by Petr Machata

[permalink] [raw]

Subject: Re: [PATCH net-next v2 3/6] net: dcb: add new rewrite table

<[email protected]> writes:

>> > + rewr = nla_nest_start_noflag(skb, DCB_ATTR_DCB_REWR_TABLE);
>> > + if (!rewr)
>> > + return -EMSGSIZE;
>>
>> This being new code, don't use _noflag please.
>
> Ack.
>
>>
>> > +
>> > + spin_lock_bh(&dcb_lock);
>> > + list_for_each_entry(itr, &dcb_rewr_list, list) {
>> > + if (itr->ifindex == netdev->ifindex) {
>> > + enum ieee_attrs_app type =
>> > + dcbnl_app_attr_type_get(itr->app.selector);
>> > + err = nla_put(skb, type, sizeof(itr->app), &itr->app);
>> > + if (err) {
>> > + spin_unlock_bh(&dcb_lock);
>>
>> This should cancel the nest started above.
>
> Yes, it should.
>
>>
>> I wonder if it would be cleaner in a separate function, so that there
>> can be a dedicated clean-up block to goto.
>
> Well yes. That would make sense if the function were reused for both APP
> and rewr.

I meant purely for to make the cleanup clear. The function would be
approximately:

static int dcbnl_ieee_fill_rewr(struct sk_buff *skb, struct net_device *netdev)
{
struct dcb_app_type *itr;
struct nlattr *rewr;
int err;

rewr = nla_nest_start_noflag(skb, DCB_ATTR_DCB_REWR_TABLE);
if (!rewr)
return -EMSGSIZE;

spin_lock_bh(&dcb_lock);
list_for_each_entry(itr, &dcb_rewr_list, list) {
if (itr->ifindex == netdev->ifindex) {
enum ieee_attrs_app type =
dcbnl_app_attr_type_get(itr->app.selector);
err = nla_put(skb, type, sizeof(itr->app), &itr->app);
if (err)
goto err_out;
}
}

spin_unlock_bh(&dcb_lock);
nla_nest_end(skb, rewr);
return 0;

err_out:
spin_unlock_bh(&dcb_lock);
nla_nest_cancel(skb, rewr);
return err;
}

Which uses an idiomatic style with the cleanup block at the end, instead
of stashing the individual cleanups before the return statement. I find
it easier to reason about.

But it's not a big deal. Your thing is readable just fine.

> Though in the APP equivalent code, nla_nest_start_noflag is used, and
> dcbnl_ops->getdcbx() is called. Is there any userspace side-effect of
> using nla_nest_start for APP too?

Yeah, the clients would be looking for code DCB_ATTR_IEEE_APP_TABLE, but
would get DCB_ATTR_IEEE_APP_TABLE | NLA_F_NESTED, and get confused.

For reuse between APP_TABLE and REWR_TABLE, you could just always call
_noflag in the helper, and pass the actual attribute in an argument.
Then the argument would be either DCB_ATTR_DCB_REWR_TABLE | NLA_F_NESTED,
or just plain DCB_ATTR_IEEE_APP_TABLE.

But that makes the code less clear, and I don't feel it brings much.

> dcbnl_ops->getdcbx() would then be left outside of the shared function.
> Does that call even have to hold the dcb_lock? Not as far as I can tell.
>
> something like:
>
> err = dcbnl_app_table_get(ndev, skb, &dcb_app_list,
> DCB_ATTR_IEEE_APP_TABLE);
> if (err)
> return -EMSGSIZE;
>
> err = dcbnl_app_table_get(ndev, skb, &dcb_rewr_list,
> DCB_ATTR_DCB_REWR_TABLE);
> if (err)
> return -EMSGSIZE;
>
> if (netdev->dcbnl_ops->getdcbx)
> dcbx = netdev->dcbnl_ops->getdcbx(netdev); <-- without lock held
> else
> dcbx = -EOPNOTSUPP;
>
> Let me hear your thoughts.

Yeah, and the dcbx stuff is the added wrinkle.

Dunno, I'd not force it. This redundancy is not great, but the code is
small and easy to understand, so I find it's not an issue.

2023-01-19 10:39:20

by Petr Machata

[permalink] [raw]

Subject: Re: [PATCH net-next v2 3/6] net: dcb: add new rewrite table

Dan Carpenter <[email protected]> writes:

> On Wed, Jan 18, 2023 at 11:54:23AM +0100, Petr Machata wrote:
>> > +
>> > + spin_lock_bh(&dcb_lock);
>> > + list_for_each_entry(itr, &dcb_rewr_list, list) {
>> > + if (itr->ifindex == netdev->ifindex) {
>> > + enum ieee_attrs_app type =
>> > + dcbnl_app_attr_type_get(itr->app.selector);
>> > + err = nla_put(skb, type, sizeof(itr->app), &itr->app);
>> > + if (err) {
>> > + spin_unlock_bh(&dcb_lock);
>>
>> This should cancel the nest started above.
>>
>
> If you see a bug like this, you may as well ask Julia or me to add a
> static checker warning for it. We're both already on the CC list but we
> might not be following the conversation closely...

I'll try to remember next time.

> In Smatch, I thought it would be easy but it turned out I need to add
> a hack around to change the second nla_nest_start_noflag() from unknown
> to valid pointer.
>
> diff --git a/check_unwind.c b/check_unwind.c
> index a397afd2346b..3128476cbeb6 100644
> --- a/check_unwind.c
> +++ b/check_unwind.c
> @@ -94,6 +94,11 @@ static struct ref_func_info func_table[] = {
>
> { "ieee80211_alloc_hw", ALLOC, -1, "$", &valid_ptr_min_sval, &valid_ptr_max_sval },
> { "ieee80211_free_hw", RELEASE, 0, "$" },
> +
> + { "nla_nest_start_noflag", ALLOC, 0, "$", &valid_ptr_min_sval, &valid_ptr_max_sval },
> + { "nla_nest_start", ALLOC, 0, "$", &valid_ptr_min_sval, &valid_ptr_max_sval },
> + { "nla_nest_end", RELEASE, 0, "$" },
> + { "nla_nest_cancel", RELEASE, 0, "$" },
> };
>
> static struct smatch_state *unmatched_state(struct sm_state *sm)
> diff --git a/smatch_data/db/kernel.return_fixes b/smatch_data/db/kernel.return_fixes
> index fa4ed4ba5f0f..4782c5e10cdb 100644
> --- a/smatch_data/db/kernel.return_fixes
> +++ b/smatch_data/db/kernel.return_fixes
> @@ -90,3 +90,4 @@ dma_resv_wait_timeout s64min-(-1),1-s64max 1-s64max[<=$3]
> mmc_io_rw_direct_host s32min-(-1),1-s32max (-4095)-(-1)
> ad3552r_transfer s32min-(-1),1-s32max (-4095)-(-1)
> adin1110_read_reg s32min-(-1),1-s32max (-4095)-(-1)
> +nla_nest_start_noflag 0-u64max 4096-ptr_max
>
> Unfortunately, there is something weird going on and only my unreleased
> version of Smatch finds the bug:
>
> net/dcb/dcbnl.c:1306 dcbnl_ieee_fill() warn: 'skb' from nla_nest_start_noflag() not released on lines: 1160,1171,1184,1197,1207,1217,1222,1232,1257.
> net/dcb/dcbnl.c:1502 dcbnl_cee_fill() warn: 'skb' from nla_nest_start_noflag() not released on lines: 1502.

Looking at a couple of those, yeah, it looks legit. Those are missing
the cancel on error returns.

> I have been working on that check recently... Both the released and
> unreleased versions of Smatch have the following complaints:
>
> net/dcb/dcbnl.c:400 dcbnl_getnumtcs() warn: 'skb' from nla_nest_start_noflag() not released on lines: 396.
> net/dcb/dcbnl.c:1061 dcbnl_build_peer_app() warn: 'skb' from nla_nest_start_noflag() not released on lines: 1061.
> net/dcb/dcbnl.c:1359 dcbnl_cee_pg_fill() warn: 'skb' from nla_nest_start_noflag() not released on lines: 1324,1342.

Likewise. Strange that each version reports a different subset. Or is
that just selective quoting?