Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1761489AbYFIOYF (ORCPT ); Mon, 9 Jun 2008 10:24:05 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1760918AbYFIOUa (ORCPT ); Mon, 9 Jun 2008 10:20:30 -0400 Received: from www.tglx.de ([62.245.132.106]:41045 "EHLO www.tglx.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757914AbYFIOU2 (ORCPT ); Mon, 9 Jun 2008 10:20:28 -0400 Date: Mon, 9 Jun 2008 16:20:07 +0200 From: "Hans J. Koch" To: Uwe =?utf-8?Q?Kleine-K=C3=B6nig?= Cc: "Hans J. Koch" , Magnus Damm , "linux-kernel@vger.kernel.org" , "gregkh@suse.de" , "akpm@linux-foundation.org" , "lethal@linux-sh.org" , "tglx@linutronix.de" Subject: Re: [PATCH] uio_pdrv: Unique IRQ Mode Message-ID: <20080609142007.GH3223@local> References: <20080604101144.GA3207@local> <20080605090916.GA3198@local> <20080605112744.GB3198@local> <20080608205455.GA3191@local> <20080609075701.GA20656@digi.com> <20080609095408.GB3223@local> <20080609123204.GA27803@digi.com> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: <20080609123204.GA27803@digi.com> User-Agent: Mutt/1.5.17+20080114 (2008-01-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 6952 Lines: 149 On Mon, Jun 09, 2008 at 02:32:04PM +0200, Uwe Kleine-König wrote: > Hi Hans, > > Hans J. Koch wrote: > > On Mon, Jun 09, 2008 at 09:57:01AM +0200, Uwe Kleine-König wrote: > > > Hello Hans, > > > > > > > Did you notice that in this thread nobody spoke up to support your > > > > patch? > > > Actually I like what the patch tries to achieve. I'd like to have it a > > > bit more explicit tough: > > > > > > - Provide the irq disabling handler in uio_pdrv.c (or even uio.c) with a > > > prototype in an adequate header. Then the platforms that want this > > > kind of handling can request it explicitly. > > > > You could provide an irqcontrol() function in uio_pdrv that calls a function > > defined in board support. If no function is defined there, it returns > > -ENOSYS. That would be consistent behaviour and not limited to > > non-shared interrupts. Note that this requires the add-write-function > > patch I recently posted. > I didn't check, but I think this is what is happening just now, though > with a different implementation: board support passed the uio_info which > might or might not include a irqcontrol() function. This is given > unchanged to uio_register. Assuming that writing without an > irqcontrol() function yields -ENOSYS we're already there. Yes, of course, you're right. > > > > - Don't use this handler automatically. > > > > > > - Provide the function named uio_pdrv_unique_irqcontrol in Magnus' patch > > > in uio_pdrv.c and in an adequate header. > > > > Why invent a new name? The approach above works with all kinds of irqs on > > all platforms. > > > > > > > > - Either rely on userspace to enable the irq before reading/polling or > > > assert that in kernel space. See also > > > http://thread.gmane.org/gmane.linux.kernel/684683/focus=689635 > > > (I asked tglx about the race condition via irc, but without a response > > > so far.) > > > > There are two problems: > > 1) If the hardware is designed in such a broken way that userspace needs > > a read-modify-write operation on a combined irq mask/status register to > > re-enable the irq, then this is racy against a new interrupt that occurs > > simultaneously. We have seen this on two devices so far. > You didn't understand what I want. (Probably because I choosed a poor > wording.) > > IMHO it should be asserted that irqs are on before waiting for the irq > in poll and read. So I suggest to call irqcontrol(ON) before doing so. > This should allow to work with that kind of hardware, right? Yes. But userspace can simply write() a 1 to /dev/uioX to achieve the same result. This would clearly show what's happening. Remember, this is only needed for certain (broken) hardware. If we hide some automagic irq enabling in the kernel, it'll become less obvious and might even have some bad side effects. I want to avoid this kind of trickery, especially as it is not needed. Userspace should use write() to control irqs. It's like this with any normal UIO driver, and we shouldn't have a different handling in uio_pdrv. Think of a chip that's directly connected to the bus on some embedded board. You use uio_pdrv to handle it. Then the very same chip appears on a PCI card in a normal PC. You write a normal UIO driver for it. The userspace part of both drivers could be exactly the same. But if uio_pdrv automagically reenabled the irq, we would need different handling in userspace, without reasons obvious to the user. Everything within the UIO framework (uio_pdrv is just becoming a part of it) should be seen from such a general point of view. I can understand that you or Magnus are mainly concerned with the embedded boards on your desk. But we want to support ALL architectures that can use UIO (in fact, all except S390). Please consider examples like the one above when you think about changes. > > > 2) If we wanted to make sure the interrupt is enabled in read() and > > poll(), we would have the problem that userspace usually calls poll() > > and then read() immediately afterwards. This would enable the irq twice, > > which can lead to two interrupts being seen in some cases. > OK, for this case a pending flag would be needed. (This doesn't mean I > suggest to implement it that way.) I'll think about that a bit. > > > For both reasons, we decided that introducing the write() function to > > enable and disable irqs is the best solution. Greg already added that > > patch to his tree, so it should appear in one of the next kernels. > > > > > Currently the former is done, but if we decide to let it as it is, I'd > > > like to have it documented. (I.e. something like: "Before > > > polling/reading /dev/uioX assert that irqs are enabled.") > > > > We cannot do this, at least not in a clean way. > We cannot document it in a clean way? (Probably not, I assume "this" > still refers to "enable the irq in read and poll"?) Yes ;-) > > > > The last point is a bit independent from that mode, but applies to > > > devices that have a irqcontrol function in general. > > > > > > Apart from the general things above, I'd change a few things in the > > > implementation: > > > > > > - call dev_info->irqcontrol(OFF) in the handler (instead of > > > disable_irq()) and demand that calling this is idempotent. > > > With this change it isn't uio_pdrv specific any more and could go to > > > uio.c. > > > > Why should we want to do this? You save five lines of irq handler code > > by introducing the need for an irqcontrol() function. > Taking Magnus' patch there is a default irqcontrol() function that does > the right thing in this case. This should probably go to uio_pdrv.c. Just doing irq_disable() limits it to irqs that are not shared. If there was a huge advantage, I'd think about it. But as it is, I'll never accept that. Magnus' patch is not needed, not even by himself. > > > I already said that in the discussion with Magnus, I don't see any > > advantage in this. Magnus cannot tell me either, he just keeps telling > > me "but we can do it" over and over again. > I think the benefit is to add some code to uio_pdrv and/or uio and in > turn save some code in board support code. Yes, but the savings (if any) are small compared with the disadvantages. > In fact this is similar to > the whole uio_pdrv driver. No. uio_pdrv is a nice idea, cleanly implemented. Lots of people can use it on different platforms. You could implement a driver for a PCI card as a platform device if you wanted to. It does one thing, and does it well. I don't want to spoil that by adding if..thens that introduce special behaviour for some cornercases, especially if this doesn't even add a significant advantage. Thanks, Hans -- 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/