В письме Thu, 29 Nov 2007 12:09:35 -0800, Matthew Dillon
написал:

quoted text
> 
> :I'm playing a bit with mechanism (register_swi(), unregister_swi() and
> :sched_ithd()). It work, but it not applicable for widely usage, because
> :number of SWI_XXX slots very limited. To example, someone want create
> :some driver and use for sending data to device function, called by
> :software interrupt. Today, this someone must change kernel sources at
> :least in <sys/interrupt.h> to add define of new SWI_XXX. If not to do
> :this, there is risk, to use number of softint, used someone else...
> 
>     Well, the main reason for that is that the whole concept of a
>     software interrupt isn't really applicable any more.  It's easier
>     to just create a high priority kernel thread and send a message to
>     the thread.  There are a ton of places in DragonFly where we do that
>     now.
> 

But it not easy way to implement such things... taskqueue_XXX in this
case is more comfortable way, then create thread, which will be in dry run
most of time.