Re: Regarding PCIe peek/poke in NetBSD

[Gnanasekar Loganathan <l.gnanasekar%gmail.com@localhost>]

Correct me if my understanding was wrong.
Conclusion is that we cannot map the PCIe regions from the separate
applicaiton which is already mapped by the driver.
The only way is the correspoding device driver should give an ioctl
command to do read/write in their address space.


Thanks,
Gnana


On Tue, Jan 4, 2011 at 9:14 PM, Masao Uebayashi 
<uebayasi%tombi.co.jp@localhost> wrote:
> On Tue, Jan 04, 2011 at 03:19:26PM +0000, Jonathan A. Kollasch wrote:
>> On Tue, Jan 04, 2011 at 01:45:24PM +0100, Manuel Bouyer wrote:
>> > On Tue, Jan 04, 2011 at 10:47:30AM +0530, Gnanasekar Loganathan wrote:
>> > > Hi Manuel,
>> > >
>> > > The requirement was, one of my customer want to do read/write on any
>> > > address/offset from any PCIe devices present in the system. He don't
>> > > want to use the corresponding device drivers, need separate
>> > > application that can map the PCIe device region of any devices by bus
>> > > no, device no and function. The application meant for testing purpose
>> > > only.
>> > >
>> > > Using the 'pcictl' utility i can able to do read/write operations on
>> > > PCIe configuration space, but not able to do in io/mem/prefetchable
>> > > space. ÂAble to map the region by 'bus_space_map', if the region was
>> > > not already mapped by the driver. bus_space_map fails if it was
>> > > already mapped.
>> > >
>> > > Is there any way Âto map the region which was already mapped by the 
>> > > driver?
>
> Mappings to kernel (driver) and userspace are very different. ÂThis
> is not clearly documented, but my understanding is that you can't
> mmap devices that are already mapped to kernel (as unmanaged).
>
>> >
>> > Well, this is not so easy.
>> > For I/O space, I guess they can be acceeded using the i386_iopl(2) call
>> > (assuming you're on x86 hardware). For memory-mapped space, maybe
>> > they can be acceeded using /dev/mem, but I never tried this.
>>
>> Also, you can mmap(4) /dev/pci now ...
>
> Yes, and it's backend is sooooo MD.
>