From: holler@ahsoftware.de (Alexander Holler) Date: Fri, 24 Jun 2011 13:40:27 +0200 Subject: [PATCH] USB: ehci: use packed, aligned(4) instead of removing the packed attribute In-Reply-To: References: Message-ID: <4E0477AB.3080400@ahsoftware.de> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org Am 23.06.2011 16:25, schrieb Alan Stern: > On Thu, 23 Jun 2011, Alexander Holler wrote: > >> Sorry, I never wanted to talk about the issue itself (I've already said >> that), I just wanted to bring in some additional clarity for people >> looking at the code. >> >> I think if there is a packed,aligned(4) most people reading that are >> able to imaging how the struct looks like, whereas nothing (without >> packed) might leave doubts which than requires to read compiler docs or >> the generated code, if one searches a problem in that area. > > I disagree. If there are no annotations at all (no packed), there > should be no doubts. The compiler will add padding wherever it is > needed for internal alignment and perhaps also at the end of the > structure. Nowhere else. I agree to disagree but I assume thats ok. ;) Let me finally add some maybe interesting or informational points for those who are working or examing the issue and/or who might be involved in other discussions on the reason for removing the packed: - I didn't have any problems booting from ehci with kernels compiled with gcc 4.6 on armv5 (or x86*). - 2.6.38.4 (and below) compiled with gcc 4.6 booted from ehci (on a classic beagleboard c4, armv7), whereas everything from 2.6.38.5 upwards didn't (same compiler, same config). I've discovered that before having seen that this might be the issue with the packed, therefor I haven't tested if 2.6.38.5 might work without a packed and have just used gcc 4.5.x for 2.6.38.x. I have tested that a 2,6,39.x compiled with gcc 4.6 and with a removed packed boots from ehci on the beagleboard, so the patch which removes the packed might be a candidate for the stable tree. The reason why booting from ehci stopped with 2.6.38.5+ (gcc 4.6) might be interesting for someone. Looking at the git log I haven't seen something special and I don't know why anything below 2.6.38.5 worked with gcc 4.6 and the packed. - I don't like the idea that every member of every packed struct (without an aligned) might be handled byte by byte. It might be necessary but I still don't like it and would prefer the old behaviour of gcc. I've added this point just to express my personal humble opinion and I don't want to get involved in a discussion on that topic. ;) I've just got involved on that topic by accident and never have had a real reason to do something there (I've done that just for fun). Therefore I now prefer to disappear, which means there is absolutely no reason to respond (to me) or to explain anything to me. Regards and sorry if I wasted someones time, Alexander