On 2/28/24 18:01, Kees Cook wrote:

On Wed, Feb 28, 2024 at 02:41:48PM -0800, Jakub Kicinski wrote:

quoted

On Wed, 28 Feb 2024 13:46:10 -0800 Kees Cook wrote:

quoted

I really don't like hiding these trailing allocations from the compiler.
Why can't something like this be done (totally untested):

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 118c40258d07..dae6df4fb177 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h

@@ -2475,6 +2475,8 @@ struct net_device {
  	/** @page_pools: page pools created for this netdevice */
  	struct hlist_head	page_pools;
  #endif
+	u32			priv_size;
+	u8			priv_data[] __counted_by(priv_size) __aligned(NETDEV_ALIGN);

I like, FWIW, please submit! :)

So, I found several cases where struct net_device is included in the
middle of another structure, which makes my proposal more awkward. But I
also don't understand why it's in the _middle_. Shouldn't it always be
at the beginning (with priv stuff following it?)
Quick search and examined manually: git grep 'struct net_device [a-z0-9_]*;'

struct rtw89_dev
struct ath10k
etc.

Some even have two included (?)

But I still like the idea -- Gustavo has been solving these cases with
having two structs, e.g.:

struct net_device {
	...unchanged...
};

struct net_device_alloc {
	struct net_device	dev;
	u32			priv_size;
	u8			priv_data[] __counted_by(priv_size) __aligned(NETDEV_ALIGN);
};

And internals can use struct net_device_alloc...

Yep, we should really consider going with the above, otherwise we would
have to do something like the following, to avoid having the flexible-array
member nested in the middle of other structs:

struct net_device {
	struct_group_tagged(net_device_hdr, hdr,
		...
		u32			priv_size;
	);
	u8			priv_data[] __counted_by(priv_size) __aligned(NETDEV_ALIGN);
}

We are grouping together the members in `struct net_device`, except the
flexible-array member, into a tagged `struct net_device_hdr`. This allows
us to exclude the flex array from its inclusion in any other struct
that contains `struct net_device` as a member without having to create
a completely separate struct definition.

And let's take as example `struct hfi1_netdev_rx`, where `struct net_device` is
included in the beginning:

drivers/infiniband/hw/hfi1/netdev.h:
struct hfi1_netdev_rx {

-	struct net_device rx_napi;
+       struct net_device_hdr rx_napi;


         struct hfi1_devdata *dd;
         struct hfi1_netdev_rxq *rxq;
         int num_rx_q;
         int rmt_start;
         struct xarray dev_tbl;
         /* count of enabled napi polls */
         atomic_t enabled;
         /* count of netdevs on top */
         atomic_t netdevs;
};

Of course we would also have to update the code that access `struct net_device`
members through `rx_napi` in `struct hfi1_netdev_rx`.

I'm currently working on the above solution for all the cases where having two
separate structs is not currently feasible. And with that we are looking to enable
`-Wflex-array-member-not-at-end`

So, if we can prevent this from the beginning it'd be really great. :)

--
Gustavo