Re: [PATCH v3 5/6] staging: r8188eu: add error handling of rtw_read32

[Pavel Skripkin]

On 8/26/21 12:22 PM, Pavel Skripkin wrote:
> On Thu, 26 Aug 2021 08:51:23 +0000
> David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> > From: Pavel Skripkin <paskripkin@xxxxxxxxx>
> > > Sent: 24 August 2021 08:28
> > > 
> > > _rtw_read32 function can fail in case of usb transfer failure. But
> > > previous function prototype wasn't designed to return an error to
> > > caller. It can cause a lot uninit value bugs all across the driver
> > > code, since rtw_read32() returns local stack variable to caller.
> > > 
> > > Fix it by changing the prototype of this function. Now it returns an
> > > int: 0 on success, negative error value on failure and callers
> > > should pass the pointer to storage location for register value.
> >
> > Pretty horrid API interface.
> > Functions like readl() - which can fail just return ~0u and let
> > the caller worry about whether that causes serious grief.
> >
> > You could make all the read functions return __u64 and return ~0ull
> > on error.
> > Testing for (value & 1ull << 63) will be reasonable even on 32bit.
>
> I am not the best at API related questions, so can you, please,
> explain why your approach is better?
>
> As I can see, most of the drivers in usb/ directory use smth like this
> interface for private reading funcions. We anyway creating tmp variable
> (but 64 bit _always_) and checking for mistery error, which we cannot
> pass up to callers.
>
> Sorry, if it's _too_ dumb question, but I really can't get your
> point....
>
>
>
>   
>
> > ...
> > > -static u32 usb_read32(struct intf_hdl *pintfhdl, u32 addr)
> > > +static int usb_read32(struct intf_hdl *pintfhdl, u32 addr, u32
> > > *data) {
> > >  	u8 requesttype;
> > >  	u16 wvalue;
> > >  	u16 len;
> > > -	__le32 data;
> > > +	int res;
> > > +	__le32 tmp;
> > > +
> > > +	if (WARN_ON(unlikely(!data)))
> > > +		return -EINVAL;
> > > 
> >
> > Kill the NULL check - it is a silly coding error.
> > An OOPS is just as easy to debug.
>
>
> I don't think that one single driver should kill the whole system. It's
> 100% an error, but kernel can recover from it (for example disconnect
> the driver, since it's broken).
>
>
> AFIAK, Greg and Linus do not like BUG_ONs in recoverable state...
> Correct me, if I am wrong
Oops, I thought about BUG_ON() instead of WARN_ON(), sorry for 
confusion. My point is "we should not let the box boom".





With regards,
Pavel Skripkin