Re: PCIe 32-bit MMIO exhaustion

[Daniel J Blueman <daniel@numascale.com>]

Hi Bjorn, Jiang,

On 29/01/2015 23:23, Bjorn Helgaas wrote:
> Hi Daniel,
>
> On Wed, Jan 28, 2015 at 2:42 AM, Daniel J Blueman <daniel@numascale.com> wrote:
>> With systems with a large number of PCI devices, we're seeing lack of 32-bit
>> MMIO space, eg one quad-port NetXtreme-2 adapter takes 128MB of space [1].
>>
>> An errata to the PCIe 2.1 spec provides guidance on limitations with 64-bit
>> non-prefetchable BARs (since bridges have only 32-bit non-prefetchable
>> ranges) stating that vendors can enable the prefetchable bit in BARs under
>> certain circumstances to allow 64-bit allocation [2].
>>
>> The problem with that, is that vendors can't know apriori what hosts their
>> products will be in, so can't just advertise prefetchable 64-bit BARs. What
>> can be done, is system firmware can use the 64-bit prefetchable BAR in
>> bridges, and assign a 64-bit non-prefetchable device BAR into that area,
>> where it is safe to do so (following the guidance).
>>
>> At present, linux denies such allocations [3] and disables the BARs. It
>> seems a practical solution to allow them if the firmware believes it is
>> safe.
>
> This particular message ([3]):
>
>> pci 0002:01:00.0: BAR 0: [mem size 0x00002000 64bit] conflicts with PCI Bus
>> 0002:00 [mem 0x10020000000-0x10027ffffff pref]
>
> is misleading at best and likely a symptom of a bug.  We printed the
> *size* of BAR 0, not an address, which means we haven't assigned space
> for the BAR.  That means it should not conflict with anything.
>
> We already do revert to firmware assignments in some situations when
> Linux can't figure out how to assign things itself.  But apparently
> not in *this* situation.
>
> Without seeing the whole picture, it's hard for me to figure out
> what's going on here.  Could you open a bug report at
> http://bugzilla.kernel.org (category drivers/PCI) and attach a
> complete dmesg and "lspci -vv" output?  Then we can look at what
> firmware did and what Linux thought was wrong with it.

Done a while back:
https://bugzilla.kernel.org/show_bug.cgi?id=92671

An interesting question popped up: I find the kernel doesn't accept IO 
BARs and bridge windows after address 0xffff, though the PCI spec and 
modern hardware allows 32-bit decode.

Thus for practical reasons, our NumaConnect firmware doesn't setup IO 
BARs/windows beyond the first PCI domain (which is the only one with 
legacy support, and no drivers seem to require IO their BARs anyway), 
and we get conflicts and warnings [1]:

pnp 00:00: disabling [io  0x0061] because it overlaps 0001:05:00.0 BAR 0 
[io  0x0000-0x00ff]
pci 0001:03:00.0: BAR 13: no space for [io  size 0x1000]
pci 0001:03:00.0: BAR 13: failed to assign [io  size 0x1000]

Is there a cleaner way of dealing with this, in our firmware and/or the 
kernel? Eg, I guess if IO BARs aren't assigned (value 0) on PCI domains 
without IO bridge windows in the ACPI AML, no need to conflict/attempt 
assignment?

Many thanks!
   Daniel

[1] https://bugzilla.kernel.org/attachment.cgi?id=165831
-- 
Daniel J Blueman
Principal Software Engineer, Numascale