[RFCv2 usb-next 0/3] initialize (multiple) PHYs in xhci-plat

Tue Jul 18 01:19:21 PDT 2017

Hi,

On Tue, Jul 18, 2017 at 3:40 AM, Chunfeng Yun <chunfeng.yun at mediatek.com> wrote:
> Hi,
> On Mon, 2017-07-17 at 11:27 +0200, Martin Blumenstingl wrote:
>> Hi,
>>
>> On Mon, Jul 17, 2017 at 9:21 AM, Chunfeng Yun <chunfeng.yun at mediatek.com> wrote:
>> > Hi,
>> > On Sat, 2017-07-15 at 14:11 +0200, Martin Blumenstingl wrote:
>> >> Hi,
>> >>
>> >> On Sat, Jul 15, 2017 at 11:33 AM, Chunfeng Yun
>> >> <chunfeng.yun at mediatek.com> wrote:
>> >> > Hi Martin,
>> >> >
>> >> > On Thu, 2017-07-13 at 12:59 +0200, Martin Blumenstingl wrote:
>> >> >> This series is the outcome of a discussion with Felipe Balbi,
>> >> >> see [0] and [1].
>> >> >> The quick-summary of this is:
>> >> >> - dwc3 already takes one USB2 and one USB3 PHY and initializes these
>> >> >>   correct
>> >> >> - some other HCI platform drivers (like ehci-platform.c, xhci-mtk.c and
>> >> >>   ohci-platform.c) do not have a limitation on the number of PHYs - they
>> >> >>   support one PHY per actual host port
>> >> >> - Amlogic Meson GXL and GXM SoCs come with a dwc3 IP block which has two
>> >> >>   or three USB2 ports enabled on the internal root-hub. The SoCs also
>> >> >>   provide separate USB2 PHYs, one per port. All USB2 PHYs (which are
>> >> >>   internally "connected" to the dwc3 roothub) need to be powered on,
>> >> >>   otherwise USB devices cannot be enumerated (even if just one PHY is
>> >> >>   disabled and if the device is plugged into another, enabled port)
>> >> >>
>> >> >> In my first attempt to get USB supported on the GXL and GXM SoCs I tried
>> >> >> to work-around the problem that I could not pass multiple PHYs to the
>> >> >> dwc3 controller.
>> >> >> This was rejected by Rob Herring (which was definitely the thing to do in
>> >> >> my opinion), see [2]
>> >> >>
>> >> >> This series adds a new "platform-roothub". This can be configured through
>> >> >> devicetree by passing a child-node with "reg = <0>" to the USB
>> >> >> controller. Additionally there has to be a child-node for each port on
>> >> >> the root-hub. Each of the child-nodes takes a "phys" and "phy-names"
>> >> >> property. This allows modeling the root-hub in devicetree similar to the
>> >> >> USB device binding (documented in devicetree/bindings/usb/usb-device.txt)
>> >> >> This avoids and backwards-compatibility problems (which was a concern
>> >> >> regardless of the solution, see [3]) since the binding for the root-hub
>> >> >> was previously not specified (and we're not using the "phys" property of
>> >> >> the controller, which might have served different purposes before,
>> >> >> depending on the drivers).
>> >> >>
>> >> >> Additionally this integrates the new platform-roothub into xhci-plat.c
>> >> >> which automatically enables it for the dwc3 driver (in host-mode).
>> >> >>
>> >> > How to handle the phy0(one u2phy and one u3phy) when port1 support
>> >> > dual-role mode? leave them to peripheral side as felipe suggested
>> >> > before? If so, no port1 node for roothub, is there any problem when
>> >> > change the port1 to host-only mode?
>> >> on Amlogic Meson GXL we have the following IP blocks:
>> >> - 2x USB2 PHYs, some external component has to tell them which mode
>> >> (host/device) they should operate in
>> >> - there is an additional (1x) USB3 PHY with built-in OTG detection logic
>> >>
>> >> on Amlogic Meson GXL it could work like this:
>> >> USB2 and USB3 phy0 can be passed to the root-hub. Additionally the
>> >> USB2 phy0 could be passed to the USB3 PHY. The USB3 PHY would then
>> >> tell the USB2 PHY in which mode it should operate.
>> >>
>> >> please note that device mode and OTG support on Amlogic Meson GXL is
>> >> more complicated, as it uses dwc2 and dwc3 controllers in combination:
>> >> - dwc3 is reponsible for host-only mode
>> >> - dwc2 is responsible for device-only mode
>> >> - OTG detection is done by the USB3 PHY
>> >>
>> >> would you mind sharing a short overview of host/device/OTG support
>> >> works on Mediatek SoCs? I assume that the Amlogic Meson GXL
>> >> implementation is quite special, so comparing this with another
>> >> implementation (for example the Mediatek one) may help spotting
>> >> potential issues.
>> >>
>> > MTK's mtu3 IP supports at most 5x USB2 phys and 4x USB3 phys. They work
>> > as following:
>> thank you for sharing this!
>>
>> > 1. device mode works as HS only:
>> >
>> > u2phy0 --- dual-role/OTG
>> >
>> > u2phy1 ---|
>> >           +  U3 host-only
>> > u3phy0 ---|
>> >
>> > ...
>> > u2phy4 ---|
>> >           +  U3 host-only
>> > u3phy3 ---|
>> > (e.g. MT8173 supports 2x u2phys and 1x u3phy, u2phy0 can work as
>> > dual-role mode, u2phy1 & u3phy0 are host-only)
>> >
>> > 2. device mode works as HS & SS, or host only:
>> >
>> > u2phy0 ---|
>> >           + dual-role or host-only
>> > u3phy0 ---|
>> >
>> > ...
>> > u2phy3 ---|
>> >           + U3 host-only
>> > u3phy3 ---|
>> >
>> > u2phy4 --- U2 host-only
>> > (e.g. on MT2701, u2phy0 and u3phy0 work as host-only mode)
>> OK, so in both cases only one port (with one u2phy and one u3phy) is
>> dual-role capable
> Yes
>>
>> > mtu3 driver supports host-only, device-only and dual-role mode(use IDDIG
>> > pin), and will take all phys it needed, include host-only phys;
>> > But if just host-only mode is supported, we can skip mtu3 driver and
>> > make use of xhci-mtk driver directly, then xhci-mtk will take all phys.
>> I see, in your example it's the mtu3
>> (Documentation/devicetree/bindings/usb/mt8173-mtu3.txt) which does the
>> mode switching. I assume that you're doing the host/device mode
>> switching through the extcon phandle (for example together with the
>> extcon-usb-gpio driver).
>>
>> in you example: can't we *always* describe the roothub via devicetree
>> (just like in my example: [0])?
>> this means that (as you already mentioned) USB host-only support is now covered.
>> to handle dual-role (host/device switching) we now need to pass the
>> dual-role capable PHYs to whatever IP can detect the mode it should
>> operate in (in your case: mtu3, in Amlogic's case: the u3phy with
>> built-in mode detection logic -> the driver for this IP block should
>> call phy_set_mode(phy, PHY_MODE_USB_{HOST,DEVICE,OTG} accordingly).
>> here's a skeleton (stripped-down) of how the .dts could look like:
>> mtu3: usb at 11271000 {
>>     compatible = "mediatek,mt8173-mtu3";
>> ...
>>     /* MT2701 = 2nd example, for MT8173 = 1st example we would skip
>> the u3phy0 */
>>     /* only list the dual role capable PHYs here */
>>     phys = <&u3phy0>, <&u2phy0>;
>>     phy-names = "usb3-phy", "usb2-phy";
>>
>>     usb_host: xhci at 11270000 {
>>         compatible = "mediatek,mt8173-xhci";
>>         ...
>>         roothub at 0 {
>>             /* includes all PHYs, including the dual role capable ones */
> If here include dual-role capable phys which are also taken by parent
> node(11271000), will cause phy_init/_power_on them twice.
the PHY framework does ref-counting for all calls [0], so
phy_ops.init/power_on will only be called for the first time
phy_{init,power_on} is called
also phy_ops.exit/power_off is only called when the last consumer
calls phy_{exit,power_off}

> it seems ok if here include host-only phys, but I need test it.
> If here include host-only phys, there will not be port1 for roothub
> node, maybe it's a problem.
on Amlogic's GXL SoC it also makes sense to pass the OTG capable USB2
PHY to the roothub (remember: the USB3 PHY does mode-detection for the
USB2 PHY)
this way host-only mode works even if the USB3 PHY driver is not
loaded (or the kernel module is unloaded). and it gets even worse: if
one of the USB2 PHYs is powered off then host-only mode is broken for
the other PHYs as well!

additionally the device-tree should describe the hardware - and the if
the dual-role capable PHYs are controlled by two different IP blocks:
why not add them to both (and let the PHY framework or the IP block -
like the Mediatek USB PHY IP does - handle this)

>>         };
>>     };
>> };
>>
>> do you think that this would work for the Mediatek SoCs?
>> I've seen that the phy-mt65xx-usb3.c PHY driver does not have any
>> .set_mode callback - I assume it's simply because it doesn't need it
>> (as this is either managed by the hardware/IP block internally, or
>> through some firmware/mailbox mechanism).
> Hardware can automatically switch host/device mode by IDDIG pin, but
> also we can manually switch them. Phy driver makes use of auto way.
OK, that makes it easy for the driver developers :)

>>
>>
>> Regards,
>> Mart
>>
>> [0] http://lists.infradead.org/pipermail/linux-amlogic/2017-July/004305.html
>
>

Regards,
Martin

[0] http://elixir.free-electrons.com/linux/latest/source/drivers/phy/phy-core.c#L218