Skip to contentSkip to conversation

dm: uclass: don't assign aliased seq numbers

PATCH New Michael Walle 3 participants 5 messages Dec 18, 2019 – Dec 19, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dm: uclass: don't assign aliased seq numbers

If there are aliases for an uclass, set the base for the "dynamically" allocated numbers next to the highest alias.

Please note, that this might lead to holes in the sequences, depending on the device tree. For example if there is only an alias "ethernet1", the next device seq number would be 2.

In particular this fixes a problem with boards which are using ethernet aliases but also might have network add-in cards like the E1000. If the board is started with the add-in card and depending on the order of the drivers, the E1000 might occupy the first ethernet device and mess up all the hardware addresses, because the devices are now shifted by one.

As a side effect, this should also make the following commits superfluous:
- 7f3289bf6d ("dm: device: Request next sequence number")
- 61607225d1 ("i2c: Fill req_seq in i2c_post_bind()") Although I don't understand the root cause of the said problem.

Thomas, Michal, could you please test this and then I'd add a second patch removing the old code.

Cc Thomas Fitzsimmons <[email protected]>
Cc Michal Simek <[email protected]>
---
 drivers/core/uclass.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)
drivers/core/uclass.c +14 -6
index c520ef113a..c3b325141a 100644
@@ -675,13 +675,14 @@ int uclass_unbind_device(struct udevice *dev)
 
 int uclass_resolve_seq(struct udevice *dev)
 {
+	struct uclass *uc = dev->uclass;
+	struct uclass_driver *uc_drv = uc->uc_drv;
 	struct udevice *dup;
-	int seq;
+	int seq = 0;
 	int ret;
 
 	assert(dev->seq == -1);
-	ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, dev->req_seq,
-					false, &dup);
+	ret = uclass_find_device_by_seq(uc_drv->id, dev->req_seq, false, &dup);
 	if (!ret) {
 		dm_warn("Device '%s': seq %d is in use by '%s'\n",
 			dev->name, dev->req_seq, dup->name);
@@ -693,9 +694,16 @@ int uclass_resolve_seq(struct udevice *dev)
 		return ret;
 	}
 
-	for (seq = 0; seq < DM_MAX_SEQ; seq++) {
-		ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, seq,
-						false, &dup);
+	if (CONFIG_IS_ENABLED(DM_SEQ_ALIAS) &&
+	    (uc_drv->flags & DM_UC_FLAG_SEQ_ALIAS)) {
+		/* dev_read_alias_highest_id() will return -1 if there no
+		 * alias. Thus we can always add one.
+		 */
+		seq = dev_read_alias_highest_id(uc_drv->name) + 1;
+	}
+
+	for (; seq < DM_MAX_SEQ; seq++) {
+		ret = uclass_find_device_by_seq(uc_drv->id, seq, false, &dup);
 		if (ret == -ENODEV)
 			break;
 		if (ret)
Signature · 1 line
2.20.1
#1 lore raw

PATCH 1/1 dm: uclass: don't assign aliased seq numbers

Hi Michael,

Quoted 53 lines · Michael Walle

If there are aliases for an uclass, set the base for the "dynamically" allocated numbers next to the highest alias.

Please note, that this might lead to holes in the sequences, depending on the device tree. For example if there is only an alias "ethernet1", the next device seq number would be 2.

In particular this fixes a problem with boards which are using ethernet aliases but also might have network add-in cards like the E1000. If the board is started with the add-in card and depending on the order of the drivers, the E1000 might occupy the first ethernet device and mess up all the hardware addresses, because the devices are now shifted by one.

As a side effect, this should also make the following commits superfluous:
- 7f3289bf6d ("dm: device: Request next sequence number")
- 61607225d1 ("i2c: Fill req_seq in i2c_post_bind()") Although I don't understand the root cause of the said problem.

Thomas, Michal, could you please test this and then I'd add a second patch removing the old code.

Cc Thomas Fitzsimmons <[email protected]>
Cc Michal Simek <[email protected]>
---
  drivers/core/uclass.c | 20 ++++++++++++++------
  1 file changed, 14 insertions(+), 6 deletions(-)
drivers/core/uclass.c +4 -3
index c520ef113a..c3b325141a 100644
@@ -675,13 +675,14 @@ int uclass_unbind_device(struct udevice *dev)
 
  int uclass_resolve_seq(struct udevice *dev)
  {
+	struct uclass *uc = dev->uclass;
+	struct uclass_driver *uc_drv = uc->uc_drv;
  	struct udevice *dup;
-	int seq;
+	int seq = 0;
  	int ret;
 
  	assert(dev->seq == -1);
-	ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, dev->req_seq,
-					false, &dup);
+	ret = uclass_find_device_by_seq(uc_drv->id, dev->req_seq, false, &dup);
  	if (!ret) {
  		dm_warn("Device '%s': seq %d is in use by '%s'\n",
  			dev->name, dev->req_seq, dup->name);

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 699

Quoted 9 lines
drivers/core/uclass.c +3 -3
@@ -693,9 +694,16 @@ int uclass_resolve_seq(struct udevice *dev)
  		return ret;
  	}
 
-	for (seq = 0; seq < DM_MAX_SEQ; seq++) {
-		ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, seq,
-						false, &dup);
+	if (CONFIG_IS_ENABLED(DM_SEQ_ALIAS) &&
+	    (uc_drv->flags & DM_UC_FLAG_SEQ_ALIAS)) {
+		/* dev_read_alias_highest_id() will return -1 if there no

nits: there /is/ no and multi-line comment starts with just /* on the 1st line

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 709

Quoted 11 lines
drivers/core/uclass.c +7 -0
+		 * alias. Thus we can always add one.
+		 */
+		seq = dev_read_alias_highest_id(uc_drv->name) + 1;
+	}
+
+	for (; seq < DM_MAX_SEQ; seq++) {
+		ret = uclass_find_device_by_seq(uc_drv->id, seq, false, &dup);
  		if (ret == -ENODEV)
  			break;
  		if (ret)

This looks nice.
One thing I noticed is that 'dm tree' indexes now don't match dev->seq which can be a bit confusing. The index is produced by dev_get_uclass_index, which in effect counts devices in the uclass. Any issue if 'dm tree' prints dev->seq instead and maybe call the column Seq rather than Index?

Thanks!
Alex

#2 lore raw

PATCH 1/1 dm: uclass: don't assign aliased seq numbers

Am 2019-12-18 21:00, schrieb Alexandru Marginean:

Quoted 63 lines

Hi Michael,

Quoted 60 lines · Michael Walle

If there are aliases for an uclass, set the base for the "dynamically" allocated numbers next to the highest alias.

Please note, that this might lead to holes in the sequences, depending on the device tree. For example if there is only an alias "ethernet1", the next device seq number would be 2.

In particular this fixes a problem with boards which are using ethernet
aliases but also might have network add-in cards like the E1000. If the
board is started with the add-in card and depending on the order of the
drivers, the E1000 might occupy the first ethernet device and mess up all the hardware addresses, because the devices are now shifted by one.

As a side effect, this should also make the following commits superfluous:
- 7f3289bf6d ("dm: device: Request next sequence number")
- 61607225d1 ("i2c: Fill req_seq in i2c_post_bind()") Although I don't understand the root cause of the said problem.

Thomas, Michal, could you please test this and then I'd add a second patch removing the old code.

Cc Thomas Fitzsimmons <[email protected]>
Cc Michal Simek <[email protected]>
---
  drivers/core/uclass.c | 20 ++++++++++++++------
  1 file changed, 14 insertions(+), 6 deletions(-)
drivers/core/uclass.c +4 -3
index c520ef113a..c3b325141a 100644
@@ -675,13 +675,14 @@ int uclass_unbind_device(struct udevice *dev)
    int uclass_resolve_seq(struct udevice *dev)
  {
+	struct uclass *uc = dev->uclass;
+	struct uclass_driver *uc_drv = uc->uc_drv;
  	struct udevice *dup;
-	int seq;
+	int seq = 0;
  	int ret;
    	assert(dev->seq == -1);
-	ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, 
 dev->req_seq,
-					false, &dup);
+	ret = uclass_find_device_by_seq(uc_drv->id, dev->req_seq, false, 
 &dup);
  	if (!ret) {
  		dm_warn("Device '%s': seq %d is in use by '%s'\n",
  			dev->name, dev->req_seq, dup->name);
@@ -693,9 +694,16 @@ int uclass_resolve_seq(struct udevice *dev)
  		return ret;
  	}
  -	for (seq = 0; seq < DM_MAX_SEQ; seq++) {
Quoted 8 lines

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 699

Quoted 5 lines
drivers/core/uclass.c +3 -2
-		ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, seq,
-						false, &dup);
+	if (CONFIG_IS_ENABLED(DM_SEQ_ALIAS) &&
+	    (uc_drv->flags & DM_UC_FLAG_SEQ_ALIAS)) {
+		/* dev_read_alias_highest_id() will return -1 if there no

nits: there /is/ no and multi-line comment starts with just /* on the 1st line

yeah, I'd prefer that style too. But checkpatch.pl would complain about it.. or so I thought.. well seems only to be the case in linux and only in "NETWORKING_BLOCK_COMMENT_STYLE" (only for net/ and drivers/net/).

I'll definitely fix that ;)

Quoted 17 lines

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 709

Quoted 12 lines
drivers/core/uclass.c +7 -0
+		 * alias. Thus we can always add one.
+		 */
+		seq = dev_read_alias_highest_id(uc_drv->name) + 1;
+	}
+
+	for (; seq < DM_MAX_SEQ; seq++) {
+		ret = uclass_find_device_by_seq(uc_drv->id, seq, false, &dup);
  		if (ret == -ENODEV)
  			break;
  		if (ret)
Quoted 6 lines

This looks nice.
One thing I noticed is that 'dm tree' indexes now don't match dev->seq which can be a bit confusing. The index is produced by dev_get_uclass_index, which in effect counts devices in the uclass. Any issue if 'dm tree' prints dev->seq instead and maybe call the column Seq rather than Index?

Mhh, did this ever match if req_seq/aliases was used? "dm uclass" dumps the seq and req_seq. So I don't know if that was ever intended to match. But I'm open to suggestions.

-michael

#3 lore raw

PATCH 1/1 dm: uclass: don't assign aliased seq numbers

Quoted 99 lines · Michael Walle

Am 2019-12-18 21:00, schrieb Alexandru Marginean:

Quoted 59 lines

Hi Michael,

Quoted 56 lines · Michael Walle

If there are aliases for an uclass, set the base for the "dynamically" allocated numbers next to the highest alias.

Please note, that this might lead to holes in the sequences, depending on the device tree. For example if there is only an alias "ethernet1", the next device seq number would be 2.

In particular this fixes a problem with boards which are using ethernet aliases but also might have network add-in cards like the E1000. If the board is started with the add-in card and depending on the order of the drivers, the E1000 might occupy the first ethernet device and mess up all the hardware addresses, because the devices are now shifted by one.

As a side effect, this should also make the following commits superfluous:
  - 7f3289bf6d ("dm: device: Request next sequence number")
  - 61607225d1 ("i2c: Fill req_seq in i2c_post_bind()") Although I don't understand the root cause of the said problem.

Thomas, Michal, could you please test this and then I'd add a second patch removing the old code.

Cc Thomas Fitzsimmons <[email protected]>
Cc Michal Simek <[email protected]>

---
  drivers/core/uclass.c | 20 ++++++++++++++------

  1 file changed, 14 insertions(+), 6 deletions(-)
drivers/core/uclass.c +4 -3
index c520ef113a..c3b325141a 100644
@@ -675,13 +675,14 @@ int uclass_unbind_device(struct udevice *dev)
     int uclass_resolve_seq(struct udevice *dev)
   {
+    struct uclass *uc = dev->uclass;
+    struct uclass_driver *uc_drv = uc->uc_drv;
       struct udevice *dup;
-    int seq;
+    int seq = 0;
       int ret;
         assert(dev->seq == -1);
-    ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, 
 dev->req_seq,
-                    false, &dup);
+    ret = uclass_find_device_by_seq(uc_drv->id, dev->req_seq, false, 
 &dup);
       if (!ret) {
           dm_warn("Device '%s': seq %d is in use by '%s'\n",
               dev->name, dev->req_seq, dup->name);
@@ -693,9 +694,16 @@ int uclass_resolve_seq(struct udevice *dev)
           return ret;
       }
   -    for (seq = 0; seq < DM_MAX_SEQ; seq++) {
Quoted 8 lines

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 699

Quoted 5 lines
drivers/core/uclass.c +3 -2
-        ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, seq,
-                        false, &dup);
+    if (CONFIG_IS_ENABLED(DM_SEQ_ALIAS) &&
+        (uc_drv->flags & DM_UC_FLAG_SEQ_ALIAS)) {
+        /* dev_read_alias_highest_id() will return -1 if there no

nits: there /is/ no and multi-line comment starts with just /* on the 1st line

yeah, I'd prefer that style too. But checkpatch.pl would complain about it.. or so I thought.. well seems only to be the case in linux and only in "NETWORKING_BLOCK_COMMENT_STYLE" (only for net/ and drivers/net/).

I'll definitely fix that ;)

Quoted 17 lines

drivers/core/uclass.c int uclass_resolve_seq(struct udevice *dev) · line 709

Quoted 12 lines
drivers/core/uclass.c +7 -0
+         * alias. Thus we can always add one.
+         */
+        seq = dev_read_alias_highest_id(uc_drv->name) + 1;
+    }
+
+    for (; seq < DM_MAX_SEQ; seq++) {
+        ret = uclass_find_device_by_seq(uc_drv->id, seq, false, &dup);
           if (ret == -ENODEV)
               break;
           if (ret)
Quoted 6 lines

This looks nice.
One thing I noticed is that 'dm tree' indexes now don't match dev->seq which can be a bit confusing.  The index is produced by dev_get_uclass_index, which in effect counts devices in the uclass. Any issue if 'dm tree' prints dev->seq instead and maybe call the column Seq rather than Index?

Quoted 5 lines

Mhh, did this ever match if req_seq/aliases was used? "dm uclass" dumps the seq and req_seq. So I don't know if that was ever intended to match. But I'm open to suggestions.

-michael

It's certainly not something that should be part of this patch and since I think it should be changed I can send a patch for it.

Thanks!
Alex

#4 lore raw

PATCH 1/1 dm: uclass: don't assign aliased seq numbers

In reply to Michael Walle

Hi Michael,

Quoted 20 lines · Michael Walle

If there are aliases for an uclass, set the base for the "dynamically" allocated numbers next to the highest alias.

Please note, that this might lead to holes in the sequences, depending on the device tree. For example if there is only an alias "ethernet1", the next device seq number would be 2.

In particular this fixes a problem with boards which are using ethernet aliases but also might have network add-in cards like the E1000. If the board is started with the add-in card and depending on the order of the drivers, the E1000 might occupy the first ethernet device and mess up all the hardware addresses, because the devices are now shifted by one.

As a side effect, this should also make the following commits superfluous:
- 7f3289bf6d ("dm: device: Request next sequence number")
- 61607225d1 ("i2c: Fill req_seq in i2c_post_bind()") Although I don't understand the root cause of the said problem.

Quoted 8 lines

Thomas, Michal, could you please test this and then I'd add a second patch removing the old code.

Cc Thomas Fitzsimmons <[email protected]>
Cc Michal Simek <[email protected]>

---

This sounds so obvious that it makes me wonder why it hasn't been the behavior thus far.

Quoted 48 lines
 drivers/core/uclass.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)
drivers/core/uclass.c +14 -7
index c520ef113a..c3b325141a 100644
@@ -675,13 +675,14 @@ int uclass_unbind_device(struct udevice *dev)
 
 int uclass_resolve_seq(struct udevice *dev)
 {
+       struct uclass *uc = dev->uclass;
+       struct uclass_driver *uc_drv = uc->uc_drv;
        struct udevice *dup;
-       int seq;
+       int seq = 0;
        int ret;
 
        assert(dev->seq == -1);
-       ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, dev->req_seq,
-                                       false, &dup);
+       ret = uclass_find_device_by_seq(uc_drv->id, dev->req_seq, false, &dup);
        if (!ret) {
                dm_warn("Device '%s': seq %d is in use by '%s'\n",
                        dev->name, dev->req_seq, dup->name);
@@ -693,9 +694,16 @@ int uclass_resolve_seq(struct udevice *dev)
                return ret;
        }
 
-       for (seq = 0; seq < DM_MAX_SEQ; seq++) {
-               ret = uclass_find_device_by_seq(dev->uclass->uc_drv->id, seq,
-                                               false, &dup);
+       if (CONFIG_IS_ENABLED(DM_SEQ_ALIAS) &&
+           (uc_drv->flags & DM_UC_FLAG_SEQ_ALIAS)) {
+               /* dev_read_alias_highest_id() will return -1 if there no
+                * alias. Thus we can always add one.
+                */
+               seq = dev_read_alias_highest_id(uc_drv->name) + 1;
+       }
+
+       for (; seq < DM_MAX_SEQ; seq++) {
+               ret = uclass_find_device_by_seq(uc_drv->id, seq, false, &dup);
                if (ret == -ENODEV)
                        break;
                if (ret)
--

2.20.1