On Mon, Jan 28, 2019 at 11:34 PM Or Gerlitz <gerlitz...@gmail.com> wrote: > > On Mon, Jan 28, 2019 at 2:10 PM Tonghao Zhang <xiangxia.m....@gmail.com> > wrote: > > On Mon, Jan 28, 2019 at 12:40 AM Or Gerlitz <gerlitz...@gmail.com> wrote: > > > > > > On Sun, Jan 27, 2019 at 1:06 PM <xiangxia.m....@gmail.com> wrote: > > > > From: Tonghao Zhang <xiangxia.m....@gmail.com> > > > > > > > > In some case, we may use multiple pedit actions to modify packets. > > > > The command shown as below: the last pedit action is effective. > > > > > > > @@ -2073,7 +2076,8 @@ static int alloc_mod_hdr_actions(struct > > > > mlx5e_priv *priv, > > > > if (!parse_attr->mod_hdr_actions) > > > > return -ENOMEM; > > > > > > > > - parse_attr->num_mod_hdr_actions = max_actions; > > > > + parse_attr->max_mod_hdr_actions = max_actions; > > > > + parse_attr->num_mod_hdr_actions = 0; > > > > > > why would we want to do this zeroing? what purpose does it serve? > > Because we use the num_mod_hdr_actions to store the number of actions > > we have parsed, > > and when we alloc it, we init it 0 as default. > > > > > On a probably related note, I suspect that the patch broke the caching > > > we do for modify header contexts, see mlx5e_attach_mod_hdr where we > > > look if a given set of modify header operations already has hw modify > > > header > > > context and we use it. > > > > > > To test that, put two tc rules with different matching but same set of > > > modify header > > > (pedit) actions and see that only one modify header context is used. > > > The patch does't break the cache, I think that different matching may > > share the same set of pedit actions. > > I suspect it does break it.. at [1] we have this code for the cache lookup: > > num_actions = parse_attr->num_mod_hdr_actions; > [..] > key.actions = parse_attr->mod_hdr_actions; > key.num_actions = num_actions; > > hash_key = hash_mod_hdr_info(&key); > > so we are doing the cached insertion and lookup with > parse_attr->num_mod_hdr_actions > which was zeroed along the way and not accounting for the full set of > pedit actions, agree? not really, before calling the mlx5e_attach_mod_hdr, we have call the offload_pedit_fields that will update the attr->num_mod_hdr_actions that indicate how many pedit action we parsed.
> [1] > https://elixir.bootlin.com/linux/latest/source/drivers/net/ethernet/mellanox/mlx5/core/en_tc.c#L179
