Re: [PATCH] add checksum selftest

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 24 June 2009, Mike Frysinger wrote:
> >> +static unsigned char __initdata do_csum_data1[] = {
> >> +     0x20,
> >> +};
> >> +static unsigned char __initdata do_csum_data2[] = {
> >> +     0x0d, 0x0a,
> >> +};
> >> +static unsigned char __initdata do_csum_data3[] = {
> >> +     0xff, 0xfb, 0x01,
> >> +};
> >
> > You define separate test vectors for each of the three
> > cases, which looks like it could be optimized by reusing
> > the same test vectors for each case.
> 
> i'm not really familiar with the interfaces to figure out how to do
> this ... i just added some printks to dump arguments/buffers and then
> copied & pasted ones that looked pretty different

I just mean you can consolidate

+struct do_csum_data {
+       unsigned short ret;
+       unsigned char *buff;
+       int len;
+};
+#define DO_CSUM_DATA(_num, _ret) \
+{ \
+       .ret = _ret, \
+       .buff = do_csum_data##_num, \
+       .len = ARRAY_SIZE(do_csum_data##_num), \
+}

and 

+struct csum_partial_data {
+       __wsum ret;
+       const void *buff;
+       int len;
+       __wsum wsum;
+};
+#define CSUM_PARTIAL_DATA(_num, _ret, _wsum) \
+{ \
+       .ret = _ret, \
+       .buff = csum_partial_data##_num, \
+       .len = ARRAY_SIZE(csum_partial_data##_num), \
+       .wsum = _wsum, \
+}

into something like

struct csum_combined_check_data {
	const char *buff;
	int len;
	unsigned short do_csum_ret;
	__wsum wsum;
	unsigned short csum_partial_ret;
};

#define CSUM_COMBINED_TEST_DATA(_num, _do_csum_ret,	\
			 _csum_partial_ret, _wsum)	\
{							\
	.buff = csum_partial_data##_num,		\
	.len = ARRAY_SIZE(csum_partial_data##_num),	\
	.do_csum_ret = _do_csum_ret,			\
	.wsum = _wsum,					\
	.csum_partial_ret = csum_partial_ret,		\
};

This could cut down the length of the module significantly,
without changing any of the semantics.

> >> +static struct csum_partial_data __initdata csum_partial_data[] = {
> >> +     CSUM_PARTIAL_DATA(1,  0x00000074, 0x0),
> >> +     CSUM_PARTIAL_DATA(2,  0x00000a0d, 0x0),
> >> +     CSUM_PARTIAL_DATA(3,  0x0000fe00, 0x0),
> >> +     CSUM_PARTIAL_DATA(5,  0x00005084, 0x0),
> >> +     CSUM_PARTIAL_DATA(8,  0x1101eefe, 0x11016a80),
> >> +     CSUM_PARTIAL_DATA(8b, 0x00008781, 0x847e),
> >> +     CSUM_PARTIAL_DATA(9,  0x1101eefe, 0x11016b80),
> >> +};
> >
> > For partial checksums, the result has to be folded into a 16-bit
> > number using csum_fold(), because csum_partial and other functions
> > return a 32-bit __wsum that can take many equivalent values taht
> > are all correct.
> 
> i hear your words, but i understand them not ;)

The problem is that IP checksumming is only defined for 16-bit
words. We use __wsum (32 bits) as an intermediate in the networking
stack so we can consolidate the folding in one place. If you have
a test vector that results in checksum 0xffff (as a well-formed
packet should), the __wsum could be one of 0x0000ffff, 0xffff0000,
0xffffffff, 0x1234edcb, for any other value x where
(((x >> 16) + (x & 0xffff)) >> 16 + ((x >> 16) + (x & 0xffff)))
& 0xffff = 0xffff. The specific __wsum returned by csum_partial()
is implementation specific, so you cannot compare it to a
precomputed value unless sending it through csum_fold().

	Arnd <><