scsi_ch issues

[Kenneth D. Merry]

On Tue, Mar 22, 2005 at 10:33:55 -0600, Henry Miller wrote:
> 
> I posted a summery of this to -questions earlier this morning (you
> can read it there if you want background), I've dug into it more, and
> I think there is a bug in  
> cam/scsi/scsi_ch.c
> function chgetelemstatus.
> 
> The short summery is my tape changer says that PVOLTAG in pg_hdr is 
> FALSE, (or zeroed latter) in the first request of 
> scsi_read_element_status() (which freeBSD uses only to get the length
> of the data) on one of my changers, and thus chio(1) will not show
> the bar code label.   pg_hdr is not updated to reflect the location
> of result of the second call, when this flag is set.  Worse, it looks
> like the results of the first call is a dead pointer by the time it
> checks anyway.  Unless free/malloc work different in the kernel from
> userland this is a problem.
> 
> I've sniped out a lot of extra code so you can see what I mean.  
> 
> 
> data = (caddr_t)malloc(1024, M_DEVBUF, M_WAITOK);
>   --- Fill data to get size
> pg_hdr = \* data + some offset *\
>         /*
>          * Reallocate storage for descriptors and get them from the
>          * device.
>          */
>         free(data, M_DEVBUF);
>         data = (caddr_t)malloc(size, M_DEVBUF, M_WAITOK);
>   --- Fill data 
> copy_element_status(softc, pg_hdr->flags, desc, ces);
> 
> Note that pg_hdr is not valid at this point anyway.

Yes, that looks like a problem.

> This more or less happens in 4.x, 5.x and current if I read the code
> correctly.

It hasn't changed in a while, so I'm sure it's the same in every branch.

> Can someone who knows kernel code better than I look at this.  
> Should I send a pr?
> 
> Here is a patch against 4.9 (where I first found this problem) that
> might help to understand the fix.  I think it needs to be ported to
> all versions.  (and perhaps netbsd if there is where we get the
> code?)   
> 
> 
> *** scsi_ch.c.old       Tue Mar 22 11:16:15 2005
> --- scsi_ch.c   Tue Mar 22 11:09:13 2005
> ***************
> *** 1184,1189 ****
> --- 1184,1192 ----
>          * Fill in the user status array.
>          */
>         st_hdr = (struct read_element_status_header *)data;
> +         pg_hdr = (struct read_element_status_page_header
> *)((uintptr_t)st_hdr
> +
> +                   sizeof(struct read_element_status_header));
> +
>         avail = scsi_2btoul(st_hdr->count);
> 
>         if (avail != cesr->cesr_element_count) {
> 

When you apply the above patch, does everything work as expected?

If so, I can check in the patch.  (It looks correct.)  If not, we'll need
to do some more debugging to track down the problem.

Ken
-- 
Kenneth Merry
ken at kdm.org