Re: net: add seq_before/seq_after functions

On Wednesday 18 May 2011 14:38:39 Antonio Quartulli wrote:

Introduce two operations to handle comparison between packet sequence
numbers taking into account overflow/wraparound. Batman-adv uses
these functions already to check for successor packet even in case of
overflow.

Thanks for your efforts to bring that to the kernel. But when you prepare a 
patch then you have to add a signoff. And also David S. Miller is the 
maintainer for this header - it would be interesting to ask him first when we 
want to change that file.

I suggested yesterday (probably too late) that it would be good to check the
type of both parameters (similar to the min and max functions in
include/linux/kernel.h

#define seq_before(x, y) ({typeof(x) _d1 = (x); \
			  typeof(y) _d2 = (y); \
			  (void) (&_d1 == &_d2); \
			  typeof(x) _dummy = (_d1 - _d2); \
			  _dummy > smallest_signed_int(_dummy); })


And your seq_before/after conflicts with the one defined in ppp_generic.c

drivers/net/ppp_generic.c:232:0: warning: "seq_before" redefined [enabled by 
default]
include/linux/net.h:312:0: note: this is the location of the previous 
definition
drivers/net/ppp_generic.c:233:0: warning: "seq_after" redefined [enabled by 
default]
include/linux/net.h:314:0: note: this is the location of the previous 
definition

The definition there is only for u32 - thus you would have to remove it and 
check that it always gives the same result:
#define seq_before(a, b)        ((s32)((a) - (b)) < 0)
#define seq_after(a, b)         ((s32)((a) - (b)) > 0)

But I would say that they have a different definition of seq_before. Changing 
that behaviour for batman-adv would not be that problematic, but maybe for 
ppp.

A defintion which should fulfil the requirements for ppp could be:

#define seq_after(x, y) ({typeof(x) _d1 = (x); \
			  typeof(y) _d2 = (y); \
			  (void) (&_d1 == &_d2); \
			  typeof(x) _dummy = (_d2 - _d1); \
			  _dummy > smallest_signed_int(_dummy); })
#define seq_before(x, y) ({typeof(x) _d1 = (x); \
			  typeof(y) _d2 = (y); \
			  (void) (&_d1 == &_d2); \
			  typeof(x) _dummy = (_d1 - _d2); \
			  _dummy >= smallest_signed_int(_dummy); })

Of course the comment above the seq_before/seq_after would be wrong.

/* Checks if a sequence number x is a predecessor/successor of y.
 * they handle overflows/underflows and can correctly check for a
 * predecessor/successor unless the variable sequence number has grown by
 * more then 2**(bitwidth(x)-1).
 * This means that for a uint8_t with the maximum value 255, it would think:
 *  - when adding nothing - it is neither a predecessor nor a successor
 *  - before adding more than 128 to the starting value - it is a predecessor,
 *  - after adding more than 127 to the starting value - it is a successor */

I think there could be more candidates which would like to use this abstract 
functionality. Maybe some one else on linux-kernel or netdev has a suggestion.

Kind regards,
	Sven