dtrace discuss - Oct 2006 - Using hm_pad2 confuses dscript ??

I am chasing down a hment_remove panic on an amd64 system.
As I am suspecting hment corruption, I thought of tracking
the PFNs in hm_pad2 field of ''struct hment'' as this might
improve my chances of detecting any inconsisteny in hme list
earlier.

With that in mind, I added the following line to
hment_insert() :

    #ifdef __amd64
        hm->hm_pad2 = (uint32_t) pp->p_pagenum;
    #endif

immediately after :
    pp->p_mapping = hm;

One of the probes in my D-script is:

hment_remove:entry
/self->rel/
{
        printf("ts: %lu pp: %p emb: %x ht: %p", timestamp, arg0,
args[0]->p_embed, arg1);
}

When I run DTrace on the amd64 system with the above change
(both unix and genunix built and copied), the page structure
pointer that DTrace prints is off by a few bytes from the actual
value (once it was 0x20bytes; once it was way off ~0x100bytes).

I haven''t checked whether other structure pointers are ok(for
example: htable).

If I get rid off the hm_pad2 assignment, it works fine.
What am I missing here?

thanks,
surya
PS: System is s10 118855-19

It could be that the CTF data is corrupt since you''re copying some, but
not
all kernel modules. You can use mdb -k to examine the offsets for page_t to
confirm that they''re what you expect. You can also run DTrace with the
-S
flag to see the DIF disassembly output; check this for the correct page_t
offsets.

Adam

On Fri, Oct 27, 2006 at 05:32:04PM +0530, Surya.Prakki at Sun.COM
wrote:
> I am chasing down a hment_remove panic on an amd64 system.
> As I am suspecting hment corruption, I thought of tracking
> the PFNs in hm_pad2 field of ''struct hment'' as this might
> improve my chances of detecting any inconsisteny in hme list
> earlier.
> 
> With that in mind, I added the following line to
> hment_insert() :
> 
>     #ifdef __amd64
>         hm->hm_pad2 = (uint32_t) pp->p_pagenum;
>     #endif
> 
> immediately after :
>     pp->p_mapping = hm;
> 
> One of the probes in my D-script is:
> 
> hment_remove:entry
> /self->rel/
> {
>         printf("ts: %lu pp: %p emb: %x ht: %p", timestamp, arg0,
> args[0]->p_embed, arg1);
> }
> 
> When I run DTrace on the amd64 system with the above change
> (both unix and genunix built and copied), the page structure
> pointer that DTrace prints is off by a few bytes from the actual
> value (once it was 0x20bytes; once it was way off ~0x100bytes).
> 
> I haven''t checked whether other structure pointers are ok(for
> example: htable).
> 
> If I get rid off the hm_pad2 assignment, it works fine.
> What am I missing here?
> 
> thanks,
> surya
> PS: System is s10 118855-19
> 
> 
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org
-- 
Adam Leventhal, Solaris Kernel Development       http://blogs.sun.com/ahl