On Wed, 2023-08-30 at 17:18 +0100, Jonathan Cameron wrote: > On Wed, 30 Aug 2023 17:11:18 +0100 > Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote: > > > On Mon, 7 Aug 2023 13:21:07 +0200 > > Paul Cercueil <paul@xxxxxxxxxxxxxxx> wrote: > > > > > [V3 was: "iio: new DMABUF based API, v3"][1] > > > > > > Hi Jonathan, > > > > > > This is a subset of my patchset that introduced a new interface based on > > > DMABUF objects [1]. It adds write() support to the IIO buffer > > > infrastructure. > > > > > > The reason it is not the full IIO-DMABUF patchset, is because you > > > requested performance benchmarks - and our current numbers are barely > > > better (~ +10%) than the fileio interface. There is a good reason for > > > that: V3 of the patchset switched from having the IIO core creating the > > > DMABUFs backed by physically contiguous memory, to having the IIO core > > > being a simple DMABUF importer, and having the DMABUFs created > > > externally. We now use the udmabuf driver to create those, and they are > > > allocated from paged memory. While this works perfectly fine, our > > > buffers are now cut in 4 KiB chunks (pages), non-contiguous in memory, > > > which causes the DMA hardware to create an IRQ storm, as it raises an > > > interrupt after each 4 KiB in the worst case scenario. > > > > Interesting. I'm guessing you don't necessarily need contiguous memory > > and huge pages would get rid of most of that overhead? > > > > Given embedded target those huge pages are hard to get so you need > > hugetlb support to improve the chances of it working. Some quick searching > > suggests there is possible support on the way. > > https://lore.kernel.org/linux-mm/20230817064623.3424348-1-vivek.kasireddy@xxxxxxxxx/ > > > > > > > > > > Anyway, this is not directly a problem of the IIO-DMABUF code - but I > > > can't really upstream a shiny new interface that I claim is much faster, > > > without giving numbers. > > > > > > So while we fix this (either by updating the DMA IP and driver to > > > support scatter-gather) > > > > Long run you almost always end up needing that unless contig requirements > > are small and you want a robust solution. I'm guessing no IOMMU to pretend > > it's all contiguous... > > > > > or by hacking something quick to give us > > > physically contiguous DMABUFs just for the benchmark), I thought it > > > would make sense to upstream the few patches of the V3 patchset that are > > > needed for the IIO-DMABUF interface but aren't directly related. > > > > Good idea. > > > > > > > > As for write() support, Nuno (Cc'd) said he will work on upstreaming the > > > DAC counterpart of adc/adi-axi-adc.c in the next few weeks, so there > > > will be a user for the buffer write() support. I hope you are okay with > > > this - otherwise, we can just wait until this work is done and submit it > > > all at once. > > > > Absolutely fine, though I won't pick this up without the user also being > > ready to go. > > > Having looked through these again, they are straight forward so no changes > requested from me. Nuno, if you can add this set into appropriate > point in your series that will make use of it that will make my life easier > and ensure and minor rebasing etc happens without having to bother Paul. > Sure... - Nuno Sá >
