Re: [PATCH 0/5] Modularize PCI_DW related drivers.

From: Arnd Bergmann
Date: Thu Feb 25 2016 - 03:36:50 EST


On Thursday 25 February 2016 13:43:48 Kishon Vijay Abraham I wrote:
> Hi Arnd,
>
> On Wednesday 24 February 2016 02:34 PM, Arnd Bergmann wrote:
> > On Wednesday 24 February 2016 11:39:26 Kishon Vijay Abraham I wrote:
> >> Hi,
> >>
> >> On Monday 08 February 2016 05:30 AM, Paul Gortmaker wrote:
> >>> In a recent patch series that aimed to remove code related to module
> >>> unload for PCI support that was simply non modular, the discussion
> >>> led to people wanting to keep the code and push towards taking the
> >>> steps needed to support moving it towards tristate instead[1].
> >>>
> >>> Here, we take step one, which is simply making the Kconfig change
> >>> and then dealing with any build fallout or modpost fallout. What
> >>> amounts to essentially a sanity build test. To be clear, these
> >>> have not been runtime validated; that will need to be done by those
> >>> with access to real hardware. However, the changes are not anything
> >>> that should disrupt any existing built-in validation, so real world
> >>> users should not be impacted by this change.
> >>>
> >>> We start with a smaller family of drivers; those that actively select
> >>> PCI_DW, as a nice self contained group to test the waters and see if
> >>> everyone is still good with this approach before investing more time
> >>> on a wider scale to other pci/host/ code blocks.
> >>>
> >>> As such the drivers here share a dependency on having the same group
> >>> of functions exported in order to successfully complete modpost.
> >>>
> >>> In addition, we have to stray outside drivers/pci to add exports
> >>> in two places; once for an ARM fault handler, and once for an OF
> >>> variable.
> >>>
> >>> The pci-keystone-dw.c instance was handled separately because it
> >>> consists of two source files that need their own group of driver
> >>> specific exports above and beyond the "shared" ones.
> >>>
> >>> Then we convert the Kconfig for all remaining at once; we could have
> >>> done it on a per driver basis for ease of revert if anyone really
> >>> objects, but since it would be a one line change, that seemed like
> >>> not a real concern.
> >>>
> >>> Build testing was done on the linux-next tree for arm allmodconfig.
> >>
> >> I took these patches and gave a test with DRA7xx board. As expected there was
> >> no issues when the driver was built-in. However when I tried to rmmod/modprobe
> >> I got this error [2].
> >
> > Thanks for testing this!
> >
> >> [2] -> http://pastebin.ubuntu.com/15185894/
> >
> > It looks like you are hitting the BUG_ON() in ioremap_pte_range()
> > that checks if a virtual address already has a page table entry,
> > which in turn is probably a result of dw_pcie_host_init()
> > calling pci_remap_iospace() again for the same memory area
> > it has called the last time, and no cleanup done inbetween.
> >
> > Could you try adding a pci_unmap_iospace() and calling that
> > in the device remove function? Let me know if you need help
> > implementing it.
>
> That didn't look straight forward to me :-( I'll try to see this next week. Any
> help from you will make it simpler for me.

I tried writing the function now, and I think it's actually quite easy:

void pci_unmap_iospace(const struct resource *res)
{
#if defined(PCI_IOBASE) && defined(CONFIG_MMU)
return iounmap(PCI_IOBASE + res->start);
#endif
}

You just need to pass the same resource in here htat you pass into
pci_remap_iospace().

Arnd