Re: [PATCH] fakephp: Allocate PCI resources before adding the device

[Trent Piepho <xyzzy@speakeasy.org>]

On Wed, 26 Nov 2008, Darrick J. Wong wrote:
> On Wed, Nov 26, 2008 at 01:56:45AM -0800, Trent Piepho wrote:
> > Ok, that makes sense.  The device I'm using fakephp for doesn't have a
> > kernel driver so I wouldn't have noticed that.
> >
> > Have you tested this with a device that isn't present at boot?  I found
> > that I needed to a call to pci_enable_device() after assigning resources,
> > otherwise the BARs wouldn't be enabled.  This only happened if the device
> > wasn't present at boot time.
>
> Yes, I was actually using this driver to <cough> turn the ioatdma
> controller on after turning it off in the BIOS.

Maybe it's different on powerpc then?  My pseudo-hotplugable device is also
the only thing connected to the PCI-E host bus controller.  At boot the
controller is empty and so I think some code to enable its BARs gets
skipped.  But without the pci_enable_device(), I get this:

01:00.0 Signal processing controller: Freescale Semiconductor Inc Aurora Nexus Trace Interface
        Flags: fast devsel, IRQ 255
        Memory at 40000000 (64-bit, prefetchable) [disabled] [size=4K]
        Capabilities: [40] Power Management version 3
        Capabilities: [48] Message Signalled Interrupts: Mask+ 64bit+ Queue=0/0 Enable-
        Capabilities: [60] Express Endpoint IRQ 0
        Capabilities: [100] Device Serial Number 00-00-00-00-00-00-00-00

> > might not be feasible.  It also looks a previous patch by Alex Chiang
> > completely changed the sysfs interface for fakephp.  I thought sysfs
> > interfaces were supposed to be stable?!  Also looks like it made fakephp
>
> I doubt that, sysfs interfaces change all the time.  I think only the
> syscall interface has any sort of stability guarantee.

The ones meant to be used from userspace usually don't.  At least that
doesn't seem to be what Linus is saying here http://lwn.net/Articles/172986/

I've written software that uses an established interface, which has been
the same for years, and I see someone went and broke it, no warning.  That
hardly seems reasonable.

> > useless.  How are you supposed to figure out which "fake-n" directory is
> > the right one to disable the device you want?
>
> cat /sys/bus/pci/slots/fake*/address

Ahh, my kernel doesn't have an address attribute in the slot directories.

Still, this is much more inefficient than the previous way.  It also has a
race condition.  After scanning each fake* device to find the one with the
correct address there is a window before the power attribute is written.
The fake* device might change in that time and one could write to the wrong
power attribute.