Re: [PATCH 2.6-git] SPI core refresh

[Vitaly Wool]

David Brownell wrote:

> On Wednesday 30 November 2005 11:17 pm, Vitaly Wool wrote:
>  
>
> > Mark Underwood wrote:
> >
> >    
> >
> > > > However, there also are some advantages of our core compared to David's I'd like to mention
> > > >
> > > > - it can be compiled as a module
> > > >        
> > > So can David's. You can use BIOS tables in which case you must compile the SPI core into the
> > > kernel but you can also use spi_new_device which allows the SPI core to be built as a module (and
> > > is how I am using it).
> > >      
> > You limit the functionality, so it's not the case.
> >    
>
> As noted in my comparison of last week (you're still ignoring that):
>
> - Mine lets board-specific device tables be declared in the
>   relevant arch_setup() thing (board-*.c).  Both frameworks allow
>   later board specific code to dynamically declare the devices,
>   with binary (Dave's) or parsed-text (Dmitry's) descriptions. 
>
> What Mark said was that in this case he used the "late" init.  You seem
> to be saying he's not allowed to do that.  Which is nonsense; there are
> distinct mechanisms for the good reason that "late" init doesn't work
> so well without dynamic discovery ... which SPI itself doesn't support.
> Hence the need for board-specific "this hardware exists" tables.
>
>  
Can you please clarify what you mean here? Better even if Mark describes 
what he does. The ideal situation would be if he posted a patch.

>  
>
> > If there's more than one SPI controller onboard, spi_write_then_read 
> > will serialize the transfers ...
> >    
>
> Which, as has been pointed out, would be a trivial thing to fix
> if anyone were actually to have a problem.  Sure it'd incur the
> cost of a kmalloc on at least some paths -- serializing in the
> slab layer instead! -- but that's one price of using convenience
> helpers not performance oriented calls.
>  
Well, most of the drivers will use that helpers I guess.
The thing however is that if you try to implement this in a "clean" way 
you will come to a sport of framework we've developed for memeory 
allocations, as I've saild previously.

>  
>
> > 	 Moreover, if, say, two  
> > kernel threads with different priorities are working with two SPI 
> > controllers respectively *priority inversion* will happen.
> >    
>
> That characteristic being inherited from semaphores (or were they
> updated with RT_PREEMPT?), and being in common with most I/O queues
> in the system.  Not something to blame on any line of code I wrote.
>  
I think they weren't.
The whole thing doesn't seem thought out nicely to me. The solution 
exists, of course, and that is -- do somthing similar to what we did there.

> Oh, and I noticed a priority inversion in your API which shows
> up with one SPI controller managing two devices.  Whoops!  I'd
> far rather have such inversions be implementation artifacts; it's
> easy to patch an implementation, hard to change all API users.
>  
Not sure if I understood you. Can you please describe the situation when 
this prio inversion happens?
What priorities are you talking about? One controller is one thread, so 
it's _one_ priority, consequently there's nothing to invert.
As for your second statement, I don't argue. The fact however is that if 
you implement the mehtod which corrects priority inverstion problems 
your core will not be either so lightweight or so flexible. :)

>  
>
> > > > - it's more adapted for use in real-time environments
> > > > - it's not so lightweight, but it leaves less effort for the bus driver developer.
> > > >        
> > > But also less flexibility. A core layer shouldn't _force_ a policy
> > >      
> > Nope, it's just a default policy.
> >    
>
> One that every driver pays the price for.  Allocating a task even
> when it doesn't need it; every call going through a midlayer that
> wants to take over queue management policy; and more.  (Unless you
> made a big un-remarked change in a patch you called "refresh"...)
>  
It's not obvious that this price is high.
Anyway, it's a point I should agree with; this functionality better be a 
config option. Feel free to submit a patch, as you like to say.

>  
>
> > > on a bus driver. I am currently developing an adapter driver for David's system and I wouldn't say
> > > that the core is making me do things I think the core should do. Please could you provide examples
> > > of where you think Davids SPI core requires 'effort'.
> > >      
> > Main are
> > - the need to call 'complete' in controller driver
> >    
>
> So you think it's better to have consistent semantics be optional?
>
> That seems to be the notion behind your spi_transfer() call, which
> can't decide whether it's going to be synchronous or asynchronous.
> Instead, it decided to be error prone and be both.  :)
>
>
>  
Not sure if I understood you here, sorry.

> > - the need to implement policy in controller driver
> >    
>
> The "policy" in question is something that sometimes needs to
> be board-specific -- priority to THAT device, synch with THIS
> external signal, etc -- which is why I see it as a drawback
> that you insist the core implement one policy.
>  
Again, the policy can be overridden.

Vitaly
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/