On Wed, Aug 18, 2021 at 06:33:39PM +0900, Vincent MAILHOL wrote:

On Wed. 18 Aug 2021 at 18:03, Kees Cook [off-list ref] wrote:

quoted

On Wed, Aug 18, 2021 at 04:55:20PM +0900, Vincent MAILHOL wrote:

quoted

At the end, the only goal of raw_msg[] is to have a tag pointing
to the beginning of the union. It would be virtually identical to
something like:
|    u8 raw_msg[];
|    union {
|        /* ... */
|    } __packed ;

I had a look at your work and especially at your struct_group() macro.
Do you think it would make sense to introduce a union_group()?

Result would look like:

|    union_group_attr(urb_msg, __packed, /* raw_msg renamed to urb_msg */
|        struct es58x_fd_tx_conf_msg tx_conf_msg;
|        u8 tx_can_msg_buf[ES58X_FD_TX_BULK_MAX * ES58X_FD_CANFD_TX_LEN];
|        u8 rx_can_msg_buf[ES58X_FD_RX_BULK_MAX * ES58X_FD_CANFD_RX_LEN];
|        struct es58x_fd_echo_msg echo_msg[ES58X_FD_ECHO_BULK_MAX];
|        struct es58x_fd_rx_event_msg rx_event_msg;
|        struct es58x_fd_tx_ack_msg tx_ack_msg;
|        __le64 timestamp;
|        __le32 rx_cmd_ret_le32;
|    );

And I can then use urb_msg in place of the old raw_msg (might
need a bit of rework here and there but I can take care of it).

This is the most pretty way I can think of to remove this zero length array.
Keeping the raw_msg[] but with another size seems odd to me.

Or maybe I would be the only one using this feature in the full
tree? In that case, maybe it would make sense to keep the
union_group_attr() macro local to the etas_es58x driver?

I actually ended up with something close to this idea, but more
generalized for other cases in the kernel. There was a sane way to
include a "real" flexible array in a union (or alone in a struct), so
I've proposed this flex_array() helper:
https://lore.kernel.org/lkml/20210818081118.1667663-2-keescook@chromium.org/ (local)

and then it's just a drop-in replacement for all the places that need
this fixed, including etas_es58x:
https://lore.kernel.org/lkml/20210818081118.1667663-3-keescook@chromium.org/#Z30drivers:net:can:usb:etas_es58x:es581_4.h (local)

Hopefully this will work out; I think it's as clean as we can get for
now. :)

The __flex_array itself is a nasty hack :D

Indeed. ;)

but the rest is clean.

Thanks!

Is this compliant to the C standard? Well, I guess that as long
as both GCC and LLVM supports it, it is safe to add it to the
kernel.

The kernel already uses a bunch of compiler extensions, none of which
were legal under the C standard to begin with. :) So, really, this is
about normalizing what we're already doing and finding a single hack
that helps the code base for readability and robustness.

I like the final result. I will do a bit more testing and give my
acknowledgement if everything goes well.

Great; thank you!

-- 
Kees Cook