[PATCH v3 2/4] PCI: of: Relax the condition for warning about non-prefetchable memory aperture size

Punit Agrawal punitagrawal at gmail.com
Thu Jun 10 07:11:10 PDT 2021


Hi Bjorn,

Bjorn Helgaas <helgaas at kernel.org> writes:

> On Wed, Jun 09, 2021 at 12:36:08AM +0530, Vidya Sagar wrote:
>> On 6/7/2021 4:58 PM, Punit Agrawal wrote:
>> > 
>> > Commit fede8526cc48 ("PCI: of: Warn if non-prefetchable memory
>> > aperture size is > 32-bit") introduced a warning for non-prefetchable
>> > resources that need more than 32bits to resolve. It turns out that the
>> > check is too restrictive and should be applicable to only resources
>> > that are limited to host bridge windows that don't have the ability to
>> > map 64-bit address space.
>>
>> I think the host bridge windows having the ability to map 64-bit address
>> space is different from restricting the non-prefetchable memory aperture
>> size to 32-bit.
>
>> Whether the host bridge uses internal translations or not to map the
>> non-prefetchable resources to 64-bit space, the size needs to be programmed
>> in the host bridge's 'Memory Limit Register (Offset 22h)' which can
>> represent sizes only fit into 32-bits.
>
>> Host bridges having the ability to map 64-bit address spaces gives
>> flexibility to utilize the vast 64-bit space for the (restrictive)
>> non-prefetchable memory (i.e. mapping non-prefetchable BARs of endpoints to
>> the 64-bit space in CPU's view) and get it translated internally and put a
>> 32-bit address on the PCIe bus finally.
>
> The vastness of the 64-bit space in the CPU view only helps with
> non-prefetchable memory if you have multiple host bridges with
> different CPU-to-PCI translations.  Each root bus can only carve up
> 4GB of PCI memory space for use by its non-prefetchable memory
> windows.
>
> Of course, if we're willing to give up the performance, there's
> nothing to prevent us from using non-prefetchable space for
> *prefetchable* resources, as in my example below.
>
> I think the fede8526cc48 commit log is incorrect, or at least
> incomplete:
>
>   As per PCIe spec r5.0, sec 7.5.1.3.8 only 32-bit BAR registers are defined
>   for non-prefetchable memory and hence a warning should be reported when
>   the size of them go beyond 32-bits.
>
> 7.5.1.3.8 is talking about non-prefetchable PCI-to-PCI bridge windows,
> not BARs.  AFAIK, 64-bit BARs may be non-prefetchable.  The warning is
> in pci_parse_request_of_pci_ranges(), which isn't looking at
> PCI-to-PCI bridge windows; it's looking at PCI host bridge windows.
> It's legal for a host bridge to have only non-prefetchable windows,
> and prefetchable PCI BARs can be placed in them.
>
> For example, we could have the following:
>
>   pci_bus 0000:00: root bus resource [mem 0x80000000-0x1_ffffffff] (6GB)
>   pci 0000:00:00.0: PCI bridge to [bus 01-7f]
>   pci 0000:00:00.0:   bridge window [mem 0x80000000-0xbfffffff] (1GB)
>   pci 0000:00:00.0:   bridge window [mem 0x1_00000000-0x1_7fffffff 64bit pref] (2GB)
>   pci 0000:00:00.1: PCI bridge to [bus 80-ff]
>   pci 0000:00:00.1:   bridge window [mem 0xc0000000-0xffffffff] (1GB)
>   pci 0000:00:00.1:   bridge window [mem 0x1_80000000-0x1_ffffffff 64bit pref] (2GB)
>
> Here the host bridge window is 6GB and is not prefetchable.  The
> PCI-to-PCI bridge non-prefetchable windows are 1GB each and the bases
> and limits fit in 32 bits.  The prefetchable windows are 2GB each, and
> we're allowed but not required to put these in prefetchable host
> bridge windows.
>
> So I'm not convinced this warning is valid to begin with.  It may be
> that this host bridge configuration isn't optimal, and we might want
> an informational message, but I think it's *legal*.

By "optimal" - are you referring to the use of non-prefetchable space
for prefetchable window?

Also, if the warning doesn't apply to PCI host bridge windows, should I
drop it in the next update? Or leave out this and the next patch to be
dealt with separately.

Thanks,
Punit

[...]




More information about the Linux-rockchip mailing list