[Lse-tech] Re: [RFC][PATCH 00/11] Process Events modularization using Global Task Notifiers

[Jes S. <je...@sg...>]

Alan Stern wrote:
> On Wed, 17 May 2006, Jes Sorensen wrote:
> 
>>  If task_notifier_call_chain() etc. were inline'd we'd save one
>> level of indirection without growing the code size given that these
>> functions only continue calling into the next level of notifier
>> functions.
> 
> The way Matt did it is quite common; lots of other notifier chains in the
> kernel work this way.  Keeping those routines out-of-line rather than
> in-line has two advantages:
> 
>      1: It makes the kernel a little bit smaller, because each caller has 
> 	to submit only two arguments instead of three.  (Of course, traded
> 	off against this is the need to include the code for the
> 	out-of-line routine itself.  The more callers, the bigger the
> 	advantage.)

Hi Alan,

In this case I think 1) is a wash, since it's a function call straight
to another function call, only dereferencing a pointer once. The cost of
the call by far outweighs it.

>      2: It allows you to hide the underlying mechanism (the chain head)
> 	and expose only an API.  This simplifies later changes, such as
> 	the move from an atomic chain to a blocking chain.

The users have to be recompiled for any kernel update anyway so I don't
feel this is applicable either.

If someone finds it makes sense the register/unregister calls could be
kept out-lined, but I think the call chain should be inlined for
performance reasons.

Cheers,
Jes