Message-ID: <aG3IWdZIhnk01t2A@google.com>
Date: Tue, 8 Jul 2025 18:39:37 -0700
From: Brian Norris <briannorris@...omium.org>
To: Manivannan Sadhasivam <mani@...nel.org>
Cc: Bartosz Golaszewski <brgl@...ev.pl>,
	Bjorn Helgaas <bhelgaas@...gle.com>,
	Jingoo Han <jingoohan1@...il.com>,
	Lorenzo Pieralisi <lpieralisi@...nel.org>,
	Krzysztof Wilczyński <kwilczynski@...nel.org>,
	Rob Herring <robh@...nel.org>, linux-pci@...r.kernel.org,
	linux-kernel@...r.kernel.org, linux-arm-msm@...r.kernel.org,
	Krishna Chaitanya Chundru <krishna.chundru@....qualcomm.com>
Subject: Re: [PATCH RFC 0/3] PCI/pwrctrl: Allow pwrctrl framework to control
 PERST# GPIO if available

Hi Manivannan,

Thanks for tackling this!

On Mon, Jul 07, 2025 at 11:48:37PM +0530, Manivannan Sadhasivam wrote:
> Hi,
> 
> This series is an RFC to propose pwrctrl framework to control the PERST# GPIO
> instead of letting the controller drivers to do so (which is a mistake btw).
> 
> Right now, the pwrctrl framework is controlling the power supplies to the
> components (endpoints and such), but it is not controlling PERST#. This was
> pointed out by Brian during a related conversation [1]. But we cannot just move
> the PERST# control from controller drivers due to the following reasons:
> 
> 1. Most of the controller drivers need to assert PERST# during the controller
> initialization sequence. This is mostly as per their hardware reference manual
> and should not be changed.
> 
> 2. Controller drivers still need to toggle PERST# when pwrctrl is not used i.e.,
> when the power supplies are not accurately described in PCI DT node. This can
> happen on unsupported platforms and also for platforms with legacy DTs.
> 
> For this reason, I've kept the PERST# retrieval logic in the controller drivers
> and just passed the gpio descriptors (for each slot) to the pwrctrl framework.

How sure are we that GPIOs (and *only* GPIOs) are sufficient for this
feature? I've seen a few drivers that pair a GPIO with some kind of
"internal" reset too, and it's not always clear that they can/should be
operated separately.

For example, drivers/pci/controller/dwc/pci-imx6.c /
imx_pcie_{,de}assert_core_reset(), and pcie-tegra194.c's
APPL_PINMUX_PEX_RST. The tegra case especially seems pretty clear that
its non-GPIO "pex_rst" is resetting an endpoint.

> This will allow both the controller drivers and pwrctrl framework to share the
> PERST# (which is ugly but can't be avoided). But care must be taken to ensure
> that the controller drivers only assert PERST# and not deassert when pwrctrl is
> used. I've added the change for the Qcom driver as a reference. The Qcom driver
> is a slight mess because, it now has to support both new DT binding (PERST# and
> PHY in Root Port node) and legacy (both in Host Bridge node). So I've allowed
> the PERST# control only for the new binding (which is always going to use
> pwrctrl framework to control the component supplies).
> 
> Testing
> =======
> 
> This series is tested on Lenovo Thinkpad T14s laptop (with out-of-tree patch
> enabling PCIe WLAN card) and on RB3 Gen2 with TC9563 switch (also with the not
> yet merged series [2]). A big take away from this series is that, it is now
> possible to get rid of the controversial {start/stop}_link() callback proposed
> in the above mentioned switch pwrctrl driver [3].

This is a tiny bit tangential to the PERST# discussion, but I believe
there are other controller driver features that don't fit into the
sequence model of:

1. start LTSSM (controller driver)
2. pwrctrl eventually turns on power + delay per spec
3. pwrctrl deasserts PERST#
4. pwrctrl delays a fixed amount of time, per the CEM spec
5. pwrctrl rescans bus

For example, tegra_pcie_dw_start_link() notes some cases where it needs
to take action and retry when the link doesn't come up. Similarly, I've
seen drivers with retry loops for cases where the link comes up, but not
at the expected link rate. None of this is possible if the controller
driver only gets to take care of #1, and has no involvement in between
#3 and #5.

(Yes, I acknowledge that DWC's .start_link() is being abused in
tegra_pcie_dw_start_link(). But it's still reality, with a
seemingly-good use case.)

Brian

> 
> - Mani
> 
> [1] https://lore.kernel.org/linux-pci/Z_6kZ7x7gnoH-P7x@google.com/
> [2] https://lore.kernel.org/linux-pci/20250412-qps615_v4_1-v5-0-5b6a06132fec@oss.qualcomm.com/ 
> [3] https://lore.kernel.org/linux-pci/20250412-qps615_v4_1-v5-4-5b6a06132fec@oss.qualcomm.com/
> 
> Signed-off-by: Manivannan Sadhasivam <mani@...nel.org>
> ---
> Manivannan Sadhasivam (3):
>       PCI/pwrctrl: Move pci_pwrctrl_init() before turning ON the supplies
>       PCI/pwrctrl: Allow pwrctrl core to control PERST# GPIO if available
>       PCI: qcom: Allow pwrctrl framework to control PERST#
> 
>  drivers/pci/controller/dwc/pcie-designware-host.c |  1 +
>  drivers/pci/controller/dwc/pcie-designware.h      |  1 +
>  drivers/pci/controller/dwc/pcie-qcom.c            | 26 ++++++++++++++-
>  drivers/pci/pwrctrl/core.c                        | 39 +++++++++++++++++++++++
>  drivers/pci/pwrctrl/pci-pwrctrl-pwrseq.c          |  4 +--
>  drivers/pci/pwrctrl/slot.c                        |  4 +--
>  include/linux/pci-pwrctrl.h                       |  2 ++
>  include/linux/pci.h                               |  2 ++
>  8 files changed, 74 insertions(+), 5 deletions(-)
> ---
> base-commit: 00f0defc332be94b7f1fdc56ce7dcb6528cdf002
> change-id: 20250707-pci-pwrctrl-perst-bdc6e36a335c
> 
> Best regards,
> -- 
> Manivannan Sadhasivam <mani@...nel.org>
> 

Powered by blists - more mailing lists