mbox series

[v8,00/17] power: sequencing: implement the subsystem and add first users

Message ID 20240528-pwrseq-v8-0-d354d52b763c@linaro.org
Headers show
Series power: sequencing: implement the subsystem and add first users | expand

Message

Bartosz Golaszewski May 28, 2024, 7:03 p.m. UTC
Note: I am resending this series in its entirety once more for
discussions and reviews. If there won't be any major objections, I'll
then start sending individual bits and pieces to appropriate trees.

Merging strategy: The DT binding and DTS changes are a no-brainer, they
can go through the wireless, regulator and arm-msm trees separately. The
bluetooth and PCI changes have a build-time dependency on the power
sequencing code. The bluetooth changes also have a run-time dependency on
the PCI pwrctl part. In order to get it into next I plan to pick up the
power sequencing code into my own tree and maintain it. I can then
provide an immutable tag for the BT and PCI trees to pull. I wouldn't
stress about the BT runtime dependency as it will be fixed once all
changes are in next.

The actual cover letter follows:

--

Problem statement #1: Dynamic bus chicken-and-egg problem.

Certain on-board PCI devices need to be powered up before they are can be
detected but their PCI drivers won't get bound until the device is
powered-up so enabling the relevant resources in the PCI device driver
itself is impossible.

Problem statement #2: Sharing inter-dependent resources between devices.

Certain devices that use separate drivers (often on different busses)
share resources (regulators, clocks, etc.). Typically these resources
are reference-counted but in some cases there are additional interactions
between them to consider, for example specific power-up sequence timings.

===

The reason for tackling both of these problems in a single series is the
fact the the platform I'm working on - Qualcomm RB5 - deals with both and
both need to be addressed in order to enable WLAN and Bluetooth support
upstream.

The on-board WLAN/BT package - QCA6391 - has a Power Management Unit that
takes inputs from the host and exposes LDO outputs consumed by the BT and
WLAN modules which can be powered-up and down independently. However
a delay of 100ms must be respected between enabling the BT- and
WLAN-enable GPIOs.

A similar design with a discreet PMU is also employed in other models of
the WCN family of chips although we can often do without the delays. With
this series we add support for the WCN7850 as well.

===

We introduce a new subsystem here - the power sequencing framework. The
qcom-wcn driver that we add is its first user. It implements the power-up
sequences for QCA6390 and WCN7850 chips. However - we only use it to
power-up the bluetooth module in the former. We use it to driver the WLAN
modules in both. The reason for this is that for WCN7850 we have
comprehensive bindings already upstream together with existing DT users.
Porting them to using the pwrseq subsystem can be done separately and in
an incremental manner once the subsystem itself is upstream. We will also
have to ensure backward DT compatibility. To avoid overcomplicating this
series, let's leave it out for now.

===

This series is logically split into several sections. I'll go
patch-by-patch and explain each step.

Patches 1/16-5/16:

These contain all relevant DT bindings changes. We add new documents for
the QCA6390 & WCN7850 PMUs and ATH12K devices as well as extend the bindings
for the Qualcomm Bluetooth and ATH11K modules with regulators used by them
in QCA6390.

Patches 6/16-8/16:

These contain changes to device-tree sources for the three platforms we
work with in this series. We model the PMUs of the WLAN/BT chips as
top-level platform devices on the device tree. In order to limit the scope
of this series and not introduce an excessive amount of confusion with
deprecating DT bindings, we leave the Bluetooth nodes on sm8650 and sm8550
as is (meaning: they continue to consumer the GPIOs and power inputs from
the host). As the WCN7850 module doesn't require any specific timings, we can
incrementally change that later.

In both cases we add WLAN nodes that consume the power outputs of the PMU.
For QCA6390 we also make the Bluetooth node of the RB5 consume the outputs
of the PMU - we can do it as the bindings for this chip did not define any
supply handles prior to this series meaning we are able to get this correct
right away.

Patches 9/16-12/16:

These contain the bulk of the PCI changes for this series. We introduce
a simple framework for powering up PCI devices before detecting them on
the bus.

The general approach is as follows: PCI devices that need special
treatment before they can be powered up, scanned and bound to their PCI
drivers must be described on the device-tree as child nodes of the PCI
port node. These devices will be instantiated on the platform bus. They
will in fact be generic platform devices with the compatible of the form
used for PCI devices already upstream ("pci<vendor ID>,<device ID">). We
add a new directory under drivers/pci/pwrctl/ that contains PCI pwrctl
drivers. These drivers are platform drivers that will now be matched
against the devices instantiated from port children just like any other
platform pairs.

Both the power control platform device *AND* the associated PCI device
reuse the same OF node and have access to the same properties. The goal
of the platform driver is to request and bring up any required resources
and let the pwrctl framework know that it's now OK to rescan the bus and
detect the devices. When the device is bound, we are notified about it
by the PCI bus notifier event and can establish a device link between the
power control device and the PCI device so that any future extension for
power-management will already be able to work with the correct hierachy.

The reusing of the OF node is the reason for the small changes to the PCI
OF core: as the bootloader can possibly leave the relevant regulators on
before booting linux, the PCI device can be detected before its platform
abstraction is probed. In this case, we find that device first and mark
its OF node as reused. The pwrctl framework handles the opposite case
(when the PCI device is detected only after the platform driver
successfully enabled it).

Patch 13/16 - 14/16:

These add a relatively simple power sequencing subsystem and the first
driver using it: the pwrseq module for the PMUs on the WCN family of chips.

I'm proposing to add a subsystem that allows different devices to use a shared
power sequence split into consumer-specific as well as common "units".

A power sequence provider driver registers a set of units with pwrseq
core. Each unit can be enabled and disabled and contains an optional list
of other units which must be enabled before it itself can be. A unit
represents a discreet chunk of the power sequence.

It also registers a list of targets: a target is an abstraction wrapping
a unit which allows consumers to tell pwrseq which unit they want to
reach. Real-life example is the driver we're adding here: there's a set
of common regulators, two PCIe-specific ones and two enable GPIOs: one
for Bluetooth and one for WLAN.

The Bluetooth driver requests a descriptor to the power sequencer and
names the target it wants to reach:

    pwrseq = devm_pwrseq_get(dev, "bluetooth");

The pwrseq core then knows that when the driver calls:

    pwrseq_power_on(pwrseq);

It must enable the "bluetooth-enable" unit but it depends on the
"regulators-common" unit so this one is enabled first. The provider
driver is also in charge of assuring an appropriate delay between
enabling the BT and WLAN enable GPIOs. The WLAN-specific resources are
handled by the "wlan-enable" unit and so are not enabled until the WLAN
driver requests the "wlan" target to be powered on.

Another thing worth discussing is the way we associate the consumer with
the relevant power sequencer. DT maintainers have expressed a discontent
with the existing mmc pwrseq bindings and have NAKed an earlier
initiative to introduce global pwrseq bindings to the kernel[1].

In this approach, we model the existing regulators and GPIOs in DT but
the pwrseq subsystem requires each provider to provide a .match()
callback. Whenever a consumer requests a power sequencer handle, we
iterate over the list of pwrseq drivers and call .match() for each. It's
up to the driver to verify in a platform-specific way whether it deals
with its consumer and let the core pwrseq code know.

The advantage of this over reusing the regulator or reset subsystem is
that it's more generalized and can handle resources of all kinds as well
as deal with any kind of power-on sequences: for instance, Qualcomm has
a PCI switch they want a driver for but this switch requires enabling
some resources first (PCI pwrctl) and then configuring the device over
I2C (which can be handled by the pwrseq provider).

Patch 15:

This patch makes the Qualcomm Bluetooth driver get and use the power
sequencer for QCA6390.

Patch 16:

While tiny, this patch is possibly the highlight of the entire series.
It uses the two abstraction layers we introduced before to create an
elegant power sequencing PCI power control driver and supports the ath11k
module on QCA6390 and ath12k on WCN7850.

With this series we can now enable BT and WLAN on several new Qualcomm
boards upstream.

Tested on RB5, sm8650-qrd, sm8650-hdk and sm8550-qrd.

Changelog:

Since v7:
- added DTS changes for sm8650-hdk
- added circular dependency detection for pwrseq units
- fixed a KASAN reported use-after-free error in remove path
- improve Kconfig descriptions
- fix typos in bindings and Kconfig
- fixed issues reported by smatch
- fix the unbind path in PCI pwrctl
- lots of minor improvements to the pwrseq core

Since v6:
- kernel doc fixes
- drop myself from the DT bindings maintainers list for ath12k
- wait until the PCI bridge device is fully added before creating the
  PCI pwrctl platform devices for its sub-nodes, otherwise we may see
  sysfs and procfs attribute failures (due to duplication, we're
  basically trying to probe the same device twice at the same time)
- I kept the regulators for QCA6390's ath11k as required as they only
  apply to this specific Qualcomm package

Since v5:
- unify the approach to modelling the WCN WLAN/BT chips by always exposing
  the PMU node on the device tree and making the WLAN and BT nodes become
  consumers of its power outputs; this includes a major rework of the DT
  sources, bindings and driver code; there's no more a separate PCI
  pwrctl driver for WCN7850, instead its power-up sequence was moved
  into the pwrseq driver common for all WCN chips
- don't set load_uA from new regulator consumers
- fix reported kerneldoc issues
- drop voltage ranges for PMU outputs from DT
- many minor tweaks and reworks

v1: Original RFC:

https://lore.kernel.org/lkml/20240104130123.37115-1-brgl@bgdev.pl/T/

v2: First real patch series (should have been PATCH v2) adding what I
    referred to back then as PCI power sequencing:

https://lore.kernel.org/linux-arm-kernel/2024021413-grumbling-unlivable-c145@gregkh/T/

v3: RFC for the DT representation of the PMU supplying the WLAN and BT
    modules inside the QCA6391 package (was largely separate from the
    series but probably should have been called PATCH or RFC v3):

https://lore.kernel.org/all/CAMRc=Mc+GNoi57eTQg71DXkQKjdaoAmCpB=h2ndEpGnmdhVV-Q@mail.gmail.com/T/

v4: Second attempt at the full series with changed scope (introduction of
    the pwrseq subsystem, should have been RFC v4)

https://lore.kernel.org/lkml/20240201155532.49707-1-brgl@bgdev.pl/T/

v5: Two different ways of handling QCA6390 and WCN7850:

https://lore.kernel.org/lkml/20240216203215.40870-1-brgl@bgdev.pl/

Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
---
Bartosz Golaszewski (16):
      regulator: dt-bindings: describe the PMU module of the QCA6390 package
      regulator: dt-bindings: describe the PMU module of the WCN7850 package
      dt-bindings: net: bluetooth: qualcomm: describe regulators for QCA6390
      dt-bindings: net: wireless: qcom,ath11k: describe the ath11k on QCA6390
      dt-bindings: net: wireless: describe the ath12k PCI module
      arm64: dts: qcom: sm8550-qrd: add the Wifi node
      arm64: dts: qcom: sm8650-qrd: add the Wifi node
      arm64: dts: qcom: qrb5165-rb5: add the Wifi node
      power: sequencing: implement the pwrseq core
      power: pwrseq: add a driver for the PMU module on the QCom WCN chipsets
      PCI: hold the rescan mutex when scanning for the first time
      PCI/pwrctl: reuse the OF node for power controlled devices
      PCI/pwrctl: create platform devices for child OF nodes of the port node
      PCI/pwrctl: add PCI power control core code
      PCI/pwrctl: add a PCI power control driver for power sequenced devices
      Bluetooth: qca: use the power sequencer for QCA6390

Neil Armstrong (1):
      arm64: dts: qcom: sm8650-hdk: add the Wifi node

 .../bindings/net/bluetooth/qualcomm-bluetooth.yaml |   17 +
 .../bindings/net/wireless/qcom,ath11k-pci.yaml     |   46 +
 .../bindings/net/wireless/qcom,ath12k.yaml         |   99 ++
 .../bindings/regulator/qcom,qca6390-pmu.yaml       |  185 ++++
 MAINTAINERS                                        |    8 +
 arch/arm64/boot/dts/qcom/qrb5165-rb5.dts           |  103 +-
 arch/arm64/boot/dts/qcom/sm8250.dtsi               |    2 +-
 arch/arm64/boot/dts/qcom/sm8550-qrd.dts            |   97 ++
 arch/arm64/boot/dts/qcom/sm8550.dtsi               |    2 +-
 arch/arm64/boot/dts/qcom/sm8650-hdk.dts            |   89 ++
 arch/arm64/boot/dts/qcom/sm8650-qrd.dts            |   89 ++
 arch/arm64/boot/dts/qcom/sm8650.dtsi               |    2 +-
 drivers/bluetooth/hci_qca.c                        |   74 +-
 drivers/pci/Kconfig                                |    1 +
 drivers/pci/Makefile                               |    1 +
 drivers/pci/bus.c                                  |    9 +
 drivers/pci/of.c                                   |   14 +-
 drivers/pci/probe.c                                |    2 +
 drivers/pci/pwrctl/Kconfig                         |   17 +
 drivers/pci/pwrctl/Makefile                        |    6 +
 drivers/pci/pwrctl/core.c                          |  137 +++
 drivers/pci/pwrctl/pci-pwrctl-pwrseq.c             |   89 ++
 drivers/pci/remove.c                               |    3 +-
 drivers/power/Kconfig                              |    1 +
 drivers/power/Makefile                             |    1 +
 drivers/power/sequencing/Kconfig                   |   28 +
 drivers/power/sequencing/Makefile                  |    6 +
 drivers/power/sequencing/core.c                    | 1105 ++++++++++++++++++++
 drivers/power/sequencing/pwrseq-qcom-wcn.c         |  336 ++++++
 include/linux/pci-pwrctl.h                         |   51 +
 include/linux/pwrseq/consumer.h                    |   56 +
 include/linux/pwrseq/provider.h                    |   75 ++
 32 files changed, 2717 insertions(+), 34 deletions(-)
---
base-commit: 6dc544b66971c7f9909ff038b62149105272d26a
change-id: 20240527-pwrseq-76fc025248a2

Best regards,

Comments

Bjorn Helgaas June 4, 2024, 5:19 p.m. UTC | #1
On Tue, May 28, 2024 at 09:03:08PM +0200, Bartosz Golaszewski wrote:
> Note: I am resending this series in its entirety once more for
> discussions and reviews. If there won't be any major objections, I'll
> then start sending individual bits and pieces to appropriate trees.
> 
> Merging strategy: The DT binding and DTS changes are a no-brainer, they
> can go through the wireless, regulator and arm-msm trees separately. The
> bluetooth and PCI changes have a build-time dependency on the power
> sequencing code. The bluetooth changes also have a run-time dependency on
> the PCI pwrctl part. In order to get it into next I plan to pick up the
> power sequencing code into my own tree and maintain it. I can then
> provide an immutable tag for the BT and PCI trees to pull. I wouldn't
> stress about the BT runtime dependency as it will be fixed once all
> changes are in next.
> ...

> ---
> base-commit: 6dc544b66971c7f9909ff038b62149105272d26a
> change-id: 20240527-pwrseq-76fc025248a2

What does this apply to?  I don't know what 6dc544b66971 is; it
doesn't seem to be in upstream or linux-next.
Bartosz Golaszewski June 4, 2024, 6:24 p.m. UTC | #2
On Tue, Jun 4, 2024 at 7:19 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Tue, May 28, 2024 at 09:03:08PM +0200, Bartosz Golaszewski wrote:
> > Note: I am resending this series in its entirety once more for
> > discussions and reviews. If there won't be any major objections, I'll
> > then start sending individual bits and pieces to appropriate trees.
> >
> > Merging strategy: The DT binding and DTS changes are a no-brainer, they
> > can go through the wireless, regulator and arm-msm trees separately. The
> > bluetooth and PCI changes have a build-time dependency on the power
> > sequencing code. The bluetooth changes also have a run-time dependency on
> > the PCI pwrctl part. In order to get it into next I plan to pick up the
> > power sequencing code into my own tree and maintain it. I can then
> > provide an immutable tag for the BT and PCI trees to pull. I wouldn't
> > stress about the BT runtime dependency as it will be fixed once all
> > changes are in next.
> > ...
>
> > ---
> > base-commit: 6dc544b66971c7f9909ff038b62149105272d26a
> > change-id: 20240527-pwrseq-76fc025248a2
>
> What does this apply to?  I don't know what 6dc544b66971 is; it
> doesn't seem to be in upstream or linux-next.

It's next-20240528 but it also applies to today's next without
conflicts. What do you want me to base the PCI part when resending?

Bart
Bjorn Helgaas June 5, 2024, 5:56 p.m. UTC | #3
On Tue, Jun 04, 2024 at 08:24:34PM +0200, Bartosz Golaszewski wrote:
> On Tue, Jun 4, 2024 at 7:19 PM Bjorn Helgaas <helgaas@kernel.org> wrote:
> >
> > On Tue, May 28, 2024 at 09:03:08PM +0200, Bartosz Golaszewski wrote:
> > > Note: I am resending this series in its entirety once more for
> > > discussions and reviews. If there won't be any major objections, I'll
> > > then start sending individual bits and pieces to appropriate trees.
> > >
> > > Merging strategy: The DT binding and DTS changes are a no-brainer, they
> > > can go through the wireless, regulator and arm-msm trees separately. The
> > > bluetooth and PCI changes have a build-time dependency on the power
> > > sequencing code. The bluetooth changes also have a run-time dependency on
> > > the PCI pwrctl part. In order to get it into next I plan to pick up the
> > > power sequencing code into my own tree and maintain it. I can then
> > > provide an immutable tag for the BT and PCI trees to pull. I wouldn't
> > > stress about the BT runtime dependency as it will be fixed once all
> > > changes are in next.
> > > ...
> >
> > > ---
> > > base-commit: 6dc544b66971c7f9909ff038b62149105272d26a
> > > change-id: 20240527-pwrseq-76fc025248a2
> >
> > What does this apply to?  I don't know what 6dc544b66971 is; it
> > doesn't seem to be in upstream or linux-next.
> 
> It's next-20240528 but it also applies to today's next without
> conflicts. What do you want me to base the PCI part when resending?

For PCI, we usually apply things on topic branches based on -rc1, so
that's the easiest.
Bjorn Helgaas June 11, 2024, 10:54 p.m. UTC | #4
On Tue, May 28, 2024 at 09:03:08PM +0200, Bartosz Golaszewski wrote:
> Note: I am resending this series in its entirety once more for
> discussions and reviews. If there won't be any major objections, I'll
> then start sending individual bits and pieces to appropriate trees.
> 
> Merging strategy: The DT binding and DTS changes are a no-brainer, they
> can go through the wireless, regulator and arm-msm trees separately. The
> bluetooth and PCI changes have a build-time dependency on the power
> sequencing code. The bluetooth changes also have a run-time dependency on
> the PCI pwrctl part. In order to get it into next I plan to pick up the
> power sequencing code into my own tree and maintain it. I can then
> provide an immutable tag for the BT and PCI trees to pull. I wouldn't
> stress about the BT runtime dependency as it will be fixed once all
> changes are in next.

The PCI changes are very self-contained and any conflicts will be
trivial, so rather than messing with an immutable tag, how about if I
just ack them and you can include them in your tree directly?
Bartosz Golaszewski June 12, 2024, 7:10 a.m. UTC | #5
On Wed, Jun 12, 2024 at 12:54 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> On Tue, May 28, 2024 at 09:03:08PM +0200, Bartosz Golaszewski wrote:
> > Note: I am resending this series in its entirety once more for
> > discussions and reviews. If there won't be any major objections, I'll
> > then start sending individual bits and pieces to appropriate trees.
> >
> > Merging strategy: The DT binding and DTS changes are a no-brainer, they
> > can go through the wireless, regulator and arm-msm trees separately. The
> > bluetooth and PCI changes have a build-time dependency on the power
> > sequencing code. The bluetooth changes also have a run-time dependency on
> > the PCI pwrctl part. In order to get it into next I plan to pick up the
> > power sequencing code into my own tree and maintain it. I can then
> > provide an immutable tag for the BT and PCI trees to pull. I wouldn't
> > stress about the BT runtime dependency as it will be fixed once all
> > changes are in next.
>
> The PCI changes are very self-contained and any conflicts will be
> trivial, so rather than messing with an immutable tag, how about if I
> just ack them and you can include them in your tree directly?

Sure, if you're convinced that eventual conflicts in PCI core won't be
an issue then it's even better.

Bart