On Mon, Nov 04, 2019 at 10:21:14PM +0100, Simon Goldschmidt wrote:
Tom Rini trini@konsulko.com schrieb am Mo., 4. Nov. 2019, 22:15:
On Mon, Nov 04, 2019 at 09:28:51PM +0100, Simon Goldschmidt wrote:
Am 04.11.2019 um 20:34 schrieb Heinrich Schuchardt:
struct ip_udp_hdr is naturally packed. There is no point in adding a __packed attribute. With the attribute the network stack does not
compile
using GCC 9.2.1:
Is this commit message correct? In lwIP, we *do* need to pack all these network header structs as they can come in unaligned. Especially the IP header is normally 16-bit aligned if the incoming Ethernet frame is
32-bit
aligned (which is a must for many DMA engines).
This is also the reason why the below code works, I guess: it is rarely totally unaligned, but nearly always at least 16-bit aligned, so dereferencing the checksum pointer as aligned u16 just works.
Nevertheless, the code is formally wrong and your patch is correct. I
just
don't like the commit message saying 'packed' is not required.
Perhaps we should fix the underlying code problem then? Or does that men "rewrite the whole file" ?
This patch fixes the code problem. If there are more problems: any assignment to an u16 pointer from an address of a packed struct issues a warning (provided that appropriate warning settings are used). If we fix all of these warnings (e.g. like we do here, by using alignment agnostic byte accesses), we should be good.
I just think the misleading commit message should be fixed before giving my RB.
Ah, I get it now I hope, thanks.