> Am Freitag, 24. November 2006 06:00 schrieb Nikita V. Youshchenko: > > > However, I've located one more problem which I don't know yet how to > > > fix correctly. > > > AFAIR, device may pass several incoming frames in one urb, and > > > second and later frames are 2-byte-aligned. > > > This may cause situations when IP layer gets IP packet not aligned > > > at 32-bit word boundary and occasionally crash because of unligned > > > access (hardware and/or codebase I'm working with really dislikes > > > unaligned accesses from kernel). > > > > > > Looks the only way to fix that is to detect such cases, and copy > > > data? But this will hurt performance on systems that can handle > > > unaligned accesses ... > > > > One more attempt, with typo fixed ... > > > > Attached is a patch that handles this, based on additional Kconfig > > key. I've tested this on hardware and kernel I'm working with, and > > then forward-ported to 2.6.19-pre6 > > Don't introduce a config option. If you do this by ifdef, use ifdef on > the arches that you used as default. On those arches the memmove is > definitely a win. > You could even test it unconditionally on i386 or x86_64. A copy may be > faster than keeping it in place, as you touch the data many times, but > copy it only once.
Do you thing that copy of whole frames (up to 1.5 kilobytes!) is faster than unaligned accesses to header on x86/x86_64 (which is definitly used by vast majority of this driver users)? ------------------------------------------------------------------------- Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT & business topics through brief surveys - and earn cash http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV _______________________________________________ linux-usb-devel@lists.sourceforge.net To unsubscribe, use the last form field at: https://lists.sourceforge.net/lists/listinfo/linux-usb-devel
