Re: [PATCH 03/12] nvmem: mtk-efuse: fix different address space warnings of sparse

[Masahiro Yamada]

2017-10-20 23:49 GMT+09:00 Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>:
> On Fri, Oct 20, 2017 at 10:55:04PM +0900, Masahiro Yamada wrote:
>> 2017-10-20 22:34 GMT+09:00 Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>:
>> > On Mon, Oct 09, 2017 at 03:26:32PM +0200, srinivas.kandagatla@xxxxxxxxxx wrote:
>> >> From: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
>> >>
>> >> Fix the following sparse warnings:
>> >>
>> >> drivers/nvmem/mtk-efuse.c:24:30: warning: incorrect type in initializer (different address spaces)
>> >> drivers/nvmem/mtk-efuse.c:24:30:    expected void [noderef] <asn:2>*base
>> >> drivers/nvmem/mtk-efuse.c:24:30:    got void *context
>> >> drivers/nvmem/mtk-efuse.c:37:30: warning: incorrect type in initializer (different address spaces)
>> >> drivers/nvmem/mtk-efuse.c:37:30:    expected void [noderef] <asn:2>*base
>> >> drivers/nvmem/mtk-efuse.c:37:30:    got void *context
>> >> drivers/nvmem/mtk-efuse.c:69:23: warning: incorrect type in assignment (different address spaces)
>> >> drivers/nvmem/mtk-efuse.c:69:23:    expected void *priv
>> >> drivers/nvmem/mtk-efuse.c:69:23:    got void [noderef] <asn:2>*[assigned] base
>> >>
>> >> The type of nvmem_config->priv is (void *), so sparse complains
>> >> about assignment of the base address with (void __iomem *) type.
>> >>
>> >> Even if we cast it out, sparse still warns:
>> >> warning: cast removes address space of expression
>> >>
>> >> Of course, we can shut up the sparse by marking __force, but a more
>> >> correct way is to put the base address into driver private data.
>> >
>> > Ick, no, really?
>> >
>> > That's horrid.  Do the __force cast as that is what you are really doing
>> > here, and it is good to see the bad use of it (hint, perhaps the api
>> > needs to be fixed up so you do not have to do that???)
>>
>>
>> Other drivers take private data from the "void *context".
>>
>> In this driver, the priv just happens to have one member.
>> What's the matter?
>>
>>
>>
>>
>> >> diff --git a/drivers/nvmem/mtk-efuse.c b/drivers/nvmem/mtk-efuse.c
>> >> index fa7a0f66b37e..c4058b598703 100644
>> >> --- a/drivers/nvmem/mtk-efuse.c
>> >> +++ b/drivers/nvmem/mtk-efuse.c
>> >> @@ -18,15 +18,19 @@
>> >>  #include <linux/nvmem-provider.h>
>> >>  #include <linux/platform_device.h>
>> >>
>> >> +struct mtk_efuse_priv {
>> >> +     void __iomem *base;
>> >> +};
>> >> +
>> >>  static int mtk_reg_read(void *context,
>> >
>> >
>> > Make the read call take a __iomem *, not a random void *.  That should
>> > solve this issue, right?
>> >
>>
>> NACK.
>>
>> See drivers/nvmem/imx-iim.c
>>
>> The ->reg_read hook needs more private data than __iomem *.
>
> Ok, that makes a bit more sense, but it still feels a bit of a hack,
> don't you agree?


But, I have no choice but __force.

I disagree with using __force to hide things.

In my opinion, __force should be used only where it is really legitimate.

This driver is not the case, as we can fix it
by writing correct code.


-- 
Best Regards
Masahiro Yamada