From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Message-ID: <54FDE1EC.9040207@ti.com> Date: Mon, 9 Mar 2015 14:09:48 -0400 From: Murali Karicheri MIME-Version: 1.0 To: Guenter Roeck CC: Bjorn Helgaas , Fengguang Wu , LKP , "linux-pci@vger.kernel.org" , "linux-kernel@vger.kernel.org" Subject: Re: [PCI] BUG: unable to handle kernel References: <20150306060631.GD28187@wfg-t540p.sh.intel.com> <54F9C407.5020602@ti.com> <54F9CC6B.5070803@ti.com> <20150306165504.GA30094@roeck-us.net> <54F9EAA8.30007@ti.com> <54FDAB8B.3010404@ti.com> <54FDC1FC.2030807@ti.com> <54FDC52B.1070602@roeck-us.net> <54FDD277.2060406@ti.com> <54FDD995.1080000@roeck-us.net> In-Reply-To: <54FDD995.1080000@roeck-us.net> Content-Type: text/plain; charset="UTF-8"; format=flowed Sender: linux-kernel-owner@vger.kernel.org List-ID: On 03/09/2015 01:34 PM, Guenter Roeck wrote: > On 03/09/2015 10:03 AM, Murali Karicheri wrote: >> On 03/09/2015 12:07 PM, Guenter Roeck wrote: >>> On 03/09/2015 08:53 AM, Murali Karicheri wrote: >>>> On 03/09/2015 10:44 AM, Bjorn Helgaas wrote: >>>>> On Mon, Mar 9, 2015 at 9:17 AM, Murali Karicheri >>>>> wrote: >>>>>> On 03/06/2015 12:58 PM, Murali Karicheri wrote: >>>>>>> >>>>>>> On 03/06/2015 11:55 AM, Guenter Roeck wrote: >>>>>>>> >>>>>>>> On Fri, Mar 06, 2015 at 10:48:59AM -0500, Murali Karicheri wrote: >>>>>>>> [ ... ] >>>>>>>> >>>>>>>>>> From 098b4f5e4ab9407fbdbfcca3a91785c17e25cf03 Mon Sep 17 >>>>>>>>>> 00:00:00 2001 >>>>>>>>> From: Murali Karicheri >>>>>>>>> Date: Fri, 6 Mar 2015 10:23:08 -0500 >>>>>>>>> Subject: [PATCH] pci: of : fix kernel crash >>>>>>>>> >>>>>>>>> This is a debug patch to root cause the kernel crash >>>>>>>>> >>>>>>>>> commit 0b2af171520e5d5e7d5b5f479b90a6a5014d9df6 >>>>>>>>> >>>>>>>>> PCI: Update DMA configuration from DT >>>>>>>>> >>>>>>>>> Signed-off-by: Murali Karicheri >>>>>>>>> --- >>>>>>>>> drivers/of/of_pci.c | 8 ++++++++ >>>>>>>>> drivers/pci/host-bridge.c | 5 +++++ >>>>>>>>> 2 files changed, 13 insertions(+) >>>>>>>>> >>>>>>>>> diff --git a/drivers/of/of_pci.c b/drivers/of/of_pci.c >>>>>>>>> index 86d3c38..5a59fb8 100644 >>>>>>>>> --- a/drivers/of/of_pci.c >>>>>>>>> +++ b/drivers/of/of_pci.c >>>>>>>>> @@ -129,6 +129,14 @@ void of_pci_dma_configure(struct pci_dev >>>>>>>>> *pci_dev) >>>>>>>>> struct device *dev =&pci_dev->dev; >>>>>>>>> struct device *bridge = pci_get_host_bridge_device(pci_dev); >>>>>>>>> >>>>>>>>> + if (!bridge || !bridge->parent) { >>>>>>>>> + if (!bridge) >>>>>>>>> + pr_err("PCI bridge not found\n"); >>>>>>>>> + if (!bridge->parent) >>>>>>>>> + pr_err("PCI bridge parent not found\n"); >>>>>>>> >>>>>>>> >>>>>>>> You'll see a crash here if bridge is NULL. Maybe add an else before >>>>>>>> the second >>>>>>>> if statement ? Also, dev_err might be a bit more useful and >>>>>>>> would be >>>>>>>> available. >>>>>>>> >>>>>>> Fixed and attached. >>>>>>> >>>>>>> Murali >>>>>>>> >>>>>>>> Thanks, >>>>>>>> Guenter >>>>>>>> >>>>>>>>> + return; >>>>>>>>> + } >>>>>>>>> + >>>>>>>>> of_dma_configure(dev, bridge->parent->of_node); >>>>>>>>> pci_put_host_bridge_device(bridge); >>>>>>>>> } >>>>>>>>> diff --git a/drivers/pci/host-bridge.c b/drivers/pci/host-bridge.c >>>>>>>>> index 3e5bbf9..ef2ab51 100644 >>>>>>>>> --- a/drivers/pci/host-bridge.c >>>>>>>>> +++ b/drivers/pci/host-bridge.c >>>>>>>>> @@ -28,6 +28,11 @@ struct device >>>>>>>>> *pci_get_host_bridge_device(struct >>>>>>>>> pci_dev *dev) >>>>>>>>> struct pci_bus *root_bus = find_pci_root_bus(dev->bus); >>>>>>>>> struct device *bridge = root_bus->bridge; >>>>>>>>> >>>>>>>>> + if (!bridge) { >>>>>>>>> + pr_err("PCI: bridge not found\n"); >>>>>>>>> + return NULL; >>>>>>>>> + } >>>>>>>>> + >>>>>>>>> kobject_get(&bridge->kobj); >>>>>>>>> return bridge; >>>>>>>>> } >>>>>>>>> -- >>>>>>>>> 1.7.9.5 >>>>>>>>> >>>>>>>> >>>>>>> >>>>>> BJorn, >>>>>> >>>>>> Any chance of applying the attached debug patch to see if this fixes >>>>>> and >>>>>> provide some additional information on this BUG? Not sure who will >>>>>> pick this >>>>>> one and apply. >>>>> >>>>> The change that caused the oops (0b2af171520e ("PCI: Update DMA >>>>> configuration from DT")) only exists on my pci/iommu branch, so I'm >>>>> the one to apply it. >>>>> >>>>> It's much easier for me to deal with plain text patches (not >>>>> attachments). >>>>> >>>>> I'm hesitating because I don't want to encourage use of the 0-day >>>>> testing robot as a tool at which we can just throw debug patches. The >>>>> robot is a service that costs somebody real money, and I want to be a >>>>> good neighbor when using it. >>>> >>>> Thanks for the clarification as I don't have much information on the >>>> testing robot. At the same time the question is how similar incidence >>>> in the past have been handled. I am a newbie w.r.t to this. This is >>>> first time I have introduced a patch that impacts multiple >>>> arch/machines. >>>> >>>>> >>>>> Was the information in the robot's report enough to reproduce the >>>>> oops? If not, is there additional information we could add to the >>>>> report that would enable you to reproduce it? Even if we can't >>>>> reproduce the oops, the report seems detailed enough that we should be >>>>> able to deduce the problem and produce a fix in which we have high >>>>> confidence. >>>> >>>> The BUG report essentially indicates the crash happened in >>>> of_pci_dma_configure(). The machine specific log make sense to a >>>> person familiar with this arch and I am not familiar with the same. So >>>> anyone can help narrow down the root cause of this? >>>> >>>> Looking at the code, there are two ptr variables that are accessed >>>> without checking for NULL as initial thinking was that these can never >>>> be NULL. So the debug patch is just adding addition check before >>>> accessing the ptr. I can send this patch without debug prints if that >>>> make sense. I was thinking to get confirmation that this is indeed the >>>> case before adding the check. What do you think the right approach >>>> here? Send a patch for this to the ML for adding the check as a >>>> potential fix? Or someone can help me investigate the crash dump and >>>> root cause it? or if we can use test robot to confirm this, I can >>>> re-send the patch ASIS to the list. Please clarify. >>>> >>> If the assumption is that the pointers can never be NULL, >>> wouldn't it be important to see a call trace and to find out >>> if the NULL pointers can actually be seen by design, >>> or if there is some other bug ? >> >> Call trace shows >> >> [ 0.576666] [<7976c1ac>] pci_device_add+0xbc/0x820 >> [ 0.576666] [<7976c1ac>] pci_device_add+0xbc/0x820 >> >> >> And BUG seems to be in of_pci_dma_configure() as shown in the BUG report. >> >> of_pci_dma_configure() calls newly added API call to >> pci_get_host_bridge_device(). Seems like this has succeeded which >> means bridge is non NULL IMO. However in this function it passes >> bridge->parent->of_node to of_dma_configure(). So I suspect >> bridge->parent is NULL for some reason. Is there a chance for parent >> being NULL in this or any other platform? >> > > Can bridge be the root bridge ? Going by the code below, bridge is assigned the ptr to bridge on the root bus. +struct device *pci_get_host_bridge_device(struct pci_dev *dev) +{ + struct pci_bus *root_bus = find_pci_root_bus(dev->bus); + struct device *bridge = root_bus->bridge; + + kobject_get(&bridge->kobj); + return bridge; +} + So to answer your question, yes it is the root bridge. -- Murali Karicheri Linux Kernel, Texas Instruments