driver core: Add a guard() definition for the device_lock()

Commit Message

Dan Williams Dec. 13, 2023, 11:02 p.m. UTC

At present there are ~200 usages of device_lock() in the kernel. Some of
those usages lead to "goto unlock;" patterns which have proven to be
error prone. Define a "device" guard() definition to allow for those to
be cleaned up and prevent new ones from appearing.

Link: http://lore.kernel.org/r/[e-mail]
Link: http://lore.kernel.org/r/[e-mail]
Cc: Vishal Verma <[e-mail]>
Cc: Ira Weiny <[e-mail]>
Cc: Peter Zijlstra <[e-mail]>
Cc: Greg Kroah-Hartman <[e-mail]>
Cc: Andrew Morton <[e-mail]>
Signed-off-by: Dan Williams <[e-mail]>
---
Hi Greg,

I wonder if you might include this change in v6.7-rc to ease some patch
sets alternately going through my tree and Andrew's tree. Those
discussions are linked above. Alternately I can can just take it through
my tree with your ack and the other use case can circle back to it in
the v6.9 cycle.

I considered also defining a __free() helper similar to __free(mutex),
but I think "__free(device)" would be a surprising name for something
that drops a lock. Also, I like the syntax of guard(device) over
something like guard(device_lock) since a 'struct device *' is the
argument, not a lock type, but I'm open to your or Peter's thoughts on
the naming.

 include/linux/device.h |    2 ++
 1 file changed, 2 insertions(+)

Comments

Ira Weiny Dec. 13, 2023, 11:47 p.m. UTC | #1

Dan Williams wrote:
> At present there are ~200 usages of device_lock() in the kernel. Some of
> those usages lead to "goto unlock;" patterns which have proven to be
> error prone. Define a "device" guard() definition to allow for those to
> be cleaned up and prevent new ones from appearing.
> 
> Link: http://lore.kernel.org/r/[e-mail]
> Link: http://lore.kernel.org/r/[e-mail]
> Cc: Vishal Verma <[e-mail]>
> Cc: Ira Weiny <[e-mail]>

Reviewed-by: Ira Weiny <[e-mail]>

> Cc: Peter Zijlstra <[e-mail]>
> Cc: Greg Kroah-Hartman <[e-mail]>
> Cc: Andrew Morton <[e-mail]>
> Signed-off-by: Dan Williams <[e-mail]>
> ---
> Hi Greg,
> 
> I wonder if you might include this change in v6.7-rc to ease some patch
> sets alternately going through my tree and Andrew's tree. Those
> discussions are linked above. Alternately I can can just take it through
> my tree with your ack and the other use case can circle back to it in
> the v6.9 cycle.
> 
> I considered also defining a __free() helper similar to __free(mutex),
> but I think "__free(device)" would be a surprising name for something
> that drops a lock. Also, I like the syntax of guard(device) over
> something like guard(device_lock) since a 'struct device *' is the
> argument, not a lock type, but I'm open to your or Peter's thoughts on
> the naming.
> 
>  include/linux/device.h |    2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/include/linux/device.h b/include/linux/device.h
> index d7a72a8749ea..6c83294395ac 100644
> --- a/include/linux/device.h
> +++ b/include/linux/device.h
> @@ -1007,6 +1007,8 @@ static inline void device_unlock(struct device *dev)
>  	mutex_unlock(&dev->mutex);
>  }
>  
> +DEFINE_GUARD(device, struct device *, device_lock(_T), device_unlock(_T))
> +
>  static inline void device_lock_assert(struct device *dev)
>  {
>  	lockdep_assert_held(&dev->mutex);
>

Verma, Vishal L Dec. 14, 2023, 12:01 a.m. UTC | #2

On Wed, 2023-12-13 at 15:02 -0800, Dan Williams wrote:
> At present there are ~200 usages of device_lock() in the kernel. Some of
> those usages lead to "goto unlock;" patterns which have proven to be
> error prone. Define a "device" guard() definition to allow for those to

"Define a definition" sounds a bit awkward, perhaps  "Add a .."?

> be cleaned up and prevent new ones from appearing.
> 
> Link: http://lore.kernel.org/r/[e-mail]
> Link: http://lore.kernel.org/r/[e-mail]
> Cc: Vishal Verma <[e-mail]>
> Cc: Ira Weiny <[e-mail]>
> Cc: Peter Zijlstra <[e-mail]>
> Cc: Greg Kroah-Hartman <[e-mail]>
> Cc: Andrew Morton <[e-mail]>
> Signed-off-by: Dan Williams <[e-mail]>

Other than that, looks good,

Reviewed-by: Vishal Verma <[e-mail]>

> ---
> Hi Greg,
> 
> I wonder if you might include this change in v6.7-rc to ease some patch
> sets alternately going through my tree and Andrew's tree. Those
> discussions are linked above. Alternately I can can just take it through
> my tree with your ack and the other use case can circle back to it in
> the v6.9 cycle.
> 
> I considered also defining a __free() helper similar to __free(mutex),
> but I think "__free(device)" would be a surprising name for something
> that drops a lock. Also, I like the syntax of guard(device) over
> something like guard(device_lock) since a 'struct device *' is the
> argument, not a lock type, but I'm open to your or Peter's thoughts on
> the naming.
> 
>  include/linux/device.h |    2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/include/linux/device.h b/include/linux/device.h
> index d7a72a8749ea..6c83294395ac 100644
> --- a/include/linux/device.h
> +++ b/include/linux/device.h
> @@ -1007,6 +1007,8 @@ static inline void device_unlock(struct device *dev)
>         mutex_unlock(&dev->mutex);
>  }
>  
> +DEFINE_GUARD(device, struct device *, device_lock(_T), device_unlock(_T))
> +
>  static inline void device_lock_assert(struct device *dev)
>  {
>         lockdep_assert_held(&dev->mutex);
>

Greg KH Dec. 14, 2023, 3:33 p.m. UTC | #3

On Wed, Dec 13, 2023 at 03:02:35PM -0800, Dan Williams wrote:
> At present there are ~200 usages of device_lock() in the kernel. Some of
> those usages lead to "goto unlock;" patterns which have proven to be
> error prone. Define a "device" guard() definition to allow for those to
> be cleaned up and prevent new ones from appearing.
> 
> Link: http://lore.kernel.org/r/[e-mail]
> Link: http://lore.kernel.org/r/[e-mail]
> Cc: Vishal Verma <[e-mail]>
> Cc: Ira Weiny <[e-mail]>
> Cc: Peter Zijlstra <[e-mail]>
> Cc: Greg Kroah-Hartman <[e-mail]>
> Cc: Andrew Morton <[e-mail]>
> Signed-off-by: Dan Williams <[e-mail]>
> ---
> Hi Greg,
> 
> I wonder if you might include this change in v6.7-rc to ease some patch
> sets alternately going through my tree and Andrew's tree. Those
> discussions are linked above. Alternately I can can just take it through
> my tree with your ack and the other use case can circle back to it in
> the v6.9 cycle.

Sure, I'll queue it up now for 6.7-final, makes sense to have it now for
others to build off of, and for me to fix up some places in the driver
core to use it as well.

> I considered also defining a __free() helper similar to __free(mutex),
> but I think "__free(device)" would be a surprising name for something
> that drops a lock. Also, I like the syntax of guard(device) over
> something like guard(device_lock) since a 'struct device *' is the
> argument, not a lock type, but I'm open to your or Peter's thoughts on
> the naming.

guard(device); makes sense to me, as that's what you are doing here, so
I'm good with it.

thanks,

greg k-h

Dave Jiang Dec. 14, 2023, 3:49 p.m. UTC | #4

On 12/13/23 16:02, Dan Williams wrote:
> At present there are ~200 usages of device_lock() in the kernel. Some of
> those usages lead to "goto unlock;" patterns which have proven to be
> error prone. Define a "device" guard() definition to allow for those to
> be cleaned up and prevent new ones from appearing.
> 
> Link: http://lore.kernel.org/r/[e-mail]
> Link: http://lore.kernel.org/r/[e-mail]
> Cc: Vishal Verma <[e-mail]>
> Cc: Ira Weiny <[e-mail]>
> Cc: Peter Zijlstra <[e-mail]>
> Cc: Greg Kroah-Hartman <[e-mail]>
> Cc: Andrew Morton <[e-mail]>
> Signed-off-by: Dan Williams <[e-mail]>

Reviewed-by: Dave Jiang <[e-mail]>
> ---
> Hi Greg,
> 
> I wonder if you might include this change in v6.7-rc to ease some patch
> sets alternately going through my tree and Andrew's tree. Those
> discussions are linked above. Alternately I can can just take it through
> my tree with your ack and the other use case can circle back to it in
> the v6.9 cycle.
> 
> I considered also defining a __free() helper similar to __free(mutex),
> but I think "__free(device)" would be a surprising name for something
> that drops a lock. Also, I like the syntax of guard(device) over
> something like guard(device_lock) since a 'struct device *' is the
> argument, not a lock type, but I'm open to your or Peter's thoughts on
> the naming.
> 
>  include/linux/device.h |    2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/include/linux/device.h b/include/linux/device.h
> index d7a72a8749ea..6c83294395ac 100644
> --- a/include/linux/device.h
> +++ b/include/linux/device.h
> @@ -1007,6 +1007,8 @@ static inline void device_unlock(struct device *dev)
>  	mutex_unlock(&dev->mutex);
>  }
>  
> +DEFINE_GUARD(device, struct device *, device_lock(_T), device_unlock(_T))
> +
>  static inline void device_lock_assert(struct device *dev)
>  {
>  	lockdep_assert_held(&dev->mutex);
> 
>

Patch

diff --git a/include/linux/device.h b/include/linux/device.h
index d7a72a8749ea..6c83294395ac 100644
--- a/include/linux/device.h
+++ b/include/linux/device.h
@@ -1007,6 +1007,8 @@  static inline void device_unlock(struct device *dev)
 	mutex_unlock(&dev->mutex);
 }
 
+DEFINE_GUARD(device, struct device *, device_lock(_T), device_unlock(_T))
+
 static inline void device_lock_assert(struct device *dev)
 {
 	lockdep_assert_held(&dev->mutex);