misaligned data

Two more tidbits on the misaligned data problem in 2.5.53:
o using -g instead of -O2 compiles just fine

o The generating the warning is the "tv_add" call in sctp_pack_cookie():

	...
        /* Set an expiration time for the cookie.  */
        do_gettimeofday(&cookie->c.expiration);
        tv_add(&asoc->cookie_life, &cookie->c.expiration);
	...

But I'm under the impression the do_gettimeofday() above would puke when trying to access the same misaligned timeval_t as well. tv_add() catches the problem since it is a macro and code gets generated to directly reference the struct instead of an anonymous address.

The asm output is:
...

        ldw 5144(%r8),%r19
        ldw 8(%r6),%r20
        add,l %r20,%r19,%r20
        ldd 5136(%r8),%r21
        ldd 52(%r3),%r19		# line 2347 is misaligned.
        extrd,s %r20,63,32,%r20
        add,l %r19,%r21,%r25
        ldil L'999999,%r19
        cmpb,*>= %r19,%r20,.L1873
	copy %r4,%r27

	ldil L'-1000000,%r19
	ldo 1(%r25),%r25
	ldo R'-1000000(%r19),%r19
	add,l %r20,%r19,%r20
.L1873:
	std %r25,52(%r3)		# also misaligned

Based on the math, looks like %r25 has the "seconds" field and %r19 the microseconds.

Oh, only one definition and usage of tv_add().
The offending field is "&cookie->c.expiration".
cookie is type "sctp_signed_cookie_t". (include/net/sctp/structs.h)

/* The format of our cookie that we send to our peer. */

typedef struct sctp_signed_cookie {
        __u8 signature[SCTP_SECRET_SIZE];
	sctp_cookie_t c;
} sctp_signed_cookie_t;

struct sctp_cookie mostly uses __u32 except for "struct timeval expiration". "struct timeval" is declared in include/linux/time.h.
Not sure if it's safe to just pad "struct sctp_cookie" or if something else needs to be done.

If this isn't a parisc toolchain bug, then other arches (eg ia64) will need a fix as well.

cheers,
grant

I'd wonder more about the packed attribute on sctp_init_chunk_t and sctp_inithdr_t. What version of gcc?

Dave

J. David Anglin                                  [email protected]
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

sorry...I overlooked the attribute.

gcc version 3.0.4

I don't recall if I tried gcc-3.2.

thanks,
grant

On Fri, Dec 27, 2002 at 09:49:18PM -0500, John David Anglin wrote:
...

I think that's the problem.

I've wimped out and just disabled CONFIG_IP_SCTP.
I really want to chase final problems with 2.4 and get $#@&% XF86 working on the OB600CT again (yes, XF86 4.0 worked, 4.2 does not :^( )

In reflecting on it a bit more, I'm pretty sure I tried gcc-3.2 and came to the same result.
("gcc version 3.2.2 20021212 (Debian prerelease)")

grant

I am not aware of any specific code in the backend related to the handling of packed structures. So, this is likely a generic problem, maybe on big endian machines.

There are some fixes in the passing of small structs in 3.3 (less than 8 bytes). This affects varargs and library functions that pass small structs by value.

Dave

J. David Anglin                                  [email protected]
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)