Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S262507AbVAZXnN (ORCPT ); Wed, 26 Jan 2005 18:43:13 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S262346AbVAZXlF (ORCPT ); Wed, 26 Jan 2005 18:41:05 -0500 Received: from rproxy.gmail.com ([64.233.170.204]:27954 "EHLO rproxy.gmail.com") by vger.kernel.org with ESMTP id S261322AbVAZS0j (ORCPT ); Wed, 26 Jan 2005 13:26:39 -0500 DomainKey-Signature: a=rsa-sha1; q=dns; c=nofws; s=beta; d=gmail.com; h=received:message-id:date:from:reply-to:to:subject:cc:in-reply-to:mime-version:content-type:content-transfer-encoding:references; b=HY5jtgflnrrO9qDD8tRast6Evqg2O+/ejnUcmj8idC1McYEQMn5Yij145D7kxaSD+T5lx0iT2dREdo32Ehm1huutyhb0FMt6MDqdOuQ0Gpw9NaF62gFldO2yOQGUYVUQ7xfNY8uihkfffWYH6jKrQ0D6YNf/3pPxK3Midj2b8OM= Message-ID: Date: Wed, 26 Jan 2005 13:26:38 -0500 From: Dmitry Torokhov Reply-To: dtor_core@ameritech.net To: johnpol@2ka.mipt.ru Subject: Re: 2.6.11-rc2-mm1 Cc: Christoph Hellwig , Andrew Morton , greg@kroah.com, linux-kernel@vger.kernel.org In-Reply-To: <1106761176.5257.246.camel@uganda> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit References: <20050124021516.5d1ee686.akpm@osdl.org> <1106727902.5257.109.camel@uganda> <1106751547.5257.162.camel@uganda> <1106754848.5257.189.camel@uganda> <1106757974.5257.229.camel@uganda> <1106761176.5257.246.camel@uganda> Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 8145 Lines: 173 On Wed, 26 Jan 2005 20:39:36 +0300, Evgeniy Polyakov wrote: > On Wed, 2005-01-26 at 11:55 -0500, Dmitry Torokhov wrote: > > On Wed, 26 Jan 2005 19:46:14 +0300, Evgeniy Polyakov > > wrote: > > > On Wed, 2005-01-26 at 11:25 -0500, Dmitry Torokhov wrote: > > > > On Wed, 26 Jan 2005 18:54:08 +0300, Evgeniy Polyakov > > > > wrote: > > > > > On Wed, 2005-01-26 at 10:26 -0500, Dmitry Torokhov wrote: > > > > > > On Wed, 26 Jan 2005 17:59:07 +0300, Evgeniy Polyakov > > > > > > wrote: > > > > > > > > > > > > > Each superio chip has the same logical devices inside. > > > > > > > With your approach we will have following schema: > > > > > > > > > > > > > > bus: > > > > > > > superio1 - voltage, temp, gpio, rtc, wdt, acb > > > > > > > superio2 - voltage, temp, gpio, rtc, wdt, acb > > > > > > > superio3 - voltage, temp, gpio, rtc, wdt, acb > > > > > > > superio4 - voltage, temp, gpio, rtc, wdt, acb > > > > > > > > > > > > > > Each logical device driver (for existing superio schema) > > > > > > > is about(more than) 150 lines of code. > > > > > > > With your approach above example will be 150*6*4 + > > > > > > > 4*superio_chip_driver_size bytes > > > > > > > of the code. > > > > > > > > > > > > ????? Let's count again, shall we? Suppose we have: > > > > > > > superio1 - voltage, temp, gpio, rtc, wdt, acb > > > > > > > superio2 - voltage, temp, gpio, rtc, wdt, acb > > > > > > superio1 is driven by scx200 hardware, superio2 is driven by pc8736x > > > > > > or whatever. So here, you have 2 drivers to manage chips. Plus you > > > > > > have 6 superio interface drivers - gpio, acb, etc... > > > > > > It is exactly the same as your cheme size-wise. > > > > > > > > > > > > There is no need to many-to-many relationship. Each chip is bound to > > > > > > only one chip driver which registers several interfaces. Each > > > > > > interface is bound to only one interface driver. Interface driver is > > > > > > not chip specific, it knows how to run a specific interface (gpio, > > > > > > temp) and relies on chip driver to properly manage access to the > > > > > > hardware. Each combination of inetrface + interface driver produce one > > > > > > class_device (call it sio, serio, whatever). Class device provides > > > > > > unified view of the interface to the userspace. > > > > > > > > > > > > What am I missing? > > > > > > > > > > Since each logical device does not know who use it, it can not be, > > > > > for example, removed optimally. > > > > > You need to run through whole superio device set to find those one that > > > > > has reference to logical device being removed. > > > > > > > > What do you mean by removing optimally? Consider teher 2 scenarios: > > > > - you unload logical device driver which knows all the logical > > > > devices (interfaces) it is bound to and it releases those devices. > > > > > > It does not know chip drivers, which reference it. > > > > Of course it does, every logical device has one parent. Individual > > GPIO pins are not spread across your motherboard. Every one of them > > lives on some chip. They can all be driven by the same driver but they > > are different devices. It is like all your IDE drives are driven by > > the same driver but (unless LVM which is diffrent layer comes to play) > > they are still different drives. > > No, there is no "parent". > Each logical device is "shared" between several superio chips. > It is presented design. Ok, I presume that userspace although you saying that logical device is "shared" between chips different GPIO pins still visible to userspace as separate entities (correct me if I am wrong). That means that you chose in your implementation not to abstract out individual parts of SuperIO container as individual devices and instead have several logical drivers bound directly to a same chip. And that is why you need n-m relationship and inability to fit to the driver model and potential issues with power management. Instead of this / - - \ you need this / - - - \ - This will decouple chip drivers from the logical drivers. I argue that present design is sub-optimal and will hurt in the long run. > > > > We need to scan all superio chips to find that. > > > > > > > - you unload chip driver which knows what logical devices it has > > > > registered and removes them. Logical devices in turn unbind themselves > > > > from the logical drivers thay are bound to (only one for each device > > > > and they have link). > > > > Seems pretty clean to me. > > > > > > Yes, since there is sc->ldev relation. > > > > > > > > And situation become much worse, when we have so called logical device > > > > > clones - > > > > > it is virtual logical device that emulates standard one, but inside > > > > > performs > > > > > in a different way. Clone creates inside superio device(for example such > > > > > device > > > > > can live at different address). > > > > > When you do not have ldev->sc relation, you must traverse through each > > > > > logical device > > > > > in each superio chip to find clones. > > > > > Thus you will need to traverse through each superio chip, > > > > > then through each logical device it references, just to remove one > > > > > logical device. > > > > > > > > I am confused and need a concrete example. I would think that cases > > > > such as these should not require eny special handling, hardware access > > > > should be hidden from logical drivers by the chip driver. Or if only > > > > difference is the port address then it probably should be set by the > > > > chip driver as attribute of logical device. > > > > > > Consider sc.c and GPIO logical device in SC1100 (scx driver). > > > > > > /* > > > * SuperIO core is registering logical device. > > > * SuperIO chip knows where it must live. > > > * If logical device being registered lives at the different location > > > * (for example when it was registered for all devices, > > > * but has address(index) corresponding to only one SuperIO chip) > > > * then we will register new logical device with the same name > > > * but with the different location(index). > > > * > > > * It is called clone. > > > */ > > > > > > > In this sense every device is a clone. Consider I have 2 EtherExpress > > cards in my box. They surely live at different addresses but PCI > > system does not mess with "clones". And network layer does not care > > whether both of them driven by e100 or one is e100 and the other is > > RTL. > > It is not the same as two different devices on the bus. > > Consider sc1100 - GPIO lives in PCI aperture, absolutely not where > superio lives. > (And it is actully not logical device in this case, but it fits well > into design). It does not matter where it "lives". The chip driver should not expose these details to interface driver. Although you said serio examples are not applicable here let's take another look there: i8042 and pcips2 modules. One provides serio ports that live in legacy IO space, the other is PCI-based. psmouse would not care less which one is behind the port it is bound to. > > You skipped my examples, but presented design is very convenient to the > situation, > when you have two different sets of devices, or classes of devices, each > of which > can have link to any device from the other set. They were from different domains so irrelevant. -- Dmitry - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/