Re: [PATCH v7 4/5] iommu/of: Require reg for iommu-addresses
From: Vishnu Reddy
Date: Wed Nov 18 2026 - 12:07:14 EST
On 11/18/2026 5:57 PM, Krzysztof Kozlowski wrote:
> On 17/11/2026 17:24, Vishnu Reddy wrote:
>> Hi Krzysztof,
>>
>> Thanks for review.
>>
>> On 11/13/2026 6:41 PM, Krzysztof Kozlowski wrote:
>>> On 06/11/2026 18:14, Vikash Garodia wrote:
>>>> The "iommu-addresses" property describes an identity 1:1 mapping between
>>>> an IOVA range and the physical memory specified by "reg" in the same
>>>> reserved-memory node. Without "reg", there is no physical memory to map.
>>> This looks like ABI break.
>> The original discussion was to avoid keeping iommu-addresses in the
>> reserved-memory, if not backed by reg. In such cases, it is more meaningful
>> to move that entry to corresponding device node.
>> Given this direction, the schema was defined accordingly and the code is now
>> updated to match the dt-schema. At least on Linux, I do not see any client
>> using iommu-addresses without reg.
> Yeah, but ABI is for all users, not only in-tree.
To avoid an ABI break, it would be better to preserve the existing
iommu-addresses behavior and avoid this change.
>
>
>>>> The DT schema now requires "reg" whenever "iommu-addresses" is present.
>>>> Align the kernel implementation with this requirement.
>>>> https://github.com/devicetree-org/dt-schema/pull/207
>>>>
>>>> An identity mapping requires matching IOVA and "reg" ranges. Reject
>>>> mismatched ranges and report an error.
>>>>
>>>> Suggested-by: Rob Herring <robh@xxxxxxxxxx>
>>>> Co-developed-by: Vishnu Reddy <busanna.reddy@xxxxxxxxxxxxxxxx>
>>>> Signed-off-by: Vishnu Reddy <busanna.reddy@xxxxxxxxxxxxxxxx>
>>>> Signed-off-by: Vikash Garodia <vikash.garodia@xxxxxxxxxxxxxxxx>
>>>> ---
>>>> drivers/iommu/of_iommu.c | 32 +++++++++++++++-----------------
>>>> 1 file changed, 15 insertions(+), 17 deletions(-)
>>>>
>>>> diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
>>>> index 6675fc900e6c2d55b34c61dc9effaeb48e24f18f..00b02c19f27e195fc36c4bc1d91253255147c9ab 100644
>>>> --- a/drivers/iommu/of_iommu.c
>>>> +++ b/drivers/iommu/of_iommu.c
>>>> @@ -174,19 +174,12 @@ iommu_resv_region_get_type(struct device *dev,
>>>> {
>>>> phys_addr_t end = start + length - 1;
>>>>
>>>> - /*
>>>> - * IOMMU regions without an associated physical region cannot be
>>>> - * mapped and are simply reservations.
>>>> - */
>>>> - if (phys->start >= phys->end)
>>>> - return IOMMU_RESV_RESERVED;
>>>> -
>>>> /* may be IOMMU_RESV_DIRECT_RELAXABLE for certain cases */
>>>> if (start == phys->start && end == phys->end)
>>>> return IOMMU_RESV_DIRECT;
>>>>
>>>> - dev_warn(dev, "treating non-direct mapping [%pr] -> [%pap-%pap] as reservation\n", phys,
>>>> - &start, &end);
>>>> + dev_err(dev, "non-direct mapping [%pr] -> [%pap-%pap] is not supported\n", phys,
>>>> + &start, &end);
>>>> return IOMMU_RESV_RESERVED;
>>>> }
>>>>
>>>> @@ -211,12 +204,14 @@ void of_iommu_get_resv_regions(struct device *dev, struct list_head *list)
>>>> struct resource phys;
>>>> int size;
>>>>
>>>> - memset(&phys, 0, sizeof(phys));
>>>> + maps = of_get_property(it.node, "iommu-addresses", &size);
>>>> + if (!maps)
>>>> + continue;
>>>>
>>>> /*
>>>> - * The "reg" property is optional and can be omitted by reserved-memory regions
>>>> - * that represent reservations in the IOVA space, which are regions that should
>>>> - * not be mapped.
>>>> + * "iommu-addresses" must be used in combination with a "reg" that provides
>>>> + * the physical address and size of this memory region, for an identity 1:1
>>>> + * IOVA mapping to that physical memory.
>>>> */
>>>> if (of_property_present(it.node, "reg")) {
>>>> err = of_address_to_resource(it.node, 0, &phys);
>>>> @@ -225,11 +220,11 @@ void of_iommu_get_resv_regions(struct device *dev, struct list_head *list)
>>>> it.node, err);
>>>> continue;
>>>> }
>>>> - }
>>>> -
>>>> - maps = of_get_property(it.node, "iommu-addresses", &size);
>>>> - if (!maps)
>>> I miss the point why you are moving this part. Irrespective to actual
>>> ABI break, why do you nee to check for iommu-addresses before?
>> It is more optimal to check for iommu-addresses and then read the reg. On the
>> contrary, earlier reg is read irrespective of iommu-addresses is present or
>> not. So in a way reading iommu-addresses first would ensure that reg is read
>> only if iommu-addresses is present.
> Sorry, I don't follow. What is here on the contrary? To what part?
The intention was only to avoid parsing reg for nodes that do not use
iommu-addresses. Given the ABI concerns, we can keep the existing behavior.
>
> Best regards,
> Krzysztof