Re: [PATCH 09/15] PCI: endpoint: Add pci-ep-dma helper for exported DMA ABI v1

[Manivannan Sadhasivam]

On Sat, Aug 22, 2026 at 12:41:32AM +0900, Koichiro Den wrote:
> On Fri, Mar 13, 2026 at 01:49:59AM +0900, Koichiro Den wrote:
> > Add a generic helper that packages controller-owned DMA resources into a
> > peer-visible BAR slice described by exported DMA ABI v1.
> > 
> > pci_epf_alloc_dma() queries EPC auxiliary resources, delegates the
> > requested DMA read channels, builds an ABI header in coherent memory,
> > and assembles the BAR region list covering the header, controller
> > register window, and per-channel descriptor windows. If the controller
> > control window is not already BAR-backed, map it into the exported slice
> > so the peer still sees a self-contained layout.
> > 
> > The first ABI is designed based on the DesignWare unrolled eDMA model,
> > but it is intended to be vendor-neutral. It exports delegated READ
> > channels only, which are the channels the host uses to send data from
> > host memory into the endpoint.
> > 
> > Signed-off-by: Koichiro Den <den@xxxxxxxxxxxxx>
> > ---
> >  drivers/pci/endpoint/Makefile     |   2 +-
> >  drivers/pci/endpoint/pci-ep-dma.c | 342 ++++++++++++++++++++++++++++++
> >  include/linux/pci-ep-dma.h        | 130 ++++++++++++
> >  3 files changed, 473 insertions(+), 1 deletion(-)
> >  create mode 100644 drivers/pci/endpoint/pci-ep-dma.c
> >  create mode 100644 include/linux/pci-ep-dma.h
> > 
> > diff --git a/drivers/pci/endpoint/Makefile b/drivers/pci/endpoint/Makefile
> > index b4869d52053a..94824f3ed5a1 100644
> > --- a/drivers/pci/endpoint/Makefile
> > +++ b/drivers/pci/endpoint/Makefile
> > @@ -5,5 +5,5 @@
> >  
> >  obj-$(CONFIG_PCI_ENDPOINT_CONFIGFS)	+= pci-ep-cfs.o
> >  obj-$(CONFIG_PCI_ENDPOINT)		+= pci-epc-core.o pci-epf-core.o\
> > -					   pci-epc-mem.o functions/
> > +					   pci-epc-mem.o pci-ep-dma.o functions/
> >  obj-$(CONFIG_PCI_ENDPOINT_MSI_DOORBELL)	+= pci-ep-msi.o
> > diff --git a/drivers/pci/endpoint/pci-ep-dma.c b/drivers/pci/endpoint/pci-ep-dma.c
> > new file mode 100644
> > index 000000000000..2a996f9b1424
> > --- /dev/null
> > +++ b/drivers/pci/endpoint/pci-ep-dma.c
> 
> ---[snip]---
> 
> > diff --git a/include/linux/pci-ep-dma.h b/include/linux/pci-ep-dma.h
> > new file mode 100644
> > index 000000000000..0ef6f9eb8593
> > --- /dev/null
> > +++ b/include/linux/pci-ep-dma.h
> 
> Hi Mani, (Cc: Frank, Niklas)
> 
> Perhaps it's time to reincarnate this series, with much of it reworked.
> 
> Following the discussion here:
> https://lore.kernel.org/r/fg7cma2ccvdgjgcref7qa4q7ocvvume5dteekedikszebkq7rd@uhrjfp3nmmwh/
> I am now thinking of dropping the generic pci-ep-dma helper and
> include/linux/pci-ep-dma.h altogether, as I understand the direction is to keep
> the DMA export within the vNTB/ntb_hw_epf path *as much as possible*.
> 
> The old wire layout was consumed by pci-epf-vntb.c, ntb_hw_epf.c, and
> dw-edma-aux.c (see patch 13). That is why I put pci-ep-dma.h under
> include/linux/. For the rework, I am thinking ntb_hw_epf should decode it and
> pass only the resulting resources to dw-edma-aux. This keeps the wire format
> between pci-epf-vntb and ntb_hw_epf. At that point, defining the same small
> layout in both files might even be preferable to adding another shared header in
> include/linux/.
> 
> Mani, does this direction look acceptable to you? I would also appreciate your
> view on the linked discussion when you have a chance.
> 

Sounds good to me. My only concern is exposing the remote DMA functionality as a
new EPF driver. As long as you keep the changes within the NTB drivers, I'm fine
with it.

- Mani

-- 
மணிவண்ணன் சதாசிவம்