Date:	Mon, 7 Dec 2015 14:52:00 +0200
From:	Heikki Krogerus <heikki.krogerus@...ux.intel.com>
To:	Chanwoo Choi <cw00.choi@...sung.com>
Cc:	Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
	MyungJoo Ham <myungjoo.ham@...sung.com>,
	David Cohen <david.a.cohen@...ux.intel.com>,
	Lu Baolu <baolu.lu@...ux.intel.com>,
	Mathias Nyman <mathias.nyman@...ux.intel.com>,
	Felipe Balbi <balbi@...com>, linux-usb@...r.kernel.org,
	linux-kernel@...r.kernel.org
Subject: Re: [PATCHv2 1/2] extcon: add driver for Intel USB mux

Hi,

On Mon, Dec 07, 2015 at 10:24:22AM +0900, Chanwoo Choi wrote:
> Hi,
> 
> On 2015년 12월 04일 17:51, Heikki Krogerus wrote:
> > Hi,
> > 
> >> I do never want to add some specific funtcion for only this driver.
> >> I think is not appropriate way.
> >> - intel_usb_mux_unregister
> >> - intel_usb_mux_register
> >>
> >> The client driver using extcon driver should use the standard extcon API
> >> for code consistency. Also, I'll do the more detailed review for this patch.
> > 
> > The internal mux we are controlling here is physically separate
> > device. Ideally we could populate child device for it, but since that
> > is not possible because of the resource conflict, we use the library
> > approach, which is really not that uncommon.
> 
> New added functions for only specific device driver is not library. 
> 
> The all device drivers which is included in some framework should
> connect to the other device driver through only framework API as following:
> 	--------------------          ----------------
> 	| EXTCON framework |<-------->| USB framework |
> 	--------------------          -----------------	
> 		|                            |
> 		|                            |
> 	extcon-intel-usb.c             pci-quirks.c
> 
> But, in this case, added funticon is just direct call function
> without any standard API. The below case is never appropriate implementation.
> 
> 	--------------------          ----------------
> 	| EXTCON framework |          | USB framework |
> 	--------------------          -----------------	
> 		|                            |
> 		|                            |
> 	extcon-intel-usb.c <--------  pci-quirks.c

Man.. Cal it what you want, but like I said, exposing driver specific
API is not ideal, but it is acceptable in special cases like this
where we simply are not able to populate child device. If nothing
else, then at least the fact that the code for the mux would otherwise
need to be duplicated, is enough to justify it.

> > I don't think I agree with your point even at general level. The
> > control required to handle this mux, even though simple, is enough to
> > deserve to be separated from xHCI code. xHCI should not need to care
> > about anything else expect does it have the mux, i.e. does it need to
> > register it or not. It should not need to care about how it needs to
> > be controlled or even what it is. We may decide to create something
> > else out of it instead of an extcon device later.
> > 
> > But in any case, the mux is available on all new Intel platforms, but
> > it needs to be controlled by OS only in few "special" cases. We can
> > not force xHCI (or pci-quirks.c to be more precise) to be aware of
> > these "special" cases. The only way to make it work like that would
> > bet by using ifdefs, and we really really don't want that.
> > 
> > And please also note that, though for now we only expect the mux
> > control registers to be part of xHCI MMIO, that is not always the
> > case. The control registers are part of the device controller MMIO on
> > some platforms. We do not want to duplicate the whole control of the
> > mux if/when we need the OS to be in control of it on a platform that
> > has those control registers mapped somewhere else then xHCI MMIO,
> > 
> > So I would say that we have pretty good justification for separating
> > the mux control, which means unfortunately custom API in this case.
> > 
> > But if you would prefer that we put the files somewhere else then
> > drivers/extcon/ and include/linux/extcon/ I'm fine with that. If you
> > like, we can put it to drivers/usb/host/ as that is where
> > pci-quirks.c is. That way I think we can also put the header to
> > include/usb/.
> 
> There are the two type of extcon drivers as following:
> - provider extcon driver which use the devm_extcon_dev_register() and extcon_set_cable_state().
> - client extcon driver which use the extcon_register_notifier() and extcon_set_cable_state() usually.
> The drivers/extcon directory only includes the provider extcon driver.
> 
> You make the extcon-intel-usb.c as provider extcon driver.
> At the same time, this driver is used for client extcon driver
> by using the extcon_register_notifier(). If you want to recevie
> the notification from extcon_register_notifier(), the extcon device
> should update the state of external connector throught extcon_set_cable_state().
> But, this driver don' use the extcon_set_cable_state().
> 
> I think that you should divide it according to role.
> 
> Again, the usage case of extcon have to consist of both provider extcon driver
> and client extcon driver. If there is no provider extcon driver,
> client extcon driver don't receive the any notification of external connector
> from provider extcon driver.

What you are saying is that it is OK for both "client" and "provider"
to change the state, but only client is allowed to react to a state
change? You got to admit that the roles are pretty obscure here...

In any case, I'm using the framework the way it allows itself to be
used. If you want your framework to be used in a particular way, then
you need to protect it from being used otherwise. Now anything with a
handle to the extcon_dev can use it in every way possible. You are not
documenting how you want the framework to be used. The only document
for extcon in Documentation/extcon/ is about porting stuff from some
Android specific "switch class" to extcon. Nowhere do you define what
is a "client" and what is a "provider" and what are they meant for.

If you want to limit what a "client" can do, you need to separate the
API for it. Use the gpio API as an example. Check how the consumers
and drivers are separated into their own headers in
include/linux/gpio/. Keep the include/extcon.h header as legacy and
deprecate it.

And if you do that, the framework should really be improved with a few
basic things regarding the "clients". At least start using some kind
of ref counting with them. Now nothing really prevents a "provider"
from being removed even if it has users (clients), or does it? This
basically also shows how obscure the line between a "client" and
"provider" is at the moment.

Right now with what we have I see nothing wrong with my approach.


Cheers,

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

Powered by blists - more mailing lists