Re: [PATCH v2] sparc: Fix page alignment in dma mapping

On 18.02.2026 12:09, Stian Halseth wrote:
> On Wed, 2026-02-18 at 11:28 +0100, Marek Szyprowski wrote:
>> On 17.02.2026 18:10, [email protected] wrote:
>>> From: Stian Halseth <[email protected]>
>>>
>>> 'phys' may include an offset within the page, while 'base_paddr' is
>>> already
>>> page-aligned. This caused incorrect DMA mapping in dma_4u_map_phys
>>> and
>>> dma_4v_map_phys.
>>>
>>> Fix both functions by masking phys with IO_PAGE_MASK or subtracting
>>> the
>>> page offset, covering both generic SPARC code and sun4v.
>>>
>>> Fixes: 38c0d0ebf520 ("sparc: Use physical address DMA mapping")
>>> Reported-by: Stian Halseth <[email protected]>
>>> Closes: https://protect2.fireeye.com/v1/url?k=4d8fbc00-1214850c-
>>> 4d8e374f-000babff3563-ee1be24be576e072&q=1&e=0f8967ce-f558-4339-
>>> bddb-
>>> f324ec46c035&u=https%3A%2F%2Fgithub.com%2Fsparclinux%2Fissues%2Fiss
>>> ues%2F75
>>> Suggested-by: Marek Szyprowski <[email protected]>
>>> Signed-off-by: Stian Halseth <[email protected]>
>>> ---
>>>    arch/sparc/kernel/iommu.c     | 2 ++
>>>    arch/sparc/kernel/pci_sun4v.c | 2 +-
>>>    2 files changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/arch/sparc/kernel/iommu.c b/arch/sparc/kernel/iommu.c
>>> index 46ef88bc9c26..f3755a388ac7 100644
>>> --- a/arch/sparc/kernel/iommu.c
>>> +++ b/arch/sparc/kernel/iommu.c
>>> @@ -312,6 +312,8 @@ static dma_addr_t dma_4u_map_phys(struct device
>>> *dev, phys_addr_t phys,
>>>     if (direction != DMA_TO_DEVICE)
>>>     iopte_protection |= IOPTE_WRITE;
>>>    
>>> + phys -= offset_in_page(phys);
>>> +
>> I've just noticed that IO_PAGE_SIZE might not equal to PAGE_SIZE (not
>> sure if this is true in case of SPARCs), so it would be better to
>> rely
>> on IO_PAGE_MASK or IO_PAGE_SIZE only. Just unify the fix for the both
>> affected functions either by masking phys with IO_PAGE_SIZE where it
>> is
>> used or by subtracting (phys & ~IO_PAGE_MASK) from it.
>>
> Good point, and thanks for the review.
>
> I failed to consider that the offset_in_page is based on PAGE_SIZE, not
> IO_PAGE_SIZE.
>
> Suggestion, subtract IO_PAGE_MASK in both functions.
>
> arch/sparc/kernel/iommu.c:
>   -phys -= offset_in_page(phys)
>   +phys &= IO_PAGE_MASK;
>
>
> arch/sparc/kernel/pci_sun4v.c (better to subtract before loop?):
>   +phys &= IO_PAGE_MASK;
>
>   for (i = 0; i < npages; i++, phys += IO_PAGE_SIZE) {
>           -long err = iommu_batch_add(phys & IO_PAGE_MASK, mask);
>           +long err = iommu_batch_add(phys, mask);
>           if (unlikely(err < 0L))
>                   goto iommu_map_fail;
>   }
>
> Is this acceptable, and in line with kernel coding style?

Yes, that's fine.

> If yes, I can submit patch v3 with this proposed change.
>
> And Patch v3, should be sent using the message ID of this review
> response, right? Still learning :)

Usually the next version is being sent as a new thread. Just add a few 
lines of history and links to the previous attempt(s) to it, like for 
example here: 
https://lore.kernel.org/all/[email protected]/

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland