Skip to contentSkip to conversation

iio: light: adjd_s311: convert to device-managed functions

PATCH In review Alexandru Ardelean 4 participants 7 messages Jun 28, 2021 – Jul 12, 2021 IIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iio: light: adjd_s311: convert to device-managed functions

This one is a little easier to convert to device-managed, now with the devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register() can be converted to their devm_ variants. And devm_krealloc() can be used to (re)alloc the buffer. When the driver unloads, this will also be free'd.

---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
drivers/iio/light/adjd_s311.c +5 -29
index 17dac8d0e11d..19d60d6986a1 100644
@@ -230,8 +230,8 @@ static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,
 {
 	struct adjd_s311_data *data = iio_priv(indio_dev);
 
-	kfree(data->buffer);
-	data->buffer = kmalloc(indio_dev->scan_bytes, GFP_KERNEL);
+	data->buffer = devm_krealloc(indio_dev->dev.parent, data->buffer,
+				     indio_dev->scan_bytes, GFP_KERNEL);
 	if (data->buffer == NULL)
 		return -ENOMEM;
 
@@ -256,7 +256,6 @@ static int adjd_s311_probe(struct i2c_client *client,
 		return -ENOMEM;
 
 	data = iio_priv(indio_dev);
-	i2c_set_clientdata(client, indio_dev);
 	data->client = client;
 
 	indio_dev->info = &adjd_s311_info;
@@ -265,34 +264,12 @@ static int adjd_s311_probe(struct i2c_client *client,
 	indio_dev->num_channels = ARRAY_SIZE(adjd_s311_channels);
 	indio_dev->modes = INDIO_DIRECT_MODE;
 
-	err = iio_triggered_buffer_setup(indio_dev, NULL,
-		adjd_s311_trigger_handler, NULL);
+	err = devm_iio_triggered_buffer_setup(&client->dev, indio_dev, NULL,
+					      adjd_s311_trigger_handler, NULL);
 	if (err < 0)
 		return err;
 
-	err = iio_device_register(indio_dev);
-	if (err)
-		goto exit_unreg_buffer;
-
-	dev_info(&client->dev, "ADJD-S311 color sensor registered\n");
-
-	return 0;
-
-exit_unreg_buffer:
-	iio_triggered_buffer_cleanup(indio_dev);
-	return err;
-}
-
-static int adjd_s311_remove(struct i2c_client *client)
-{
-	struct iio_dev *indio_dev = i2c_get_clientdata(client);
-	struct adjd_s311_data *data = iio_priv(indio_dev);
-
-	iio_device_unregister(indio_dev);
-	iio_triggered_buffer_cleanup(indio_dev);
-	kfree(data->buffer);
-
-	return 0;
+	return devm_iio_device_register(&client->dev, indio_dev);
 }
 
 static const struct i2c_device_id adjd_s311_id[] = {
@@ -306,7 +283,6 @@ static struct i2c_driver adjd_s311_driver = {
 		.name	= ADJD_S311_DRV_NAME,
 	},
 	.probe		= adjd_s311_probe,
-	.remove		= adjd_s311_remove,
 	.id_table	= adjd_s311_id,
 };
 module_i2c_driver(adjd_s311_driver);
Signature · 1 line
2.31.1
maintainer
#1 lore raw
Quoted 17 lines · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now with the devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register() can be converted to their devm_ variants. And devm_krealloc() can be used to (re)alloc the buffer. When the driver unloads, this will also be free'd.

---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
drivers/iio/light/adjd_s311.c +0 -0
index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 7 lines
drivers/iio/light/adjd_s311.c +2 -2
@@ -230,8 +230,8 @@ static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,
 {
 	struct adjd_s311_data *data = iio_priv(indio_dev);
 
-	kfree(data->buffer);
-	data->buffer = kmalloc(indio_dev->scan_bytes, GFP_KERNEL);
+	data->buffer = devm_krealloc(indio_dev->dev.parent, data->buffer,
+				     indio_dev->scan_bytes, GFP_KERNEL);

I got some complaints about exactly this trick in a review recently so I'll pass them on.

Whilst devm_krealloc() usage like this won't lose the original reference, its not what people expect from a realloc() case, so to not confuse people it is better to do a dance where you use a local variable, then only set data->buffer to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that were a normal realloc in which case you would just have leaked the original allocation.

More interestingly, why are we bothering with resizing the buffer dependent on what is enabled? Can't we just allocate a 128 byte buffer and not bother changing it as we really aren't wasting that much space? Just embed it in the adjd_s311_data structure directly and don't worry about the allocations. Will need to be aligned(8) though to avoid the push_to_buffer_with_timestamp() issue. Using something like

struct {
	s16 chans[4];
	s64 ts __aligned(8); /* I hate x86 32 bit */
} scan;

Inside the priv structure should work nicely.

Quoted 57 lines
drivers/iio/light/adjd_s311.c +3 -27
 	if (data->buffer == NULL)
 		return -ENOMEM;
 
@@ -256,7 +256,6 @@ static int adjd_s311_probe(struct i2c_client *client,
 		return -ENOMEM;
 
 	data = iio_priv(indio_dev);
-	i2c_set_clientdata(client, indio_dev);
 	data->client = client;
 
 	indio_dev->info = &adjd_s311_info;
@@ -265,34 +264,12 @@ static int adjd_s311_probe(struct i2c_client *client,
 	indio_dev->num_channels = ARRAY_SIZE(adjd_s311_channels);
 	indio_dev->modes = INDIO_DIRECT_MODE;
 
-	err = iio_triggered_buffer_setup(indio_dev, NULL,
-		adjd_s311_trigger_handler, NULL);
+	err = devm_iio_triggered_buffer_setup(&client->dev, indio_dev, NULL,
+					      adjd_s311_trigger_handler, NULL);
 	if (err < 0)
 		return err;
 
-	err = iio_device_register(indio_dev);
-	if (err)
-		goto exit_unreg_buffer;
-
-	dev_info(&client->dev, "ADJD-S311 color sensor registered\n");
-
-	return 0;
-
-exit_unreg_buffer:
-	iio_triggered_buffer_cleanup(indio_dev);
-	return err;
-}
-
-static int adjd_s311_remove(struct i2c_client *client)
-{
-	struct iio_dev *indio_dev = i2c_get_clientdata(client);
-	struct adjd_s311_data *data = iio_priv(indio_dev);
-
-	iio_device_unregister(indio_dev);
-	iio_triggered_buffer_cleanup(indio_dev);
-	kfree(data->buffer);
-
-	return 0;
+	return devm_iio_device_register(&client->dev, indio_dev);
 }
 
 static const struct i2c_device_id adjd_s311_id[] = {
@@ -306,7 +283,6 @@ static struct i2c_driver adjd_s311_driver = {
 		.name	= ADJD_S311_DRV_NAME,
 	},
 	.probe		= adjd_s311_probe,
-	.remove		= adjd_s311_remove,
 	.id_table	= adjd_s311_id,
 };
 module_i2c_driver(adjd_s311_driver);
#2 lore raw
Quoted 46 lines · Jonathan Cameron
Quoted 17 lines · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now with the devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register() can be converted to their devm_ variants. And devm_krealloc() can be used to (re)alloc the buffer. When the driver unloads, this will also be free'd.

---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
drivers/iio/light/adjd_s311.c +0 -0
index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 7 lines
drivers/iio/light/adjd_s311.c +2 -2
@@ -230,8 +230,8 @@ static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,
 {
      struct adjd_s311_data *data = iio_priv(indio_dev);
 
-     kfree(data->buffer);
-     data->buffer = kmalloc(indio_dev->scan_bytes, GFP_KERNEL);
+     data->buffer = devm_krealloc(indio_dev->dev.parent, data->buffer,
+                                  indio_dev->scan_bytes, GFP_KERNEL);

I got some complaints about exactly this trick in a review recently so I'll pass them on.

Whilst devm_krealloc() usage like this won't lose the original reference, its not what people expect from a realloc() case, so to not confuse people it is better to do a dance where you use a local variable, then only set data->buffer to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that were a normal realloc in which case you would just have leaked the original allocation.

More interestingly, why are we bothering with resizing the buffer dependent on what is enabled? Can't we just allocate a 128 byte buffer and not bother changing it as we really aren't wasting that much space? Just embed it in the adjd_s311_data structure directly and don't worry about the allocations. Will need to be aligned(8) though to avoid the push_to_buffer_with_timestamp() issue. Using something like

Quoted 3 lines
struct {
        s16 chans[4];
        s64 ts __aligned(8); /* I hate x86 32 bit */

do you want to me t also add this comment? :p
[just kidding]

Quoted 3 lines

} scan;

Inside the priv structure should work nicely.

i agree; will do it like this;
i hesitated a bit due to the inertia of converting things to devm_

Quoted 60 lines
Quoted 59 lines
drivers/iio/light/adjd_s311.c +3 -27
      if (data->buffer == NULL)
              return -ENOMEM;
 
@@ -256,7 +256,6 @@ static int adjd_s311_probe(struct i2c_client *client,
              return -ENOMEM;
 
      data = iio_priv(indio_dev);
-     i2c_set_clientdata(client, indio_dev);
      data->client = client;
 
      indio_dev->info = &adjd_s311_info;
@@ -265,34 +264,12 @@ static int adjd_s311_probe(struct i2c_client *client,
      indio_dev->num_channels = ARRAY_SIZE(adjd_s311_channels);
      indio_dev->modes = INDIO_DIRECT_MODE;
 
-     err = iio_triggered_buffer_setup(indio_dev, NULL,
-             adjd_s311_trigger_handler, NULL);
+     err = devm_iio_triggered_buffer_setup(&client->dev, indio_dev, NULL,
+                                           adjd_s311_trigger_handler, NULL);
      if (err < 0)
              return err;
 
-     err = iio_device_register(indio_dev);
-     if (err)
-             goto exit_unreg_buffer;
-
-     dev_info(&client->dev, "ADJD-S311 color sensor registered\n");
-
-     return 0;
-
-exit_unreg_buffer:
-     iio_triggered_buffer_cleanup(indio_dev);
-     return err;
-}
-
-static int adjd_s311_remove(struct i2c_client *client)
-{
-     struct iio_dev *indio_dev = i2c_get_clientdata(client);
-     struct adjd_s311_data *data = iio_priv(indio_dev);
-
-     iio_device_unregister(indio_dev);
-     iio_triggered_buffer_cleanup(indio_dev);
-     kfree(data->buffer);
-
-     return 0;
+     return devm_iio_device_register(&client->dev, indio_dev);
 }
 
 static const struct i2c_device_id adjd_s311_id[] = {
@@ -306,7 +283,6 @@ static struct i2c_driver adjd_s311_driver = {
              .name   = ADJD_S311_DRV_NAME,
      },
      .probe          = adjd_s311_probe,
-     .remove         = adjd_s311_remove,
      .id_table       = adjd_s311_id,
 };
 module_i2c_driver(adjd_s311_driver);
maintainer
#3 lore raw
Quoted 54 lines · On Mon, 5 Jul 2021 09:38:21 +0300 Alexandru Ardelean <[email protected]>…
Quoted 46 lines · Jonathan Cameron
Quoted 17 lines · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now with the devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register() can be converted to their devm_ variants. And devm_krealloc() can be used to (re)alloc the buffer. When the driver unloads, this will also be free'd.

---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
drivers/iio/light/adjd_s311.c +0 -0
index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 7 lines
drivers/iio/light/adjd_s311.c +2 -2
@@ -230,8 +230,8 @@ static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,
 {
      struct adjd_s311_data *data = iio_priv(indio_dev);
 
-     kfree(data->buffer);
-     data->buffer = kmalloc(indio_dev->scan_bytes, GFP_KERNEL);
+     data->buffer = devm_krealloc(indio_dev->dev.parent, data->buffer,
+                                  indio_dev->scan_bytes, GFP_KERNEL);  

I got some complaints about exactly this trick in a review recently so I'll pass them on.

Whilst devm_krealloc() usage like this won't lose the original reference, its not what people expect from a realloc() case, so to not confuse people it is better to do a dance where you use a local variable, then only set data->buffer to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that were a normal realloc in which case you would just have leaked the original allocation.

More interestingly, why are we bothering with resizing the buffer dependent on what is enabled? Can't we just allocate a 128 byte buffer and not bother changing it as we really aren't wasting that much space? Just embed it in the adjd_s311_data structure directly and don't worry about the allocations. Will need to be aligned(8) though to avoid the push_to_buffer_with_timestamp() issue. Using something like

Quoted 3 lines
struct {
        s16 chans[4];
        s64 ts __aligned(8); /* I hate x86 32 bit */  

do you want to me t also add this comment? :p
[just kidding]

Quoted 6 lines
Quoted 3 lines

} scan;

Inside the priv structure should work nicely.

i agree; will do it like this;
i hesitated a bit due to the inertia of converting things to devm_

A long discussion on rust usage in linux diverted into the issues around devm. I 'believe' that we are fine in IIO after some work Lars did a long time back to make us resilient to unbinds whilst the chardev was open, but probably worth keeping an eye on that discussion.

https://lore.kernel.org/ksummit/CANiq72nkNrekzbxMci6vW02w=Q2L-SVTk_U4KN_LT8u_b=YPgw@mail.gmail.com/T/#m6db86a574237c22a32ecf49b596b3c2917967c5e

I'm a tiny bit nervous that there might be races where we are doing the devm_realloc. I 'think' we are fine, but the 'think' and 'believe' in these statements expresses a slight lack of certainty!

Jonathan

Quoted 61 lines
Quoted 61 lines
Quoted 60 lines
drivers/iio/light/adjd_s311.c +3 -27
      if (data->buffer == NULL)
              return -ENOMEM;
 
@@ -256,7 +256,6 @@ static int adjd_s311_probe(struct i2c_client *client,
              return -ENOMEM;
 
      data = iio_priv(indio_dev);
-     i2c_set_clientdata(client, indio_dev);
      data->client = client;
 
      indio_dev->info = &adjd_s311_info;
@@ -265,34 +264,12 @@ static int adjd_s311_probe(struct i2c_client *client,
      indio_dev->num_channels = ARRAY_SIZE(adjd_s311_channels);
      indio_dev->modes = INDIO_DIRECT_MODE;
 
-     err = iio_triggered_buffer_setup(indio_dev, NULL,
-             adjd_s311_trigger_handler, NULL);
+     err = devm_iio_triggered_buffer_setup(&client->dev, indio_dev, NULL,
+                                           adjd_s311_trigger_handler, NULL);
      if (err < 0)
              return err;
 
-     err = iio_device_register(indio_dev);
-     if (err)
-             goto exit_unreg_buffer;
-
-     dev_info(&client->dev, "ADJD-S311 color sensor registered\n");
-
-     return 0;
-
-exit_unreg_buffer:
-     iio_triggered_buffer_cleanup(indio_dev);
-     return err;
-}
-
-static int adjd_s311_remove(struct i2c_client *client)
-{
-     struct iio_dev *indio_dev = i2c_get_clientdata(client);
-     struct adjd_s311_data *data = iio_priv(indio_dev);
-
-     iio_device_unregister(indio_dev);
-     iio_triggered_buffer_cleanup(indio_dev);
-     kfree(data->buffer);
-
-     return 0;
+     return devm_iio_device_register(&client->dev, indio_dev);
 }
 
 static const struct i2c_device_id adjd_s311_id[] = {
@@ -306,7 +283,6 @@ static struct i2c_driver adjd_s311_driver = {
              .name   = ADJD_S311_DRV_NAME,
      },
      .probe          = adjd_s311_probe,
-     .remove         = adjd_s311_remove,
      .id_table       = adjd_s311_id,
 };
 module_i2c_driver(adjd_s311_driver);  
reviewer
#4 lore raw
Quoted 107 lines

From: Jonathan Cameron <[email protected]>
Sent: Sunday, July 11, 2021 12:26 PM
To: Alexandru Ardelean <[email protected]>
Cc: linux-iio <[email protected]>; Linux Kernel Mailing List <[email protected]>; [email protected]
Subject: Re: [PATCH] iio: light: adjd_s311: convert to device-managed functions

Quoted 1 line · On Mon, 5 Jul 2021 09:38:21 +0300 Alexandru Ardelean <[email protected]>…

On Sat, 3 Jul 2021 at 20:47, Jonathan Cameron <[email protected]>

Quoted 5 lines · wrote:
Quoted 5 lines
Quoted 1 line · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now with

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register()

can be

Quoted 1 line
Quoted 1 line
Quoted 1 line

converted to their devm_ variants. And devm_krealloc() can be

used to

Quoted 1 line
Quoted 1 line
Quoted 1 line

(re)alloc the buffer. When the driver unloads, this will also be

free'd.

Quoted 7 lines
Quoted 7 lines

Review

Quoted 7 lines
---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
(quoted diff) +0 -0

b/drivers/iio/light/adjd_s311.c

Quoted 4 lines
Quoted 4 lines

drivers/iio/light/adjd_s311.c static int

Quoted 4 lines

index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c +0 -0
@@ -230,8 +230,8 @@ static int

adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 5 lines
Quoted 5 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 5 lines
drivers/iio/light/adjd_s311.c +0 -2
 {
      struct adjd_s311_data *data = iio_priv(indio_dev);
 
-     kfree(data->buffer);
-     data->buffer = kmalloc(indio_dev->scan_bytes,

GFP_KERNEL);

Quoted 4 lines
Quoted 1 line

drivers/iio/light/adjd_s311.c static int · line 233

Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+     data->buffer = devm_krealloc(indio_dev->dev.parent, data-

buffer,

Quoted 2 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+                                  indio_dev->scan_bytes, GFP_KERNEL);

I got some complaints about exactly this trick in a review recently

so I'll

Quoted 3 lines
Quoted 3 lines

pass them on.

Whilst devm_krealloc() usage like this won't lose the original

reference, its

Quoted 1 line
Quoted 1 line

not what people expect from a realloc() case, so to not confuse

people it is

Quoted 1 line
Quoted 1 line

better to do a dance where you use a local variable, then only set

data->buffer

Quoted 3 lines
Quoted 3 lines

to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that

were a normal

Quoted 1 line
Quoted 1 line

realloc in which case you would just have leaked the original

allocation.

Quoted 2 lines
Quoted 2 lines

More interestingly, why are we bothering with resizing the buffer

dependent on what

Quoted 1 line
Quoted 1 line

is enabled? Can't we just allocate a 128 byte buffer and not bother

changing it

Quoted 1 line
Quoted 1 line

as we really aren't wasting that much space? Just embed it in the

adjd_s311_data

Quoted 1 line
Quoted 1 line

structure directly and don't worry about the allocations. Will need

to be

Quoted 1 line
Quoted 1 line

aligned(8) though to avoid the push_to_buffer_with_timestamp()

issue.

Quoted 9 lines
Quoted 5 lines

Using something like

struct {
        s16 chans[4];
        s64 ts __aligned(8); /* I hate x86 32 bit */

do you want to me t also add this comment? :p
[just kidding]

Quoted 6 lines
Quoted 3 lines

} scan;

Inside the priv structure should work nicely.

i agree; will do it like this;
i hesitated a bit due to the inertia of converting things to devm_

A long discussion on rust usage in linux diverted into the issues around devm.
I 'believe' that we are fine in IIO after some work Lars did a long time back
to make us resilient to unbinds whilst the chardev was open, but probably
worth keeping an eye on that discussion.

https://urldefense.com/v3/__https://lore.kernel.org/ksummit/CANiq
72nkNrekzbxMci6vW02w=Q2L- [email protected]/T/*m6db86a574237c22a3 2ecf49b596b3c2917967c5e__;Iw!!A3Ni8CS0y2Y!oeM8GJzKVXb8mYa1m VJNw5fI2adsFk3FKkFzbnqyuDkUMKVTKQ3OoT0cnXP5rA$

Quoted 8 lines

I'm a tiny bit nervous that there might be races where we are doing the devm_realloc.
I 'think' we are fine, but the 'think' and 'believe' in these statements expresses
a slight lack of certainty!

Jonathan

Hi,

It's the second thread where I see you mentioning this, so this I will take the opportunity to also give a bit on though about this. I actually have in mind a RFC (hopefully sending it out this week) for this as I think we might still have some issues with open chardevs and device unbinding.

What we have in [1] is not enough to make sure the whole thing is synchronized with device unbinding... We still have the door open to races where we call 'iio_buffer_ready()' or even 'rb->access->read()' after the device gets unbinded. Maybe we are lucky and nothing bad really happens and we just error out in the next time 'read()' is done on our fd. However, during the possible race, I think it's very likely that we end up touching the same data structures concurrently. On some devices, we surely
(in theory and if all the stars align) have a path where 'iio_buffer_flush_hwfifo()' might be called with 'indio_dev->info' already set to NULL...

IMO, the only way to have this fully in sync is to use the 'info_exist_lock' as it's done in [2]. I think [2] was actually "fixed" when Alex sent his patches for multi buffer support... Naturally, for the read case, we need to make sure we are not going to sleep with the mutex held so we might need an unlock -> lock dance which is not that nice. But I'm not really seeing another way. We also need to look at other file operations and also for the events case to see if this is also a thing.

Naturally, I might be missing some subtlety and that's why I had this planned as RFC. But since is mentioned here, I thought I could bring this up as in the end I might not even need to send the patches :)

[1]: https://elixir.bootlin.com/linux/latest/source/drivers/iio/industrialio-buffer.c#L117
[2]: https://elixir.bootlin.com/linux/latest/source/drivers/iio/industrialio-core.c#L1763

- Nuno Sá

#5 lore raw
Quoted 112 lines · Nuno
Quoted 112 lines

From: Jonathan Cameron <[email protected]>
Sent: Sunday, July 11, 2021 12:26 PM
To: Alexandru Ardelean <[email protected]>
Cc: linux-iio <[email protected]>; Linux Kernel Mailing List <[email protected]>; [email protected]
Subject: Re: [PATCH] iio: light: adjd_s311: convert to device-managed functions

Quoted 1 line · On Mon, 5 Jul 2021 09:38:21 +0300 Alexandru Ardelean <[email protected]>…

On Sat, 3 Jul 2021 at 20:47, Jonathan Cameron <[email protected]>

Quoted 5 lines · wrote:
Quoted 5 lines
Quoted 1 line · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now with

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

devm_krealloc() function.

The other iio_triggered_buffer_setup() and iio_device_register()

can be

Quoted 1 line
Quoted 1 line
Quoted 1 line

converted to their devm_ variants. And devm_krealloc() can be

used to

Quoted 1 line
Quoted 1 line
Quoted 1 line

(re)alloc the buffer. When the driver unloads, this will also be

free'd.

Quoted 7 lines
Quoted 7 lines

Review

Quoted 7 lines
---
 drivers/iio/light/adjd_s311.c | 34 +++++-----------------------------
 1 file changed, 5 insertions(+), 29 deletions(-)
(quoted diff) +0 -0

b/drivers/iio/light/adjd_s311.c

Quoted 4 lines
Quoted 4 lines

drivers/iio/light/adjd_s311.c static int

Quoted 4 lines

index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c +0 -0
@@ -230,8 +230,8 @@ static int  

adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 5 lines
Quoted 5 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 5 lines
drivers/iio/light/adjd_s311.c +0 -2
 {
      struct adjd_s311_data *data = iio_priv(indio_dev);
 
-     kfree(data->buffer);
-     data->buffer = kmalloc(indio_dev->scan_bytes,  

GFP_KERNEL);

Quoted 4 lines
Quoted 1 line

drivers/iio/light/adjd_s311.c static int · line 233

Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+     data->buffer = devm_krealloc(indio_dev->dev.parent, data-  

buffer,

Quoted 2 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+                                  indio_dev->scan_bytes, GFP_KERNEL);  

I got some complaints about exactly this trick in a review recently

so I'll

Quoted 3 lines
Quoted 3 lines

pass them on.

Whilst devm_krealloc() usage like this won't lose the original

reference, its

Quoted 1 line
Quoted 1 line

not what people expect from a realloc() case, so to not confuse

people it is

Quoted 1 line
Quoted 1 line

better to do a dance where you use a local variable, then only set

data->buffer

Quoted 3 lines
Quoted 3 lines

to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that

were a normal

Quoted 1 line
Quoted 1 line

realloc in which case you would just have leaked the original

allocation.

Quoted 2 lines
Quoted 2 lines

More interestingly, why are we bothering with resizing the buffer

dependent on what

Quoted 1 line
Quoted 1 line

is enabled? Can't we just allocate a 128 byte buffer and not bother

changing it

Quoted 1 line
Quoted 1 line

as we really aren't wasting that much space? Just embed it in the

adjd_s311_data

Quoted 1 line
Quoted 1 line

structure directly and don't worry about the allocations. Will need

to be

Quoted 1 line
Quoted 1 line

aligned(8) though to avoid the push_to_buffer_with_timestamp()

issue.

Quoted 9 lines
Quoted 5 lines

Using something like

struct {
        s16 chans[4];
        s64 ts __aligned(8); /* I hate x86 32 bit */  

do you want to me t also add this comment? :p
[just kidding]

Quoted 6 lines
Quoted 3 lines

} scan;

Inside the priv structure should work nicely.

i agree; will do it like this;
i hesitated a bit due to the inertia of converting things to devm_

A long discussion on rust usage in linux diverted into the issues around devm.
I 'believe' that we are fine in IIO after some work Lars did a long time back
to make us resilient to unbinds whilst the chardev was open, but probably
worth keeping an eye on that discussion.

https://urldefense.com/v3/__https://lore.kernel.org/ksummit/CANiq
72nkNrekzbxMci6vW02w=Q2L- [email protected]/T/*m6db86a574237c22a3 2ecf49b596b3c2917967c5e__;Iw!!A3Ni8CS0y2Y!oeM8GJzKVXb8mYa1m VJNw5fI2adsFk3FKkFzbnqyuDkUMKVTKQ3OoT0cnXP5rA$

I'm a tiny bit nervous that there might be races where we are doing the devm_realloc.
I 'think' we are fine, but the 'think' and 'believe' in these statements expresses
a slight lack of certainty!

Quoted 4 lines
Quoted 2 lines

Jonathan

Hi,

+CC Lars who might recall how this all works!

Quoted 11 lines

It's the second thread where I see you mentioning this, so this I will take the opportunity to also give a bit on though about this. I actually have in mind a RFC (hopefully sending it out this week) for this as I think we might still have some issues with open chardevs and device unbinding.

What we have in [1] is not enough to make sure the whole thing is synchronized with device unbinding... We still have the door open to races where we call 'iio_buffer_ready()' or even 'rb->access->read()' after the device gets unbinded. Maybe we are lucky and nothing bad really happens and we just error out in the next time 'read()' is done on our fd.

My understanding of that test is it was only intended to ensure a smooth exit 'after' the buffer pull down has occurred. From vague memory rather than careful analysis, the reason it is needed is we only send the break out signal once for a given buffer, so we need to be sure that userspace doesn't call read() then ignore the error returned due to the buffer going away mid read and call read() again. There may be races in the first time path though. In particularly I'm not sure the reference count on the buffer is raised during the read and it perhaps should be.

Quoted 4 lines

However, during the possible race, I think it's very likely that we end up touching the same data structures concurrently. On some devices, we surely
(in theory and if all the stars align) have a path where 'iio_buffer_flush_hwfifo()' might be called with 'indio_dev->info' already set to NULL...

Yeah, the hwfifo stuff is more recent, it's definitely possible there is a race around that.

Quoted 3 lines

IMO, the only way to have this fully in sync is to use the 'info_exist_lock' as it's done in [2]. I think [2] was actually "fixed" when Alex sent his patches for multi buffer support...

It's rather painful to take that lock. If we can make things safe with appropriate reference counting that's definitely preferable.

For ioctl's they are always slow path so the exist_lock route is fine.

Quoted 8 lines

Naturally, for the read case, we need to make sure we are not going to sleep with the mutex held so we might need an unlock -> lock dance which is not that nice. But I'm not really seeing another way. We also need to look at other file operations and also for the events case to see if this is also a thing.

Naturally, I might be missing some subtlety and that's why I had this planned as RFC. But since is mentioned here, I thought I could bring this up as in the end I might not even need to send the patches :)

Wise move :)

I'd suggest that any fix in this space would ideally be accompanied by a confirmed race. Heavy use of sleeps can usually open one up enough to actually hit them in a few tries.

Jonathan

Quoted 5 lines

[1]: https://elixir.bootlin.com/linux/latest/source/drivers/iio/industrialio-buffer.c#L117
[2]: https://elixir.bootlin.com/linux/latest/source/drivers/iio/industrialio-core.c#L1763

- Nuno Sá

reviewer
#6 lore raw
Quoted 173 lines

From: Jonathan Cameron <[email protected]>
Sent: Monday, July 12, 2021 12:37 PM
To: Sa, Nuno <[email protected]>
Cc: Jonathan Cameron <[email protected]>; Alexandru Ardelean <[email protected]>; linux-iio <[email protected]>; Linux Kernel Mailing List <[email protected]>; [email protected]; Lars-Peter Clausen <[email protected]>
Subject: Re: [PATCH] iio: light: adjd_s311: convert to device-managed functions

[External]

Quoted 6 lines · Nuno
Quoted 6 lines

From: Jonathan Cameron <[email protected]>
Sent: Sunday, July 11, 2021 12:26 PM
To: Alexandru Ardelean <[email protected]>
Cc: linux-iio <[email protected]>; Linux Kernel Mailing List <[email protected]>; [email protected]
Subject: Re: [PATCH] iio: light: adjd_s311: convert to device-

managed

Quoted 5 lines
Quoted 5 lines

functions

On Mon, 5 Jul 2021 09:38:21 +0300 Alexandru Ardelean <[email protected]> wrote:

Quoted 7 lines
Quoted 7 lines
Quoted 1 line

On Sat, 3 Jul 2021 at 20:47, Jonathan Cameron <[email protected]>

Quoted 5 lines · wrote:
Quoted 5 lines
Quoted 1 line · On Mon, 28 Jun 2021 16:51:32 +0300 Alexandru Ardelean <[email protected]>…

This one is a little easier to convert to device-managed, now

with

Quoted 4 lines
Quoted 4 lines

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

devm_krealloc() function.

The other iio_triggered_buffer_setup() and

iio_device_register()

Quoted 2 lines
Quoted 2 lines

can be

Quoted 1 line
Quoted 1 line
Quoted 1 line

converted to their devm_ variants. And devm_krealloc() can

be

Quoted 5 lines
Quoted 5 lines

used to

Quoted 1 line
Quoted 1 line
Quoted 1 line

(re)alloc the buffer. When the driver unloads, this will also be

free'd.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Signed-off-by Alexandru Ardelean

<[email protected]>

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
---
 drivers/iio/light/adjd_s311.c | 34 +++++--------------------------

---

Quoted 9 lines
Quoted 9 lines
Quoted 3 lines
Quoted 3 lines

Review

Quoted 3 lines
 1 file changed, 5 insertions(+), 29 deletions(-)
(quoted diff) +0 -0

b/drivers/iio/light/adjd_s311.c

Quoted 4 lines
Quoted 4 lines

drivers/iio/light/adjd_s311.c static int

Quoted 4 lines

index 17dac8d0e11d..19d60d6986a1 100644

drivers/iio/light/adjd_s311.c +0 -0
@@ -230,8 +230,8 @@ static int

adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Quoted 5 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev,

Quoted 5 lines
drivers/iio/light/adjd_s311.c +0 -2
 {
      struct adjd_s311_data *data = iio_priv(indio_dev);
 
-     kfree(data->buffer);
-     data->buffer = kmalloc(indio_dev->scan_bytes,

GFP_KERNEL);

Quoted 1 line
Quoted 1 line
Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+     data->buffer = devm_krealloc(indio_dev->dev.parent,

data-

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

buffer,

Quoted 2 lines

drivers/iio/light/adjd_s311.c static int adjd_s311_update_scan_mode(struct iio_dev *indio_dev, · line 234

Quoted 1 line
drivers/iio/light/adjd_s311.c +1 -0
+                                  indio_dev->scan_bytes, GFP_KERNEL);

I got some complaints about exactly this trick in a review

recently

Quoted 8 lines
Quoted 8 lines

so I'll

Quoted 3 lines
Quoted 3 lines

pass them on.

Whilst devm_krealloc() usage like this won't lose the original

reference, its

Quoted 1 line
Quoted 1 line

not what people expect from a realloc() case, so to not confuse

people it is

Quoted 1 line
Quoted 1 line

better to do a dance where you use a local variable, then only

set

Quoted 9 lines
Quoted 9 lines

data->buffer

Quoted 3 lines
Quoted 3 lines

to it once we know the realloc succeeded.

That avoids this looking like the anti-pattern it would be if that

were a normal

Quoted 1 line
Quoted 1 line

realloc in which case you would just have leaked the original

allocation.

Quoted 2 lines
Quoted 2 lines

More interestingly, why are we bothering with resizing the

buffer

Quoted 2 lines
Quoted 2 lines

dependent on what

Quoted 1 line
Quoted 1 line

is enabled? Can't we just allocate a 128 byte buffer and not

bother

Quoted 2 lines
Quoted 2 lines

changing it

Quoted 1 line
Quoted 1 line

as we really aren't wasting that much space? Just embed it in

the

Quoted 2 lines
Quoted 2 lines

adjd_s311_data

Quoted 1 line
Quoted 1 line

structure directly and don't worry about the allocations. Will

need

Quoted 2 lines
Quoted 2 lines

to be

Quoted 1 line
Quoted 1 line

aligned(8) though to avoid the

push_to_buffer_with_timestamp()

Quoted 10 lines
Quoted 10 lines

issue.

Quoted 9 lines
Quoted 5 lines

Using something like

struct {
        s16 chans[4];
        s64 ts __aligned(8); /* I hate x86 32 bit */

do you want to me t also add this comment? :p
[just kidding]

Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

} scan;

Inside the priv structure should work nicely.

i agree; will do it like this;
i hesitated a bit due to the inertia of converting things to devm_

A long discussion on rust usage in linux diverted into the issues

around

Quoted 2 lines
Quoted 2 lines

devm.
I 'believe' that we are fine in IIO after some work Lars did a long

time

Quoted 6 lines
Quoted 6 lines

back
to make us resilient to unbinds whilst the chardev was open, but probably
worth keeping an eye on that discussion.

https://urldefense.com/v3/__https://lore.kernel.org/ksummit/CANiq

Quoted 2 lines
Quoted 2 lines

72nkNrekzbxMci6vW02w=Q2L-

[email protected]/T/*m6db86a574237c22a3

Quoted 1 line
Quoted 1 line

2ecf49b596b3c2917967c5e__;Iw!!A3Ni8CS0y2Y!oeM8GJzKVXb8mYa1m

Quoted 5 lines
Quoted 5 lines

VJNw5fI2adsFk3FKkFzbnqyuDkUMKVTKQ3OoT0cnXP5rA$

I'm a tiny bit nervous that there might be races where we are doing the devm_realloc.
I 'think' we are fine, but the 'think' and 'believe' in these

statements

Quoted 7 lines
Quoted 5 lines

expresses
a slight lack of certainty!

Jonathan

Hi,

+CC Lars who might recall how this all works!

Quoted 2 lines

It's the second thread where I see you mentioning this, so this I will

take the

Quoted 1 line

opportunity to also give a bit on though about this. I actually have in

mind a RFC

Quoted 1 line

(hopefully sending it out this week) for this as I think we might still

have some

Quoted 3 lines

issues with open chardevs and device unbinding.

What we have in [1] is not enough to make sure the whole thing is

synchronized with

Quoted 1 line

device unbinding... We still have the door open to races where we

call 'iio_buffer_ready()'

Quoted 1 line

or even 'rb->access->read()' after the device gets unbinded. Maybe

we are lucky and

Quoted 1 line

nothing bad really happens and we just error out in the next time

'read()' is done on

Quoted 1 line

our fd.

Quoted 12 lines

My understanding of that test is it was only intended to ensure a smooth exit 'after' the
buffer pull down has occurred. From vague memory rather than careful analysis, the
reason it is needed is we only send the break out signal once for a given buffer,
so we need to be sure that userspace doesn't call read() then ignore the error returned
due to the buffer going away mid read and call read() again. There may be races in the
first time path though. In particularly I'm not sure the reference count on the buffer

Yes, that is my concern. In theory , with the current code, there's no guarantee that some user might already be in a 'read()'-> passed the 'if (iio_dev->info)' check and then get's preempted for example. In that time, the device might very well be unbonded so that when the user is scheduled again, we will be left thinking the device is there but it won't. Or maybe even worst, we get re-scheduled in the exact time the device is unbonded so there's plenty of room for data races... As you stated, this is only a thing in the first path. And probably one of those concurrent things that might have a 1 in trillion chances to happen. But when they do, they are hell to debug :).

Quoted 33 lines

is raised during the read and it perhaps should be.

Quoted 1 line

However, during the possible race, I think it's very likely that we end

up touching

Quoted 2 lines

the same data structures concurrently. On some devices, we surely (in theory and if all the stars align) have a path where

'iio_buffer_flush_hwfifo()' might

Quoted 1 line

be called with 'indio_dev->info' already set to NULL...

Yeah, the hwfifo stuff is more recent, it's definitely possible there is a race around that.

Quoted 2 lines

IMO, the only way to have this fully in sync is to use the

'info_exist_lock' as it's done

Quoted 1 line

in [2]. I think [2] was actually "fixed" when Alex sent his patches for

multi buffer support...
It's rather painful to take that lock. If we can make things safe with appropriate reference
counting that's definitely preferable.

For ioctl's they are always slow path so the exist_lock route is fine.

Quoted 1 line

Naturally, for the read case, we need to make sure we are not going

to sleep with the

Quoted 1 line

mutex held so we might need an unlock -> lock dance which is not

that nice. But I'm

Quoted 1 line

not really seeing another way. We also need to look at other file

operations and also

Quoted 3 lines

for the events case to see if this is also a thing.

Naturally, I might be missing some subtlety and that's why I had this

Quoted 8 lines

planned as RFC.

Quoted 1 line

But since is mentioned here, I thought I could bring this up as in the

end I might not

Quoted 1 line

even need to send the patches :)

Wise move :)

I'd suggest that any fix in this space would ideally be accompanied by a confirmed race.

Yeah, it makes sense but this is one of those cases where this is very hard to hit...

- Nuno Sá