Re: uclinux 2.4.32 panic

[Steve Clark <sclark@netwolves.com>]

Hi Willy,

Thanks for your analysis - see my comments below;

Steve

Willy TARREAU wrote:
> On Wed, May 24, 2006 at 05:24:33PM -0400, Steve Clark wrote:
> (...)
>
> > > > > > LR;  008727cc <rs_write+148/294>
> > > >
> > > > > > PC;  0090e5fc <memmove+25c/460>   <=====
> > > >
> > > > Trace; 0090e3a0 <memcpy+0/0>
> > > > Trace; 008727cc <rs_write+148/294>
> > > >
> > > > > > r8; 00956228 <tmp_buf+0/1000>
> > > >
> > > > Trace; 00872684 <rs_write+0/294>
>
> (...)
>
>
> > The hardware is the ActionTec DualPC Modem it has a conexant cx82100 
> > arm processor.
> > I can reproduce it at will by connecting to the internet and running 
> > nttcp thru it at the same time
> > I am scping file both ways from and to, and then finally starting a 
> > getty on /dev/ttyS0, the modem is
> > at ttyS1 and also ttyS0 is where all the kernel printk messages come out.
> >
> > When I start the getty if I have all the other traffic going it 
> > usually will panic in under a minute. IF I don't
> > have the getty running it will run for hours and not panic.
> >
> >
> >
> > 2.4.32-uc0 with patches from
> > http://www.bettina-attack.de/jonny/view.php/projects/uclinux_on_cx82100/
>
>
> Well, at least the cnxtserial.c file looks suspicious to me :
>
> static int rs_write(struct tty_struct * tty, int from_user,
>                     const unsigned char *buf, int count)
> {
>         int     c, total = 0;
>         unsigned long flags;
>         struct cnxt_serial *info = (struct cnxt_serial *)tty->driver_data;
>                                                          ^^^^^
>
>         if (serial_paranoia_check(info,tty->device, "rs_write"))
>                                        ^^^^^
>           return 0;
>
>         if (!tty || !info->xmit_buf)
>           return 0;
>
> => tty already referenced twice before the check. Either the check is
>    useless, or the person who wrote it had a good reason for it which
>    was not considered when writing the two lines above. I would suggest
>    to start from something like this :
>
> static int rs_write(struct tty_struct * tty, int from_user,
>                     const unsigned char *buf, int count)
> {
>         int     c, total = 0;
>         unsigned long flags;
>         struct cnxt_serial *info;
>
>         if (!tty)
>         	return 0;
>
>         info = (struct cnxt_serial *)tty->driver_data;
>         if (serial_paranoia_check(info, tty->device, "rs_write"))
>         	return 0;
>
>         if (!info->xmit_buf)
>         	return 0;

I did this - actually I matched to serial.c
> Further :
>
>           c = MIN(count, MIN(SERIAL_XMIT_SIZE - info->xmit_cnt - 1,
>                              SERIAL_XMIT_SIZE - info->xmit_head));
>           if (c <= 0)
>             break;
>
> => info->xmit_cnt and info->xmit_head are signed ints. If you encounter
>    memory corruption (eg: during your ethernet transfers) and those get
>    negative, nothing prevents the buffer from being overwritten past the
>    end.
>
> Further :
>
>           if (from_user) {
>             down(&tmp_buf_sem);
>             copy_from_user(tmp_buf, buf, c);
>             c = MIN(c, MIN(SERIAL_XMIT_SIZE - info->xmit_cnt - 1,
>                            SERIAL_XMIT_SIZE - info->xmit_head));
>         ^^^^^^^^^^^^^^^^^^
>             memcpy(info->xmit_buf + info->xmit_head, tmp_buf, c);
>             up(&tmp_buf_sem);
>           } else
>
> => What the hell is this ? c was assigned the same value above, so
>    we get :
>    c = MIN(MIN(count, MIN(SERIAL_XMIT_SIZE - info->xmit_cnt - 1,
>                        SERIAL_XMIT_SIZE - info->xmit_head)),
>                       MIN(SERIAL_XMIT_SIZE - info->xmit_cnt - 1,
>                        SERIAL_XMIT_SIZE - info->xmit_head));
>
>    I'm not sure this was what the developper originally intented to do,
>    but although useless, it does not seem incorrect. However, I don't
>    know if he wanted to further reduce the buffer for any reason.
>
> Also, it appears that nothing prevents any code running outside the
> loop from changing info->xmit_buf between the restore_flags() and
> the cli(). I don't know if this is functionnaly possible, but at
> least it is possible by memory corruption (eg: padding too large
> for a packet and writing zeroes past the end of one buffer).
I moved the cli() outside of the while loop - similar to serial.c

> You should definitely add printks or at least double checks
> everywhere within this loop I think.
I did and info->xmit_head is getting clobbered!
see below:
> That's all I can tell, I don't know this platform at all.
>
> Regards,
> Willy
xmit_buf+head=a2279c,tmp_buf=956228,c=250
rs_write:xmit_buf=a22000,tmp_buf=956228
xmit_buf+head=a22896,tmp_buf=956228,c=250
rs_write:xmit_buf=a22000,tmp_buf=956228
xmit_buf+head=a22990,tmp_buf=956228,c=19
rs_write:xmit_buf=170f1200,tmp_buf=956228
                              ^^^^^^
xmit_buf+head=170f1200,tmp_buf=956228,c=13
Unable to handle kernel NULL pointer dereference at virtual address 
0000003f
fault-common.c(97): start_code=0xdc4040, start_stack=0xe11f38)
Internal error: Oops: ffffffff
CPU: 0
pc : [<0090e6a0>]    lr : [<00872808>]    Not tainted
sp : 00a1fe94  ip : 00000001  fp : 00a1feb8
r10: 00000fff  r9 : 00956228  r8 : 0093d26c
r7 : 0000000d  r6 : 009922a0  r5 : 00955f00  r4 : 0000000d
r3 : 0000007e  r2 : 00000009  r1 : 009922ac  r0 : 170f120d
Flags: Nzcv  IRQs off  FIQs on  Mode SVC_32  Segment user
Control: C000107D
Process pppd (pid: 64, stackpage=00a1f000)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/