Message-ID: <CAA8EJpqnFnZX5+NCMbd2Yhq6mv9Bp0NZgk60btchS6He7y_2hA@mail.gmail.com>
Date:   Tue, 11 Jan 2022 00:03:46 +0300
From:   Dmitry Baryshkov <dmitry.baryshkov@...aro.org>
To:     Saravana Kannan <saravanak@...gle.com>
Cc:     Rob Herring <robh+dt@...nel.org>,
        Frank Rowand <frowand.list@...il.com>,
        linux-kernel@...r.kernel.org, devicetree@...r.kernel.org,
        Bjorn Andersson <bjorn.andersson@...aro.org>,
        Stephen Boyd <swboyd@...omium.org>,
        Android Kernel Team <kernel-team@...roid.com>
Subject: Re: [PATCH] of: property: stop parsing remote-endpoint graph properties

HI Saravana,

On Tue, 30 Nov 2021 at 03:40, Dmitry Baryshkov
<dmitry.baryshkov@...aro.org> wrote:
>
> On Tue, 30 Nov 2021 at 02:45, Saravana Kannan <saravanak@...gle.com> wrote:
> >
> > On Thu, Nov 25, 2021 at 4:26 PM Dmitry Baryshkov
> > <dmitry.baryshkov@...aro.org> wrote:
> > >
> > > When parsing remote-endpoint properties, two counter devlinks will be
> > > created, resulting in the circular dependency, which is later broken. In
> > > most of the cases, the order in which depency is broken does not matter
> > > (or is correct). However lately I stumbled upon the following
> > > configuration.
> > >
> > > In this case for some reason devlink code decided to break the loop by
> > > making panel depend on the bridge driver, enforcing that bridge is
> > > probed before the panel.
> >
> > Let's find and fix the "for some reason" part then instead of just
> > removing support for a property.
>
> How can I help you to debug this? I can post the resulting device tree
> or add debugging patches of your choice.
>
> How is the cycle dependency broken? Is it done by removing a single
> arc or by dropping all the arcs that form a cycle?
> If the former is true, then we know the case: it sees a circular
> dependency and just decides incorrectly, which arc should be dropped.

Could you please comment on this? The issue is still present.

If there are several graphs in place, the remote-endpoint loops end up
being broken in strange ways, making devices hard (or impossible) to
probe. For example in some cases the source can end up being dependent
on the sink, while the driver code expects the sink to be probed after
the source. Removing parsing of the property replaces this with
driver-based deferrals, which do not prevent devices from being
probed.

>
> >
> > > However in such cases the bridge will lookup next bridge or panel using
> > > drm_of_find_panel_or_bridge() in the probe callback. Thus we have a
> > > deadlock: panel is waiting for the bridge because of the devlink
> > > dependency and bridge probe() expects the panel to be available and thus
> > > returns -EPROBE_DEFER.
> > >
> > > To prevent such deadlocks, stop parsing the remote-endpoint property and
> > > let drivers decide their probe order using standard -EPROBE_DEFER
> > > returns.
> >
> > Nak.
> >
> > Removing support for a property will always be NAKed. Not because I
> > care about one specific property. It's because fw_devlink needs to get
> > the full view of the dependencies to be able to break cycles. The
> > cycle detection and fixing logic has been improving steadily. So
> > there's no reason to give up on it suddenly.
>
> Regarding the remote-endpoint.
>
> While I highly value the whole devlink idea and the way it
> eases/streamlines device probing in typical dependency cases, I still
> think that graph/remote-endpoint handling is not a proper way.
>
> Current code handles remote-endpoint links in the same way as it
> handles directional links does not look right. Generic code can not
> predict, which side of bidirectional link is the primary side for the
> link, checking for the existence of the counterpart, and which one is
> a secondary side which just gets probed (and waits for the primary
> part to find it). Always getting a circular dependency (for each graph
> link) and always breaking should have the same result, as not getting
> the circular dependency at all, Is this statement correct?
>
> In fact I can predict that creating such extra dependencies can hide
> actual dependencies between devices. Consider for example two devices
> A and B, with a graph connection between A and B and another
> dependency (for example, clocks or regulator supply) from A to B. How
> will devlink handle such a case? Will it correctly determine that A
> depends on B or will it break the cycle by removing both dependencies?
>
> The MSM DRM driver employs several graphs, and things were very
> fragile here. Dropping remote-endpoint parsing typically fixed those
> extra dependency/devlink issues. So selecting between running with
> fw_devlink turned off (to let the driver to bind at all) and just
> disabling remote-endpoint parsing I'd choose the second option.
>
> >
> > -Saravana
> >
> > >
> > > DTS except follows:
> > >
> > > / {
> > >         panel0 {
> > >                 compatible = "powertip,ph800480t013-idf02";
> > >                 power-supply = <&vreg_l11c_3p3>;
> > >                 backlight = <&lcd0_reg>;
> > >                 port {
> > >                         panel0_in: endpoint {
> > >                                 remote-endpoint = <&bridge0_out>;
> > >                         };
> > >                 };
> > >         };
> > > };
> > >
> > > &dsi0 {
> > >         #address-cells = <1>;
> > >         #size-cells = <0>;
> > >         status = "okay";
> > >
> > >         bridge@0 {
> > >                 reg = <0>;
> > >                 compatible = "toshiba,tc358762";
> > >
> > >                 ports {
> > >                         #address-cells = <1>;
> > >                         #size-cells = <0>;
> > >
> > >                         port@0 {
> > >                                 reg = <0>;
> > >                                 bridge0_in: endpoint {
> > >                                         remote-endpoint = <&dsi0_out>;
> > >                                 };
> > >                         };
> > >
> > >                         port@1 {
> > >                                 reg = <1>;
> > >                                 bridge0_out: endpoint {
> > >                                         remote-endpoint = <&panel0_in>;
> > >                                 };
> > >                         };
> > >                 };
> > >         };
> > >         ports {
> > >                 port@1 {
> > >                         endpoint {
> > >                                 remote-endpoint = <&bridge0_in>;
> > >                                 data-lanes = <0 1 2 3>;
> > >                         };
> > >                 };
> > >         };
> > >
> > > };
> > >
> > > Fixes: f7514a663016 ("of: property: fw_devlink: Add support for remote-endpoint")
> > > Cc: Bjorn Andersson <bjorn.andersson@...aro.org>
> > > Cc: Stephen Boyd <swboyd@...omium.org>
> > > Cc: Saravana Kannan <saravanak@...gle.com>
> > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@...aro.org>
> > > ---
> > >  drivers/of/property.c | 8 +-------
> > >  1 file changed, 1 insertion(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/of/property.c b/drivers/of/property.c
> > > index f7229e4030e3..83548076ee63 100644
> > > --- a/drivers/of/property.c
> > > +++ b/drivers/of/property.c
> > > @@ -1249,7 +1249,6 @@ static struct device_node *parse_##fname(struct device_node *np,       \
> > >   * @parse_prop.index: For properties holding a list of phandles, this is the
> > >   *                   index into the list
> > >   * @optional: Describes whether a supplier is mandatory or not
> > > - * @node_not_dev: The consumer node containing the property is never a device.
> > >   *
> > >   * Returns:
> > >   * parse_prop() return values are
> > > @@ -1261,7 +1260,6 @@ struct supplier_bindings {
> > >         struct device_node *(*parse_prop)(struct device_node *np,
> > >                                           const char *prop_name, int index);
> > >         bool optional;
> > > -       bool node_not_dev;
> > >  };
> > >
> > >  DEFINE_SIMPLE_PROP(interconnects, "interconnects", "#interconnect-cells")
> > > @@ -1285,7 +1283,6 @@ DEFINE_SIMPLE_PROP(pinctrl5, "pinctrl-5", NULL)
> > >  DEFINE_SIMPLE_PROP(pinctrl6, "pinctrl-6", NULL)
> > >  DEFINE_SIMPLE_PROP(pinctrl7, "pinctrl-7", NULL)
> > >  DEFINE_SIMPLE_PROP(pinctrl8, "pinctrl-8", NULL)
> > > -DEFINE_SIMPLE_PROP(remote_endpoint, "remote-endpoint", NULL)
> > >  DEFINE_SIMPLE_PROP(pwms, "pwms", "#pwm-cells")
> > >  DEFINE_SIMPLE_PROP(resets, "resets", "#reset-cells")
> > >  DEFINE_SIMPLE_PROP(leds, "leds", NULL)
> > > @@ -1388,7 +1385,6 @@ static const struct supplier_bindings of_supplier_bindings[] = {
> > >         { .parse_prop = parse_pinctrl6, },
> > >         { .parse_prop = parse_pinctrl7, },
> > >         { .parse_prop = parse_pinctrl8, },
> > > -       { .parse_prop = parse_remote_endpoint, .node_not_dev = true, },
> > >         { .parse_prop = parse_pwms, },
> > >         { .parse_prop = parse_resets, },
> > >         { .parse_prop = parse_leds, },
> > > @@ -1437,9 +1433,7 @@ static int of_link_property(struct device_node *con_np, const char *prop_name)
> > >                 while ((phandle = s->parse_prop(con_np, prop_name, i))) {
> > >                         struct device_node *con_dev_np;
> > >
> > > -                       con_dev_np = s->node_not_dev
> > > -                                       ? of_get_compat_node(con_np)
> > > -                                       : of_node_get(con_np);
> > > +                       con_dev_np = of_node_get(con_np);
> > >                         matched = true;
> > >                         i++;
> > >                         of_link_to_phandle(con_dev_np, phandle);
> > > --
> > > 2.33.0
> > >
>
> --
> With best wishes
> Dmitry



-- 
With best wishes
Dmitry

Powered by blists - more mailing lists