Re: [Patch net] net_sched: avoid generating same handle for u32 filters

[Jamal Hadi Salim <jhs@mojatatu.com>]

On 07/17/14 20:34, Cong Wang wrote:
> From: Cong Wang <cwang@twopensource.com>
>
> When kernel generates a handle for a u32 filter, it tries to start
> from the max in the bucket. So when we have a filter with the max (fff)
> handle, it will cause kernel always generates the same handle for new
> filters. This can be shown by the following command:
>
> 	tc qdisc add dev eth0 ingress
> 	tc filter add dev eth0 parent ffff: protocol ip pref 770 handle 800::fff u32 match ip protocol 1 0xff
> 	tc filter add dev eth0 parent ffff: protocol ip pref 770 u32 match ip protocol 1 0xff
> 	...
>
> we will get some u32 filters with same handle:
>
>   # tc filter show dev eth0 parent ffff:
> filter protocol ip pref 770 u32
> filter protocol ip pref 770 u32 fh 800: ht divisor 1
> filter protocol ip pref 770 u32 fh 800::fff order 4095 key ht 800 bkt 0
>    match 00010000/00ff0000 at 8
> filter protocol ip pref 770 u32 fh 800::fff order 4095 key ht 800 bkt 0
>    match 00010000/00ff0000 at 8
> filter protocol ip pref 770 u32 fh 800::fff order 4095 key ht 800 bkt 0
>    match 00010000/00ff0000 at 8
> filter protocol ip pref 770 u32 fh 800::fff order 4095 key ht 800 bkt 0
>    match 00010000/00ff0000 at 8
>
> handles should be unique. This patch fixes it by looking up a bitmap,
> so that can guarantee the handle is as unique as possible. For compatibility,
> we still start from 0x800.
>

No fundamental objection - question though: What happens if i was to 
specify the handle when i provision the filter?
Side note: these handles + other info are used to "route" around
hash trees for filters.

I apologize in advance if i am slow in responding; i am on travel mode.

cheers,
jamal