Re: [PATCH][next] wifi: mac80211: ieee80211_i: Avoid dozens of -Wflex-array-member-not-at-end warnings

["Gustavo A. R. Silva" <gustavo@embeddedor.com>]

On 25/10/24 15:16, Johannes Berg wrote:
> On Fri, 2024-10-25 at 15:10 -0600, Gustavo A. R. Silva wrote:
>>
>> On 25/10/24 14:48, Johannes Berg wrote:
>>> On Fri, 2024-10-25 at 14:36 -0600, Gustavo A. R. Silva wrote:
>>>>>>
>>>>>> Yeah, I was actually going to mention this commit, as it's the one that introduced
>>>>>> that `bool radar_detected` to the flex struct. However, it wasn't obvious to me
>>>>>> how `struct ieee80211_chanctx_conf conf` could overwrite `radar_detected` as I didn't
>>>>>> see `conf->drv_priv` being accessed through `struct struct ieee80211_chanctx_conf`.
>>>>>
>>>>> You have to look at the drivers, see hwsim_clear_chanctx_magic() for
>>>>> example; I wonder why hwsim_check_chanctx_magic() never caught this.
>>>>
>>>> Sorry, I actually meant through `struct ieee80211_chanctx`. Something like:
>>>>
>>>> struct ieee80211_chanctx *foo;
>>>> ...
>>>>
>>>> foo->conf.drv_priv[i] = something;
>>>>
>>>> or
>>>>
>>>> struct bar *ptr = (void *)foo->conf->drv_priv;
>>>> then write something into *ptr...
>>>>
>>>> In the above cases the code will indeed overwrite `radar_detected`.
>>>
>>> Right, that's what it does though, no? Except it doesn't have, in the
>>> driver, "foo->conf." because mac80211 only gives it a pointer to conf,
>>> so it's only "conf->drv_priv" (and it's the "struct bar" example.)
>>
>> OK, so do you mean that pointer to `conf` is actually coming from
>> `foo->conf`?
> 
> Well depends what code you're looking at? I guess we should get more
> concrete now. Let's say hwsim:
> 
> struct hwsim_chanctx_priv {
>          u32 magic;
> };
> 
> ...
> 
> static inline void hwsim_set_chanctx_magic(struct ieee80211_chanctx_conf *c)
> {
>          struct hwsim_chanctx_priv *cp = (void *)c->drv_priv;
>          cp->magic = HWSIM_CHANCTX_MAGIC;
> }
> 
> probably shouldn't be marked 'inline' now that I look at it :-)
> 
> This is being called in hwsim itself, of course:
> 
> static int mac80211_hwsim_add_chanctx(struct ieee80211_hw *hw,
>                                        struct ieee80211_chanctx_conf *ctx)
> {
>          hwsim_set_chanctx_magic(ctx);
> ...
> 
> which is only referenced as a function pointer in the ops:
> 
> static const struct ieee80211_ops mac80211_hwsim_mchan_ops = {
> ...
> 	.add_chanctx = mac80211_hwsim_add_chanctx,
> 
> (via some macros)
> 
> 
> And that's called by mac80211:
> 
> static inline int drv_add_chanctx(struct ieee80211_local *local,
>                                    struct ieee80211_chanctx *ctx)
> {
>          int ret = -EOPNOTSUPP;
> 
>          might_sleep();
>          lockdep_assert_wiphy(local->hw.wiphy);
> 
>          trace_drv_add_chanctx(local, ctx);
>          if (local->ops->add_chanctx)
>                  ret = local->ops->add_chanctx(&local->hw, &ctx->conf);
> 
> 
> so you see that the struct ieee80211_chanctx is never passed to the
> driver, but instead &ctx->conf, which is the struct with the flex array
> for driver priv.
> 
> So in this example, struct hwsim_chanctx_priv::magic overlaps the
> radar_detected value.
> 
> 
> (The allocation happens via chanctx_data_size.)

Ah, I see now. Thanks so much for this!

It really obscures the whole thing when pointers to flex structs
are passed to functions, and then the flex-array member is finally
accessed after a few more calls.

This is precisely why -Wfamnae is needed, because it's not that
obvious for people to immediately notice when they are introducing
this kinds of bugs.

OK, I'll send v2, shortly.

Thanks!
--
Gustavo