Re: [RFC PATCH 0/3] PCI: Use Configuration RRS to wait for device ready

[Bjorn Helgaas <helgaas@kernel.org>]

On Wed, Aug 28, 2024 at 04:24:01PM -0500, Mario Limonciello wrote:
> On 8/27/2024 18:48, Bjorn Helgaas wrote:
> > From: Bjorn Helgaas <bhelgaas@google.com>
> > 
> > After a device reset, pci_dev_wait() waits for a device to become
> > completely ready by polling the PCI_COMMAND register.  The spec envisions
> > that software would instead poll for the device to stop responding to
> > config reads with Completions with Request Retry Status (RRS).
> > 
> > Polling PCI_COMMAND leads to hardware retries that are invisible to
> > software and the backoff between software retries doesn't work correctly.
> > 
> > Root Ports are not required to support the Configuration RRS Software
> > Visibility feature that prevents hardware retries and makes the RRS
> > Completions visible to software, so this series only uses it when available
> > and falls back to PCI_COMMAND polling when it's not.
> > 
> > This is completely untested and posted for comments.
> > 
> > Bjorn Helgaas (3):
> >    PCI: Wait for device readiness with Configuration RRS
> >    PCI: aardvark: Correct Configuration RRS checking
> >    PCI: Rename CRS Completion Status to RRS
> > 
> >   drivers/bcma/driver_pci_host.c             | 10 ++--
> >   drivers/pci/controller/dwc/pcie-tegra194.c | 18 +++---
> >   drivers/pci/controller/pci-aardvark.c      | 64 +++++++++++-----------
> >   drivers/pci/controller/pci-xgene.c         |  6 +-
> >   drivers/pci/controller/pcie-iproc.c        | 18 +++---
> >   drivers/pci/pci-bridge-emul.c              |  4 +-
> >   drivers/pci/pci.c                          | 41 +++++++++-----
> >   drivers/pci/pci.h                          | 11 +++-
> >   drivers/pci/probe.c                        | 33 +++++------
> >   include/linux/bcma/bcma_driver_pci.h       |  2 +-
> >   include/linux/pci.h                        |  1 +
> >   include/uapi/linux/pci_regs.h              |  6 +-
> >   12 files changed, 117 insertions(+), 97 deletions(-)
> 
> Although this looks like a useful series, I'm sorry to say but this doesn't
> solve the issue that Gary and I raised.  We double checked today and found
> that reading the vendor ID works just fine at this time.

Thanks for testing that.

> I think that we're still better off polling PCI_PM_CTRL to "wait" for D0
> after the state change from D3cold.

Is there some spec justification for polling PCI_PM_CTRL?  I'm dubious
about doing that just because "it works" in this situation, unless we
have some better understanding about *why* it works and whether all
devices are supposed to work that way.

Bjorn