lkml.org 
[lkml]   [2022]   [Aug]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH RESEND v5 24/24] PCI: dwc: Add DW eDMA engine support
On Thu, Aug 25, 2022 at 11:04:43AM -0500, Bjorn Helgaas wrote:
> On Thu, Aug 25, 2022 at 08:16:14AM +0300, Serge Semin wrote:
> > On Wed, Aug 24, 2022 at 01:17:54PM -0500, Bjorn Helgaas wrote:
> > > On Wed, Aug 24, 2022 at 09:13:19PM +0300, Serge Semin wrote:
> > > > On Wed, Aug 24, 2022 at 11:51:18AM -0500, Bjorn Helgaas wrote:
> > > > > On Mon, Aug 22, 2022 at 09:53:32PM +0300, Serge Semin wrote:
> > >
> > > > > > + val = dw_pcie_readl_dbi(pci, PCIE_DMA_VIEWPORT_BASE + PCIE_DMA_CTRL);
> > > > > > + if (val == 0xFFFFFFFF && pci->edma.reg_base) {
> > > > > > + pci->edma.mf = EDMA_MF_EDMA_UNROLL;
> > > > > > +
> > > > > > + val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL);
> > > > > > + } else if (val != 0xFFFFFFFF) {
> > > > >
> > > >
> > > > > Consider PCI_POSSIBLE_ERROR() as an annotation about the meaning of
> > > > > 0xFFFFFFFF and something to grep for.
> > > >
> > > > In this case FFs don't mean an error but a special value, which
> > > > indicates that the eDMA is mapped via the unrolled CSRs space. The
> > > > similar approach has been implemented for the iATU legacy/unroll setup
> > > > auto-detection. So I don't see much reasons to have it grepped, so as
> > > > to have a macro-based parametrization since the special value will
> > > > unluckily change while having the explicit literal utilized gives a
> > > > better understanding of the way the algorithm works.
> >
> > > If 0xFFFFFFFF is the result of a successful PCIe Memory Read,
> >
> > Right. It is.
> >
> > > and not
> > > something synthesized by the host bridge when it handles an
> > > Unsupported Request completion,
> >
> > No it isn't. To be clear 0xFFs don't indicate some PCIe bus/controller
> > malfunction, but they are a result of reading the
> > DMA_CTRL_VIEWPORT_OFF register which doesn't exist. The manual
> > explicitly says: "Note - When register does not exist, value is fixed
> > to 32'hFFFF_FFFF". The register doesn't exist if either eDMA is
> > unavailable or the eDMA CSRs are mapped via the unrolled state.
>

> OK. I don't think that's worded very well in the manual. A register
> that does not exist does not have a value, and attempts to read it
> should fail.

No. The manual explicitly says that this particular CSR
(DMA_CTRL_VIEWPORT_OFF) value is tied to 32'hFFFF_FFFF if the register
doesn't exist. There is no such text mentioned for any other
non-existing CSR.

> If they want to say the register always exists and
> contains 0xFFFFFFFF for versions earlier than X, that would make
> sense. Wouldn't be the first time a manual is ambiguous ;)

They say, that the register doesn't exist if either eDMA isn't
available or it's mapped via the unrolled CSR space. There is no
reference to the IP-core version.

Anyway basically you are right they indeed imply that the register
always exists.

>
> If the device itself, i.e., not the Root Complex, is fabricating this
> 0xFFFFFFFF value, reading it should not cause any AER or other error
> status bits to be set.
>
> If the Root Complex fabricates 0xFFFFFFFF upon receipt of a Completion
> with Unsupported Request status, I would expect bits like Received
> Master Abort to be set in the Root Port's Secondary Status register.

The device CSRs are accessed by means of the controller DBI-interface.
Even though the whole CSRs space do look as the extended PCIe config
space, the DBI-based access isn't tracked by the standard PCIe
capabilities like AER. So in case of the eDMA auto-detection procedure
introduced in this patch (and iATU auto-detection, which is already
available in the DW PCIe driver ) we won't have any bus error status
raised.

-Sergey

\
 
 \ /
  Last update: 2022-08-25 19:07    [W:0.083 / U:0.140 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site