Patchwork of: reduce the number of PROBE_DEFERs

login
register
mail settings
Submitter Jean-Francois Moine
Date Aug. 20, 2013, 10:01 a.m.
Message ID <20130820120105.55a4c0ce@armhf>
Download mbox | patch
Permalink /patch/268428/
State New
Headers show

Comments

Jean-Francois Moine - Aug. 20, 2013, 10:01 a.m.
This patch populates the platform from the device tree into two steps:
the first step creates the nodes that are referenced by a phandle,
the second step creates the other nodes.

This permits to reduce the number of PROBE_DEFERs.

Signed-off-by: Jean-Francois Moine <moinejf@free.fr>
---
A better way to reduce probe deferral could be sorting the nodes
according to their phandle level in the DT blob at compilation time ...
---
 drivers/of/platform.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
Grant Likely - Aug. 20, 2013, 10:13 a.m.
On Tue, Aug 20, 2013 at 11:01 AM, Jean-Francois Moine <moinejf@free.fr> wrote:
> This patch populates the platform from the device tree into two steps:
> the first step creates the nodes that are referenced by a phandle,
> the second step creates the other nodes.
>
> This permits to reduce the number of PROBE_DEFERs.
>
> Signed-off-by: Jean-Francois Moine <moinejf@free.fr>
> ---
> A better way to reduce probe deferral could be sorting the nodes
> according to their phandle level in the DT blob at compilation time ...

Have you got measurements or statistics that show this making a
difference? I suspect you'll find for boot time it will have little to
no affect since the device driver probe order is more closely related
to the kernel link order than the order that devices were registered.

g.

> ---
>  drivers/of/platform.c | 17 ++++++++++++++++-
>  1 file changed, 16 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/of/platform.c b/drivers/of/platform.c
> index e0a6514..a2ea858 100644
> --- a/drivers/of/platform.c
> +++ b/drivers/of/platform.c
> @@ -382,8 +382,23 @@ static int of_platform_bus_create(struct device_node *bus,
>         if (!dev || !of_match_node(matches, bus))
>                 return 0;
>
> +       /* first step: create the nodes that are referenced by phandle */
>         for_each_child_of_node(bus, child) {
> -               pr_debug("   create child: %s\n", child->full_name);
> +               if (child->phandle == 0)
> +                       continue;
> +               pr_debug("   create child 1: %s\n", child->full_name);
> +               rc = of_platform_bus_create(child, matches, lookup, &dev->dev, strict);
> +               if (rc) {
> +                       of_node_put(child);
> +                       return rc;
> +               }
> +       }
> +
> +       /* second step: create the other nodes */
> +       for_each_child_of_node(bus, child) {
> +               if (child->phandle != 0)
> +                       continue;
> +               pr_debug("   create child 2: %s\n", child->full_name);
>                 rc = of_platform_bus_create(child, matches, lookup, &dev->dev, strict);
>                 if (rc) {
>                         of_node_put(child);
>
>
> --
> Ken ar c'hentaƱ |             ** Breizh ha Linux atav! **
> Jef             |               http://moinejf.free.fr/
Jean-Francois Moine - Aug. 22, 2013, 8 a.m.
On Tue, 20 Aug 2013 11:13:24 +0100
Grant Likely <grant.likely@linaro.org> wrote:

> On Tue, Aug 20, 2013 at 11:01 AM, Jean-Francois Moine <moinejf@free.fr> wrote:
> > This patch populates the platform from the device tree into two steps:
> > the first step creates the nodes that are referenced by a phandle,
> > the second step creates the other nodes.
> >
> > This permits to reduce the number of PROBE_DEFERs.
> >
> > Signed-off-by: Jean-Francois Moine <moinejf@free.fr>
> > ---
> > A better way to reduce probe deferral could be sorting the nodes
> > according to their phandle level in the DT blob at compilation time ...  
> 
> Have you got measurements or statistics that show this making a
> difference? I suspect you'll find for boot time it will have little to
> no affect since the device driver probe order is more closely related
> to the kernel link order than the order that devices were registered.

With the device tree and most drivers as modules, the kernel link order
does not matter.

I admit that the gain may be small: I just get none or just one probe
deferral instead of 3 on my cubox with this patch.
Grant Likely - Aug. 22, 2013, 10:20 a.m.
On Thu, Aug 22, 2013 at 9:00 AM, Jean-Francois Moine <moinejf@free.fr> wrote:
> On Tue, 20 Aug 2013 11:13:24 +0100
> Grant Likely <grant.likely@linaro.org> wrote:
>
>> On Tue, Aug 20, 2013 at 11:01 AM, Jean-Francois Moine <moinejf@free.fr> wrote:
>> > This patch populates the platform from the device tree into two steps:
>> > the first step creates the nodes that are referenced by a phandle,
>> > the second step creates the other nodes.
>> >
>> > This permits to reduce the number of PROBE_DEFERs.
>> >
>> > Signed-off-by: Jean-Francois Moine <moinejf@free.fr>
>> > ---
>> > A better way to reduce probe deferral could be sorting the nodes
>> > according to their phandle level in the DT blob at compilation time ...
>>
>> Have you got measurements or statistics that show this making a
>> difference? I suspect you'll find for boot time it will have little to
>> no affect since the device driver probe order is more closely related
>> to the kernel link order than the order that devices were registered.
>
> With the device tree and most drivers as modules, the kernel link order
> does not matter.
>
> I admit that the gain may be small: I just get none or just one probe
> deferral instead of 3 on my cubox with this patch.

If you've only got 3 probe deferrals, then is this really an issue
worth complicating the code over? I'd like to see a more significant
impact before taking this step.

g.

Patch

diff --git a/drivers/of/platform.c b/drivers/of/platform.c
index e0a6514..a2ea858 100644
--- a/drivers/of/platform.c
+++ b/drivers/of/platform.c
@@ -382,8 +382,23 @@  static int of_platform_bus_create(struct device_node *bus,
 	if (!dev || !of_match_node(matches, bus))
 		return 0;
 
+	/* first step: create the nodes that are referenced by phandle */
 	for_each_child_of_node(bus, child) {
-		pr_debug("   create child: %s\n", child->full_name);
+		if (child->phandle == 0)
+			continue;
+		pr_debug("   create child 1: %s\n", child->full_name);
+		rc = of_platform_bus_create(child, matches, lookup, &dev->dev, strict);
+		if (rc) {
+			of_node_put(child);
+			return rc;
+		}
+	}
+
+	/* second step: create the other nodes */
+	for_each_child_of_node(bus, child) {
+		if (child->phandle != 0)
+			continue;
+		pr_debug("   create child 2: %s\n", child->full_name);
 		rc = of_platform_bus_create(child, matches, lookup, &dev->dev, strict);
 		if (rc) {
 			of_node_put(child);