Message-ID: <Z_WUgPMNzFAftLeE@google.com>
Date: Tue, 8 Apr 2025 14:26:24 -0700
From: Brian Norris <briannorris@...omium.org>
To: Bartosz Golaszewski <brgl@...ev.pl>,
	Bjorn Helgaas <bhelgaas@...gle.com>,
	Jingoo Han <jingoohan1@...il.com>,
	Manivannan Sadhasivam <manivannan.sadhasivam@...aro.org>,
	Lorenzo Pieralisi <lpieralisi@...nel.org>,
	Krzysztof Wilczyński <kw@...ux.com>
Cc: Krishna Chaitanya Chundru <krishna.chundru@....qualcomm.com>,
	Rob Herring <robh@...nel.org>, linux-pci@...r.kernel.org,
	linux-kernel@...r.kernel.org, linux-arm-msm@...r.kernel.org,
	dmitry.baryshkov@...aro.org, Tsai Sung-Fu <danielsftsai@...gle.com>,
	Jim Quinlan <jim2101024@...il.com>,
	Nicolas Saenz Julienne <nsaenz@...nel.org>,
	Florian Fainelli <florian.fainelli@...adcom.com>,
	Broadcom internal kernel review list <bcm-kernel-feedback-list@...adcom.com>
Subject: Re: [RFC] PCI: pwrctrl and link-up dependencies

+ adding pcie-brcmstb.c folks

On Tue, Apr 08, 2025 at 12:59:39PM -0700, Brian Norris wrote:
> TL;DR: PCIe link-up may depend on pwrctrl; however, link-startup is
> often run before pwrctrl gets involved. I'm exploring options to resolve
> this.

Apologies if a quick self-reply is considered nosiy or rude, but I
nearly forgot that I previously was looking at "pwrctrl"-like
functionality and noticed that drivers/pci/controller/pcie-brcmstb.c has
had a portion of the same "pwrctrl" functionality for some time (commit
93e41f3fca3d ("PCI: brcmstb: Add control of subdevice voltage
regulators")).

Notably, it performs its power sequencing before starting its link, for
(I believe) exactly the same reasons as I mention below. While I'm sure
it could theoretically be nice for them to be able to use
drivers/pci/pwrctrl/, I expect they cannot today, for the same reasons.

Brian

P.S. My original post is here, in case you're interested:
https://lore.kernel.org/linux-pci/Z_WAKDjIeOjlghVs@google.com/
I also leave the rest of the message intact below.

> Hi all,
> 
> I'm currently looking at reworking how some (currently out-of-tree, but I'm
> hoping to change that) pcie-designware based drivers integrate power sequencing
> for their endpoint devices, as well as the corresponding start_link()
> functionality.
> 
> For power sequencing, drivers/pci/pwrctrl/ looks like a very good start at what
> we need, since we have various device-specific regulators, GPIOs, and
> sequencing requirements, which we'd prefer not to encode directly in the
> controller driver.
> 
> For link startup, pcie-designware-host.c currently
> (a) starts the link via platform-specific means (dw_pcie::ops::start_link()) and
> (b) waits for the link training to complete.
> 
> However, (b) will fail if the other end of the link is not powered up --
> e.g., if the appropriate pwrctrl driver has not yet loaded, or its
> device hasn't finished probing. Today, this can mean the designware
> driver will either fail to probe, or at least waste time for a condition
> that we can't achieve (link up), depending on the HW/driver
> implementation.
> 
> I'm wondering how any designware-based platforms (on which I believe pwrctrl
> was developed) actually support this, and how I should look to integrate
> additional platforms/drivers. From what I can tell, the only way things would
> work today would either be if:
> (1) a given platform uses the dw_pcie_rp::use_linkup_irq==true functionality,
>     which means pcie-designware-host will only start the link, but not wait for
>     training to succeed. (And presumably the controller will receive its
>     link-up IRQ after power sequencing is done, at which point both pwrctrl and
>     the IRQ may rescan the PCI bus.) Or:
> (2) pci/pwrctrl sequencing only brings up some non-critical power rails for the
>     device in question, so link-up can actually succeed even without
>     pwrctrl.
> 
> My guess is that (1) is the case, and specifically that the relevant folks are
> using the pcie-qcom.c, with its "global" IRQ used for link-up events.
> 
> So how should I replicate this on other designware-based platforms? I suppose
> if the platform in question has a link-up IRQ, I can imitate (1). But what if
> it doesn't? (Today, we don't validate and utilize a link-up IRQ, although it's
> possible there is one available. Additionally, we use various retry mechanisms
> today, which don't trivially fit into this framework, as we won't know when
> precisely to retry, if power sequencing is controlled entirely by some other
> entity.)
> 
> Would it make sense to introduce some sort of pwrctrl -> start_link()
> dependency? For example, I see similar work done in this series [1], for
> slightly different reasons. In short, that series adds new
> pci_ops::{start,stop}_link() callbacks, and teaches a single pwrctrl driver to
> stop and restart the bridge link before/after powering things up.
> 
> I also see that Manivannan has a proposal out [2] to add semi-generic
> link-down + retraining support to core code. It treads somewhat similar
> ground, and I could even imagine that its pci_ops::retrain_link()
> callback could even be reimplemented in terms of the aforementioned
> pci_ops::{start,stop}_link(), or possibly vice versa.
> 
> Any thoughts here? Sorry for a lot of text and no patch, but I didn't just want
> to start off by throwing a 3rd set of patches on top of the existing ones that
> tread similar ground[1][2].
> 
> Regards,
> Brian
> 
> [1] [PATCH v4 00/10] PCI: Enable Power and configure the TC956x PCIe switch
> https://lore.kernel.org/linux-pci/20250225-qps615_v4_1-v4-0-e08633a7bdf8@oss.qualcomm.com/
> 
> [PATCH v4 03/10] PCI: Add new start_link() & stop_link function ops
> https://lore.kernel.org/linux-pci/20250225-qps615_v4_1-v4-3-e08633a7bdf8@oss.qualcomm.com/
> 
> [...]
> [
> [PATCH v4 08/10] PCI: pwrctrl: Add power control driver for tc956x
> https://lore.kernel.org/linux-pci/20250225-qps615_v4_1-v4-8-e08633a7bdf8@oss.qualcomm.com/
> 
> [2] [PATCH 0/2] PCI: Add support for handling link down event from host bridge drivers
> https://lore.kernel.org/linux-pci/20250221172309.120009-1-manivannan.sadhasivam@linaro.org/

Powered by blists - more mailing lists