From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Rafael J. Wysocki" Subject: Re: [RFC PATCH v11 4/5] PCI / PM: Add support for the PCIe WAKE# signal for OF Date: Thu, 28 Dec 2017 01:48:30 +0100 Message-ID: References: <20171225114742.18920-1-jeffy.chen@rock-chips.com> <9593014.U0c2JQdoEZ@aspire.rjw.lan> <5A41A0A7.5070605@rock-chips.com> <5491520.JsEYlZiECc@aspire.rjw.lan> <20171227150800.GF3875@atomide.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Return-path: In-Reply-To: <20171227150800.GF3875@atomide.com> Sender: linux-pci-owner@vger.kernel.org To: Tony Lindgren Cc: "Rafael J. Wysocki" , JeffyChen , Linux Kernel Mailing List , Bjorn Helgaas , Linux PM , Shawn Lin , Brian Norris , Doug Anderson , "devicetree@vger.kernel.org" , Linux PCI , Rob Herring , Frank Rowand List-Id: devicetree@vger.kernel.org On Wed, Dec 27, 2017 at 4:08 PM, Tony Lindgren wrote: > * Rafael J. Wysocki [171227 01:00]: >> On Tuesday, December 26, 2017 2:06:47 AM CET JeffyChen wrote: >> > Hi Rafael, >> > >> > Thanks for your reply :) >> > >> > On 12/26/2017 08:11 AM, Rafael J. Wysocki wrote: >> > >> >+ >> > >> >+ dn = pci_device_to_OF_node(ppdev); >> > >> >+ if (!dn) >> > >> >+ return 0; >> > >> >+ >> > >> >+ irq = of_irq_get_byname(dn, "wakeup"); >> > > Why is this a property of the bridge and not of the device itself? >> > >> > That is suggested by Brian, because in that way, the wakeup pin would >> > not "tied to what exact device is installed (or no device, if it's a slot)." >> >> But I don't think it works when there are two devices using different WAKE# >> interrupt lines under the same bridge. Or how does it work then? > > It won't work currently for multiple devices but adding more than > one wakeriq per device is doable. And I think we will have other > cases where multiple wakeirqs are connected to a single device, so > that issue should be sorted out sooner or later. > > And if requesting wakeirq for the PCI WAKE# lines at the PCI > controller does the job, then maybe that's all we need to start with. These are expected to be out-of-band, so not having anything to do with the Root Complex. In-band PME Messages go through the PCIe hierarchy, but that is a standard mechanism and it is supported already. WAKE# are platform-specific, pretty much by definition and I guess that on most ARM boards they are just going to be some kind of GPIO pins. > Then in addition to that, we could do the following to allow > PCI devices to request the wakeirq from the PCI controller: > > 1. PCI controller or framework implements a chained irq for > the WAKE# lines assuming it can mask/unmask the WAKE# lines > > 2. PCI devices then can just request the wakeirq from the PCI > controller > > And that's about it. Optionally we could leave out the dependency > to having PCI devices implement PM runtime and just resume the > parent (PCI controller) if PCI devices has not implemented > PM runtime. So if my understanding is correct, DT should give you the WAKE# IRQ for the given endpoint PCI device and you only are expected to request it. The rest should just follow from the other pieces of information in the DT. With the quite obvious caveat that the same IRQ may be used as WAKE# for multiple endpoint devices (which BTW need not be under the same bridge even). >> > >> >+ if (irq == -EPROBE_DEFER) >> > > Braces here, please. >> > ok, will fix in the next version. >> > >> > > >> > >> >+ return irq; >> > >> >+ /* Ignore other errors, since a missing wakeup is non-fatal. */ >> > >> >+ else if (irq < 0) { >> > >> >+ dev_info(&pdev->dev, "cannot get wakeup interrupt: %d\n", irq); >> > >> >+ return 0; >> > >> >+ } >> > >> >+ >> > >> >+ device_init_wakeup(&pdev->dev, true); >> > > Why do you call this before dev_pm_set_dedicated_wake_irq()? >> > >> > hmmm, i thought so too, but it turns out the dedicated wake irq >> > framework requires device_init_wakeup(dev, true) before attach the wake irq: >> > >> > int device_wakeup_attach_irq(struct device *dev, >> > struct wake_irq *wakeirq) >> > { >> > struct wakeup_source *ws; >> > >> > ws = dev->power.wakeup; >> > if (!ws) { >> > dev_err(dev, "forgot to call device_init_wakeup?\n"); >> > return -EINVAL; >> > >> >> Well, that's a framework issue, fair enough. >> >> That said, what if user space removes the wakeup source from under you >> concurrently via sysfs? Tony? > > Hmm sounds racy, need to take a look. Not only racy, as I don't see anything to prevent user space from making the dev->power.wakeup wakeup source go away via sysfs at any time *after* the IRQ has been requested. Pretty much right after dev_pm_set_dedicated_wake_irq() has returned, device_wakeup_disable() may be called on the device via wakeup_store() and it doesn't even check if the device has a wakeup irq. > I think the only reason > to have the wakeirq pointer there was to save memory. It might > make sense to remove the wakeirq dependency here. Well, that looks necessary to be honest. Thanks, Rafael