[RFC/PATCH 2/7] iommu-api: Add map_range/unmap_range functions

Rob Clark robdclark at gmail.com
Tue Jul 8 16:49:59 PDT 2014


On Tue, Jul 8, 2014 at 5:53 PM, Olav Haugan <ohaugan at codeaurora.org> wrote:
> Hi Hiroshi,
>
> On 7/3/2014 9:29 PM, Hiroshi Doyu wrote:
>> Hi Olav,
>>
>> Olav Haugan <ohaugan at codeaurora.org> writes:
>>
>>> Mapping and unmapping are more often than not in the critical path.
>>> map_range and unmap_range allows SMMU driver implementations to optimize
>>> the process of mapping and unmapping buffers into the SMMU page tables.
>>> Instead of mapping one physical address, do TLB operation (expensive),
>>> mapping, do TLB operation, mapping, do TLB operation the driver can map
>>> a scatter-gatherlist of physically contiguous pages into one virtual
>>> address space and then at the end do one TLB operation.
>>>
>>> Additionally, the mapping operation would be faster in general since
>>> clients does not have to keep calling map API over and over again for
>>> each physically contiguous chunk of memory that needs to be mapped to a
>>> virtually contiguous region.
>>>
>>> Signed-off-by: Olav Haugan <ohaugan at codeaurora.org>
>>> ---
>>>  drivers/iommu/iommu.c | 24 ++++++++++++++++++++++++
>>>  include/linux/iommu.h | 24 ++++++++++++++++++++++++
>>>  2 files changed, 48 insertions(+)
>>>
>>> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
>>> index e5555fc..f2a6b80 100644
>>> --- a/drivers/iommu/iommu.c
>>> +++ b/drivers/iommu/iommu.c
>>> @@ -898,6 +898,30 @@ size_t iommu_unmap(struct iommu_domain *domain, unsigned long iova, size_t size)
>>>  EXPORT_SYMBOL_GPL(iommu_unmap);
>>>
>>>
>>> +int iommu_map_range(struct iommu_domain *domain, unsigned int iova,
>>> +                struct scatterlist *sg, unsigned int len, int prot)
>>> +{
>>> +    if (unlikely(domain->ops->map_range == NULL))
>>> +            return -ENODEV;
>>> +
>>> +    BUG_ON(iova & (~PAGE_MASK));
>>> +
>>> +    return domain->ops->map_range(domain, iova, sg, len, prot);
>>> +}
>>> +EXPORT_SYMBOL_GPL(iommu_map_range);
>>
>> We have the similar one internally, which is named, "iommu_map_sg()",
>> called from DMA API.
>
> Great, so this new API will be useful to more people!
>
>>> +int iommu_unmap_range(struct iommu_domain *domain, unsigned int iova,
>>> +                  unsigned int len)
>>> +{
>>> +    if (unlikely(domain->ops->unmap_range == NULL))
>>> +            return -ENODEV;
>>> +
>>> +    BUG_ON(iova & (~PAGE_MASK));
>>> +
>>> +    return domain->ops->unmap_range(domain, iova, len);
>>> +}
>>> +EXPORT_SYMBOL_GPL(iommu_unmap_range);
>>
>> Can the existing iommu_unmap() do the same?
>
> I believe iommu_unmap() behaves a bit differently because it will keep
> on calling domain->ops->unmap() until everything is unmapped instead of
> letting the iommu implementation take care of unmapping everything in
> one call.
>
> I am abandoning the patch series since our driver was not accepted.
> However, if there are no objections I will resubmit this patch (PATCH
> 2/7) as an independent patch to add this new map_range API.

+1 for map_range().. I've seen for gpu workloads, at least, it is the
downstream map_range() API is quite beneficial.   It was worth at
least a few fps in xonotic.

And, possibly getting off the subject a bit, but I was wondering about
the possibility of going one step further and batching up mapping
and/or unmapping multiple buffers (ranges) at once.  I have a pretty
convenient sync point in drm/msm to flush out multiple mappings before
kicking gpu.

BR,
-R

> Thanks,
>
> Olav Haugan
>
> --
> The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> hosted by The Linux Foundation
> --
> To unsubscribe from this list: send the line "unsubscribe linux-arm-msm" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html



More information about the linux-arm-kernel mailing list