sleeping is not allowed in uma->fini
>>>>
>>>
>>> I have not tested if this change will hurt libpthread performance,
>>> but it is obvious that benifit of uma cache will be reduced by
>>> this change.
>>
>>
>>
>> I am aware of that and I don't like it either.
>>
>> The problem is that the uma_mutex is held over the uma->fini() call
>> and I'm surprised that we set so tight constraints on these functions
>> but I did not manage to corner the UMA crew and talk to them about it
>> yesterday evening.
>>
>> One obvious solution is to taskq the release of the unit number.
>>
>
> Since unit number release isn't a time-critical operation (right?), it
> sounds like a perfect candidate for a taskqueue.
>
> Scott
>
>
Thread creation should be fast. I like the unit allocator because it can
save duplicated code, but if things become complex, we should restore
orignal code which is straight forward.

David Xu