Re: [hid-picolcd] Avoid compile warning/error triggered by copy_from_user()

["Bruno Prémont" <bonbons@linux-vserver.org>]

Hi Jiri,

On Thu, 5 May 2011 11:27:54 wrote:
> On Wed, 4 May 2011, Bruno Prémont wrote:
> 
> > With CONFIG_DEBUG_STRICT_USER_COPY_CHECKS=y compilation of PicoLCD
> > driver fails on copy_from_user(), without it a warning is generated:
> > 
> >   CC [M]  drivers/hid/hid-picolcd.o
> > In file included from /usr/src/linux-2.6/arch/x86/include/asm/uaccess.h:571,
> >                  from /usr/src/linux-2.6/arch/x86/include/asm/sections.h:5,
> >                  from /usr/src/linux-2.6/arch/x86/include/asm/hw_irq.h:26,
> >                  from /usr/src/linux-2.6/include/linux/irq.h:359,
> >                  from /usr/src/linux-2.6/arch/x86/include/asm/hardirq.h:5,
> >                  from /usr/src/linux-2.6/include/linux/hardirq.h:7,
> >                  from /usr/src/linux-2.6/include/linux/interrupt.h:12,
> >                  from /usr/src/linux-2.6/include/linux/usb.h:15,
> >                  from /usr/src/linux-2.6/drivers/hid/hid-picolcd.c:25:
> > In function 'copy_from_user',
> >     inlined from 'picolcd_debug_eeprom_write' at drivers/hid/hid-picolcd.c:1592:
> > arch/x86/include/asm/uaccess_32.h:212: error: call to 'copy_from_user_overflow' declared with attribute error: copy_from_user() buffer size is not provably correct
> > 
> > gcc-4.4.5 is not able to track size calculation when it is stored into
> > a variable, thus tell copy_from_user() maximum size via
> > min(*max-size*, *effective-size*) explicitly and inline how much to copy
> > at most.
> > 
> > Signed-off-by: Bruno Prémont <bonbons@linux-vserver.org>
> > --
> > 
> > diff --git a/drivers/hid/hid-picolcd.c b/drivers/hid/hid-picolcd.c
> > index b2f56a1..9d8710f 100644
> > --- a/drivers/hid/hid-picolcd.c
> > +++ b/drivers/hid/hid-picolcd.c
> > @@ -1585,11 +1585,11 @@ static ssize_t picolcd_debug_eeprom_write(struct file *f, const char __user *u,
> >  	memset(raw_data, 0, sizeof(raw_data));
> >  	raw_data[0] = *off & 0xff;
> >  	raw_data[1] = (*off >> 8) & 0xff;
> > -	raw_data[2] = s < 20 ? s : 20;
> > +	raw_data[2] = min((size_t)20, s);
> >  	if (*off + raw_data[2] > 0xff)
> >  		raw_data[2] = 0x100 - *off;
> >  
> > -	if (copy_from_user(raw_data+3, u, raw_data[2]))
> > +	if (copy_from_user(raw_data+3, u, min((u8)20, raw_data[2])))
> 
> Hmm ... this is quite an obfuscation just for the sake of making gcc 
> happy.

As long as CONFIG_DEBUG_STRICT_USER_COPY_CHECKS is not set it's just
making gcc happy, when it is set compilation fails, which is at least
annoying.

> Do other versions of gcc get this right? (i.e. is this gcc bug?)

Hm, don't remember exactly but older gcc versions triggered the same
error. I didn't check with more recent version.

I can try with gcc-4.5.1/gcc-4.5.2 or gcc-4.6.0 but it will take
some time to first compile those versions of gcc and then check how
they handle this case.

> Don't we have similar problems all over the place in the kernel?

I have not seen any similar warnings turning up in my compiles so
probably only a limited amount of places around the kernel are
affected, if at all (they might have worked around as well).
As my configs are rather optimized for the target machines I'm not
compiling most of the drivers.
Maybe some of Ingo's/Randy's (and whoever else is building them)
randconfig/allyesconfig/allmodconfig show other affected code (not
sure where to find their build logs though).

I don't think that many users of copy_from_user() have variable counts
of data to copy (most often a struct the sizeof which is then passed).

Bruno


> >  		return -EFAULT;
> >  	resp = picolcd_send_and_wait(data->hdev, REPORT_EE_WRITE, raw_data,
> >  			sizeof(raw_data));
> > 
>