Re: [PATCHv3 0/7] dma mapping optimisations

[Keith Busch <kbusch@kernel.org>]

On Thu, Aug 11, 2022 at 09:22:32AM +0200, Christoph Hellwig wrote:
> On Wed, Aug 10, 2022 at 12:05:05PM -0600, Keith Busch wrote:
> > The functions are implemented under 'include/linux/', indistinguishable from
> > exported APIs. I think I understand why they are there, but they look the same
> > as exported functions from a driver perspective.
> 
> swiotlb.h is not a driver API.  There's two leftovers used by the drm
> code I'm trying to get fixed up, but in general the DMA API is the
> interface and swiotlb is just an implementation detail.
> 
> > Perhaps I'm being daft, but I'm totally missing why I should care if swiotlb
> > leverages this feature. If you're using that, you've traded performance for
> > security or compatibility already. If this idea can be used to make it perform
> > better, then great, but that shouldn't be the reason to hold this up IMO.
> 
> We firstly need to make sure that everything actually works on swiotlb, or
> any other implementation that properly implements the DMA API.
> 
> And the fact that I/O performance currently sucks and we can fix it on
> the trusted hypervisor is an important consideration.  At least as
> importantant as micro-optimizing performance a little more on setups
> not using them.  So not taking care of both in one go seems rather silly
> for a feature that is in its current form pretty intrusive and thus needs
> a really good justification.

Sorry for the delay response; I had some trouble with test setup.

Okay, I will restart developing this with swiotlb in mind.

In the mean time, I wanted to share some results with this series because I'm
thinking this might be past the threshold for when we can drop the "micro-"
prefix on optimisations.

The most significant data points are these:

  * submission latency stays the same regardless of the transfer size or depth
  * IOPs is always equal or better (usually better) with up to 50% reduced
    cpu cost

Based on this, I do think this type of optimisation is worth having a something
like a new bio type. I know this introduces some complications in the io-path,
but it is pretty minimal and doesn't add any size penalties to common structs
for drivers that don't use them.

Test details:

  fio with ioengine=io_uring
    'none': using __user void*
    'bvec': using buf registered with IORING_REGISTER_BUFFERS
    'dma': using buf registered with IORING_REGISTER_MAP_BUFFERS (new)

  intel_iommu=on

Results:

(submission latency [slat] in nano-seconds)
Q-Depth 1:

 Size |  Premap  |   IOPs  |  slat  | sys-cpu%
 .....|..........|.........|........|.........
 4k   |    none  |  41.4k  |  2126  |   16.47%
      |    bvec  |  43.8k  |  1843  |   15.79%
      |     dma  |  46.8k  |  1504  |   14.94%
 16k  |    none  |  33.3k  |  3279  |   17.78%
      |    bvec  |  33.9k  |  2607  |   14.59%
      |     dma  |  40.2k  |  1490  |   12.57%
 64k  |    none  |  18.7k  |  6778  |   18.22%
      |    bvec  |  20.0k  |  4626  |   13.80%
      |     dma  |  22.6k  |  1586  |    7.58%

Q-Depth 16:

 Size |  Premap  |   IOPs  |  slat  | sys-cpu%
 .....|..........|.........|........|.........
 4k   |    none  |   207k  |  3657  |   72.81%
      |    bvec  |   219k  |  3369  |   71.55%
      |     dma  |   310k  |  2237  |   60.16%
 16k  |    none  |   164k  |  5024  |   78.38%
      |    bvec  |   177k  |  4553  |   76.29%
      |     dma  |   186k  |  1880  |   43.56%
 64k  |    none  |  46.7k  |  4424  |   30.51%
      |    bvec  |  46.7k  |  4389  |   29.42%
      |     dma  |  46.7k  |  1574  |   15.61%