Skip to contentSkip to conversation

fdtdec: fdtdec_get_aliases_highest_id: skip aliases to disabled nodes

PATCH New Tim Harvey 3 participants 9 messages Apr 16, 2021 – Jul 4, 2021 Library code

lore
Share & tools
lore

Tim Harvey

author
#0 lore raw

PATCH 1/1 fdtdec: fdtdec_get_aliases_highest_id: skip aliases to disabled nodes

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
 lib/fdtdec.c | 2 ++
 1 file changed, 2 insertions(+)
lib/fdtdec.c +2 -0
index 864589193b..d47195525a 100644
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
 		if (*prop != '/' || prop[len - 1] ||
 		    strncmp(name, base, base_len))
 			continue;
+		if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+			continue;
 
 		val = trailing_strtol(name);
 		if (val > max) {
Signature · 1 line
2.17.1
#1 lore raw

PATCH 1/1 fdtdec: fdtdec_get_aliases_highest_id: skip aliases to disabled nodes

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
 lib/fdtdec.c | 2 ++
 1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                if (*prop != '/' || prop[len - 1] ||
                    strncmp(name, base, base_len))
                        continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Quoted 6 lines
lib/fdtdec.c +0 -1
                val = trailing_strtol(name);
                if (val > max) {
--

2.17.1

Regards,
Simon

author
#2 lore raw

PATCH 1/1 fdtdec: fdtdec_get_aliases_highest_id: skip aliases to disabled nodes

Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
 lib/fdtdec.c | 2 ++
 1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                if (*prop != '/' || prop[len - 1] ||
                    strncmp(name, base, base_len))
                        continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

In general it is just wrong to reserve id's for disabled devices.

Tim

author
#3 lore raw
Quoted 42 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
 lib/fdtdec.c | 2 ++
 1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                if (*prop != '/' || prop[len - 1] ||
                    strncmp(name, base, base_len))
                        continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Quoted 2 lines

In general it is just wrong to reserve id's for disabled devices.

Simon,

Do you have any additional feedback on this? It has not been picked up yet and does cause issues on IMX boards using dm that do not use the internal SoC's fec ethernet.

Perhaps there is a better way to resolve this?

Best regards,

Tim

#4 lore raw
Quoted 35 lines · Tim Harvey
Quoted 34 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
  lib/fdtdec.c | 2 ++
  1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                 if (*prop != '/' || prop[len - 1] ||
                     strncmp(name, base, base_len))
                         continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

Quoted 7 lines
Quoted 7 lines

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Is this due getting mac address from the OTP? As I understand it, the OTP mac addresses are for fec0/fec1, and not for whatever ethernet0 happens to be.

--Sean

Quoted 16 lines
Quoted 3 lines

In general it is just wrong to reserve id's for disabled devices.

Simon,

Do you have any additional feedback on this? It has not been picked up yet and does cause issues on IMX boards using dm that do not use the internal SoC's fec ethernet.

Perhaps there is a better way to resolve this?

Best regards,

Tim

author
#5 lore raw
Quoted 47 lines · Sean Anderson
Quoted 35 lines · Tim Harvey
Quoted 34 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
  lib/fdtdec.c | 2 ++
  1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                 if (*prop != '/' || prop[len - 1] ||
                     strncmp(name, base, base_len))
                         continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

Quoted 7 lines
Quoted 7 lines

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Quoted 4 lines

Is this due getting mac address from the OTP? As I understand it, the OTP mac addresses are for fec0/fec1, and not for whatever ethernet0 happens to be.

Sean,

No, the issue is that U-Boot thinks there is an fec device simply because an alias points to it even though the fec node pointed to is disabled so when U-Boot finds an ethernet device that is enumerated at runtime (think pcie or usb based mac) those start at index 1 instead of index 0 and end up wanting to be assigned mac's from eth1addr instead of ethaddr.

Tim

#6 lore raw
Quoted 53 lines · Tim Harvey
Quoted 47 lines · Sean Anderson
Quoted 35 lines · Tim Harvey
Quoted 34 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
  lib/fdtdec.c | 2 ++
  1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                 if (*prop != '/' || prop[len - 1] ||
                     strncmp(name, base, base_len))
                         continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

Quoted 7 lines
Quoted 7 lines

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Quoted 4 lines

Is this due getting mac address from the OTP? As I understand it, the OTP mac addresses are for fec0/fec1, and not for whatever ethernet0 happens to be.

Quoted 8 lines

Sean,

No, the issue is that U-Boot thinks there is an fec device simply because an alias points to it even though the fec node pointed to is disabled so when U-Boot finds an ethernet device that is enumerated at runtime (think pcie or usb based mac) those start at index 1 instead of index 0 and end up wanting to be assigned mac's from eth1addr instead of ethaddr.

Ok, and this behavior was different in the past? E.g. it used to be that these interfaces were assigned to ethernet0, and so their mac address was stored in ethaddr. And then something changed and now they are ethernet1 and are trying to get the incorrect mac address?

Do you know whether it would be possible to remove the alias at the same time you mark the fec as disabled?

--Sean

author
#7 lore raw
Quoted 66 lines · Sean Anderson
Quoted 53 lines · Tim Harvey
Quoted 47 lines · Sean Anderson
Quoted 35 lines · Tim Harvey
Quoted 34 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
  lib/fdtdec.c | 2 ++
  1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                 if (*prop != '/' || prop[len - 1] ||
                     strncmp(name, base, base_len))
                         continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

Quoted 7 lines
Quoted 7 lines

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Quoted 4 lines

Is this due getting mac address from the OTP? As I understand it, the OTP mac addresses are for fec0/fec1, and not for whatever ethernet0 happens to be.

Quoted 8 lines

Sean,

No, the issue is that U-Boot thinks there is an fec device simply because an alias points to it even though the fec node pointed to is disabled so when U-Boot finds an ethernet device that is enumerated at runtime (think pcie or usb based mac) those start at index 1 instead of index 0 and end up wanting to be assigned mac's from eth1addr instead of ethaddr.

Quoted 5 lines

Ok, and this behavior was different in the past? E.g. it used to be that these interfaces were assigned to ethernet0, and so their mac address was stored in ethaddr. And then something changed and now they are ethernet1 and are trying to get the incorrect mac address?

Sean,

Yes, in the past before using driver-model this did not occur as everything was statically declared in board files.

Quoted 3 lines

Do you know whether it would be possible to remove the alias at the same time you mark the fec as disabled?

The dt's are all merged together. The base SoC dt declares the alias to internal fec ethernet device and that alias can be overridden by any dts that includes that base soc dt but if you have a gbe that is discovered and not declared in dt there will be no alias to override it. When I submitted this patch I looked at what the kernel code did and noticed it did not reserve indexes/slots for disabled devices so I figure that's what the bootloader should do also.

Tim

#8 lore raw

Hi Tim,

Quoted 83 lines · Tim Harvey
Quoted 66 lines · Sean Anderson
Quoted 53 lines · Tim Harvey
Quoted 47 lines · Sean Anderson
Quoted 35 lines · Tim Harvey
Quoted 34 lines · Tim Harvey
Quoted 27 lines · Simon Glass

Hi Tim,

Quoted 14 lines · Tim Harvey

When looking for an alias with the highest id skip aliases for nodes that are disabled.

---
  lib/fdtdec.c | 2 ++
  1 file changed, 2 insertions(+)
lib/fdtdec.c +0 -0
index 864589193b..d47195525a 100644

lib/fdtdec.c int fdtdec_get_alias_highest_id(const void *blob, const char *base) · line 550

Quoted 5 lines
lib/fdtdec.c +2 -0
@@ -546,6 +546,8 @@ int fdtdec_get_alias_highest_id(const void *blob, const char *base)
                 if (*prop != '/' || prop[len - 1] ||
                     strncmp(name, base, base_len))
                         continue;
+               if (!fdtdec_get_is_enabled(blob, fdt_path_offset(blob, prop)))
+                       continue;

We really can't do this here. It is quite an expensive operation to locate the node for a path.

Quoted 2 lines

Why is this needed? It seems odd to have an alias pointing to a disabled device.

Simon,

Quoted 7 lines
Quoted 7 lines

The issue I ran into here was with an imx6 based board that does not use the FEC ethernet on the SoC. In this case imx6qdl.dtsi delcares an alias 'thernet0 = &fec' yet the fec node is not enabled. However because fdtdec_get_alias_highest_id does not skip this alias to a disabled device any enumerated ethernet devices get an index of 1 instead of 0 which is incorrect and causes the mac addresses to be misaligned.

Quoted 4 lines

Is this due getting mac address from the OTP? As I understand it, the OTP mac addresses are for fec0/fec1, and not for whatever ethernet0 happens to be.

Quoted 8 lines

Sean,

No, the issue is that U-Boot thinks there is an fec device simply because an alias points to it even though the fec node pointed to is disabled so when U-Boot finds an ethernet device that is enumerated at runtime (think pcie or usb based mac) those start at index 1 instead of index 0 and end up wanting to be assigned mac's from eth1addr instead of ethaddr.

Quoted 5 lines

Ok, and this behavior was different in the past? E.g. it used to be that these interfaces were assigned to ethernet0, and so their mac address was stored in ethaddr. And then something changed and now they are ethernet1 and are trying to get the incorrect mac address?

Sean,

Yes, in the past before using driver-model this did not occur as everything was statically declared in board files.

Quoted 3 lines

Do you know whether it would be possible to remove the alias at the same time you mark the fec as disabled?

Quoted 7 lines

The dt's are all merged together. The base SoC dt declares the alias to internal fec ethernet device and that alias can be overridden by any dts that includes that base soc dt but if you have a gbe that is discovered and not declared in dt there will be no alias to override it. When I submitted this patch I looked at what the kernel code did and noticed it did not reserve indexes/slots for disabled devices so I figure that's what the bootloader should do also.

Can you try enabled OF_LIVE? It is possible that it might be different there. Even if not, it should be a cheap fix.

As mentioned the problem with your patch is that the path operation is very expensive. I'm fine with it if you want to add it as a new CONFIG option, so it doesn't affect all platforms. But if you do that, please do update test-fdt.c for the sandbox tests.

Regards,
Simon