diff mbox series

dm: make uclass_find_first_device() return error when no defice is found

Message ID 20200225065803.30275-1-yamada.masahiro@socionext.com
State New
Headers show
Series dm: make uclass_find_first_device() return error when no defice is found | expand

Commit Message

Masahiro Yamada Feb. 25, 2020, 6:58 a.m. UTC
uclass_find_first_device() succeeds even if it cannot find any device.
So, the caller must check the return code and also *devp is not NULL.

Returning -ENODEV will be sensible in this case.

Signed-off-by: Masahiro Yamada <yamada.masahiro at socionext.com>
---

If this patch is acceptable, I want to fold this
into my pull request because it need it
for my another patch:
http://patchwork.ozlabs.org/patch/1238000/

 drivers/core/uclass.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Simon Glass Feb. 26, 2020, 3:33 p.m. UTC | #1
Hi Masahiro,

On Mon, 24 Feb 2020 at 23:58, Masahiro Yamada
<yamada.masahiro at socionext.com> wrote:
>
> uclass_find_first_device() succeeds even if it cannot find any device.
> So, the caller must check the return code and also *devp is not NULL.
>
> Returning -ENODEV will be sensible in this case.
>
> Signed-off-by: Masahiro Yamada <yamada.masahiro at socionext.com>
> ---
>
> If this patch is acceptable, I want to fold this
> into my pull request because it need it
> for my another patch:
> http://patchwork.ozlabs.org/patch/1238000/
>
>  drivers/core/uclass.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)

I sort-of agree and have thought about this a lot.

But what are you doing in the case of uclass_find_next_device()?

Also what about the tests and other code that expects the current behaviour?

>
> diff --git a/drivers/core/uclass.c b/drivers/core/uclass.c
> index 58b19a421091..3580974f3b85 100644
> --- a/drivers/core/uclass.c
> +++ b/drivers/core/uclass.c
> @@ -227,7 +227,7 @@ int uclass_find_first_device(enum uclass_id id, struct udevice **devp)
>         if (ret)
>                 return ret;
>         if (list_empty(&uc->dev_head))
> -               return 0;
> +               return -ENODEV;
>
>         *devp = list_first_entry(&uc->dev_head, struct udevice, uclass_node);
>
> --
> 2.17.1
>

Regards,
Simon
Masahiro Yamada Feb. 27, 2020, 5:56 p.m. UTC | #2
Hi Simon,

On Thu, Feb 27, 2020 at 12:33 AM Simon Glass <sjg at chromium.org> wrote:
>
> Hi Masahiro,
>
> On Mon, 24 Feb 2020 at 23:58, Masahiro Yamada
> <yamada.masahiro at socionext.com> wrote:
> >
> > uclass_find_first_device() succeeds even if it cannot find any device.
> > So, the caller must check the return code and also *devp is not NULL.
> >
> > Returning -ENODEV will be sensible in this case.
> >
> > Signed-off-by: Masahiro Yamada <yamada.masahiro at socionext.com>
> > ---
> >
> > If this patch is acceptable, I want to fold this
> > into my pull request because it need it
> > for my another patch:
> > http://patchwork.ozlabs.org/patch/1238000/
> >
> >  drivers/core/uclass.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
>
> I sort-of agree and have thought about this a lot.
>
> But what are you doing in the case of uclass_find_next_device()?
>
> Also what about the tests and other code that expects the current behaviour?


OK, I will change my caller side.



I would probably design these functions as follows:

struct udevice *uclass_find_first_device(enum uclass_id id);

struct udevice *uclass_find_next_device(struct udevice *dev);


Return the pointer to the device, or NULL if not found.









> >
> > diff --git a/drivers/core/uclass.c b/drivers/core/uclass.c
> > index 58b19a421091..3580974f3b85 100644
> > --- a/drivers/core/uclass.c
> > +++ b/drivers/core/uclass.c
> > @@ -227,7 +227,7 @@ int uclass_find_first_device(enum uclass_id id, struct udevice **devp)
> >         if (ret)
> >                 return ret;
> >         if (list_empty(&uc->dev_head))
> > -               return 0;
> > +               return -ENODEV;
> >
> >         *devp = list_first_entry(&uc->dev_head, struct udevice, uclass_node);
> >
> > --
> > 2.17.1
> >
>
> Regards,
> Simon



--
Best Regards
Masahiro Yamada
Simon Glass Feb. 27, 2020, 11:40 p.m. UTC | #3
Hi Masahiro,

On Thu, 27 Feb 2020 at 09:57, Masahiro Yamada <masahiroy at kernel.org> wrote:
>
> Hi Simon,
>
> On Thu, Feb 27, 2020 at 12:33 AM Simon Glass <sjg at chromium.org> wrote:
> >
> > Hi Masahiro,
> >
> > On Mon, 24 Feb 2020 at 23:58, Masahiro Yamada
> > <yamada.masahiro at socionext.com> wrote:
> > >
> > > uclass_find_first_device() succeeds even if it cannot find any device.
> > > So, the caller must check the return code and also *devp is not NULL.
> > >
> > > Returning -ENODEV will be sensible in this case.
> > >
> > > Signed-off-by: Masahiro Yamada <yamada.masahiro at socionext.com>
> > > ---
> > >
> > > If this patch is acceptable, I want to fold this
> > > into my pull request because it need it
> > > for my another patch:
> > > http://patchwork.ozlabs.org/patch/1238000/
> > >
> > >  drivers/core/uclass.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > I sort-of agree and have thought about this a lot.
> >
> > But what are you doing in the case of uclass_find_next_device()?
> >
> > Also what about the tests and other code that expects the current behaviour?
>
>
> OK, I will change my caller side.
>
>
>
> I would probably design these functions as follows:
>
> struct udevice *uclass_find_first_device(enum uclass_id id);
>
> struct udevice *uclass_find_next_device(struct udevice *dev);
>
>
> Return the pointer to the device, or NULL if not found.

I think that is fine, since we don't support any other error than -ENODEV.

I did intend to add checks that the structs are correct (by adding
type information into each struct conditional on a DM_DEBUG option)
and returning a special error when something goes wrong. But even then
I think we could just return NULL and -ENODEV in the caller would be
good enough.

Regards,
Simon
diff mbox series

Patch

diff --git a/drivers/core/uclass.c b/drivers/core/uclass.c
index 58b19a421091..3580974f3b85 100644
--- a/drivers/core/uclass.c
+++ b/drivers/core/uclass.c
@@ -227,7 +227,7 @@  int uclass_find_first_device(enum uclass_id id, struct udevice **devp)
 	if (ret)
 		return ret;
 	if (list_empty(&uc->dev_head))
-		return 0;
+		return -ENODEV;
 
 	*devp = list_first_entry(&uc->dev_head, struct udevice, uclass_node);