Skip to contentSkip to conversation

Avoid regmap debugfs collisions in qcom llcc driver

PATCH In review Stephen Boyd 3 participants 10 messages Oct 8, 2019 – Oct 10, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v2 cover Avoid regmap debugfs collisions in qcom llcc driver

Now a two part series. These patches fix a debugfs name collision for the llcc regmaps and moves the config to a local variable to save on image size.

Changes from v1 (https://lkml.kernel.org/r/[email protected]):
* New second patch
* Dropped static
* See range-diff below!

Stephen Boyd (2):
soc: qcom: llcc: Name regmaps to avoid collisions
soc: qcom: llcc: Move regmap config to local variable

 drivers/soc/qcom/llcc-slice.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
Cc Venkata Narendra Kumar Gutta <[email protected]>
Cc Evan Green <[email protected]>

Range-diff against v1:

-:  ------------ > 1:  07bc0e8bdb6e soc: qcom: llcc: Name regmaps to avoid collisions
1:  0c54fc8a7ed6 ! 2:  5c4446e36783 soc: qcom: llcc: Name regmaps to avoid collisions

@@ Metadata
Author: Stephen Boyd <[email protected]>

## Commit message ##
- soc: qcom: llcc: Name regmaps to avoid collisions
+ soc: qcom: llcc: Move regmap config to local variable

- We'll end up with debugfs collisions if we don't give names to the
- regmaps created inside this driver. Copy the template config over into
- this function and give the regmap the same name as the resource name.
+ This is now a global variable that we're modifying to fix the name.
+ That isn't terribly thread safe and it's not necessary to be a global so
+ let's just move this to a local variable instead. This saves space in
+ the symtab and actually reduces kernel image size because the regmap
+ config is large and we can replace the initialization of that structure
+ with a memset and a few member assignments.

- Fixes: 7f9c136216c7 ("soc: qcom: Add broadcast base for Last Level Cache Controller (LLCC)")
- Cc: Venkata Narendra Kumar Gutta <[email protected]>
- Cc: Evan Green <[email protected]>

## drivers/soc/qcom/llcc-slice.c ##
@@ drivers/soc/qcom/llcc-slice.c

static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

--static const struct regmap_config llcc_regmap_config = {
+-static struct regmap_config llcc_regmap_config = {
- .reg_bits = 32,
- .reg_stride = 4,
- .val_bits = 32,
@@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
{
struct resource *res;
void __iomem *base;
-+ static struct regmap_config llcc_regmap_config = {
++ struct regmap_config llcc_regmap_config = {
+ .reg_bits = 32,
+ .reg_stride = 4,
+ .val_bits = 32,
@@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct

res = platform_get_resource_byname(pdev, IORESOURCE_MEM, name);
if (!res)
-@@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct platform_device *pdev,
- if (IS_ERR(base))
- return ERR_CAST(base);
-
-+ llcc_regmap_config.name = name;
- return devm_regmap_init_mmio(&pdev->dev, base, &llcc_regmap_config);
- }
-

base-commit: 8b0eed9f6e36a5488967b0acc51444d658dd711b

Signature · 1 line
Sent by a computer through tubes
#1 lore raw

PATCH v2 1/2 soc: qcom: llcc: Name regmaps to avoid collisions

We'll end up with debugfs collisions if we don't give names to the regmaps created by this driver. Change the name of the config before registering it so we don't collide in debugfs.

Fixes 7f9c136216c7 ("soc: qcom: Add broadcast base for Last Level Cache Controller (LLCC)")
Cc Venkata Narendra Kumar Gutta <[email protected]>
Cc Evan Green <[email protected]>
---
 drivers/soc/qcom/llcc-slice.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/soc/qcom/llcc-slice.c +2 -1
index 9090ea12eaf3..4a6111635f82 100644
@@ -48,7 +48,7 @@
 
 static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;
 
-static const struct regmap_config llcc_regmap_config = {
+static struct regmap_config llcc_regmap_config = {
 	.reg_bits = 32,
 	.reg_stride = 4,
 	.val_bits = 32,
@@ -323,6 +323,7 @@ static struct regmap *qcom_llcc_init_mmio(struct platform_device *pdev,
 	if (IS_ERR(base))
 		return ERR_CAST(base);
 
+	llcc_regmap_config.name = name;
 	return devm_regmap_init_mmio(&pdev->dev, base, &llcc_regmap_config);
 }
 
Signature · 1 line
Sent by a computer through tubes
Reviewed-by on v2 1/2
#2 lore raw

Re: [PATCH v2 1/2] soc: qcom: llcc: Name regmaps to avoid collisions

Quoted 9 lines · Stephen Boyd

We'll end up with debugfs collisions if we don't give names to the regmaps created by this driver. Change the name of the config before registering it so we don't collide in debugfs.

Fixes 7f9c136216c7 ("soc: qcom: Add broadcast base for Last Level Cache Controller (LLCC)")
Cc Venkata Narendra Kumar Gutta <[email protected]>
Cc Evan Green <[email protected]>
#3 lore raw

PATCH v2 2/2 soc: qcom: llcc: Move regmap config to local variable

In reply to Stephen Boyd

This is now a global variable that we're modifying to fix the name. That isn't terribly thread safe and it's not necessary to be a global so let's just move this to a local variable instead. This saves space in the symtab and actually reduces kernel image size because the regmap config is large and we can replace the initialization of that structure with a memset and a few member assignments.

Cc Venkata Narendra Kumar Gutta <[email protected]>
Cc Evan Green <[email protected]>
---
 drivers/soc/qcom/llcc-slice.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)
drivers/soc/qcom/llcc-slice.c +6 -7
index 4a6111635f82..50aea3f0be41 100644
@@ -48,13 +48,6 @@
 
 static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;
 
-static struct regmap_config llcc_regmap_config = {
-	.reg_bits = 32,
-	.reg_stride = 4,
-	.val_bits = 32,
-	.fast_io = true,
-};
-
 /**
  * llcc_slice_getd - get llcc slice descriptor
  * @uid: usecase_id for the client
@@ -314,6 +307,12 @@ static struct regmap *qcom_llcc_init_mmio(struct platform_device *pdev,
 {
 	struct resource *res;
 	void __iomem *base;
+	struct regmap_config llcc_regmap_config = {
+		.reg_bits = 32,
+		.reg_stride = 4,
+		.val_bits = 32,
+		.fast_io = true,
+	};
 
 	res = platform_get_resource_byname(pdev, IORESOURCE_MEM, name);
 	if (!res)
Signature · 1 line
Sent by a computer through tubes
#4 lore raw
In reply to Stephen Boyd
Quoted 49 lines · Stephen Boyd

Now a two part series. These patches fix a debugfs name collision for the llcc regmaps and moves the config to a local variable to save on image size.

Changes from v1 (https://lkml.kernel.org/r/[email protected]):
* New second patch
* Dropped static
* See range-diff below!

Stephen Boyd (2):
soc: qcom: llcc: Name regmaps to avoid collisions
soc: qcom: llcc: Move regmap config to local variable

 drivers/soc/qcom/llcc-slice.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
Cc Venkata Narendra Kumar Gutta <[email protected]>
Cc Evan Green <[email protected]>

Range-diff against v1:

-:  ------------ > 1:  07bc0e8bdb6e soc: qcom: llcc: Name regmaps to avoid collisions
1:  0c54fc8a7ed6 ! 2:  5c4446e36783 soc: qcom: llcc: Name regmaps to avoid collisions

@@ Metadata
Author: Stephen Boyd <[email protected]>

## Commit message ##
- soc: qcom: llcc: Name regmaps to avoid collisions
+ soc: qcom: llcc: Move regmap config to local variable

- We'll end up with debugfs collisions if we don't give names to the
- regmaps created inside this driver. Copy the template config over into
- this function and give the regmap the same name as the resource name.
+ This is now a global variable that we're modifying to fix the name.
+ That isn't terribly thread safe and it's not necessary to be a global so
+ let's just move this to a local variable instead. This saves space in
+ the symtab and actually reduces kernel image size because the regmap
+ config is large and we can replace the initialization of that structure
+ with a memset and a few member assignments.

- Fixes: 7f9c136216c7 ("soc: qcom: Add broadcast base for Last Level Cache Controller (LLCC)")
- Cc: Venkata Narendra Kumar Gutta <[email protected]>
- Cc: Evan Green <[email protected]>

## drivers/soc/qcom/llcc-slice.c ##
@@ drivers/soc/qcom/llcc-slice.c

static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

Quoted 11 lines

--static const struct regmap_config llcc_regmap_config = {
+-static struct regmap_config llcc_regmap_config = {
- .reg_bits = 32,
- .reg_stride = 4,
- .val_bits = 32,
@@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
{
struct resource *res;
void __iomem *base;
-+ static struct regmap_config llcc_regmap_config = {
++ struct regmap_config llcc_regmap_config = {

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Regards,
Bjorn

Quoted 20 lines
     +		.reg_bits = 32,
     +		.reg_stride = 4,
     +		.val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      
      	res = platform_get_resource_byname(pdev, IORESOURCE_MEM, name);
      	if (!res)
    -@@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct platform_device *pdev,
    - 	if (IS_ERR(base))
    - 		return ERR_CAST(base);
    - 
    -+	llcc_regmap_config.name = name;
    - 	return devm_regmap_init_mmio(&pdev->dev, base, &llcc_regmap_config);
    - }
    - 

base-commit: 8b0eed9f6e36a5488967b0acc51444d658dd711b

Signature · 1 line
Sent by a computer through tubes
#5 lore raw

Quoting Bjorn Andersson (2019-10-08 16:55:04)

Quoted 17 lines · Stephen Boyd
Quoted 7 lines
    @@ drivers/soc/qcom/llcc-slice.c
      
      static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;
      
    --static const struct regmap_config llcc_regmap_config = {
    +-static struct regmap_config llcc_regmap_config = {
     -        .reg_bits = 32,
Quoted 8 lines
     -        .reg_stride = 4,
     -        .val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      {
              struct resource *res;
              void __iomem *base;
    -+        static struct regmap_config llcc_regmap_config = {
    ++        struct regmap_config llcc_regmap_config = {
Quoted 4 lines

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Well I split it into bug fix and micro-optimization so backport choices can be made. But yeah, I hope Evan is happy enough to provide a reviewed-by tag!

#6 lore raw
Quoted 19 lines · Stephen Boyd

Quoting Bjorn Andersson (2019-10-08 16:55:04)

Quoted 17 lines · Stephen Boyd
Quoted 15 lines
    @@ drivers/soc/qcom/llcc-slice.c

      static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

    --static const struct regmap_config llcc_regmap_config = {
    +-static struct regmap_config llcc_regmap_config = {
     -        .reg_bits = 32,
     -        .reg_stride = 4,
     -        .val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      {
              struct resource *res;
              void __iomem *base;
    -+        static struct regmap_config llcc_regmap_config = {
    ++        struct regmap_config llcc_regmap_config = {
Quoted 8 lines
Quoted 4 lines

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Well I split it into bug fix and micro-optimization so backport choices can be made. But yeah, I hope Evan is happy enough to provide a reviewed-by tag!

It's definitely better without the static local since it no longer has the cognitive trap, but I still don't really get why we're messing with the global v. local aspect of it. We're now inconsistent with every other caller of this function, and for what exactly? We've traded some data space for a call to memset() and some instructions. I would have thought anecdotally that memory was the cheaper thing (ie cpu speeds stopped increasing awhile ago, but memory is still getting cheaper).

But either way it's correct, so really it's fine if you ignore me :) -Evan

#7 lore raw
Quoted 29 lines · Evan Green
Quoted 19 lines · Stephen Boyd

Quoting Bjorn Andersson (2019-10-08 16:55:04)

Quoted 17 lines · Stephen Boyd
Quoted 15 lines
    @@ drivers/soc/qcom/llcc-slice.c

      static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

    --static const struct regmap_config llcc_regmap_config = {
    +-static struct regmap_config llcc_regmap_config = {
     -        .reg_bits = 32,
     -        .reg_stride = 4,
     -        .val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      {
              struct resource *res;
              void __iomem *base;
    -+        static struct regmap_config llcc_regmap_config = {
    ++        struct regmap_config llcc_regmap_config = {
Quoted 8 lines
Quoted 4 lines

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Well I split it into bug fix and micro-optimization so backport choices can be made. But yeah, I hope Evan is happy enough to provide a reviewed-by tag!

Quoted 9 lines

It's definitely better without the static local since it no longer has the cognitive trap, but I still don't really get why we're messing with the global v. local aspect of it. We're now inconsistent with every other caller of this function, and for what exactly? We've traded some data space for a call to memset() and some instructions. I would have thought anecdotally that memory was the cheaper thing (ie cpu speeds stopped increasing awhile ago, but memory is still getting cheaper).

The reason for making the structure local is because it's being modified per instance, meaning it would still work as long as qcom_llcc_init_mmio() is never called concurrently for two llcc instances. But the correctness outweighs the performance degradation of setting it up on the stack in my view.

Or am I missing something?

Regards,
Bjorn

Quoted 2 lines

But either way it's correct, so really it's fine if you ignore me :) -Evan

Reviewed-by
#8 lore raw
Quoted 42 lines · Bjorn Andersson
Quoted 29 lines · Evan Green
Quoted 19 lines · Stephen Boyd

Quoting Bjorn Andersson (2019-10-08 16:55:04)

Quoted 17 lines · Stephen Boyd
Quoted 15 lines
    @@ drivers/soc/qcom/llcc-slice.c

      static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

    --static const struct regmap_config llcc_regmap_config = {
    +-static struct regmap_config llcc_regmap_config = {
     -        .reg_bits = 32,
     -        .reg_stride = 4,
     -        .val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      {
              struct resource *res;
              void __iomem *base;
    -+        static struct regmap_config llcc_regmap_config = {
    ++        struct regmap_config llcc_regmap_config = {
Quoted 8 lines
Quoted 4 lines

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Well I split it into bug fix and micro-optimization so backport choices can be made. But yeah, I hope Evan is happy enough to provide a reviewed-by tag!

Quoted 9 lines

It's definitely better without the static local since it no longer has the cognitive trap, but I still don't really get why we're messing with the global v. local aspect of it. We're now inconsistent with every other caller of this function, and for what exactly? We've traded some data space for a call to memset() and some instructions. I would have thought anecdotally that memory was the cheaper thing (ie cpu speeds stopped increasing awhile ago, but memory is still getting cheaper).

Quoted 6 lines

The reason for making the structure local is because it's being modified per instance, meaning it would still work as long as qcom_llcc_init_mmio() is never called concurrently for two llcc instances. But the correctness outweighs the performance degradation of setting it up on the stack in my view.

I hadn't considered the concurrency aspect of the change, since I had anchored myself on the static local. I'm convinced. Might be worth mentioning that in the commit message.

For the series:

#9 lore raw
Quoted 51 lines · Evan Green
Quoted 42 lines · Bjorn Andersson
Quoted 29 lines · Evan Green
Quoted 19 lines · Stephen Boyd

Quoting Bjorn Andersson (2019-10-08 16:55:04)

Quoted 17 lines · Stephen Boyd
Quoted 15 lines
    @@ drivers/soc/qcom/llcc-slice.c

      static struct llcc_drv_data *drv_data = (void *) -EPROBE_DEFER;

    --static const struct regmap_config llcc_regmap_config = {
    +-static struct regmap_config llcc_regmap_config = {
     -        .reg_bits = 32,
     -        .reg_stride = 4,
     -        .val_bits = 32,
    @@ drivers/soc/qcom/llcc-slice.c: static struct regmap *qcom_llcc_init_mmio(struct
      {
              struct resource *res;
              void __iomem *base;
    -+        static struct regmap_config llcc_regmap_config = {
    ++        struct regmap_config llcc_regmap_config = {
Quoted 8 lines
Quoted 4 lines

Now that this isn't static I like the end result better. Not sure about the need for splitting it in two patches, but if Evan is happy I'll take it.

Well I split it into bug fix and micro-optimization so backport choices can be made. But yeah, I hope Evan is happy enough to provide a reviewed-by tag!

Quoted 9 lines

It's definitely better without the static local since it no longer has the cognitive trap, but I still don't really get why we're messing with the global v. local aspect of it. We're now inconsistent with every other caller of this function, and for what exactly? We've traded some data space for a call to memset() and some instructions. I would have thought anecdotally that memory was the cheaper thing (ie cpu speeds stopped increasing awhile ago, but memory is still getting cheaper).

Quoted 6 lines

The reason for making the structure local is because it's being modified per instance, meaning it would still work as long as qcom_llcc_init_mmio() is never called concurrently for two llcc instances. But the correctness outweighs the performance degradation of setting it up on the stack in my view.

Quoted 6 lines

I hadn't considered the concurrency aspect of the change, since I had anchored myself on the static local. I'm convinced. Might be worth mentioning that in the commit message.

For the series:

Thank you, patches applied.

Regards,
Bjorn