Re: [PATCH 00/17] dmaengine: dw-edma: Support dynamic LL appends

[Vinod Koul <vkoul@kernel.org>]

On 23-06-26, 14:56, Koichiro Den wrote:
> Hi Niklas, Mani,
> 
> On Mon, Jun 22, 2026 at 04:18:01PM +0200, Niklas Cassel wrote:
> > On Mon, Jun 22, 2026 at 04:38:49PM +0900, Koichiro Den wrote:
> > > On Tue, Jun 16, 2026 at 12:40:54AM +0900, Koichiro Den wrote:
> > > 
> > > Hi Frank, Niklas, all,
> > > 
> > > I am looking for a good way to stress PCIe controller DMA engines, such as
> > > eDMA/HDMA, and measure their upper-bound throughput.
> > > 
> > > nvmet_pci_epf is useful since it is a real in-tree consumer, but it is not a
> > > very direct benchmark for the DMA engine itself. So I wonder if
> > > pci_endpoint_test would be a reasonable place to add an opt-in DMA performance
> > > mode.
> > > 
> > > One possible option I have in mind is:
> > > 
> > >   - a new fixture, pci_ep_dma_perf
> > >   - opt-in execution, for example with PCITEST_PERF=1 environment variable
> > >   - a few variants such as single and sg, possibly with a few knobs:
> > >      - PCITEST_PERF_NUM_WORKERS, to use multiple EP-side workers
> > >      - PCITEST_PERF_NUM_CHANS, to use multiple DMA channels
> > >      - perhaps other knobs for SG entry size, number of entries, etc.
> > >   - the new tests: READ_PERF_TEST and WRITE_PERF_TEST
>        `--- (A)
> 
> > > 
> > > For the other possible places I could think of, this still seems to fit best in
> > > pci_endpoint_test. For example, extending dmatest does not seem to fit well
>                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > because this needs both EP and RC side setup. A separate kselftest also feels
>     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>                          `--- (B)
> 
> > > like it would duplicate a lot of pci_endpoint_test code. That said, I might be
> > > missing something.
> > > 
> > > What do you think? Any thoughts or suggestions would be much appreciated.
> > 
> > There are two existing (out-of-tree) tests for eDMA that I know of:
> > 
> > 1)
> > https://patchwork.kernel.org/project/linux-pci/patch/cc195ac53839b318764c8f6502002cd6d933a923.1547230339.git.gustavo.pimentel@synopsys.com/
> > 
> > But as you can see, the comment was to use dmatest instead.
> > AFAICT, dmatest currently only supports DMA_MEMCPY, which, by hardware design,
> > cannot be supported by DWC eDMA HW (since it only allows remote to local, or
> > local to remote, and remote has to be a PCI address, while local is local
> > physical address).
> > 
> > Perhaps it is possible to add DMA_SLAVE support to dmatest.
> 
> Thanks for the pointers, Niklas.
> 
> The first one looks like a host-side test on top of dw-edma-pcie, where the
> RC-side driver programs the eDMA through BARs. That is useful, but it is a bit
> different from what I had in mind here.
> 
> What I am looking for is closer to pci_endpoint_test READ_TEST/WRITE_TEST, but
> with a perf/stress mode: the RC side provides the buffers, while EP Linux
> drives the EP-local DMA engine and reports the throughput.
> 
> I agree that, if we extend dmatest with DMA_SLAVE support, then Vinod should be
> involved early. In theory that could cover this too, if dmatest also had a way
> to set up RC-side buffers and pass their PCI addresses to the EP side. That is
> the part that makes me unsure dmatest is the right fit here (see (B) above).
> Before going too far down that path, I wanted to check whether the PCI endpoint
> test stack would be an acceptable home for this EP-driven case.
> 
> > 
> > 
> > 2)
> > https://github.com/rockchip-linux/kernel/blob/develop-6.1/drivers/pci/controller/dwc/pcie-dw-dmatest.c
> > https://github.com/rockchip-linux/kernel/blob/develop-6.1/drivers/pci/controller/rockchip-pcie-dma.h
> > 
> > 
> > Anyway, since Vinod is the maintainer, it is probably him you need to talk
> > to come up with a way forward. To not waste your time, I would talk to him
> > before you spend a lot of time implementing something :)
> 
> Yes, that makes sense. I will keep Vinod in the loop. :)

Thanks, I am not sure it makes sense for dmatest to support slave
transfers. We need a hardware signalling for these dma transfers to
work and data is pushed/pulled from hardware.

I know it can be made to work for some kind of slave transfers like
possibly pcie etc and(there was another attempt to add a
driver for slave tests) but it is not really right as it would confuse
folks who might want to test it for slave transfers which typically cant
be done

So I am not is support of supporting this

-- 
~Vinod