[dpdk-dev] [PATCH] vfio: map contiguous areas in one go

[Burakov, Anatoly]

On 25-Feb-20 1:49 PM, Ray Kinsella wrote:
> Hi Anatoly,
> 
> On 25/02/2020 13:24, Anatoly Burakov wrote:
>> Currently, when we are creating DMA mappings for memory that's
>> either external or is backed by hugepages in IOVA as PA mode, we
>> assume that each page is necessarily discontiguous. This may not
>> actually be the case, especially for external memory, where the
>> user is able to create their own IOVA table and make it
>> contiguous. This is a problem because VFIO has a limited number
>> of DMA mappings, and it does not appear to concatenate them and
>> treats each mapping as separate, even when they cover adjacent
>> areas.
>>> Fix this so that we always map contiguous memory in a single
>> chunk, as opposed to mapping each segment separately.
> 
> Can I confirm my understanding.
> 
> We are essentially correcting user errant behavior,
> trading off startup/mapping time to save IOMMU resources?
> 

That's not quite what we're doing.

First of all, in terms of "trading startup/mapping time", i think this 
will actually be faster because the DMA map is the more 
resource-intensive part of this loop by far, and we're doing _less_ of 
those (because we're concatenating). We're also doing the same number of 
loop iterations as before.

To be perfectly clear: we're not reordering segments here - segments 
have to be VA- and IOVA-contiguous in the first place, otherwise we're 
breaking them up. It's just that previously, we were also breaking up 
contiguous segments into separate, per-page mappings, but now we 
concatenate them.

-- 
Thanks,
Anatoly