Skip to contentSkip to conversation

IOMMU/DMA API inquiry

DISCUSSION Mark Hounschell 2 participants 3 messages Feb 17, 2015 – Feb 19, 2015

lore
Share & tools
lore
#0 lore raw

I've searched the Doc tree and web to no avail. I was hoping I might get some answers to a couple of questions that have arisen as a result of actually trying to take advantage of the IOMMU in our out of tree GPL drivers. The AMD IOMMU in particular. I'm currently using the 3.18.x kernel versions. Both 32 and 64 bit.

I understand that AMD IOMMU support is not available for 32-bit kernels.
I believe the INTEL IOMMU is supported there. Not knowing why, I was
curious if that is going to remain that way?

I've learned that the AMD IOMMU does not play well with the kernels "Contiguous Memory Allocator" (CMA). I also believe, but could be mistaken, that the INTEL IOMMU does. Again, not knowing why, I was curious if that is going to remain that way also?

Is the fact that the AMD IOMMU is not supported on 32 bit kernels the reason that dma_map_page always returns 0 on 32 bit kernels. I found this strange because the dma_alloc_coherent function returns a page_to_phys address in the dma_handle when no IOMMU is enabled. I was sort of expecting dma_map_page to do the same. Other wise I seem to have to KNOW if an enabled and supported IOMMU is present. How does a driver tell if an IOMMU is present? I suspect it should not have to?

I've read the DMA-API-HOWTO.txt concerning dma_map_sg and at first I thought that maybe dma_map_sg could be used to get around the fact that AMD IOMMU doesn't work with CMA. But it looks as though I was mistaken and I would actually have to do a DMA for_each_sg(sglist, sg, count, i). Is that correct or can dma_map_sg somehow enable you to do a single DMA using a single address for the entire sglist?

Sorry if these questions seem ignorant. It was not intentional.

Thanks for any insight.

Regards
Mark

#1 lore raw

Hi Mark,

Quoted 3 lines · Mark Hounschell

I understand that AMD IOMMU support is not available for 32-bit kernels. I believe the INTEL IOMMU is supported there. Not knowing why, I was curious if that is going to remain that way?

Yes, I have no plan on making the AMD IOMMU driver available on 32bit. But I would not be resistant to patches enabling the driver there.

Quoted 4 lines

I've learned that the AMD IOMMU does not play well with the kernels "Contiguous Memory Allocator" (CMA). I also believe, but could be mistaken, that the INTEL IOMMU does. Again, not knowing why, I was curious if that is going to remain that way also?

No, I will queue a patch for the next merge window to enable CMA use in the AMD IOMMU driver. So there will be support for this.

Quoted 2 lines

Is the fact that the AMD IOMMU is not supported on 32 bit kernels the reason that dma_map_page always returns 0 on 32 bit kernels.

There is no particular reason for not supporting it on 32 bit kernels, it just didn't seem to be important yet. At least not important enough to justify the work.

Quoted 6 lines

I've read the DMA-API-HOWTO.txt concerning dma_map_sg and at first I thought that maybe dma_map_sg could be used to get around the fact that AMD IOMMU doesn't work with CMA. But it looks as though I was mistaken and I would actually have to do a DMA for_each_sg(sglist, sg, count, i). Is that correct or can dma_map_sg somehow enable you to do a single DMA using a single address for the entire sglist?

The map_sg functions can't be used with the AMD IOMMU driver to work around missing CMA support. Depending on what you want it might work with the Intel IOMMU driver, as this one allocates a single IOVA region for the entire sg_list.

Joerg

#2 lore raw

Hi Joerg, Thanks for the response.

Quoted 10 lines · On 02/18/2015 01:19 PM, [email protected] >> Joerg Roedel wrote:

Hi Mark,

Quoted 3 lines · Mark Hounschell

I understand that AMD IOMMU support is not available for 32-bit kernels. I believe the INTEL IOMMU is supported there. Not knowing why, I was curious if that is going to remain that way?

Yes, I have no plan on making the AMD IOMMU driver available on 32bit. But I would not be resistant to patches enabling the driver there.

OK. I might take a look. Should it not "just work" more or less?

Quoted 9 lines
Quoted 5 lines

I've learned that the AMD IOMMU does not play well with the kernels "Contiguous Memory Allocator" (CMA). I also believe, but could be mistaken, that the INTEL IOMMU does. Again, not knowing why, I was curious if that is going to remain that way also?

No, I will queue a patch for the next merge window to enable CMA use in the AMD IOMMU driver. So there will be support for this.

Great! That's a couple of kernels away then. I would be happy to test if you have a need.

Quoted 8 lines
Quoted 3 lines

Is the fact that the AMD IOMMU is not supported on 32 bit kernels the reason that dma_map_page always returns 0 on 32 bit kernels.

There is no particular reason for not supporting it on 32 bit kernels, it just didn't seem to be important yet. At least not important enough to justify the work.

Understood.

Quoted 7 lines
Quoted 6 lines

I've read the DMA-API-HOWTO.txt concerning dma_map_sg and at first I thought that maybe dma_map_sg could be used to get around the fact that AMD IOMMU doesn't work with CMA. But it looks as though I was mistaken and I would actually have to do a DMA for_each_sg(sglist, sg, count, i). Is that correct or can dma_map_sg somehow enable you to do a single DMA using a single address for the entire sglist?

Quoted 5 lines

The map_sg functions can't be used with the AMD IOMMU driver to work around missing CMA support. Depending on what you want it might work with the Intel IOMMU driver, as this one allocates a single IOVA region for the entire sg_list.

Thanks Joerg

Regards
Mark