Re: [PATCH v12 01/12] lib: introduce copy_struct_{to,from}_user helpers

[Al Viro]

On Fri, Sep 06, 2019 at 12:49:44AM +0100, Al Viro wrote:
> On Fri, Sep 06, 2019 at 09:00:03AM +1000, Aleksa Sarai wrote:
> > > > +			return -EFAULT;
> > > > +	}
> > > > +	/* Copy the interoperable parts of the struct. */
> > > > +	if (__copy_to_user(dst, src, size))
> > > > +		return -EFAULT;
> > > 
> > > Why not simply clear_user() and copy_to_user()?
> > 
> > I'm not sure I understand what you mean -- are you asking why we need to
> > do memchr_inv(src + size, 0, rest) earlier?
> 
> I'm asking why bother with __ and separate access_ok().
> 
> > > 	if ((unsigned long)addr & 1) {
> > > 		u8 v;
> > > 		if (get_user(v, (__u8 __user *)addr))
> > > 			return -EFAULT;
> > > 		if (v)
> > > 			return -E2BIG;
> > > 		addr++;
> > > 	}
> > > 	if ((unsigned long)addr & 2) {
> > > 		u16 v;
> > > 		if (get_user(v, (__u16 __user *)addr))
> > > 			return -EFAULT;
> > > 		if (v)
> > > 			return -E2BIG;
> > > 		addr +=2;
> > > 	}
> > > 	if ((unsigned long)addr & 4) {
> > > 		u32 v;
> > > 		if (get_user(v, (__u32 __user *)addr))
> > > 			return -EFAULT;
> > > 		if (v)
> > > 			return -E2BIG;
> > > 	}
> > > 	<read the rest like you currently do>
> 
> Actually, this is a dumb way to do it - page size on anything
> is going to be a multiple of 8, so you could just as well
> read 8 bytes from an address aligned down.  Then mask the
> bytes you don't want to check out and see if there's anything
> left.
> 
> You can have readability boundaries inside a page - it's either
> the entire page (let alone a single word) being readable, or
> it's EFAULT for all parts.
> 
> > > would be saner, and things like x86 could trivially add an
> > > asm variant - it's not hard.  Incidentally, memchr_inv() is
> > > an overkill in this case...
> > 
> > Why is memchr_inv() overkill?
> 
> Look at its implementation; you only care if there are
> non-zeroes, you don't give a damn where in the buffer
> the first one would be.  All you need is the same logics
> as in "from userland" case
> 	if (!count)
> 		return true;
> 	offset = (unsigned long)from & 7
> 	p = (u64 *)(from - offset);
> 	v = *p++;
> 	if (offset) {	// unaligned
> 		count += offset;
> 		v &= ~aligned_byte_mask(offset); // see strnlen_user.c
> 	}
> 	while (count > 8) {
> 		if (v)
> 			return false;
> 		v = *p++;
> 		count -= 8;
> 	}
> 	if (count != 8)
> 		v &= aligned_byte_mask(count);
> 	return v == 0;
> 
> All there is to it...

... and __user case would be pretty much this with
	if (user_access_begin(from, count)) {
		....
		user_access_end();
	}
wrapped around the damn thing - again, see strnlen_user.c, with
	unsafe_get_user(v, p++, efault);
instead of those
	v = *p++;

Calling conventions might need some thinking - it might be
	* all read, all zeroes
	* non-zero found
	* read failed
so we probably want to map the "all zeroes" case to 0,
"read failed" to -EFAULT and "non-zero found" to something
else.  Might be positive, might be some other -E.... - not
sure if E2BIG (or EFBIG) makes much sense here.  Need to
look at the users...