[PATCH] iommu/arm-smmu-v3: fix probe device bug due to duplicated stream IDS.

[Robin Murphy]

On 2025-10-23 7:12 pm, Jason Gunthorpe wrote:
> On Sat, Oct 11, 2025 at 10:30:03AM +0800, Reaper Li wrote:
>> From: Reaper <reaperlioc at glenfly.com>
>>
>> Commit 9246b487ab3c ("PCI: Add function 0 DMA alias quirk for Glenfly Arise
>> chip ") add quirk to fix hda dma request issue, but IORT logic populaties
>> two identical IDs into the fwspec->ids array via DMA aliasing in
>> iort_pci_iommu_init() called by pci_for_each_dma_alias().
> 
> I'd rather we not have duplicate IDs in the same fwspec, can we avoid
> that at the source?

The original reason for not doing that is that there are multiple 
sources - the drivers' own .of_xlate routines for DT, but the 
IORT/VIOT/RIMT/whatever code for ACPI. Yes, iommu_fwspec_add_ids() is in 
a common path there, but that's only responsible for appending an opaque 
block of data to another opaque block of data - only the producer or 
consumer of that data know how to interpret it (e.g. for SMMUv2, 
[0x00000002, 0x00000003] may or may not be considered equivalent to 
[0x00010002]).

Anyway, this patch is wrong regardless, and it is definitely not fixing 
any bug. SMMUv3 has explicitly never supported StreamID aliasing between 
*different* devices, because doing that correctly is a challenge. It 
needs custom group assignment based on StreamID-to-group lookup (like 
SMMUv2) - we can't just assume it's OK for PCI devices, since as soon as 
we allow aliasing at all then we also allow it outside the PCI hierarchy 
where pci_device_group() can't see it (like on the Arm Juno platform). 
These days we do have arm_smmu_find_master() which makes *that* 
feasible, but conversely means we now also have to be concerned with 
disallowing ATS/PRI and working out what to do with anything else that 
is subtly broken by no longer having the assumed 1:1 correspondence of 
StreamID:device.

Thanks,
Robin.