Re: pci_get_subsys: GFP_KERNEL allocations with IRQs disabled

[Bjorn Helgaas <bhelgaas@google.com>]

On Wed, Aug 22, 2012 at 12:49 AM, Feng Tang <feng.tang@intel.com> wrote:
> Hi Fengguang,
>
>
> On Wed, 22 Aug 2012 10:50:08 +0800
> Fengguang Wu <fengguang.wu@intel.com> wrote:
>
>> Feng,
>>
>> > I think it's pci_get_subsys() triggered this assert:
>> >
>> >         /*
>> >          * Oi! Can't be having __GFP_FS allocations with IRQs disabled.
>> >          */
>> >         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
>> >                 return;
>>
>> It's bisected down to this commit:
>>
>> commit 55c844a4dd16a4d1fdc0cf2a283ec631a02ec448
>> Author:     Feng Tang <feng.tang@intel.com>
>> AuthorDate: Wed May 30 23:15:41 2012 +0800
>> Commit:     Ingo Molnar <mingo@kernel.org>
>> CommitDate: Wed Jun 6 12:03:23 2012 +0200
>>
>>     x86/reboot: Fix a warning message triggered by stop_other_cpus()
>>
>> Thanks,
>> Fengguang
>
> Thanks for the bisection.
>
> Revert my commit should be a solution, but can we simply make the pci_device_id
> a local on stack one instead of using sleepable kmalloc for it, as this
> sounds fragile when pci_get_subsys get called in a late system reboot stage?

I think this is a great idea.  Can you make this a real patch, with a
changelog and Signed-off-by?

We should also remove the obsolete comment about early boot.  I'm not
sure the no_pci_devices() check is needed, either.  And we can make
the same simplification in pci_get_class().

> ------------
> diff --git a/drivers/pci/search.c b/drivers/pci/search.c
> index 993d4a0..e5ccede 100644
> --- a/drivers/pci/search.c
> +++ b/drivers/pci/search.c
> @@ -246,7 +246,7 @@ struct pci_dev *pci_get_subsys(unsigned int vendor, unsigned int device,
>                                struct pci_dev *from)
>  {
>         struct pci_dev *pdev;
> -       struct pci_device_id *id;
> +       struct pci_device_id id;
>
>         /*
>          * pci_find_subsys() can be called on the ide_setup() path,
> @@ -257,17 +257,12 @@ struct pci_dev *pci_get_subsys(unsigned int vendor, unsigned int device,
>         if (unlikely(no_pci_devices()))
>                 return NULL;
>
> -       id = kzalloc(sizeof(*id), GFP_KERNEL);
> -       if (!id)
> -               return NULL;
> -       id->vendor = vendor;
> -       id->device = device;
> -       id->subvendor = ss_vendor;
> -       id->subdevice = ss_device;
> -
> -       pdev = pci_get_dev_by_id(id, from);
> -       kfree(id);
> +       id.vendor = vendor;
> +       id.device = device;
> +       id.subvendor = ss_vendor;
> +       id.subdevice = ss_device;
>
> +       pdev = pci_get_dev_by_id(&id, from);

No need for "pdev" here, since we don't have to free anything.

>         return pdev;
>  }