From: arnd@arndb.de (Arnd Bergmann) Date: Mon, 20 Jun 2011 22:39:31 +0200 Subject: [PATCH] USB: ehci: use packed, aligned(4) instead of removing the packed attribute In-Reply-To: References: <201106202207.23104.arnd@arndb.de> Message-ID: <201106202239.31997.arnd@arndb.de> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On Monday 20 June 2011 22:28:49 Nicolas Pitre wrote: > On Mon, 20 Jun 2011, Arnd Bergmann wrote: > > > On Monday 20 June 2011 19:39:34 Alexander Holler wrote: > > > That packed without an additional aligned() caused errors on ARM with > > > gcc 4.6 is another problem which got (currently) fixed by removing packed. > > > > Packed caused errors because it is *wrong*. The code as it was used undefined > > behavior in the language. > > I wouldn't call this issue as such, but this is a Red herring. > > Could you please provide a pointer to the structure definition so a > second opinion to the usefulness of __packed there could be provided? The structures in question are ehci_caps, ehci_regs and ehci_dbg_port. The patch that remove the __packed attribute was 139540170 "USB: ehci: remove structure packing from ehci_def". The reason why I consider it a bug is that an access to a register using readl/writel on the structure requires casting a pointer with byte alignment to a pointer with word alignment, which is undefined in C. Gcc just tries to be helpful and work around this by turning the access into bytewise load/store instructions. In older gcc versions, it would not do that if you happen to also case from non-volatile to volatile pointer, but according to Uli that was not an intentional feature of gcc but the ARM code just worked by pure coincidence. > If it is not matching any of the fairly limited cases where having > __packed is relevant then we can just confirm that it should go. It's already gone. Arnd