Re: [PATCH v7 08/21] iommu/dma: support PCI P2PDMA pages in dma-iommu map_sg

From: Logan Gunthorpe
Date: Wed Jun 29 2022 - 11:57:48 EST





On 2022-06-29 06:07, Robin Murphy wrote:
> On 2022-06-15 17:12, Logan Gunthorpe wrote:
>> When a PCI P2PDMA page is seen, set the IOVA length of the segment
>> to zero so that it is not mapped into the IOVA. Then, in finalise_sg(),
>> apply the appropriate bus address to the segment. The IOVA is not
>> created if the scatterlist only consists of P2PDMA pages.
>>
>> A P2PDMA page may have three possible outcomes when being mapped:
>>    1) If the data path between the two devices doesn't go through
>>       the root port, then it should be mapped with a PCI bus address
>>    2) If the data path goes through the host bridge, it should be mapped
>>       normally with an IOMMU IOVA.
>>    3) It is not possible for the two devices to communicate and thus
>>       the mapping operation should fail (and it will return -EREMOTEIO).
>>
>> Similar to dma-direct, the sg_dma_mark_pci_p2pdma() flag is used to
>> indicate bus address segments. On unmap, P2PDMA segments are skipped
>> over when determining the start and end IOVA addresses.
>>
>> With this change, the flags variable in the dma_map_ops is set to
>> DMA_F_PCI_P2PDMA_SUPPORTED to indicate support for P2PDMA pages.
>>
>> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
>> Reviewed-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
>> ---
>>   drivers/iommu/dma-iommu.c | 68 +++++++++++++++++++++++++++++++++++----
>>   1 file changed, 61 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>> index f90251572a5d..b01ca0c6a7ab 100644
>> --- a/drivers/iommu/dma-iommu.c
>> +++ b/drivers/iommu/dma-iommu.c
>> @@ -21,6 +21,7 @@
>>   #include <linux/iova.h>
>>   #include <linux/irq.h>
>>   #include <linux/list_sort.h>
>> +#include <linux/memremap.h>
>>   #include <linux/mm.h>
>>   #include <linux/mutex.h>
>>   #include <linux/pci.h>
>> @@ -1062,6 +1063,16 @@ static int __finalise_sg(struct device *dev,
>> struct scatterlist *sg, int nents,
>>           sg_dma_address(s) = DMA_MAPPING_ERROR;
>>           sg_dma_len(s) = 0;
>>   +        if (is_pci_p2pdma_page(sg_page(s)) && !s_iova_len) {
>
> Logically, should we not be able to use sg_is_dma_bus_address() here? I
> think it should be feasible, and simpler, to prepare the p2p segments
> up-front, such that at this point all we need to do is restore the
> original length (if even that, see below).

Per my previous email, no, because sg_is_dma_bus_address() is not set
yet and not meant to tell you something about the page. That flag will
be set below by pci_p2pdma_map_bus_segment() and then checkd in
iommu_dma_unmap_sg() to determine if the dma_address in the segment
needs to be unmapped.

>
>> +            if (i > 0)
>> +                cur = sg_next(cur);
>> +
>> +            pci_p2pdma_map_bus_segment(s, cur);
>> +            count++;
>> +            cur_len = 0;
>> +            continue;
>> +        }
>> +
>>           /*
>>            * Now fill in the real DMA data. If...
>>            * - there is a valid output segment to append to
>> @@ -1158,6 +1169,8 @@ static int iommu_dma_map_sg(struct device *dev,
>> struct scatterlist *sg,
>>       struct iova_domain *iovad = &cookie->iovad;
>>       struct scatterlist *s, *prev = NULL;
>>       int prot = dma_info_to_prot(dir, dev_is_dma_coherent(dev), attrs);
>> +    struct dev_pagemap *pgmap = NULL;
>> +    enum pci_p2pdma_map_type map_type;
>>       dma_addr_t iova;
>>       size_t iova_len = 0;
>>       unsigned long mask = dma_get_seg_boundary(dev);
>> @@ -1193,6 +1206,35 @@ static int iommu_dma_map_sg(struct device *dev,
>> struct scatterlist *sg,
>>           s_length = iova_align(iovad, s_length + s_iova_off);
>>           s->length = s_length;
>>   +        if (is_pci_p2pdma_page(sg_page(s))) {
>> +            if (sg_page(s)->pgmap != pgmap) {
>> +                pgmap = sg_page(s)->pgmap;
>> +                map_type = pci_p2pdma_map_type(pgmap, dev);
>> +            }
>
> There's a definite code smell here, but per above and below I think we
> *should* actually call the new helper instead of copy-pasting half of it.


>
>> +
>> +            switch (map_type) {
>> +            case PCI_P2PDMA_MAP_BUS_ADDR:
>> +                /*
>> +                 * A zero length will be ignored by
>> +                 * iommu_map_sg() and then can be detected
>
> If that is required behaviour then it needs an explicit check in
> iommu_map_sg() to guarantee (and document) it. It's only by chance that
> __iommu_map() happens to return success for size == 0 *if* all the other
> arguments still line up, which is a far cry from a safe no-op.

What should such a check look like? I could certainly add some comments
to iommu_map_sg(), but I don't see what the code would need to check for...

> However, rather than play yet more silly tricks, I think it would make
> even more sense to make iommu_map_sg() properly aware and able to skip
> direct p2p segments on its own. Once it becomes normal to pass mixed
> scatterlists around, it's only a matter of time until one ends up being
> handed to a driver which manages its own IOMMU domain, and then what?

I suppose we can add another call to is_pci_p2pdma_page() inside
iommu_map_sg() if you think that is cleaner. Seems like more work on the
fast path to me, but I'm not opposed to it.

>> +                 * in __finalise_sg() to actually map the
>> +                 * bus address.
>> +                 */
>> +                s->length = 0;
>> +                continue;
>> +            case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
>> +                /*
>> +                 * Mapping through host bridge should be
>> +                 * mapped with regular IOVAs, thus we
>> +                 * do nothing here and continue below.
>> +                 */
>> +                break;
>> +            default:
>> +                ret = -EREMOTEIO;
>> +                goto out_restore_sg;
>> +            }
>> +        }
>> +
>>           /*
>>            * Due to the alignment of our single IOVA allocation, we can
>>            * depend on these assumptions about the segment boundary mask:
>> @@ -1215,6 +1257,9 @@ static int iommu_dma_map_sg(struct device *dev,
>> struct scatterlist *sg,
>>           prev = s;
>>       }
>>   +    if (!iova_len)
>> +        return __finalise_sg(dev, sg, nents, 0);
>> +
>>       iova = iommu_dma_alloc_iova(domain, iova_len, dma_get_mask(dev),
>> dev);
>>       if (!iova) {
>>           ret = -ENOMEM;
>> @@ -1236,7 +1281,7 @@ static int iommu_dma_map_sg(struct device *dev,
>> struct scatterlist *sg,
>>   out_restore_sg:
>>       __invalidate_sg(sg, nents);
>>   out:
>> -    if (ret != -ENOMEM)
>> +    if (ret != -ENOMEM && ret != -EREMOTEIO)
>>           return -EINVAL;
>>       return ret;
>>   }
>> @@ -1244,7 +1289,7 @@ static int iommu_dma_map_sg(struct device *dev,
>> struct scatterlist *sg,
>>   static void iommu_dma_unmap_sg(struct device *dev, struct
>> scatterlist *sg,
>>           int nents, enum dma_data_direction dir, unsigned long attrs)
>>   {
>> -    dma_addr_t start, end;
>> +    dma_addr_t end, start = DMA_MAPPING_ERROR;
>
> There are several things I don't like about this logic, I'd rather have
> "end = 0" here...

Ok, I think that should work.

>>       struct scatterlist *tmp;
>>       int i;
>>   @@ -1260,14 +1305,22 @@ static void iommu_dma_unmap_sg(struct device
>> *dev, struct scatterlist *sg,
>>        * The scatterlist segments are mapped into a single
>>        * contiguous IOVA allocation, so this is incredibly easy.
>>        */
>
> [ This comment rather stops being true :( ]

Not exactly. Sure there are some segments in the SGL that have bus
addresses, but all the regular IOVAs still have a single contiguous
allocation and only require one call to __iommu_dma_unmap(). The only
trick issues is finding the first and last actual IOVA SG to get the range.

>
>> -    start = sg_dma_address(sg);
>> -    for_each_sg(sg_next(sg), tmp, nents - 1, i) {
>
> ...then generalise the first-element special case here into a dedicated
> "walk to the first non-p2p element" loop...

Ok, I'll see what I can do for that.

Logan