Re: [PATCH v1] vfio/pci: Avoid mapping BARs for devices with non-mappable BARs

[Farhan Ali <alifm@linux.ibm.com>]

On 8/3/2026 12:54 PM, Alex Williamson wrote:
> On Mon, 3 Aug 2026 09:39:19 -0700
> Farhan Ali <alifm@linux.ibm.com> wrote:
>
>> On 7/29/2026 2:50 PM, Alex Williamson wrote:
>>> On Wed, 29 Jul 2026 14:32:45 -0700
>>> Farhan Ali <alifm@linux.ibm.com> wrote:
>>>   
>>>> On 7/29/2026 1:36 PM, Alex Williamson wrote:
>>>>> On Wed, 29 Jul 2026 13:28:46 -0700
>>>>> Farhan Ali <alifm@linux.ibm.com> wrote:
>>>>>      
>>>>>> On 7/29/2026 12:41 PM, Matthew Rosato wrote:
>>>>>>> On 7/29/26 2:11 PM, Farhan Ali wrote:
>>>>>>>> vfio_pci_core_map_bars() calls pci_iomap() to set up BAR resources, but not
>>>>>>>> all devices support having their BARs mapped by the CPU. The
>>>>>>>> non_mappable_bars flag indicates that a PCI device's BARs cannot be
>>>>>>>> accessed by the CPU. The ISM device on s390 is one such device. The BAR
>>>>>>>> size for an ISM device is 256 TiB, and attempting to map the BAR will lead
>>>>>>>> to warnings:
>>>>>>>>
>>>>>>>> vmalloc_node_range for size 281474976714752 failed: Address range
>>>>>>>> restricted to 0x2110bab00000 - 0x21903ab00000
>>>>>>>>
>>>>>>>> Use pdev->non_mappable_bars to skip pci_iomap() for such devices. This flag
>>>>>>>> is set by the PCI core at enumeration time and already serves the same
>>>>>>>> purpose in vfio_pci_probe_mmaps().
>>>>>>>>
>>>>>>>> Fixes: 05f2a68b407a ("vfio/pci: Set up BAR resources and maps in vfio_pci_core_enable()")
>>>>>>>> Reported-by: Christian Borntraeger <borntraeger@linux.ibm.com>
>>>>>>>> Signed-off-by: Farhan Ali <alifm@linux.ibm.com>
>>>>>>>> ---
>>>>>>>>      drivers/vfio/pci/vfio_pci_core.c | 3 +++
>>>>>>>>      1 file changed, 3 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/drivers/vfio/pci/vfio_pci_core.c b/drivers/vfio/pci/vfio_pci_core.c
>>>>>>>> index 3f11a9624b9c..6a184588ff23 100644
>>>>>>>> --- a/drivers/vfio/pci/vfio_pci_core.c
>>>>>>>> +++ b/drivers/vfio/pci/vfio_pci_core.c
>>>>>>>> @@ -554,6 +554,9 @@ static void vfio_pci_core_map_bars(struct vfio_pci_core_device *vdev)
>>>>>>>>      
>>>>>>>>      		vdev->barmap[bar] = IOMEM_ERR_PTR(-ENODEV);
>>>>>>>>      
>>>>>>>> +		if (pdev->non_mappable_bars)
>>>>>>>> +			continue;
>>>>>>>> +
>>>>>>> This would work for the ISM case at least, but I wonder: should we check
>>>>>>> vdev->bar_mmap_supported[bar] instead?
>>>>>>>
>>>>>>> My question boils down to: do we still want messages for some of the
>>>>>>> cases where we set vdev->bar_mmap_supported[bar] = false in
>>>>>>> vfio_pci_probe_mmaps()?
>>>>>> AFAIU vfio_pci_probe_mmaps() is only called in
>>>>>> vfio_pci_core_finish_enable(). Since vfio_pci_core_map_bars() is called
>>>>>> in vfio_pci_core_enable(), and before vfio_pci_core_finish_enable(),
>>>>>> bar_mmap_supported would be false here for all devices. So I don't think
>>>>>> it would work here, unless I missed something?
>>>>> Also IO Port and sub-page MMIO BARs are things that do exist.  Thanks,
>>>> Just to clarify, are you suggesting we expand this check to also avoid
>>>> mapping IO port and sub-page MMIO BARs?
>>> Sorry, no, I'm not.  I think we're conflating that the barmap is
>>> related to mmap access.  The barmap itself is holding the iomap of the
>>> BAR, used for read/write.  The only real relation to the mmap is that
>>> we request the resource via this path as well.
>>>
>>> Therefore not only is the ordering of setting up bar_mmap_supported
>>> wrong, it's flagging entirely the wrong thing here and keying on it
>>> would entirely break IO port and sub-page MMIO BAR access.  Thanks,
>>>
>>> Alex
>> Hi Alex,
>>
>> I wanted some guidance on how we should proceed with this patch? The
>> warning messages are a regression on s390 for ISM devices, so we would
>> like to fix it.
> I think the original proposal is probably the correct one.  The
> non_mmapable_bars flag doesn't restrict its application to specific BAR
> types or access, at least not beyond "can't be mapped to CPU or peers."
>
> If we can't map the BAR to the CPU, then we don't need to request the
> region or perform the pci_iomap(), which is what I understand explodes
> here.  Therefore we really only need to establish the errno in the
> barmap here.

Yes, its the call to pci_iomap() that leads to the warning.

>
> The bar_mmap_supported flag describes something else, whether the BAR
> can be mapped into the user address space.  My intention was only to
> point out that there are BARs that cannot be mapped to the user address
> space because either they're not MMIO or we can't safely map the full
> page, therefore bar_mmap_supported is an invalid test for whether we
> should request the region or iomap the BAR.
>
> Is there still a gap with the original proposal that I'm missing?

AFAICT this fix should be sufficient. I think Matt's question was to see 
if there were other cases for which bar_mmap_supported were set to 
false, that we need to consider here. But I will let him clarify to see 
if he has any concerns.

Thanks

Farhan