Re: [PATCH 1/4] rtw88: Add packed attribute to the eFuse structs

[Ping-Ke Shih]

On Sun, 2023-01-01 at 11:54 +0000, David Laight wrote:
> From: Ping-Ke Shih
> > Sent: 01 January 2023 11:42
> > 
> > On Sat, 2022-12-31 at 16:57 +0000, David Laight wrote:
> > > From: Ping-Ke Shih
> > > > Sent: 29 December 2022 09:25
> > > > 
> > > > > -----Original Message-----
> > > > > From: Martin Blumenstingl <martin.blumenstingl@xxxxxxxxxxxxxx>
> > > > > Sent: Wednesday, December 28, 2022 9:36 PM
> > > > > To: linux-wireless@xxxxxxxxxxxxxxx
> > > > > Cc: tony0620emma@xxxxxxxxx; kvalo@xxxxxxxxxx; Ping-Ke Shih <pkshih@xxxxxxxxxxx>;
> > > > tehuang@xxxxxxxxxxx;
> > > > > s.hauer@xxxxxxxxxxxxxx; netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; Martin
> > > > > Blumenstingl
> > > > > <martin.blumenstingl@xxxxxxxxxxxxxx>
> > > > > Subject: [PATCH 1/4] rtw88: Add packed attribute to the eFuse structs
> > > > > 
> > > > > The eFuse definitions in the rtw88 are using structs to describe the
> > > > > eFuse contents. Add the packed attribute to all structs used for the
> > > > > eFuse description so the compiler doesn't add gaps or re-order
> > > > > attributes.
> > > > > 
> > > > > Also change the type of the res2..res3 eFuse fields to u16 to avoid the
> > > > > following warning, now that their surrounding struct has the packed
> > > > > attribute:
> > > > >   note: offset of packed bit-field 'res2' has changed in GCC 4.4
> > > > > 
> > > > > Fixes: e3037485c68e ("rtw88: new Realtek 802.11ac driver")
> > > > > Fixes: ab0a031ecf29 ("rtw88: 8723d: Add read_efuse to recognize efuse info from map")
> > > > > Fixes: 769a29ce2af4 ("rtw88: 8821c: add basic functions")
> > > > > Fixes: 87caeef032fc ("wifi: rtw88: Add rtw8723du chipset support")
> > > > > Fixes: aff5ffd718de ("wifi: rtw88: Add rtw8821cu chipset support")
> > > > > Signed-off-by: Martin Blumenstingl <martin.blumenstingl@xxxxxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/net/wireless/realtek/rtw88/main.h     |  6 +++---
> > > > >  drivers/net/wireless/realtek/rtw88/rtw8723d.h |  6 +++---
> > > > >  drivers/net/wireless/realtek/rtw88/rtw8821c.h | 20 +++++++++----------
> > > > >  drivers/net/wireless/realtek/rtw88/rtw8822b.h | 20 +++++++++----------
> > > > >  drivers/net/wireless/realtek/rtw88/rtw8822c.h | 20 +++++++++----------
> > > > >  5 files changed, 36 insertions(+), 36 deletions(-)
> > > > > 
> > > > 
> > > > [...]
> > > > 
> > > > > @@ -43,13 +43,13 @@ struct rtw8821ce_efuse {
> > > > >  	u8 link_cap[4];
> > > > >  	u8 link_control[2];
> > > > >  	u8 serial_number[8];
> > > > > -	u8 res0:2;			/* 0xf4 */
> > > > > -	u8 ltr_en:1;
> > > > > -	u8 res1:2;
> > > > > -	u8 obff:2;
> > > > > -	u8 res2:3;
> > > > > -	u8 obff_cap:2;
> > > > > -	u8 res3:4;
> > > > > +	u16 res0:2;			/* 0xf4 */
> > > > > +	u16 ltr_en:1;
> > > > > +	u16 res1:2;
> > > > > +	u16 obff:2;
> > > > > +	u16 res2:3;
> > > > > +	u16 obff_cap:2;
> > > > > +	u16 res3:4;
> > > > 
> > > > These should be __le16. Though bit fields are suitable to efuse layout,
> > > > we don't access these fields for now. It would be well.
> > 
> > Uh. I typo the sentence. Originally, I would like to type
> > "...bit fields are NOT suitable...".
> > 
> > In this driver, efuse is read into a u8 array, and cast to this struct
> > pointer to access the field.
> 
> Then define it as such.
> The 16bit endianness and bit-order dependant bitfields serve
> no purpose. 
> 
> > > IIRC the assignment of actual bits to bit-fields is (at best)
> > > architecturally defined - so isn't really suitable for anything
> > > where the bits have to match a portable memory buffer.
> > > The bit allocation isn't tied to the byte endianness.
> > 
> > Yes, this kind of struct has endian problem. Fortunately, we don't
> > actually access values via bit fields.
> > 
> > > To get an explicit layout you have to do explicit masking.
> > 
> > If we actually want to access these values, we will define masks
> > and use {u8,_le16,le32}_get_bits() or bare '&' bit operation to access
> > them.
> 
> But you can't take the address of bitfield members.
> Define the data properly.

Yes, it should not use bit filed. Instead, use a __le16 for all fields, such as

struct rtw8821ce_efuse {
    ...
    __le16 caps;
    ...
}


#define CAPS_RES0 GENMASK(1, 0)
#define CAPS_LTR_EN BIT(2)
#define CAPS_RES1 GENMASK(4, 3)
#define CAPS_OBFF GENMASK(6, 5)
...


Assume the pointer of efuse content is 'const u8 *efuse_raw;'

   const struct rtw8821ce_efuse *efuse = (const struct rtw8821ce_efuse *)efuse_raw;

Then, get ltr_en

   ltr_en = le16_get_bits(efuse->caps, CAPS_LTR_EN);


> 
> > > You also don't need __packed unless the 'natural' alignment
> > > of fields would need gaps or the actual structure itself might
> > > be misaligned in memory.
> > > While C compilers are allowed to add arbitrary padding the Linux kernel
> > > requires that they don't.
> > > I'm also pretty sure that compilers are not allowed to reorder fields.
> > > 
> > > Specifying __packed can add considerable run-time (and code size)
> > > overhead on some architectures - it should only be used if actually
> > > needed.
> > > 
> > 
> > Understood. We only add __packed to the struct which is used to
> > access predefined format, like efuse content defined by vendor.
> 
> No - that doesn't mean you need to use __packed.
> It does mean that you shouldn't use bitfields.
> Look at all the hardware drivers, they use structs to map device
> registers and absolutely require the compile not add padding.
> 

I think the original struct has two problem -- endian and __packed.

I mentioned endian part above. 

Taking below as example to explain why I need __packed, 

struct rtw8821ce_efuse {
   ...
   u8 data1;       // offset 0x100
   __le16 data2;   // offset 0x101-0x102
   ...
} __packed;

Without __packed, compiler could has pad between data1 and data2,
and then get wrong result.

In this patch, struct isn't to map registers. Instead it is used to
access efuse content of a u8 array existing in memory.

Ping-Ke