Skip to contentSkip to conversation

net: mdiobus: reset deassert delay

PATCH In review Bruno Thomsen 3 participants 6 messages Jul 28, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 net: mdiobus: reset deassert delay

The current reset logic only has a delay during assert. This reuses the delay value as deassert delay to ensure PHYs are ready for commands. Delays are typically needed when external hardware slows down reset release with a RC network. This solution does not need any new device tree bindings.
It also improves handling of long delays (>20ms) by using the generic fsleep() for selecting appropriate delay function.

---
 drivers/net/phy/mdio_bus.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/phy/mdio_bus.c +2 -1
index 6ceee82b2839..84d5ab07fe16 100644
@@ -627,8 +627,9 @@ int __mdiobus_register(struct mii_bus *bus, struct module *owner)
 		bus->reset_gpiod = gpiod;
 
 		gpiod_set_value_cansleep(gpiod, 1);
-		udelay(bus->reset_delay_us);
+		fsleep(bus->reset_delay_us);
 		gpiod_set_value_cansleep(gpiod, 0);
+		fsleep(bus->reset_delay_us);
 	}
 
 	if (bus->reset) {

base-commit: 92ed301919932f777713b9172e525674157e983d

Signature · 1 line
2.26.2
#1 lore raw

PATCH 2/2 dt-bindings: net: mdio: update reset-delay-us description

Property is now used for both assert and deassert delay so PHYs can be reset before type id auto detection.

---
 Documentation/devicetree/bindings/net/mdio.yaml | 2 ++
 1 file changed, 2 insertions(+)
Documentation/devicetree/bindings/net/mdio.yaml +2 -0
index d6a3bf8550eb..8385960e3b3e 100644
@@ -38,6 +38,8 @@ properties:
       RESET pulse width in microseconds. It applies to all MDIO devices
       and must therefore be appropriately determined based on all devices
       requirements (maximum value of all per-device RESET pulse widths).
+      Additional it also provides a reset deassert delay of the same width
+      to ensure devices are ready for communication.
 
   clock-frequency:
     description:
Signature · 1 line
2.26.2
#2 lore raw
In reply to Bruno Thomsen

Hi Bruno,

Quoted 10 lines · Bruno Thomsen

The current reset logic only has a delay during assert. This reuses the delay value as deassert delay to ensure PHYs are ready for commands. Delays are typically needed when external hardware slows down reset release with a RC network. This solution does not need any new device tree bindings.
It also improves handling of long delays (>20ms) by using the generic fsleep() for selecting appropriate delay function.

It seems that this patch should be split in two:

One that changes from udelay() to sleep()
and another one that adds the delay after the reset line is de-asserted.

Quoted 11 lines
---
 drivers/net/phy/mdio_bus.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/phy/mdio_bus.c +0 -0
index 6ceee82b2839..84d5ab07fe16 100644

drivers/net/phy/mdio_bus.c int __mdiobus_register(struct mii_bus *bus, struct module *owner) · line 632

Quoted 7 lines
drivers/net/phy/mdio_bus.c +2 -1
@@ -627,8 +627,9 @@ int __mdiobus_register(struct mii_bus *bus, struct module *owner)
                bus->reset_gpiod = gpiod;
 
                gpiod_set_value_cansleep(gpiod, 1);
-               udelay(bus->reset_delay_us);
+               fsleep(bus->reset_delay_us);
                gpiod_set_value_cansleep(gpiod, 0);
+               fsleep(bus->reset_delay_us);

Shouldn't it use the value passed in the reset-deassert-us property instead?

#3 lore raw

Hi Fabio

Den tir. 28. jul. 2020 kl. 14.32 skrev Fabio Estevam <[email protected]>:

Quoted 12 lines · Bruno Thomsen
Quoted 10 lines

The current reset logic only has a delay during assert. This reuses the delay value as deassert delay to ensure PHYs are ready for commands. Delays are typically needed when external hardware slows down reset release with a RC network. This solution does not need any new device tree bindings.
It also improves handling of long delays (>20ms) by using the generic fsleep() for selecting appropriate delay function.

Quoted 4 lines

It seems that this patch should be split in two:

One that changes from udelay() to sleep()
and another one that adds the delay after the reset line is de-asserted.

Okay, I will split it in version 2.

Quoted 10 lines
Quoted 10 lines
 drivers/net/phy/mdio_bus.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/phy/mdio_bus.c +0 -0
index 6ceee82b2839..84d5ab07fe16 100644
@@ -627,8 +627,9 @@ int __mdiobus_register(struct mii_bus *bus, struct module *owner)
                bus->reset_gpiod = gpiod;
 
Quoted 7 lines

drivers/net/phy/mdio_bus.c int __mdiobus_register(struct mii_bus *bus, struct module *owner) · line 632

Quoted 5 lines
drivers/net/phy/mdio_bus.c +2 -1
                gpiod_set_value_cansleep(gpiod, 1);
-               udelay(bus->reset_delay_us);
+               fsleep(bus->reset_delay_us);
                gpiod_set_value_cansleep(gpiod, 0);
+               fsleep(bus->reset_delay_us);

Shouldn't it use the value passed in the reset-deassert-us property instead?

This place does not use the per PHY reset properties as they are only used after auto type ID detection. When you need a reset before auto type ID detection,
it can only be done using the MDIO reset, and it only has "reset-delay-us" property.
It will be a rather big change if all PHY nodes can choose if reset should happen
before or after type ID detection.

/Bruno

maintainer
#4 lore raw
In reply to Fabio Estevam
Quoted 12 lines · Fabio Estevam
Quoted 12 lines
---
 drivers/net/phy/mdio_bus.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/phy/mdio_bus.c +0 -0
index 6ceee82b2839..84d5ab07fe16 100644
@@ -627,8 +627,9 @@ int __mdiobus_register(struct mii_bus *bus, struct module *owner)
                bus->reset_gpiod = gpiod;
 
Quoted 9 lines

drivers/net/phy/mdio_bus.c int __mdiobus_register(struct mii_bus *bus, struct module *owner) · line 632

Quoted 5 lines
drivers/net/phy/mdio_bus.c +2 -1
                gpiod_set_value_cansleep(gpiod, 1);
-               udelay(bus->reset_delay_us);
+               fsleep(bus->reset_delay_us);
                gpiod_set_value_cansleep(gpiod, 0);
+               fsleep(bus->reset_delay_us);

Shouldn't it use the value passed in the reset-deassert-us property instead?

Hi Fabio

As Bruno pointed out, that property is not relevant here. But i agree with you in principle. Bruno, please add a new optional property for the delay after releasing the reset.

Andrew

maintainer
#5 lore raw
In reply to Bruno Thomsen
Quoted 11 lines · Bruno Thomsen

The current reset logic only has a delay during assert. This reuses the delay value as deassert delay to ensure PHYs are ready for commands. Delays are typically needed when external hardware slows down reset release with a RC network. This solution does not need any new device tree bindings.
It also improves handling of long delays (>20ms) by using the generic fsleep() for selecting appropriate delay function.

Hi Bruno

In netdev, it is normal to include a [patch 0/2] for a patchset which explains the big picture. Also, please use [patch net-next ..] to indicate this is for the net-next tree. If you think these are fixes for stable, please base them on net, and use [patch net ..], and include Fixes: tags.

Andrew