[LARTC] Problem setting shift value in tcindex filter on big endian

[LARTC] Problem setting shift value in tcindex filter on big endian

[Ole Reinartz]

Hi all,

I'm trying to get some DiffServ QoS shaping to work on an XScale 
machine, running big endian. I'm setting it up with tc. Using the 
tcindex filter I found that regardless what shift value I enter, only 
'0' is returned when I list the filters afterwards. The very same rules 
work fine on my (little endian) PC.
Looking at the code (iproute2-2.6.18-061002) I found that tc (in 
tc/f_tcindex.c, line 72 and after) sends the shift value to the kernel 
as an int. The kernel, however, expects it as a 'u16' 
(net/sched/cls_tcindex.c, around line 250 depending on the exact kernel 
version). I checked 2.6 kernel versions back until 2.6.11.

So... do we have a type mismatch here? As 'shift' is the last parameter 
in the buffer, this works still very well on a little endian machine, 
however on a big endian machine allways 0 is received in the kernel. To 
check that I changed the type of the shift value to unsigned short in 
tc, and that fixed it for me.
Someone interested in a patch?


Regards
Ole

_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Re: [LARTC] Problem setting shift value in tcindex filter on big

[Patrick McHardy]

Please send bugreports to netdev@vger.kernel.org.

Ole Reinartz wrote:
> I'm trying to get some DiffServ QoS shaping to work on an XScale
> machine, running big endian. I'm setting it up with tc. Using the
> tcindex filter I found that regardless what shift value I enter, only
> '0' is returned when I list the filters afterwards. The very same rules
> work fine on my (little endian) PC.
> Looking at the code (iproute2-2.6.18-061002) I found that tc (in
> tc/f_tcindex.c, line 72 and after) sends the shift value to the kernel
> as an int. The kernel, however, expects it as a 'u16'
> (net/sched/cls_tcindex.c, around line 250 depending on the exact kernel
> version). I checked 2.6 kernel versions back until 2.6.11.


It appears this was broken during some ->change operation fixes
in 2.6.11.

> So... do we have a type mismatch here? As 'shift' is the last parameter
> in the buffer, this works still very well on a little endian machine,
> however on a big endian machine allways 0 is received in the kernel. To
> check that I changed the type of the shift value to unsigned short in
> tc, and that fixed it for me.
> Someone interested in a patch?


Yes, but its the kernel that needs to be fixed to expect a u32.

_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Re: [LARTC] Problem setting shift value in tcindex filter on big endian machine

[Patrick McHardy]

Please send bugreports to netdev@vger.kernel.org.

Ole Reinartz wrote:
> I'm trying to get some DiffServ QoS shaping to work on an XScale
> machine, running big endian. I'm setting it up with tc. Using the
> tcindex filter I found that regardless what shift value I enter, only
> '0' is returned when I list the filters afterwards. The very same rules
> work fine on my (little endian) PC.
> Looking at the code (iproute2-2.6.18-061002) I found that tc (in
> tc/f_tcindex.c, line 72 and after) sends the shift value to the kernel
> as an int. The kernel, however, expects it as a 'u16'
> (net/sched/cls_tcindex.c, around line 250 depending on the exact kernel
> version). I checked 2.6 kernel versions back until 2.6.11.


It appears this was broken during some ->change operation fixes
in 2.6.11.

> So... do we have a type mismatch here? As 'shift' is the last parameter
> in the buffer, this works still very well on a little endian machine,
> however on a big endian machine allways 0 is received in the kernel. To
> check that I changed the type of the shift value to unsigned short in
> tc, and that fixed it for me.
> Someone interested in a patch?


Yes, but its the kernel that needs to be fixed to expect a u32.