On 7/26/2019 5:01 PM, Marcelo Ricardo Leitner wrote:
> On Fri, Jul 26, 2019 at 08:39:43PM +0800, wenxu wrote:
>>
>> 在 2019/7/26 20:19, Or Gerlitz 写道:
>>> On Fri, Jul 26, 2019 at 12:24 AM Saeed Mahameed <sae...@mellanox.com> wrote:
>>>> On Thu, 2019-07-25 at 19:24 +0800, we...@ucloud.cn wrote:
>>>>> From: wenxu <we...@ucloud.cn>
>>>>>
>>>>> The flow_cls_common_offload prio is zero
>>>>>
>>>>> It leads the invalid table prio in hw.
>>>>>
>>>>> Error: Could not process rule: Invalid argument
>>>>>
>>>>> kernel log:
>>>>> mlx5_core 0000:81:00.0: E-Switch: Failed to create FDB Table err -22
>>>>> (table prio: 65535, level: 0, size: 4194304)
>>>>>
>>>>> table_prio = (chain * FDB_MAX_PRIO) + prio - 1;
>>>>> should check (chain * FDB_MAX_PRIO) + prio is not 0
>>>>>
>>>>> Signed-off-by: wenxu <we...@ucloud.cn>
>>>>> ---
>>>>> drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c | 4 +++-
>>>>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git
>>>>> a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
>>>>> b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
>>>>> index 089ae4d..64ca90f 100644
>>>>> --- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
>>>>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
>>>>> @@ -970,7 +970,9 @@ static int esw_add_fdb_miss_rule(struct
>>>> this piece of code isn't in this function, weird how it got to the
>>>> diff, patch applies correctly though !
>>>>
>>>>> mlx5_eswitch *esw)
>>>>> flags |= (MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT |
>>>>> MLX5_FLOW_TABLE_TUNNEL_EN_DECAP);
>>>>>
>>>>> - table_prio = (chain * FDB_MAX_PRIO) + prio - 1;
>>>>> + table_prio = (chain * FDB_MAX_PRIO) + prio;
>>>>> + if (table_prio)
>>>>> + table_prio = table_prio - 1;
>>>>>
>>>> This is black magic, even before this fix.
>>>> this -1 seems to be needed in order to call
>>>> create_next_size_table(table_prio) with the previous "table prio" ?
>>>> (table_prio - 1) ?
>>>>
>>>> The whole thing looks wrong to me since when prio is 0 and chain is 0,
>>>> there is not such thing table_prio - 1.
>>>>
>>>> mlnx eswitch guys in the cc, please advise.
>>> basically, prio 0 is not something we ever get in the driver, since if
>>> user space
>>> specifies 0, the kernel generates some random non-zero prio, and we support
>>> only prios 1-16 -- Wenxu -- what do you run to get this error?
>>>
>>>
>> I run offload with nfatbles(but not tc), there is no prio for each rule.
>>
>> prio of flow_cls_common_offload init as 0.
>>
>> static void nft_flow_offload_common_init(struct flow_cls_common_offload
>> *common,
>>
>> __be16 proto,
>> struct netlink_ext_ack *extack)
>> {
>> common->protocol = proto;
>> common->extack = extack;
>> }
>>
>>
>> flow_cls_common_offload
>
> Note that on
> [PATCH net-next] netfilter: nf_table_offload: Fix zero prio of
> flow_cls_common_offload
> I asked Pablo on how nftables should behave on this situation.
>
> It's the same issue as in the patch above but being fixed at a
> different level.
That's better, since the original code relied on not having prio 0 as valid,
the suggested fix (net/mlx5e: Fix zero table prio set by user) maps NFT offload
prio 0 and tc prio 1 to the same
hardware table. This is wrong and can cause issues.
